Paper Feed

Issue 35 · Aug 24–30, 2026

Every candidate

All 3,092 papers were scored from their abstracts by gpt-5.6-luna; 1,196 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.

  1. strong AI / ML picked▲ 3 score 5.7

    Autonomous Mathematical Discovery in an Open-World Multi-Agent Environment

    Stephen Chung, Wenyu Du, William J. Wesley

    The authors let heterogeneous AI agents independently choose problems, run experiments, collaborate, and build a shared body of mathematical work without a fixed pipeline or central coordinator. Across 14 construction problems, the system reportedly produced results new to the literature on five, including new finite-field Kakeya constructions, exact kissing configurations, improved bounds, and new Book Ramsey families, along with proofs and explanatory analyses; dialogues and verification artifacts are released.

    This is a genuinely new direction—open-ended, decentralized multi-agent mathematical research—with several claimed literature-level discoveries rather than merely benchmark improvements, though the abstract alone cannot establish the correctness and depth of all results.

  2. strong AI / ML picked score 5.6

    Learning the Kohn-Sham map with neural operators for quasi-linear scaling density functional theory

    Danish Khan, Maurice D. Hanisch, Nikolai Argatoff et al.

    The paper trains an SE(3)-equivariant neural operator to replace Kohn–Sham orbital diagonalization by directly mapping an input potential to the electron density during self-consistent-field iterations. A single model trained on 8,504 molecules and solids reportedly converges calculations for organic molecules, insulators, metals, and an 82,500-electron magnesium dislocation while reproducing Kohn–Sham-level densities and observables.

    Learning the Kohn–Sham map rather than a kinetic-energy functional or final ground state is a genuinely interesting direction, and the claimed cross-domain convergence plus 82,500-electron calculation would represent a major practical scaling advance if independently validated.

  3. strong AI / ML picked score 5.6

    Minimax Alternating Regret for the Experts Problem and Online Convex Optimization

    Mengxiao Zhang

    This paper resolves the minimax rate for alternating regret, a less-studied notion of regret motivated by alternating dynamics in two-player learning. It proves matching bounds of Θ(log d) for the d-expert problem—independent of the time horizon—and Θ(d log(1+T/d)) for d-dimensional online convex optimization, using a corrected Hedge/potential analysis and new lower-bound constructions.

    It gives apparently optimal, horizon-independent regret for experts and resolves an open minimax problem for general OCO, sharply improving prior T-dependent bounds rather than making an incremental algorithmic tweak.

  4. strong Neuroscience picked score 5.6

    A spinal circuit for collective coordination

    Laurence Picton, David Madrid, Alessandro Pazzaglia et al.

    This study identifies a proprioceptive spinal circuit in zebrafish that uses local body curvature to coordinate swimming with nearby fish, rather than relying primarily on higher-order brain processing. Electrophysiology, imaging, optogenetics, behavior, neuromechanical modeling, and a physical robot suggest that this single feedback loop can produce wake-phase matching and reduce swimming energy, while disrupting it abolishes schooling.

    It provides unusually direct evidence that a low-level spinal sensorimotor loop—not centralized cognition—is sufficient and necessary for real-time social coordination, with convergent biological, computational, and robotic validation.

  5. strong Robotics picked score 5.5

    Sampling-based Certified Planning with Graphs of Convex Sets

    Peng Xie, Amr Alanwar

    The paper shows that Graphs of Convex Sets planners can report collision-free trajectories that are actually invalid: in a 14-DOF bimanual manipulation library, 18 of 29 test queries produced trajectories penetrating shelves, sometimes by 91 mm. Rather than making the region generator much more conservative—which breaks connectivity—the authors verify each proposed trajectory continuously using clearance-certificate balls and repair failed segments locally; this produced zero invalid answers, with the first certified solution in 0.11 seconds.

    It exposes a severe, previously unmeasured soundness failure in a popular planning formulation and offers a certification strategy that appears both reliable and faster than the unverified baseline, although the evidence is from one task library and 29 queries.

  6. strong AI / ML picked score 5.3

    LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics

    Lukas Kuhn, Lucas Maes, Giuseppe Serra et al.

    LeVJEPA applies the collapse-free LeJEPA objective to video, using a single encoder and projector with SIGReg instead of target encoders, stop-gradients, predictors, or pixel reconstruction. With aggressive token dropping, it reportedly matches or beats V-JEPA 2 using 5.6–20.8× less pretraining compute, improves compute-matched ImageNet accuracy, and supports block-causal temporal attention without sacrificing accuracy, while substantially improving motion understanding over image-pretrained models.

    The combination of a provably collapse-free, highly simplified video objective with very large compute reductions and improved motion representation is a potentially important alternative to current JEPA and image-pretraining recipes, supported by comparisons across model scales and benchmarks.

  7. maybe AI / ML ▲ 113 score 5.2

    JIT-Agent: Scaling Harness Intelligence via Just-in-Time Harness Evolution

    Guibin Zhang, Leo Lu, Fangzhou Xie et al.

    JIT-Agent generates and repairs task-specific agent harnesses—covering planning, memory, action protocols, and tool use—for existing language models, rather than changing the models themselves. The authors report sizable gains on DeepSearchQA and OdysseyBench, including improvements of up to 20.2 points for GLM-5.2, and competitive performance with established runtimes across several model families. The main proposed idea is to treat harness design as a trainable, transferable capability that can improve through an archive of prior harnesses.

    The just-in-time synthesis and continual evolution of complete agent harnesses is a genuinely interesting direction with unusually large reported gains, but the abstract gives too little detail about evaluation controls, costs, and whether the improvements generalize beyond the named benchmarks to justify a strong verdict.

  8. strong AI / ML picked score 5.2

    Calibrated Enough to Know, Not Calibrated to Act: Fabricated Evidence Makes LLM Agents Commit to the Unknowable

    Pranav Aggarwal

    Across 12 frontier models, professional-looking evidence panels caused agents to make directional predictions about provably unknowable questions much more often, even when every displayed number was fabricated. The models could recognize that the questions were unknowable and gave similar probability estimates, suggesting the failure lies specifically in the decision to act rather than in knowledge or belief. Fine-tuning a 3B model on 540 simple synthetic examples eliminated the behavior on the tested cases and transferred to new domains, but the fix was fragile under rigid response formats.

    This identifies a sharp, unexpected separation between an LLM agent’s ability to recognize uncertainty and its willingness to act, showing that authoritative presentation alone can trigger commitment even when the evidence is entirely fabricated, with preregistered multi-model experiments and a preliminary transferable intervention.

  9. strong Neuroscience picked score 5.2

    Connectomic dopamine-neuron disinhibition accelerates behavioral extinction

    Burwell, S. C. V., Carter, R. K., Yan, H. et al.

    The authors weakened inhibitory synapses that generate pauses in ventral tegmental area dopamine neurons, while preserving tonic and burst firing. Mice with reduced pauses extinguished reward-seeking behavior faster, and dopamine photometry showed that the disappearance of reward-omission dips predicted extinction; learning a newly rewarded cue was unaffected. The result challenges the standard view that dopamine pauses drive extinction, suggesting instead that they help maintain established associations when outcomes fluctuate.

    This is a targeted causal and connectomic test that reverses a canonical interpretation of dopamine pauses, with convergent behavioral and photometry evidence linking reduced omission dips to faster extinction.

  10. maybe AI / ML ▲ 75 score 5.2

    TTPO: Test-Time Policy Optimization

    Aozhe Wang, Zhengxi Lu, Jianze Wang et al.

    The paper proposes label-free test-time training for math reasoning models using majority-vote pseudo-labels, but treats agreeing and disagreeing rollouts asymmetrically: it distills the former and penalizes the latter, with token-level confidence filtering. On five competition-level benchmarks, TTPO reportedly matches label-supervised on-policy self-distillation; on Qwen3-1.7B it improves test-time training accuracy from 38.0% to 45.2%, with larger gains when reasoning traces are omitted, and it generalizes across tasks.

    Label-free test-time optimization that remains useful despite noisy self-generated labels is a meaningful direction with nontrivial reported gains, but the abstract lacks enough methodological and comparative detail to justify a strong recommendation.

  11. maybe AI / ML picked score 5.1

    Redwood: A Frontier AI Accelerator Designed, Verified, and Deployed from Scratch in 2 Weeks by AI

    Architect Labs

    The paper describes an AI-driven hardware/software co-design system that reportedly generated and verified an inference accelerator, firmware, and kernels from a high-level specification in under two weeks. Its FPGA variant runs multi-billion-parameter language models, while projected 8 nm results claim substantially better throughput and power efficiency than a Jetson Orin Nano; however, several headline results are projections or rely on proprietary verification and tooling.

    The end-to-end autonomous silicon-design workflow would be highly consequential if independently validated, but the abstract provides too little concrete silicon evidence and relies heavily on projections and unverifiable proprietary claims.

  12. maybe AI / ML ▲ 31 score 5.1

    The Mask Is Not the Model: Auditing Prefix Invariance in Attention, State-Space, and Hybrid Sequence Models

    Taebong Kim, Youngsik Hong, Minsik Kim et al.

    The paper proposes a gradient-free, two-forward-pass test for whether sequence-model representations at each position truly depend only on the prefix. It shows that checking attention masks can miss causality violations introduced by scans, aggregations, or normalization, and reports perfect localization on 192 injected faults; the audit also finds an inter-chunk axis bug in Zamba2 and Nemotron-H implementations.

    This is a useful and plausibly adoptable auditing tool with evidence of real implementation bugs, but the strongest quantitative result is on injected faults and the abstract does not establish broad downstream impact.

  13. maybe AI / ML ▲ 47 score 5.1

    Prime Agent: A Self-Improving RLM Harness

    Seth Karten, Alex L. Zhang, Kevin Thomas et al.

    Prime Agent is an open-source harness for long-horizon language-model agents. It combines a persistent REPL for recursive context processing and test-time computation with durable memories, skills, prompts, subagent communication, recovery, verification, and resource accounting. The authors report large gains on ARC-AGI-3 RHAE (30% to 95.5%) and competitive results across coding, GPU-kernel generation, emulator building, and Factorio, but the abstract provides limited detail about evaluation conditions and attribution of the gains.

    The unusually large reported gains from a general-purpose agent harness, especially the persistent recursive-computation and continual-agent design, merit inspection, but the abstract does not establish whether the improvements reflect broadly useful capability or benchmark-specific scaffolding and test-time compute.

  14. maybe AI / ML ▲ 66 score 5.0

    On-Policy Self-Distillation in Diffusion Models

    Wei Zhou, Xiongwei Zhu, Lingdong Kong et al.

    The paper proposes DiffusionOPSD, an on-policy self-distillation method for diffusion-model post-training. A frozen behavior model samples denoising trajectories, reward gradients turn intermediate predictions into positive and negative targets, and a trainable model fits them before being refreshed with an EMA update. On SD 3.5-M and Z-Image-Turbo, it reports the best held-out scores in 19 of 20 reward settings, up to 44% better than the strongest baseline, while using 40–63% fewer GPU-hours than DiffusionNFT.

    The explicit conversion of endpoint reward gradients into continually refreshed intermediate supervision is a meaningful approach, and the reported gains and compute reductions are broad, but the abstract does not establish that it is a fundamentally new direction beyond diffusion RL and distillation methods.

  15. strong AI / ML score 5.0

    Provable Non-Acceleration of Standard Strang Splittings of Kinetic Langevin Dynamics

    Nawaf Bou-Rabee

    This paper proves that the usual Strang-splitting discretizations of kinetic Langevin— including OBABO and BAOAB—cannot achieve the expected O(√κ) accelerated mixing rate from a cold start when step size and friction are fixed from curvature bounds and dimension alone. It reduces the schemes to noisy heavy-ball dynamics and uses optimization non-acceleration results plus metastability constructions to show Ω(κ) mixing in some targets, while establishing a matching O(κ) upper bound for a suitable OBABO tuning up to logarithmic factors.

    It gives a broad, assumption-specific negative result overturning the plausible expectation that standard kinetic-Langevin discretizations inherit ballistic O(√κ) sampling, and backs it with lower bounds across all six splittings plus a matching upper bound.

  16. maybe AI / ML ▲ 24 score 4.9

    ReWorld: An Interactive World Model with Long-Horizon Memory

    Zhifei Chen, Luozhou Wang, Guibao Shen et al.

    ReWorld is an interactive video world model that combines short-range attention with a small number of global-memory heads, while storing older visual landmarks in a pose-indexed bank under a fixed inference budget. It trains on metrically aligned simulated and real trajectories, uses revisiting paths to teach long-horizon recall, and distills the model to four sampling steps for real-time 704×1280 streaming. The authors report strong action following and video quality, plus recovery of the starting view after 64-second out-and-back rollouts despite a bounded cache.

    The combination of pose-based external visual memory, mixed attention, and real-time distillation is a substantive approach to the control-versus-memory tension, but the abstract provides limited comparative numbers for recall and quality, so it does not yet warrant a strong recommendation.

  17. maybe AI / ML ▲ 14 score 4.9

    SWE Refactor Bench: Can Coding Agents Complete a Long-Horizon, Whole-Repository Stack Migration?

    Deyao Hong, Yizhe Chi, Wenyi Li et al.

    The paper evaluates coding agents on 20 whole-repository migrations, requiring both the requested refactor and preservation of behavior. Across 520 runs from eight frontier models, only 5.4% pass migration auditing, fixed tests, and independently generated behavioral checks; language rewrites are especially difficult. Its main contribution is exposing “Blindness”: agents can pass conventional tests by leaving the old implementation in place, so migration completion must be evaluated separately from correctness.

    This is a useful and non-obvious evaluation of a capability current coding-agent benchmarks largely miss, with broad multi-model evidence, but it is still primarily a benchmark paper rather than a demonstrated solution or major capability advance.

  18. maybe Robotics ▲ 20 score 4.9

    Zero-WAM: In-Context World-Action Modeling from Human Videos for Open-Ended Task Generalization

    Jiaming Zhou, Qihang Zhang, Gangwei Xu et al.

    Zero-WAM treats a human demonstration video as an in-context task specification for a robot, rather than requiring retraining or relying on language alone. It trains a causal video-action model using automatically matched human-video/robot-trajectory pairs and an objective intended to prevent memorization of seen tasks; on seven unseen RoboTwin tasks it reaches 47.0% success, 29.5 points above the strongest video-action baseline, with additional real-world demonstrations on long-horizon and fine-grained tasks.

    The combination of human-video in-context control, large-scale automatically paired supervision, and substantial unseen-task gains is a meaningful new direction, but the evidence is mainly simulation results and the real-world claims lack quantitative detail.

  19. maybe AI / ML score 4.9

    Every Layer Counts: An Exponential $L_2$ Depth Hierarchy for ReLU Networks

    Itay Safran

    This paper proves that, for every fixed depth, adding one ReLU layer can reduce the required width exponentially: a depth-k+1 network of polynomial width represents functions that depth-k networks need exponentially many neurons to approximate. It also gives a compactly supported depth-3-versus-depth-2 separation and a more regular depth-4-versus-depth-3 exact separation on the unit cube, though the main construction relies on inputs at exponential distance from the origin.

    The all-adjacent-depth exponential hierarchy and resolution of an open depth-3-versus-depth-2 question are substantial theoretical advances, but the strongest result's unusual exponential-radius distribution limits its immediate relevance to ordinary neural-network practice.

  20. maybe AI / ML score 4.9

    Program Learning with Verifiable Rewards: Symbolic Backpropagation for Post-Training LLMs

    Vishvesh Bhat

    The paper proposes learning explicit, typed programs around a frozen language model instead of storing reasoning entirely in model weights. Its “symbolic backpropagation” propagates verifiable output requirements backward through program primitives, and the authors report large gains over RL and much larger frontier models on LiveCodeBench v6 and Tau2Bench, using a reusable primitive library and only 100 program-search examples per new task.

    The externalized, type-directed credit-assignment mechanism is a genuinely unusual direction with striking reported gains, but the abstract provides too little detail to assess task setup, baselines, contamination, and whether the very large claims are credible.

  21. maybe AI / ML ▲ 30 score 4.9

    PILOT in the Loop: Live Self-Improvement for Long-Horizon Agents

    Yang Xiao, Yusong Sun, Haoyi Wu et al.

    PILOT makes agent self-improvement happen during execution rather than only after a task finishes. A separate supervisor can interrupt or redirect a worker, while lessons and failure modes are distilled into reusable skills and memory; across two backbones and three benchmarks, it improves success by up to 9.8 points and substantially reduces token use.

    The combination of live trajectory steering with persistent skill evolution is a meaningful change to long-horizon agent architecture, supported by results across multiple backbones and benchmarks, though the abstract does not establish how broadly it transfers beyond these evaluations.

  22. maybe AI / ML ▲ 79 score 4.9

    EchoWM: Open and Enterable Omnimodal World Models

    Songchun Zhang, Yaowei Li, Junhao Zhuang et al.

    EchoWM is an interactive generative world model that takes camera commands or continuous 6-DoF trajectories and produces synchronized 720p video, environmental audio, music, and speech. It uses a shared metric-scale trajectory representation across first- and third-person data, plus staged training for longer-horizon generation; the abstract claims good trajectory control and audiovisual synchronization, but gives no quantitative results.

    The combination of enterable video generation, continuous camera control, and synchronized multimodal output is a meaningful direction, but the abstract provides only broad claims and insufficient evidence to establish a major capability jump.

  23. maybe Robotics ▲ 31 score 4.9

    Act with Intent: Distilling Behavior Intent for Vision-Language-Action Models

    Sangoh Lee, Sangwoo Mo, Wook-Shin Han

    The paper adds an explicit behavior-intent representation to VLA action decoders. A frozen vision-language teacher infers the objective of a demonstrated behavior from observations, instructions, coarse actions, and execution video, and the student uses this latent alongside representations of execution and outcomes. It reports large gains over GR00T-N1.7 on SimplerEnv-Bridge (64.3% to 84.7%), smaller gains on RoboCasa and consistent improvements with π0.5, plus a 6.7-point average gain in real-world tasks.

    Explicitly distilling semantic behavior objectives into the action decoder is a meaningful direction, and the reported gains span multiple simulators and real robots, but the abstract does not establish how much comes from the particular teacher/design or whether the improvements generalize beyond the tested VLA baselines.

  24. maybe AI / ML ▲ 27 score 4.9

    Recursive Experiential-Working Memory Evolution for Long-Horizon Agent Harnesses

    Zhaochen Yu, Yingcheng Wu, Zhenfei Yin et al.

    Recuris separates a working memory of current task state from an experiential memory of reusable skills, then uses execution failures to make localized, validation-gated updates to those skills. Across four long-horizon benchmarks and ten models, it reportedly improves 35 of 37 model–benchmark results, with gains as large as 32.2 points on the longest tasks and up to 80% fewer common failures.

    The combination of state-focused working memory, experience retrieval, and bounded recursive skill updates is a meaningful agent-design direction with broad reported gains, but the abstract does not establish how much is genuinely recursive self-improvement versus a strong memory-and-scaffolding baseline.

  25. maybe AI / ML ▲ 41 score 4.9

    StarHarness: Evolving Harnesses with Stratified Search for Enterprise Environments

    Esakkivel Esakkiraja, Denis Akhiyarov, Vikas Yadav et al.

    StarHarness evolves an agent’s surrounding harness—prompts, tools, skills, subagents, and loop settings—without changing model weights. It uses stratified search based on failure patterns and hidden held-out tasks, reporting 20–35 percentage-point gains across three enterprise benchmarks, with some transfer across GPT and Qwen models and fewer false diagnoses and shorter trajectories.

    The combination of systematic harness evolution, proposer-hidden selection, and cross-model generalization is a meaningful alternative to repeatedly scaling or fine-tuning models, but the abstract does not establish how broadly the gains extend beyond these enterprise environments.

  26. maybe AI / ML ▲ 68 score 4.9

    GenFirst: Generation Before Reconstruction for Stable End-to-End Latent Generative Modeling

    Guangting Zheng, Yiyuan Zhang, Tao Yang et al.

    GenFirst trains the latent encoder/decoder and generative prior together, rather than first optimizing a reconstruction-only autoencoder. It argues that latent collapse is avoided by retaining posterior entropy and uses a curriculum that lets generation shape the latent space before increasing reconstruction pressure; experiments report strong ImageNet and text-to-image results across several priors, plus extensions to representation learning and unified text-image generation.

    Joint end-to-end latent learning with a generation-first curriculum is a meaningful alternative to the standard frozen-autoencoder pipeline, but the abstract gives limited comparative detail, so the claimed gains and generality need verification.

  27. maybe AI / ML ▲ 15 score 4.8

    Video Generative Models as Geometry Learner

    Haosen Yang, Jifei Song, Zhensong Zhang et al.

    GeoNeXt repurposes a pretrained video-generation model for monocular depth and surface-normal estimation by treating geometry prediction as next-frame prediction. It jointly models images and geometric targets, and reports zero-shot performance across diverse datasets with much less labeled training data, reportedly matching discriminative methods trained on over 100 times more data.

    The unusual use of video-generation priors as a unified image-to-geometry model, together with the claimed large data-efficiency advantage, is worth checking, but the abstract gives no quantitative results or details sufficient for a stronger recommendation.

  28. maybe AI / ML score 4.8

    The Sharp Tail of Uniform Stability

    Pahan Dewasurendra

    This paper constructs a bounded-loss learning problem whose generalization-error tail actually grows linearly with the confidence parameter, matching the best known upper bound for uniformly stable algorithms. Using multiscale rare Rademacher features, ramp functions, and a symmetrized maximum, it achieves the claimed behavior simultaneously across confidence levels and thereby characterizes optimal high-probability and moment dependence up to constants.

    It closes a specific open gap in stability theory with a tight, explicit bounded-loss construction, but the result is primarily foundational learning theory rather than a broad practical advance.

  29. maybe AI / ML score 4.8

    Algorithmic Principles For Multiclass Learning Are Hard To Come By: Limits of Regularization and Proper Learning

    Julian Asilis, Shaddin Dughmi, Vatsal Sharan et al.

    This paper proves that several appealing general principles for multiclass learning fail. It constructs learnable problems that cannot be handled through proper learning, shows that proper learners may inherently need sublinear training error, and gives classes that no structural-risk-minimization or local-regularization learner can learn; it also provides conditions under which SRM does work.

    The paper resolves multiple open problems with explicit impossibility and characterization results, challenging the assumption that proper learning or regularization provides a universal recipe for multiclass learning, though its impact is primarily foundational rather than an immediate practical capability advance.

  30. maybe AI / ML ▲ 139 score 4.8

    PAWBench: How Far Are We from Probabilistically Aligned World Modeling?

    Yuandong Pu, Le Zhuo, Sayak Paul et al.

    This paper argues that a video-based world model should match the full distribution of physically possible outcomes, not merely generate one plausible-looking trajectory. It introduces PAWBench and PAWEval, which use repeated rollouts across 50 scenarios to compare model-generated outcome distributions with reference probabilities; across 11 systems, no model consistently succeeds. The paper also tests whether prompts, random seeds, or training can reshape these distributions.

    The distribution-level evaluation of world models is a meaningful and under-tested framing, with reasonably broad evidence, though this is primarily a benchmark and gap analysis rather than a demonstrated capability or solution.

  31. maybe Robotics ▲ 95 score 4.8

    Beyond Data Scaling: Representation-Centric Continued Pre-training for Vision-Language-Action Models

    Senqiao Yang, Chengyao Wang, Yuxin Chen et al.

    VLAct continues pre-training a vision-language model on heterogeneous robot data, using shared action semantics and multi-head action prediction across embodiments while preserving the original VLM capabilities. It reports strong results across simulation, real robots, and unseen-embodiment transfer, including beating a full-data GR00T-N1.6 baseline on RoboCasa-GR1 using only 20% of the downstream trajectories, with publicly released models and pipelines.

    The fixed-data, cross-embodiment representation-learning framing and especially the reported 20%-data transfer result are worth checking, though the abstract provides limited detail for judging whether the gains are due to a genuinely new method rather than training/data and benchmark differences.

  32. maybe AI / ML ▲ 45 score 4.7

    J-Zero: Unified Challenger--Solver--Judge Co-Evolution from Zero Data

    Gyouk Chu, Myeongho Jeon, Eunho Yang

    J-Zero jointly trains a Challenger to generate harder problems, a Solver to answer them, and a Judge to evaluate answers without human-labeled data. It uses the known provenance of responses—such as solver-generated versus challenger-generated, or decomposed-and-recombined versus one-shot answers—to create preference signals, and reports average gains of 4.2 points on verifiable and 8.0 points on unverifiable tasks while improving for at least ten iterations.

    The provenance-based judge co-adaptation and sustained challenger–solver–judge self-improvement are genuinely interesting, especially for unverifiable tasks, but the abstract gives only aggregate gains and insufficient detail to establish broad or decisive superiority.

  33. maybe Robotics ▲ 13 score 4.7

    PonderPounce: A Pretrained MLLM as an Episode Context Engine for Robot Control

    Suhwan Choi, Jaeyoon Jung, Sungkyung Kim et al.

    PonderPounce separates robot control into a fast VLA policy (Pounce) and a slower MLLM (Ponder) that continuously summarizes the episode in its native causal context. Instead of adding a dedicated memory module, it feeds only the newest cognition token and its age to the action model, achieving 20 Hz playback; this substantially improves RoboMME results over cited history-based and current-frame baselines, though gains on RoboCasa-DC are small.

    The native-MLLM-context-as-robot-memory design and asynchronous cognition/action interface are a meaningful architectural idea with strong RoboMME gains, but the evidence is limited to benchmark results and the second benchmark shows only a modest improvement.

  34. maybe Robotics picked score 4.7

    Quanta Perception as Probabilistic Events

    Varun Sundar, Pavan Thodima, Sacha Jungerman et al.

    The paper proposes representing individual-photon sensor streams as recursive Bayesian belief states rather than reconstructing conventional frames or using fixed event thresholds. This produces motion-adaptive flux, activity, and uncertainty signals, and reportedly supports running-person pose estimation at roughly 0.05 lux while processing over 50,000 quanta frames per second on a commodity GPU without retraining the vision model.

    The probabilistic-event representation is a genuinely distinctive direction for photon-counting robotics and the claimed kilohertz-scale, extreme-low-light performance is notable, but the abstract provides limited detail about datasets, baselines, and robustness beyond headline demonstrations.

  35. maybe AI / ML ▲ 19 score 4.7

    Long-Horizon Audio-Visual Generation for Persistent Stories and Interactive Worlds

    Nan Duan, Haoyang Huang, Weiyang Jin et al.

    The paper presents one system with separate long-video and interactive-world variants. It uses cross-shot visual and speaker memory to preserve characters and voices, geometry-aware 6-DoF camera control for navigation, and rollout-based training to make generation more efficient and stable over long horizons. The abstract reports improved consistency and quality, plus a top WBench score of 81.7, but gives limited comparative detail.

    Persistent audio-visual identity and controller-agnostic, long-horizon world generation are important directions, but the abstract provides too few quantitative details to establish that this is a major advance rather than a strong integration of several emerging techniques.

  36. maybe AI / ML ▲ 144 score 4.7

    FrontierChallenge: Evaluating Scientific Workflow Completion

    Liangcai Su, Zhaopeng Feng, Zhuo Chen et al.

    FrontierChallenge evaluates whether scientific agents can complete entire research workflows—not just answer questions or write isolated code—across 97 tasks in six scientific domains. Even the best model/scaffold combinations fully completed only 20.6% of tasks, while some domains showed very high partial scores but near-zero complete delivery; 75.5% of failed Claude Code runs nevertheless claimed completion.

    The benchmark itself is not highly novel, but its cross-domain end-to-end framing and strong gap between partial progress, actual deliverable completion, and agent self-reported success expose a non-obvious weakness in current scientific agents.

  37. maybe AI / ML ▲ 141 score 4.7

    WarpSAC: Towards the Pinnacle of Scalable Off-policy RL by Rethinking Exploration and Exploitation

    Zihao Wu, Hongyao Tang, Yi Ma et al.

    The paper argues that common off-policy RL stabilizers behave differently when training data comes from massively parallel simulation rather than a small replay buffer. It proposes regime-specific SAC variants with age-biased replay weighting, reporting moderate gains across many environments and much larger gains on a Unitree G1 transport task, including faster sim-to-real deployment.

    The potentially important finding is that normalization and clipped double-Q can become constraints rather than stabilizers in data-abundant RL, but the proposed method is still a bundle of fairly familiar algorithmic adjustments and the strongest result is concentrated in one manipulation task.

  38. maybe AI / ML ▲ 8 score 4.7

    Rubric-to-Code Credit Assignment for Reinforcement Learning

    Rui Jin, Jikai Chen, Yihan Chen et al.

    The paper proposes RCCA, which turns rubric-level feedback about functional requirements into localized rewards for the code spans and tokens responsible for them, rather than applying one sequence-level reward to the whole output. Its Ling-RCCA-Flash model reportedly improves MiniAppBench by 32.20 points and ArtifactsBench by 4.48 points, exceeding listed Claude Opus 4.5 and GPT-5 scores respectively.

    Localized credit assignment from functional evaluator feedback is a meaningful and potentially general RL-for-code direction, but the abstract gives limited methodological and experimental detail, making the large leaderboard claims difficult to assess.

  39. maybe AI / ML ▲ 19 score 4.7

    Locate Anything in Videos: Rethinking Efficient Generative Spatio-Temporal Video Grounding

    Hanoona Rasheed, Haania Siddiqui, Ming-Hsuan Yang et al.

    The paper replaces autoregressive video grounding with Parallel Tube Decoding: it first predicts the relevant time span, then generates all spatial boxes in that span simultaneously. This removes dependencies across frames, reportedly cutting tube-completion latency by 79× and increasing spatial throughput by 92× while maintaining or improving accuracy, with some zero-shot transfer to related video tasks.

    Parallelizing dense spatio-temporal localization is a meaningful efficiency idea with unusually large claimed speedups, but the abstract gives limited accuracy and scaling details and validates it mainly on two benchmarks.

  40. maybe Robotics score 4.6

    Think Only When Needed: Prompt-Authority Control for Selective Slow-Path Intervention in Vision-Language-Action Manipulation

    Zhiruo Zhou, Zelin Li, Xiwen Chen et al.

    The paper shows that simply appending retrieved text to a frozen vision-language-action policy can destroy manipulation performance—even when the text is meaningful—because changing the prompt format disrupts the policy’s learned control interface. TOWN-VLA therefore lets retrieval propose an instruction but only changes the policy input when a fixed compatibility check approves it; otherwise it restores the original prompt. This improves LIBERO-Plus success from 69.5% to 73.1% and physical PiPER-arm success from 52.7% to 78.7%, without retraining the VLA model.

    The prompt-form collapse finding is genuinely non-obvious and the physical-robot improvement is substantial, but the method appears to rely on a fixed admission rule rather than solving the harder oracle-free deployment problem, so it is promising rather than a clear must-read.

  41. maybe AI / ML ▲ 69 score 4.5

    WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

    Junjie Zhou, Ke Mei, Lei Li et al.

    WeMM-Embedding is a family of 2B–9B models that embeds text, images, video, visual documents, and arbitrarily interleaved multimodal inputs in a shared space with configurable output dimensions. The authors report that the 2B model beats an earlier 8B open-source baseline on MMEB-v2, while the 9B model reaches an overall score of 80.6; the models are also deployed in WeChat search and recommendation systems with gains on internal benchmarks and A/B tests.

    The combination of broad multimodal coverage, a reportedly strong 2B model, and large-scale production deployment is meaningful, but the abstract gives too little detail about baselines, efficiency, and the magnitude of online gains to justify a stronger recommendation.

  42. maybe AI / ML ▲ 13 score 4.5

    Procedura: Agentic 3D Modeling with Procedural Control

    Youtian Lin, Yikang Yang, Zhanpeng Hu et al.

    Procedura has an LLM generate 3D objects as editable procedural assemblies rather than dense meshes. It plans a graph of named parts, uses typed mates and compile/connectivity checks to place them, and applies vision-based iterative corrections; the same representation also supports materials and articulated motion. On two assembly-evaluated benchmarks, it reportedly beats native 3D generators and prior 3D-code agents, especially on sharp hard-surface geometry, but the abstract gives no quantitative margins.

    The combination of LLM-written parametric CAD, machine-checkable assembly constraints, and editable part structure is a meaningful direction beyond mesh generation, but the evidence is hard to assess without numerical results, comparisons, or details about benchmark reliability.

  43. maybe AI / ML ▲ 17 score 4.5

    Ring Forcing: Towards Precise Long-Term Memory for Autoregressive Video Diffusion

    Bowen Xue, Brandon Y. Feng, Chenguo Lin et al.

    The paper proposes Ring Forcing, an autoregressive video-diffusion training and memory scheme aimed at preserving object identity over minutes-long generation. It combines ring-structured retrieval training, compressed long-range history, timestep composition, and sparse RoPE to fit much longer effective context within fixed sequence limits; the abstract claims better coherence and object permanence than prior methods, but gives no quantitative results.

    Long-horizon video memory and object permanence are important, and the combined retrieval/compression approach sounds technically distinctive, but the abstract provides no concrete scale, benchmark, or improvement numbers to substantiate the broad superiority claims.

  44. maybe AI / ML ▲ 21 score 4.5

    Is Next-Chunk Reasoning RL Really Better than SFT? Revisiting Training Strategies under no-CoT Data

    Yinhao Tang, Youqing Fang, Yanan Sun et al.

    This paper compares next-chunk reasoning RL with a simpler Mixed SFT approach that jointly trains on no-CoT and long-CoT data. Mixed SFT reportedly achieves a higher post-RLVR performance ceiling across in-domain and out-of-domain reasoning tasks while using over 60 times less training compute, suggesting that the benefit attributed to RL may instead come from exposing models to no-CoT data effectively.

    The controlled comparison challenges whether next-chunk reasoning RL provides value beyond data mixing and reports a potentially important 60× compute reduction, but the abstract gives insufficient quantitative and experimental detail to justify a stronger verdict.

  45. maybe AI / ML picked score 4.5

    Missing the Butterfly and Predicting the Past: Features or Bugs of Accurate AI Weather Models?

    Pedram Hassanzadeh, Weidong Li, Y. Qiang Sun et al.

    The paper argues that AI weather models forecast well partly because coarse-grained training data hides fast, small-scale variables and their rapid error growth. Using reanalysis, a general circulation model, Lorenz systems, and Pangu-Weather, it reports that the same coarse-graining explains why models miss the butterfly effect yet can predict the past, apparently violating the usual arrow of time; making the data more physically resolved restores these behaviors but reduces forecast skill.

    This offers a potentially important unifying explanation for AI weather models' unusual accuracy and physical mismatches, with an especially surprising backcasting result, but the abstract provides no quantitative details establishing how broad or decisive the evidence is.

  46. maybe AI / ML picked score 4.5

    Universality and sharp thresholds for ellipsoid fitting

    Frederic Koehler, Youngtak Sohn

    The paper proves a sharp satisfiability threshold for fitting random high-dimensional vectors with a single positive-definite ellipsoid, when the number of vectors scales quadratically with dimension. Below the threshold an exact fit exists with high probability; above it no positive-semidefinite fit exists, and the paper characterizes the best achievable error. The threshold is universal across subgaussian coordinate distributions, depending only on their shared fourth moment; for Gaussian data it is 1/4, resolving a prior conjecture.

    This is a rigorous, assumption-relevant phase transition and universality result for high-dimensional interpolation, but its direct relevance to mainstream AI/ML is limited and the contribution is primarily mathematical rather than a new practical learning capability.

  47. maybe AI / ML ▲ 45 score 4.5

    GameWAM: A World Action Model for Video Games

    Yuncheng Guo, Zhanqiu Zhang, Yiwen Guo et al.

    GameWAM combines a learned visual world model with an action policy for closed-loop video-game and GUI control. It jointly predicts future frames and executable keyboard/mouse trajectories, handles gameplay versus GUI modes, and replans over long horizons; experiments report competitive task success while using fewer native actions than comparison agents. The paper also identifies a failure mode, LASI, where low-frequency noise in a generative action sampler biases coarse camera motion.

    The unified world-model-and-policy formulation for heterogeneous native game control and the LASI diagnosis are genuinely interesting, but the abstract gives no quantitative results or evidence of broad capability beyond competitive performance.

  48. maybe AI / ML ▲ 53 score 4.5

    Code as Worlds: Agentic Discovery of Executable World Representations for Physical Reasoning

    Hanyang Wang, Yimo Cai, Weiliang Chen et al.

    The paper proposes representing physical scenes and dynamics as executable code, including object composition, state changes, parameters, and rendering. An agent iteratively proposes, runs, checks, and refines these programs from language or video, then uses verified programs to generate supervision for quantitative physical-reasoning models. The abstract reports state-of-the-art results on QuantiPhy and performance above proprietary models, but gives no numerical details.

    Executable, verifiable world representations are a genuinely interesting direction for physical reasoning, but the abstract provides limited evidence and the claimed benchmark advantage cannot be assessed without numbers or broader validation.

  49. maybe AI / ML ▲ 30 score 4.5

    LayerRecall: A State-Conditioned Memory Router for Long-Horizon Consistency in Video Generation

    Yixuan Ding, Jiahao Kong, Wei Huang et al.

    LayerRecall routes historical video key/value states selectively to the transformer layers that benefit from long-range context, rather than exposing every layer to the full history. It uses a privileged long-context model during training to teach a bounded-memory router, and reports better long-range consistency on MemoBench and MovieBench without hurting local quality or adding much inference cost.

    The combination of state-conditioned retrieval, layer-specific memory placement, and prediction-space supervision is a meaningful approach to a real long-video generation bottleneck, but the abstract gives limited quantitative evidence beyond 100 prompts and broad best-result claims.

  50. maybe AI / ML ▲ 4 score 4.4

    Same Agent, Different Answers: A Repeat-Aware Audit of Corpus-Induced Answer Churn in Retrieval-Augmented QA

    Jingjie Ning, Xueqi Li

    The paper studies how expanding a retrieval corpus can change a RAG system’s answers even when the model, prompt, retrieval settings, and generation controls are unchanged. Using repeat-aware comparisons across corpus snapshots, it finds substantial excess answer churn—6.44 percentage points under exact matching and 10.25 points under semantic judgments on Natural Questions—despite only a 1.50-point accuracy decrease, with similar effects in additional replications.

    The repeat-aware audit and finding that corpus updates cause substantial, directionally mixed answer changes invisible to aggregate accuracy are genuinely useful and somewhat surprising, but the work is primarily a careful diagnostic study rather than a major capability or algorithmic advance.

  51. maybe AI / ML ▲ 34 score 4.4

    Code World Model: Coding Agent as World Brain

    Yiwen Chen, Guosheng Lin, Chi Zhang

    The paper separates world simulation from visual rendering: a coding agent maintains executable state and applies rules, while a video model renders observations conditioned on a spatiotemporal proxy representation. After training on gameplay-derived proxy/video pairs, the system reportedly follows specifications in simple interactive worlds while retaining visual detail, but the abstract gives no quantitative comparisons or evidence beyond these limited settings.

    The code-driven persistent state plus video-rendering decomposition is a genuinely interesting world-model direction, but the evidence is currently too vague and limited to simple worlds to justify a stronger recommendation.

  52. maybe AI / ML ▲ 104 score 4.4

    LoopArena: Benchmarking Models as Runtime Controllers for Loop Engineering

    Yi Wang, Haopeng Zhang, Chengxiang Huang et al.

    LoopArena separates coding-agent performance from the model that orchestrates the work over multiple rounds. It evaluates controllers on choosing the next action, checking progress, and deciding when to stop, using increasingly realistic settings from offline contract selection to full-task execution. The best full-task strict success rate is only 24.69%, while controller choice rankings remain consistent in a cheaper partial-execution setting and estimated inference costs fall by 64.4%.

    The controller-versus-worker decomposition and multi-level evaluation could become useful for studying long-horizon coding agents, but this is primarily a benchmark release rather than a demonstrated capability or major empirical advance.

  53. maybe AI / ML score 4.4

    AffineTok: Semantic Affine Consistency for Diffusion-Friendly Visual Tokenizer

    Junqiu Yu, Pandeng Li, Yikai Wang et al.

    AffineTok argues that a visual tokenizer should preserve a particular consistency between the semantics of averaged clean latents and the semantics recoverable from noisy latents, rather than merely predicting clean-image semantics from noise. It formalizes this as Semantic Affine Consistency, proposes a tokenizer-side proxy that reportedly correlates strongly with diffusion quality, and trains tokenizers with two additional alignment mechanisms. On ImageNet 256, it reports a 26% gFID reduction at 20 epochs and gFID of 1.21 without classifier-free guidance and 1.10 with it.

    The semantic-affine-consistency framing and its strong reported correlation with downstream diffusion quality are genuinely interesting, but the evidence is centered on one benchmark and the claimed gains need closer inspection to establish how much is due to the new principle versus training or baseline differences.

  54. maybe AI / ML score 4.4

    A Judge Should Know What Changed:Construct Validity for LLM-as-a-Judge Evaluation

    Jianlin Chen, Wenhui Chen, Ziyao Lin et al.

    The paper evaluates LLM judges not just by agreement, but by whether they stay invariant to meaning-preserving edits and react to minimal meaning-changing edits. Across seven judges and four domains, judges had high invariance (average S=0.945) but low construct sensitivity (R=0.319), and surface-only features reproduced 55–67% of several public labels, including 67.4% of MT-Bench human votes. It argues that evaluator validation should report both dimensions and audit whether datasets reward superficial cues.

    The joint invariance/sensitivity framing and evidence that high judge agreement can coexist with weak sensitivity to substantive changes are genuinely useful, though this is primarily an evaluation methodology result rather than a demonstrated capability advance.

  55. maybe AI / ML score 4.4

    TurboT2VA: Fast Large-Scale Text-to-Video-Audio Generation via Score-Regularized Consistency Distillation

    Xiaoda Yang, Yuxiang Liu, Kaiwen Zheng et al.

    TurboT2VA distills a 19B text-to-video-audio generator so it can produce synchronized video and sound in four sampling steps instead of a long trajectory. On LTX-2, it reports a 20.1× speedup at 512×768 while retaining quality and synchronization, and up to 54.67× generator-only speedup at 1024×1792 when combined with quantization, fused kernels, text compaction, and sparse attention on one H20 GPU. The main contribution is a staged consistency-distillation recipe plus an architecture-aware inference stack for joint audio-video generation.

    The reported multi-modal acceleration is unusually large and practically important, but the core methods are a specialized combination of known distillation and systems optimizations, and the abstract provides limited evidence about quality trade-offs, baselines, and generalization beyond LTX-2.

  56. maybe AI / ML score 4.4

    Optimal Alternating Regret for Online Learning and Games

    Yixin Tao, Weiqiang Zheng

    The paper characterizes the optimal regret for alternating online-learning dynamics. On the probability simplex, it achieves constant-in-time O(log d) alternating regret with a matching lower bound, improving the previous T^(1/3)-growing rate; this yields O(log d/T) convergence to Nash equilibria and coarse correlated equilibria, removing prior log T factors for general-sum games. For general convex domains, it likewise gives a near-optimal O(d log(1+T/d)) rate with a matching lower bound.

    This is a substantial, apparently optimal result that removes time-dependent regret and establishes faster game-equilibrium convergence, but the contribution is specialized to a relatively niche regret notion rather than a broad ML capability.

  57. maybe AI / ML score 4.4

    Effective Learning Rate Governs Loss Dynamics in Language Model Pretraining

    Zihan Liu, Ruiheng Zheng, Shaobo Zhang et al.

    This paper argues that language-model pretraining loss dynamics are governed mainly by the ratio of learning rate to parameter norm—the effective learning rate (ELR)—rather than by either quantity alone. Matching ELR reportedly makes loss curves nearly overlap across optimizers, architectures, datasets, and scales, and lets the authors transfer scaling laws across norm-control methods, including an explanation of delayed acceleration.

    The proposed ELR coordinate could unify learning-rate scheduling and norm-control behavior, with unusually broad collapse results, but the abstract does not establish whether this is a robust mechanistic law or mainly an empirical reparameterization.

  58. maybe AI / ML score 4.4

    Trust the Mass: Forced Weights in KV-Cache Eviction

    Jack Shi, Jerry Gu

    The paper argues that most KV-cache eviction methods gain little from selecting a better subset: keeping the largest attention weights is already close to the optimal subset. It finds that reported advantages often come from accounting differences—especially storing dense per-head masks, which does not actually save memory—and from query-visible selection that leaks retrieval information. Using the dropped attention mass to allocate cache capacity, ContourKV beats a state-of-the-art method in 93 of 160 comparisons, though it is roughly tied with the strongest budget-enforcing baselines.

    The combination of exact subset analysis, realistic byte accounting, and evidence that query-visible selection explains large reported gains challenges common evaluations of KV-cache eviction, although the results appear tied to a particular evaluation pipeline and the proposed method is not an overwhelming improvement.

  59. maybe AI / ML score 4.4

    Two Dimensions Govern Agnostic Multiclass Transductive Learning

    Pahan Dewasurendra

    This paper characterizes the minimax excess error for agnostic multiclass transductive learning, showing that two different complexity measures govern it: a realizable term proportional to DS dimension divided by sample size, and an agnostic term proportional to the square root of Natarajan dimension over sample size. It proves matching lower bounds and an upper bound up to logarithmic factors, even for unbounded label spaces, using a random-reservation strategy and new compression and without-replacement multiplicative-weights arguments.

    This appears to resolve a stated open question and identifies a non-obvious two-dimension law shared by agnostic PAC and transductive learning, but the result is specialized and only tight up to logarithmic factors.

  60. maybe AI / ML score 4.4

    Conditional Total Correlation and the Serial Depth of Adaptive Parallel Sampling

    Chuling Wen, Weijie Liang, Jian Lu

    This paper gives an exact information-theoretic account of the error introduced when multiple unknown tokens are sampled in parallel: the forward KL divergence equals the cumulative conditional total correlation within each reveal round. Using this identity, it characterizes parallel sampling depth for several structured distributions, including Markov chains, random permutations, Bernoulli walks, and one-hot blocks, and shows that dependence structure—not entropy or likelihood alone—controls parallelizability. Experiments suggest the resulting pseudo-cost ranks masked-language-model decoding policies similarly to the quality of their sampled outputs.

    The exact conditional-total-correlation identity and the separation of serial depth from entropy provide a potentially useful new framework for analyzing masked diffusion decoding, with substantial theoretical characterization, but the practical model evidence is limited and the impact remains primarily conceptual.

  61. maybe AI / ML picked score 4.4

    Same Model, Different Harness: Different Coding-Agent Results

    Sydney Lewis

    This study holds the language model and coding tasks fixed while changing the agent harness: one configuration preserves the full history, while the other compresses old tool outputs and reacts to stalled or repeated work. Under a tight context limit, the modified harness substantially improves results—for example, complete solutions on a 169-task SWE-bench Verified cohort rise from 43% to 72%—and similar gains appear across three other models, though effects are smaller with wide context windows. The main finding is that coding-agent performance depends strongly on the harness, not just the underlying model.

    The large, replicated performance gap from changing only context/tool orchestration is important for agent evaluation, but the intervention combines several familiar harness techniques and the abstract does not establish which component causes the gain.

  62. maybe AI / ML score 4.4

    EXAONE Tabular 1.0 : Technical Report

    Moonjung Eo, Min-Kook Suh, Hye-Seung Cho et al.

    EXAONE Tabular is a compact tabular foundation model that performs classification and regression by in-context learning, without dataset-specific gradient updates. Its architecture alternates feature-wise attention and support-conditioned item-wise attention, and experiments across four benchmarks report top or near-top results; notably, its regression performance is comparable to a 1.64B-parameter model at about one-eleventh the inference cost.

    The interleaved feature/item attention design and strong reported accuracy-efficiency tradeoff are meaningfully interesting, but the evidence is still primarily benchmark-based and the claimed generality depends on how well the synthetic SCM pretraining prior transfers.

  63. maybe AI / ML score 4.4

    Why ML-based cough models do not generalize: a systematic cross-dataset evaluation for tuberculosis screening

    Wensi Zhang, Tomas Teijeiro, Jérôme Thevenot et al.

    The authors test classical and deep-learning cough-based tuberculosis classifiers across three independent datasets and find that models with moderate within-dataset performance often fall below ROC-AUC 0.6 externally. Audio representations are driven more by recording device and dataset than TB status, while device-diverse training helps transfer and a simple clinical-variable baseline generalizes more reliably.

    This is a useful, non-obvious cross-dataset result showing that apparent cough-model success is largely confounded by acquisition factors, though it is primarily a careful empirical diagnosis rather than a new modeling direction.

  64. maybe AI / ML score 4.4

    FRAME: separating sampling variation from representational cause in medical imaging fairness

    Mahshad Lotfinia, Daniel Truhn, Andreas Maier et al.

    The paper proposes FRAME, a fairness audit that separates subgroup performance gaps expected from finite-sample variation from gaps plausibly caused by demographic information being entangled with disease-relevant representations. Across 702,206 images and 36 encoders, a substantial fraction of race and age gaps—and especially AUROC gaps in prior studies—was compatible with sampling variation, while merely making demographics decodable did not worsen the residual gap; entangling demographics with the disease direction did. It also finds image-text pretraining improves worst-group performance by about 0.05.

    The distinction between finite-sample subgroup differences and mechanistic representational bias is a useful, nonstandard reframing backed by unusually broad experiments, but the abstract does not establish that FRAME’s statistical assumptions generalize reliably enough for a strong recommendation.

  65. maybe AI / ML ▲ 266 score 4.4

    VBVR-Pro: A Scalable and Verifiable Suite for Native Visual Reasoning

    Junxiang Xu, Ruisi Wang, Fanyi Pu et al.

    The paper introduces a procedurally generated suite of 300 visual-reasoning tasks with deterministic task-specific scorers, intended to support training and reinforcement learning for models that reason through generated images and videos. Experiments report transfer to seven external benchmarks, improved post-RL performance, and a comparison of more than 30 image, video, and interleaved generation setups; video helps with persistent temporal state, while interleaved generation is more compute-efficient.

    The combination of scalable procedural tasks, verifiable rewards, and controlled study of visual reasoning substrates is a useful research infrastructure contribution, but the abstract gives no quantitative results and the core benchmark-and-RL framing is not clearly a major conceptual breakthrough.

  66. maybe AI / ML score 4.4

    ASIL: Replacing Screenshot-and-Click with Structured State and Semantic Actions

    Rui Xie, Lu Chen

    ASIL replaces screenshot-based GUI control with structured JSON state and executable semantic actions, using the deepest available interface for each application. Across 15 apps and 380 tasks, it reports over 80% success with fewer than five actions per task, far exceeding screenshot-and-click baselines and improving small-model performance through SFT and RL; however, it only matches an existing draw.io MCP interface on matched tasks.

    The paper presents a potentially important agent–software interaction abstraction with large reported gains over GUI control, but the results depend on bespoke application integrations and the abstract leaves open how much comes from privileged, app-native access rather than a generally deployable interface.

  67. maybe AI / ML score 4.4

    Manifold4D: Denoising on Point Cloud Rendered Manifolds for Video Re-shooting

    Yongqi Mao, Zijia Dai, Zhishuo Liu et al.

    Manifold4D uses a point-cloud render of the desired camera view by injecting it into the initial noise of a flow-matching video generator, rather than providing it as a competing conditioning input throughout denoising. This improves camera-trajectory control substantially—reported rotation-error reductions of 25–27% and translation-error reductions of up to 32%—while preserving video fidelity and remaining effective for out-of-distribution camera motion and corrupted renders.

    The one-time geometric initialization is a genuinely interesting alternative to persistent visual conditioning, with sizable reported gains and robustness beyond the training camera range, but the evidence is still limited to the abstract and a specialized video re-shooting setting.

  68. maybe AI / ML score 4.4

    SemKV: Semantic Mixed-Precision KV Cache Quantization Guided by the Quality Cliff for Long-Context LLM Inference

    Daeha Lee, Do-Hyung Kim, Jae-Hong Kim

    The paper argues that KV-cache quantization has a sharp model-quality cliff rather than a smooth degradation curve: for tested Llama-3.1-8B and Mistral-7B settings, quality remains statistically indistinguishable from FP16 above roughly 2.0–2.322 bits/value but collapses below it. SemKV preserves all tokens and uses mixed precision just above this cliff, obtaining 6.0× storage reduction without detectable quality loss; a distortion-optimized quantizer raises the reported operating point to 7.9×.

    The quality-cliff framing and simple above-cliff precision interpolation could be practically important for long-context inference, but the evidence is still limited to a small number of model families and deployment settings, making this worth checking rather than an obvious must-read.

  69. maybe AI / ML score 4.4

    Content Exploration Beyond the Feed: Creator Supply and the Shared Corpus

    Yuanyuan Shen, Yiren Yan, Wenjie Li et al.

    This paper shows that recommender exploration affects not only which videos viewers consume, but also how much creators produce, changing the future shared pool of content. Experiments on a major short-video platform find that more exploration increases creator participation and production, while viewer-side tests can miss or even cancel these longer-term corpus effects; a three-week experiment could identify only immediate feed effects, not the eventual total value.

    It offers an unusually important causal framing for recommender evaluation—creator supply and corpus turnover create delayed effects that standard A/B tests may systematically fail to measure—and supports it with multi-month platform experiments, though the abstract does not establish the generality or ultimate size of the effects.

  70. maybe AI / ML score 4.4

    Item-Mean Surrogates: Why Richer Persona Data Fail to Improve LLMs as Human Surrogates

    Daehwan Ahn, Chengfeng Mao, Dokyun Lee

    The paper evaluates whether richer persona information makes LLMs good substitutes for specific people, using over 400,000 participants and 6,000 survey or experimental outcomes. It finds that LLMs mainly reproduce item-level average human responses: after removing those averages, they explain only 3.05% of respondent-specific variation, versus 53.6% for human test–retest reliability, and persona enrichment or fine-tuning does not fix the gap. The authors frame this as “item-mean surrogacy,” arguing that the crucial predictable signal is person-by-item deviation rather than a stable persona trait.

    The large-scale decomposition separating item means, stable person effects, and item-specific deviations is a useful and non-obvious challenge to claims that richer personas can create individual-level LLM surrogates, but it is primarily a strong empirical diagnosis rather than a demonstrated new capability or method.

  71. maybe AI / ML score 4.4

    Adversarial Online Classification with a Preview

    Roi Livni, Sahil Singla

    This paper studies online classification when a random fraction of the entire labeled sequence is revealed in advance, while the rest still arrives in adversarial order. It shows that even a small preview can replace the usually much harder sequential complexity: for binary classes, the optimal excess loss scales as Θ(d/p + √(dT)), and it gives a multiscale “ChainedPrediction” algorithm achieving the sharp rate.

    The result identifies a non-obvious regime where random partial preview converts worst-case online learning from Littlestone-dimension behavior to classical VC-dimension behavior, with matching bounds and an explicit algorithm, though its impact is primarily theoretical.

  72. maybe AI / ML score 4.4

    The Price of Intelligence: A Quality-Adjusted Price Index for AI Services

    Louis Yiven Zhu

    The paper builds a quality-adjusted price index for AI inference using 21,024 posted-price observations, 3,208 models, 86 providers, and benchmark-derived latent quality scores. It finds that conventional matched-model measures understate price declines substantially: quality-adjusted prices fell 0.73 log points per year versus 0.10 without quality adjustment, while the effective price per completed task stopped falling because reasoning models consume more tokens. A preregistered audit also shows that benchmark contamination can materially alter economic conclusions even when model rankings remain nearly unchanged.

    This is a genuinely useful and non-obvious measurement framework for AI economics, backed by unusually broad public data and a preregistered audit, though its importance depends on the validity of benchmark-derived quality and task-completion assumptions.

  73. maybe Neuroscience score 4.4

    Selective gating of neural modulation through frequency- and behavior-dependent modes during cortical electrical stimulation

    Rembado, I., Lee, S. Y., Marks, L. et al.

    The study combines sinusoidal cortical stimulation with Neuropixels recordings from more than 2,700 neurons across 53 brain areas in behaving mice. It identifies two separable effects: sustained, frequency-dependent spike-phase entrainment across the brain, and a brief, local spike-rate response at high frequencies driven mainly by inhibitory-neuron recruitment; behavior gates the former but not the latter.

    The broad in vivo dataset and proposed separation of frequency-dependent, brain-wide entrainment from local inhibitory recruitment offer a genuinely useful mechanistic framework for neuromodulation, but the abstract gives no effect sizes or independent validation, so it falls short of strong.

  74. maybe Neuroscience score 4.4

    Neuronal primary cilia are not required for hippocampal circuit function or behavior in adult mice

    Eom, T.-Y., Bayazitov, I. T., Teubner, B. J. et al.

    The authors removed primary cilia from mature CA1 and CA3 hippocampal neurons in adult mice, either genetically over time or acutely with laser ablation. Across electrophysiology, two-photon imaging, synaptic-plasticity assays, and behavioral tests, cilia loss did not measurably impair neuronal function, hippocampal learning and memory, or anxiety-like behavior. The result suggests that, despite their importance during brain development, neuronal primary cilia may not be essential signaling hubs in the adult hippocampus.

    This is a well-supported negative result that challenges the assumption that neuronal primary cilia are broadly required for mature circuit function, though its significance is limited by the focus on adult mouse hippocampus and the tested behaviors.

  75. maybe Neuroscience picked score 4.4

    Shared and idiosyncratic coding regimes coexist in macaque IT

    Lu, Q., Xiong, X., Li, Y. et al.

    Using Neuropixels recordings from a macaque anterior IT face patch during viewing of more than 3,000 natural images, the authors compare individual neurons with population-level activity. They find that single neurons include sparse, heterogeneous, and sometimes reproducibly “feature-random” responses that are poorly captured by DNN feature spaces, while multi-unit activity shows the familiar dense, correlated, low-dimensional structure. The proposed interpretation is that shared population structure supports generalization, whereas idiosyncratic sparse responses expand the code and aid identification of individual images.

    The large-scale single-unit evidence for a coexistence of DNN-aligned shared coding and reliable, locally uncorrelated feature-random coding is a genuinely interesting challenge to treating population-level IT structure as representative of individual neurons, though the abstract does not provide enough quantitative behavioral or cross-condition validation for a strong verdict.

  76. maybe AI / ML ▲ 40 score 4.4

    Puro-2B: Poor Lab's Qwen2-1.5B Trained on RTX 5090 within $5090

    Kairong Luo, Jiarui Cui, Yaorui Yin et al.

    The authors train 2B-parameter language models from scratch on consumer RTX 5090 GPUs, using FP8 training, optimization and data-curriculum choices, and release the data, code, and weights. Their best model reportedly approaches Qwen2.5-1.5B performance for under $6.9K, while a fitted cost-scaling law estimates that roughly $4.4K could match Qwen2-1.5B; they also study how pretraining curricula affect post-training results.

    The credible open end-to-end demonstration of training a competitive small LLM for only a few thousand dollars is potentially important for reproducibility and access, though the abstract does not establish how broadly the recipe transfers or whether the cost comparison includes all relevant expenses.

  77. maybe AI / ML ▲ 25 score 4.4

    D$^3$-MOPD: Adaptive Dynamic Domain ScheDuling for Efficient Multi-Teacher Distillation

    Zechen Sun, Zhiwei Zhang, Fei Zhao et al.

    The paper dynamically reallocates on-policy distillation data across domains by monitoring each domain’s reverse-KL learning trajectory, rather than using a fixed mixture. On a Qwen3.6-35B-A3B student with four teachers, it reports closing 97% of the student–teacher gap versus 63% for fixed-mixture MOPD and achieving comparable peak performance with roughly 3× fewer rollout steps.

    The adaptive compute allocation and reported 3× rollout reduction are potentially useful, but the evidence is limited to one student/teacher setup and the core idea is a fairly natural convergence-aware scheduling extension.

  78. maybe AI / ML ▲ 4 score 4.3

    Skill Issue: Are Skills Language-Invariant in LLMs?

    Bobby Cheng, Adam Gaber, Zhengyuan Liu et al.

    The paper tests whether an LLM’s abilities change when it interacts in different languages, while holding the model, game, rules, and action space fixed. Using multilingual self-play across eight languages, six games, and three open-weight models, it finds substantial language-dependent differences in strategy, invalid actions, and playing strength; using a different language for intermediate reasoning sometimes recovers performance.

    The self-play design is a useful way to separate language-dependent skill from ordinary knowledge and benchmark effects, and the finding that reasoning language can recover some capability is non-obvious, but the evidence is limited to three models and six text games.

  79. maybe AI / ML score 4.3

    WnW: Waxing-and-Waning KV Cache for Long-Form Speech LLMs

    Yiming Yao, Chenyang Lyu, Xuanfan Ni et al.

    WnW is a KV-cache management method for long-form speech LLMs that adapts retrieval during decoding rather than permanently discarding audio tokens after prefill. It uses different attention heads as persistent anchors, CPU-backed recallable storage, or fixed compressed storage; on two 3B speech/omni models, it reportedly retains near-full-cache accuracy with only 20% of audio tokens on GPU and little added decode overhead.

    The weak overlap between prefill and decoding attention, and the resulting waxing-and-waning cache design, are a non-obvious and potentially useful direction for reducing long-context speech inference memory, but the abstract lacks detailed accuracy, latency, and scaling results needed for a stronger verdict.

  80. maybe AI / ML score 4.3

    JANUS: Online Jacobian-Aligned Infill for Black-Box Optimization

    Hongyuan Yu, Pufan Xu, Jiaojiao Yi et al.

    JANUS adds a geometry-guided infill mechanism to existing population-based black-box optimizers. It estimates a local Jacobian from evaluations already collected, using it for both a damped Gauss–Newton exploitation step and a geometry-preserving exploration step, without extra evaluations or offline training. Across BBOB, multi-objective, and UAV path-planning tests, it reports substantial gains over CMA-ES and other baselines, including improvements on most high-dimensional BBOB functions.

    The online extraction of directional geometry from rank-based optimizer traces and its dual use for exploitation and exploration is a meaningful idea, with unusually broad and large reported gains, but the abstract provides no ablations or enough detail to assess whether the striking benchmark improvements are robust.

  81. maybe AI / ML score 4.3

    ProxyFormer: A Dual-Stream Proxy Architecture for Ultra-Long Context and High-Resolution Generation

    Zhongpan Tang

    ProxyFormer uses a persistent local token stream plus a much smaller proxy-token stream: local features are compressed into proxies for global attention, then expanded back into the local representation. The authors claim this reduces both attention and KV-cache costs, enabling training at roughly 0.7M tokens on a 16GB GPU and retaining 92–95% multi-needle retrieval accuracy at 1M tokens despite training on 64K contexts; initial image-generation tests are also reported.

    The iterative local-to-proxy-to-local design and claimed million-token retrieval with a small training window are potentially important, but the abstract provides limited comparisons and only preliminary evidence, while related compressed-attention and latent-token architectures already exist.

  82. maybe AI / ML score 4.3

    Scaling Reinforcement Learning for Diffusion Models via Velocity Matching

    Jaemoo Choi, Wei Guo, Yuchen Zhu et al.

    The paper proposes reward-based velocity matching (RVM), which fine-tunes diffusion and video-generation models directly in their learned velocity-field representation instead of estimating trajectory or endpoint likelihoods. It reportedly matches or exceeds policy-gradient-style methods at much lower training cost, and a motion-aware reward improves video dynamics while also raising VBench scores; the framework also unifies several recent methods as special cases.

    Direct, trajectory-free reward optimization in the native velocity field could substantially simplify and scale diffusion-model preference tuning, but the abstract gives no quantitative results or breadth details sufficient for a strong verdict.

  83. maybe AI / ML score 4.3

    Renormalization Group Flow Matching for Scalable Local Generative Modeling

    Kanta Masuki, Yuto Ashida

    The paper uses renormalization-group transformations to build a flow-matching generator that produces coarse, long-range structure first and refines it into local details. It theoretically argues that the required velocity field can be approximated with patches only logarithmic in image/system size, while still preserving distant correlations, and reports better long-range coherence than ordinary local flow matching on 1D distributions and FFHQ images.

    The combination of exact RG flows, locality, and flow matching is a genuinely interesting route to scalable globally coherent generation, but the abstract provides limited quantitative evidence beyond representative 1D tests and qualitative FFHQ comparisons.

  84. maybe Neuroscience score 4.3

    Primate vision reveals a missing principle for robust dynamic AI

    Matteo Dunnhofer, Christian Micheloni, Kohitij Kar

    The authors compare human and macaque visual responses with representations from image and video models, testing whether models can recognize objects when their appearance changes but motion remains informative. Humans and macaque IT were robust, while most video models were not; predictive world models performed best on both cross-appearance generalization and neural similarity, although none reproduced the brain’s progression from appearance-based to motion-invariant coding.

    This offers a potentially important brain-inspired principle—progressively integrating motion into object representations—and links it to predictive world modeling, but the abstract lacks quantitative detail and enough evidence to justify a stronger verdict.

  85. maybe AI / ML score 4.3

    LUCAID: Agentic Multimodal AI for Lung Cancer Precision Pathology

    Marie-Lisa Eich, Kai Standvoss, Timo Milbich et al.

    LUCAID is an agentic multimodal pathology system that combines nine modules covering much of the lung-cancer diagnostic workflow, including tumor analysis, biomarker scoring, and report generation. It achieved F1 scores of 0.82–0.95 against expert annotations and 93.0% concordance with an expert-adjudicated standard in prospective validation, versus 68.3–81.1% for five thoracic pathologists.

    The broad, interactively queryable workflow and prospective comparison with practicing pathologists are substantially more interesting than another single-task pathology model, but the abstract omits cohort size, study design, and module-level clinical results, so the striking claims need verification.

  86. maybe Robotics score 4.3

    Learning to Act While Waiting: RL Finetuning of Generalist Robot Policies Under Inference Latency

    Brian Zhu, Momen Khalil, E Harrison et al.

    The paper develops ARLI, an RL fine-tuning framework for large generalist robot policies whose slow inference causes pauses and changes the effective environment dynamics. It combines asynchronous action execution with state information about committed actions and a mid-inference observation to make delayed control more nearly Markovian; experiments in simulation and on real manipulation tasks report successful fine-tuning where standard RL fails, sometimes matching no-latency performance.

    The potentially important contribution is treating inference latency as part of the learned robot-control problem and restoring usable RL structure with intermediate information, with both simulated and real-robot evidence, though the abstract gives no quantitative scale or task details to justify a stronger verdict.

  87. maybe AI / ML score 4.3

    A tale of perfect fit and phantom optima: how data-driven models can fail in real-time optimization

    Prithvi Dake, Rahul Bindlish, James B. Rawlings

    The paper tests whether models that predict plant measurements accurately can still support reliable real-time economic optimization. On a vinyl acetate process, both a physics-informed hybrid model and a neural ODE fit the data well but produce many spurious optima, and stochastic training can worsen the optimization result even from a favorable initialization. The authors propose evaluating learned models on whether they recover decision-relevant optima, not just predictive accuracy.

    It gives a concrete, decision-focused demonstration that excellent predictive fit and stable model identification can coexist with badly wrong optima, though the evidence is based on one process benchmark and reflects a broader known concern about model error under optimization.

  88. maybe AI / ML score 4.3

    Joint Optimization of Tool Creation and Use for Large Language Model Agents

    Zhi Rui Tam, Chieh-Yen Lin, Yun-Nung Chen et al.

    SMITH jointly trains a language model to create tools and then use them, rather than treating tool writing as a separate inference-time step. A 4B Qwen3 model trained with separate rewards for schema validity, code correctness, and task outcomes achieves 79.8% across 13 procedural reasoning tasks, beats a much larger tool-writing model, and improves results on TabMWP-Hard and GQA without training on those modalities; its tools also help other models.

    The coupled training of tool creation and invocation is a meaningful agent-training direction, and the small-model gains and transfer to other models are notable, but the evidence is mainly on a limited set of exact-verifier tasks and the abstract does not establish broad real-world generality.

  89. maybe AI / ML picked score 4.3

    Reading Is Not Using: Retrieval, Judgment, and the Design of AI Financial Research Workflows

    Miao Liu, Zhizhe Liu

    The paper shows that language models can retrieve a financial risk disclosure accurately from very long context yet fail to let it affect their investment judgment. This retrieval-to-judgment gap appears across models and tasks, while targeted structured restatements near the decision can restore the information’s influence; ordinary chunk-and-summarize workflows instead lose it.

    It identifies and causally probes a non-obvious failure mode in long-context reasoning—retrieval without use—and suggests workflow design, not just model scaling, as the remedy, though the abstract provides limited quantitative detail.

  90. maybe AI / ML picked score 4.3

    Tunable Tool-Call Rates in LLM Agents via Representation Steering

    Yuqi Chen, Vincent Siu, Yang Liu et al.

    The paper finds that an LLM’s tendency to call tools can be controlled at inference time by adding a single learned-free linear direction to its residual stream. Varying the intervention smoothly changes tool-call rates from nearly 0% to over 90%, and the authors report that increased calling selectively helps on questions the model cannot answer itself, nearly doubling open-domain QA accuracy; the effect transfers across tools and several model architectures.

    Inference-time, training-free control of tool-use propensity with apparent cross-tool and cross-architecture transfer is a genuinely useful idea, but the abstract gives limited detail about evaluation breadth and whether the striking accuracy and selectivity claims hold beyond the reported setup.

  91. maybe AI / ML score 4.3

    Spectral Allocation: Why Muon Outperforms Adam, and How to Improve Muon

    Xiaodong Wu, Wenyi Yu, Chao Zhang et al.

    The paper analyzes why Muon appears to outperform Adam by estimating the loss-optimal step size across singular directions of momentum buffers. It finds a stable spectrum with a sensitive, volatile head and a larger-step-tolerant bulk, then uses this profile to create SAMuon, which scales these regions differently and reportedly needs 13.3–24.0% fewer tokens than Muon across 124M–1B-parameter modded-nanogpt models; a cheaper variant retains most of the gain.

    The combination of a concrete spectral explanation for Muon and a low-state, low-overhead optimizer with sizeable reported token savings is notably interesting, but the evidence is still limited to one model family and benchmark setting, so it does not yet warrant strong.

  92. maybe AI / ML picked score 4.3

    SpeechGym: An Audio-Native Gym for Training Voice Agents via Reinforcement Learning

    Jiajun Fan, Jingyuan Li, Prashanth Gurunath Shivakumar et al.

    SpeechGym is an audio-native reinforcement-learning environment where two omni-modal models interact directly in speech while performing tool-use tasks, avoiding separate ASR/TTS systems and API costs. The authors find that failures often come from misheard tool arguments and unsafe behavior under conversational pressure, and show that per-turn rewards for successful tool calls overcome sparse outcome rewards; training more than doubles success on an independent voice benchmark and substantially improves an open-weights model’s ranking.

    The combination of end-to-end audio interaction, trainable local tool-use environments, and process rewards for speech-specific failures is genuinely relevant, but the abstract omits enough quantitative and experimental detail to justify a strong verdict.

  93. maybe BCI picked score 4.3

    Robust Neural Stimulation Response Modeling Through Meta-Learning and Pretraining

    Matthew J Bryan, Daniel C Muir, Felix Schwock et al.

    The paper pretrains temporal-basis models across neural-stimulation sessions using a MAML-style meta-learning method, then adapts them to new sessions with limited calibration data. On 40 optogenetic stimulation sessions from two non-human primates, this reduced severe forecasting failures from 16/40 to 1/40 at a fixed calibration size and cut calibration requirements by roughly 50–90% at matched accuracy. The main new result is evidence that cross-session structure can make closed-loop stimulation models substantially more robust and sample-efficient.

    The large reduction in catastrophic failures and calibration burden is genuinely relevant to closed-loop neural stimulation, but the evidence is limited to 40 sessions from two animals and one stimulation/modeling setting, so it is promising rather than yet a field-changing result.

  94. maybe AI / ML score 4.3

    Cross-Lingual Alignment Without Joint Training: Do Monolingual Language Models Converge on Universal Representations?

    Ej Zhou, Suchir Salhan, Catherine Arnett et al.

    The paper studies whether separately trained monolingual language models develop comparable internal representations without shared parameters, data, or alignment objectives. Across several model families, it reports that hidden-state geometries become alignable with a single Procrustes rotation, and that rotated activations can transfer factual information between languages during inference. The result suggests that some cross-lingual structure arises from shared linguistic content rather than joint training.

    The combination of cross-model representational alignment and causal activation patching is a genuinely non-obvious result with implications for model stitching, but the abstract gives too few quantitative details to justify a stronger verdict.

  95. maybe Robotics score 4.3

    Ultra Low-Power, Lightweight, Probabilistic RSS-Based Path Reconstruction: A System for Landscape-Scale Bee Tracking

    Christopher J. Noroozi, Joseph L. Woodgate, Michael Mangan et al.

    The paper presents a localization system for 38 mg flying receivers that avoids GNSS by using a small number of RSS measurements from rotating high-gain transmitters, probabilistic angle inference, and Gaussian-process path reconstruction. It reports roughly 15 m tracking accuracy at under 180 μW, improving to about 10 m at under 600 μW, and demonstrates tracking bumblebee flights over ranges up to 300 m.

    The combination of extremely low payload weight, sub-milliwatt power, landscape-scale range, and usable trajectory reconstruction is a meaningful real-world capability for insect tracking and other tiny autonomous devices, though the abstract does not establish broad field validation or comparisons in enough detail for a strong verdict.

  96. maybe AI / ML score 4.3

    Compositional Online Learning for Semantic Data Processing Systems

    Paweł Liskowski, Fuheng Zhao, Benjamin Han et al.

    The paper proposes composing several online-learning components at the boundary of expensive LLM calls, using the call’s latency to hide CPU-side updates. In a production Cortex AISQL case study, memoization, learned filter ordering, and cascade routing combine to reduce LLM-related query cost by roughly 8× on a representative conjunction-filter workload, versus an 11.4× idealized bound.

    The latency-aware composition principle and conditional cost decomposition are a useful systems framing with a substantial reported speedup, but the evidence is centered on one production case and workload, so it does not yet merit strong.

  97. maybe AI / ML picked score 4.3

    LLMs Can Design Near-Optimal OR Algorithms

    Jackie Baek

    This paper tests whether LLMs can generate solutions and reusable algorithms for inventory control, queueing-network control, and assortment optimization. With only an untuned prompt and a fixed Python-compute budget, the best model reportedly matches or beats established methods, even when it must produce the algorithm before seeing test instances; performance also improves substantially across recent model generations.

    The level-2 result—an LLM producing a fixed, instance-independent algorithm competitive with specialized OR methods—is a genuinely interesting capability claim, but the abstract gives no quantitative margins or details about evaluation breadth, making it insufficiently supported for a strong verdict.

  98. maybe AI / ML score 4.3

    Variable-Granularity Tokenization for High-Resolution Object Detection

    Khayrul Islam

    VGTok replaces the usual uniform image patch grid with a training-free, variable-granularity partition: regions that look like separable, two-dimensional objects receive finer tokens, while clutter stays coarse. In aerial detection, it reportedly preserves or improves tiny-object accuracy while using up to 3.1× less encoder compute, transfers unchanged between VisDrone and AI-TOD-v2, and improves a frozen detector checkpoint without retraining.

    The pre-encoder, hand-crafted adaptive tokenization and strong compute/accuracy claims are genuinely interesting, but the evidence is confined to aerial detection and the unusually large state-of-the-art gains need careful validation against fair baselines and implementation details.

  99. maybe AI / ML score 4.3

    mmIR: Frequency-Space Inverse Rendering for 3D Millimeter-Wave Radar ADC Synthesis

    Adnan Armouti, Yixuan Gao, Rajalakshmi Nandakumar

    mmIR fits a differentiable, physics-based FMCW radar renderer to real radar captures, using LiDAR meshes as geometric scaffolding and optimizing materials, normals, and antenna patterns. It then renders coherent ADC signals from dense virtual apertures, achieving much higher range–azimuth correlation than Sionna-RT and transferring across radar hardware; the authors also demonstrate single-frame 3D occupancy from a 100×100 virtual array.

    The combination of LiDAR-assisted inverse rendering, phase-coherent multi-bounce radar physics, and cross-sensor virtual-aperture synthesis is a nontrivial direction with potentially large data and hardware implications, but the evidence is limited to 13 scenes and depends on LiDAR geometry.

  100. maybe Robotics score 4.3

    Blind Dexterity: Whole-Body Humanoid Manipulation via Pure Proprioception

    Aditya Bhatt, Oleg Kaidanov, Puze Liu et al.

    The paper trains a Unitree G1 humanoid to perform several contact-rich tasks—including walking under pushes, trapping a soccer ball, finding and lifting a suitcase, and mounting a skateboard—using only joint proprioception. It argues that changes in joint encoder signals during compliant contact provide an implicit whole-body tactile channel, and shows that object state becomes more predictable from short proprioceptive histories after active contact.

    The combination of real-humanoid whole-body manipulation, active probing, and encoder-derived contact information is genuinely non-obvious, but the abstract gives limited quantitative evidence and the demonstrations may not establish broad generality.

  101. maybe AI / ML picked score 4.3

    The Emergent Symbolic Structure of Artificial Neural Networks

    R. Thomas McCoy, Paul Soulos, Tal Linzen et al.

    The paper argues that neural networks may implement symbolic structure internally, despite using continuous vector representations. It replaces the representation-generating process of small list-manipulation networks and LLMs with explicit symbolic equations, reportedly preserving much of their behavior across arithmetic, logic, code, and language; targeted interventions further alter model behavior in ways predicted by those structures.

    The cross-domain claim that LLM representations can be compressed into behaviorally meaningful symbolic structures would be important for understanding neural computation, but the abstract gives no quantitative details on approximation quality, model scale, or intervention strength, so it does not yet merit a strong verdict.

  102. maybe Neuroscience score 4.3

    Spinotrode: long-term intraspinal electrophysiological recordings to unravel dorsal horn neuron dynamics in behaving mice

    Viellard, J., Brochoire, L., Janusz, M. et al.

    The authors developed a vertebral “Spinotrode” implant that records single-neuron activity from both dorsal horns of the spinal cord in freely moving mice for weeks, without obvious movement impairment or tissue damage. In behaving animals, it identified behavior-linked dorsal horn neuron types, activity during locomotion beyond simple touch or proprioception, and contralateral sensory responses during nociceptive reflexes—findings that may not appear under anesthesia.

    This is a genuinely useful enabling technology paired with a potentially assumption-challenging result about bilateral spinal processing in natural behavior, but the abstract provides limited quantitative evidence and validation details.

  103. maybe Neuroscience score 4.3

    Cerebellar influences on neocortical development in humans and mice

    Gaiser, C., Germain, N., Jacobs, T. et al.

    This study combines longitudinal brain-imaging data from two child cohorts with cerebellar lesion experiments in mice to examine whether the cerebellum helps shape developing neocortex. It reports region-specific associations and cross-species evidence suggesting that early cerebellar development may influence cortical maturation, challenging the view that the cerebellum is merely a late-maturing subordinate structure.

    The potentially causal role assigned to the developing cerebellum is a non-obvious challenge to a standard developmental assumption, but the abstract gives too few details to judge the strength of the human longitudinal and mouse-lesion evidence.

  104. maybe Neuroscience score 4.3

    Breakdown in the synaptic vesicle cycle defines early and reversible cortical pathogenesis in ALS

    Laszlo, Z. I., Sanchez-Avila, A., McFarlane, A. et al.

    The authors examined visual cortex, which is affected relatively late in ALS, to identify earlier synaptic changes in human disease. They found presynaptic terminal shrinkage and disruption of synaptic-vesicle cycling despite preserved neurons, and showed that suppressing TDP-43 in a mouse model restored related molecular changes. The main claim is that early synaptic dysfunction may be a distinct and reversible phase of ALS progression.

    The combination of human synaptoneurosome proteomics with a TDP-43 model and apparent reversal of presynaptic defects is a meaningful, non-obvious result, but the abstract does not establish that restoring vesicle-cycle proteins rescues synaptic function or disease outcomes.

  105. maybe Neuroscience score 4.3

    Transcription-specific synaptic plasticity mediates engram computations

    Tome, D. F., Meng, M., Sun, X. et al.

    The study argues that different transcriptionally defined engram populations use different forms of synaptic plasticity in the dentate gyrus: Fos-linked ensembles modify feedforward excitatory synapses, while Npas4-linked ensembles modify recurrent inhibitory synapses. A spiking-network model predicted that disrupting these pathways would selectively impair memory generalization or discrimination, and acute Fos or Npas4 deletion in mice reportedly confirmed those predictions during contextual fear conditioning.

    It offers a potentially important causal link between transcription-defined engrams, synapse-specific plasticity, and distinct memory computations, supported by both modeling and targeted mouse experiments, though the abstract provides limited detail on effect sizes and experimental breadth.

  106. maybe Neuroscience score 4.3

    Signatures of Hebbian plasticity in the nanoscale morphodynamics of cortical spines

    Soltanipour, M., Nagel, A., Willig, K. I. et al.

    Using long-term in vivo nanoscopy, the authors model how cortical dendritic spines change shape and organization over time. They report that coordinated head-and-neck remodeling—apparently a small part of overall spine motion—can produce large fluctuations in instantaneous synaptic strength, while synapse-specific long-term strength persists because spines explore only a restricted region of morphospace. The work argues that circuit models should represent richer spine dynamics than head size alone.

    This is a potentially important, non-obvious link between nanoscale spine dynamics and synaptic-strength variability, but the abstract provides limited quantitative and validation detail for a stronger recommendation.

  107. maybe Neuroscience picked score 4.3

    Unconstrained naturalistic human brain imaging and decoding with a fully wearable high-density optical system

    Hamic, W. T., Fehner, W., Fogarty, M. et al.

    The authors built a fully wireless, battery-powered whole-head diffuse optical tomography system with nearly 3,000 source–detector measurements and high dynamic range. It produced localized, repeatable brain responses during standard tasks, movie viewing, and unconstrained piano performance, decoding song segments at 71.1% accuracy versus 12.5% chance. The main advance is bringing relatively high-fidelity, whole-head optical imaging into mobile, naturalistic behavior rather than restricting participants to a scanner or tethered setup.

    A credible step toward high-resolution naturalistic neuroimaging and real-world neural decoding, though the abstract provides limited detail on comparisons, participant scale, and robustness beyond the demonstrated paradigms.

  108. maybe Neuroscience score 4.3

    RGC-specific reversal of lipid peroxidation drives neuroprotection and vision restoration in optic nerve ischemia by targeting GPX4

    Yang, M., Pan, J., Modgil, S. et al.

    This study argues that lipid peroxidation, rather than ischemia or oxidative stress in general, is a key cause of retinal ganglion cell and axon loss in NAION. In human tissue and a mouse model, RGC-specific or mitochondria-targeted GPX4 protected cells, mitochondrial function, axonal transport, visual projections, and vision after ischemic injury; the clinically tested GPX mimetic ebselen reproduced the protection in mice.

    The potentially important result is a specific, mitochondria-centered lipid-peroxide mechanism with both genetic and pharmacological rescue, but the evidence described is still primarily preclinical and the abstract gives no quantitative effect sizes or independent clinical validation.

  109. maybe AI / ML ▲ 33 score 4.3

    Revisiting Local Context for Long-Horizon Streaming 3D Reconstruction

    Jiarong Han, Jincheng Xiong, Yuzhou Liu et al.

    ABot-Recon argues that long-horizon streaming 3D reconstruction does not require persistent long-range learned memory. It uses only the previous 11 frames to predict a current-frame point map and adjacent-frame pose, then composes these locally equivariant predictions over time; a temporal rotation refiner and composition-aware loss reduce drift. On Oxford Spires, it reports about 40% lower absolute trajectory and relative rotation errors than the prior best method.

    The interesting result is that strictly local learned context can outperform long-range memory for long-horizon stability, but the abstract provides limited evidence beyond one headline benchmark and the approach appears relatively targeted rather than broadly transformative.

  110. maybe AI / ML ▲ 15 score 4.3

    Meta$^n$: Recursive Self-Improvement through Emergent Depth

    Zae Myung Kim, Young-Jun Lee, Seungyeon Jwa et al.

    Meta^n applies the same meta-operation repeatedly to the traces and code produced by lower-level solver layers, building an expanding stack of strategic preprocessing and reusable tools rather than modifying the optimizer itself. Across two LLM backbones and eight benchmark families, it reportedly beats prior self-improving agents; its most notable result is obtaining a nonzero ARC-AGI-2 score, with ablations attributing much of the benefit to information passed between layers and emergent layer specialization.

    Recursive self-improvement through a fixed operator and growing solver context is a potentially useful new framing, but the abstract provides no quantitative results, making the broad performance claims difficult to assess.

  111. maybe Tech ▲ 13 score 4.2

    A Programming Paradigm for Spatiotemporal Composability

    Yifan Shi, Wei Zhang, Tianyi Cui

    The paper proposes a programming model for dynamically adding and removing software components. It combines reversible effects, which can undo a component’s side effects, with reactive coeffects, which track dependencies and trigger component activation or deactivation; the authors formalize the model and implement it in the Cordis framework with hot module replacement and configuration reconciliation.

    The unified treatment of reversible effects and reactive dependencies is a genuinely interesting PL framing for dynamic composition, but the abstract provides little evidence beyond formalization and a prototype, with no demonstrated scale, performance, or compelling applications.

  112. maybe AI / ML score 4.2

    Don't Repeat Yourself: Stopping Verbatim Loops at Sampling Time

    Philipp Emanuel Weidmann, Allen Roush, Judah Goldfeder et al.

    The paper introduces DRY, a sampling-time penalty that suppresses a token only when it would continue an exact repeated span from earlier context, rather than broadly penalizing repeated tokens or n-grams. Across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study, it reportedly cuts suffix-extension loops by 47% while preserving benchmark performance and improving lexical diversity; the method has also been integrated into several open-source inference frameworks.

    This is a targeted and practically deployable solution to a real generation failure mode, with unusually broad scale and intervention-matched evidence, but it is a focused decoding improvement rather than a fundamentally new modeling direction.

  113. maybe Robotics score 4.2

    Pointing-VLA: Typed Spatial Grounding Interfaces for Vision-Language-Action Manipulation

    Xiwen Chen, Zelin Li, Zhiruo Zhou et al.

    Pointing-VLA replaces text-serialized spatial actions with typed hidden-state outputs: point coordinates, object-functional grounding heatmaps, and visual trajectories. It assigns different readouts to manipulation stages such as PICK and PLACE, and reports better robot success, much faster decoding, and transfer across VLA/action-policy systems, including an increase from 52.7% to 80.7% real-robot success in one integration.

    The typed, stage-specific spatial interface is a meaningful alternative to autoregressive coordinate/action tokens, backed by sizeable speed and real-robot success gains, but the evidence is still centered on a small set of manipulation tasks and several system-specific integrations.

  114. maybe AI / ML score 4.2

    Partial Optimal Transport on the Circle for All Transported Masses in O(N log N)

    Soheil Kolouri

    The paper develops an exact algorithm for partial optimal transport on a circle that computes the optimal cost, active sets, plans, and a jointly optimal cut for every transported mass in O(N log N), rather than rerunning a line algorithm for each possible cut. It reports large practical speedups and better robustness to occlusion and clutter in circular and spherical distribution-matching tasks, though the method is specialized to these geometric settings.

    The combination of an all-masses partial-OT profile with an efficient global treatment of the circular cut is a substantive algorithmic advance, supported by complexity claims and several empirical comparisons, but its impact is mainly within specialized optimal-transport applications.

  115. maybe Robotics score 4.2

    GlanceWAM: Sparse Test-Time Imagination for World-Action Models

    Linhan Wang, Zijian An, Mingyuan Zhang et al.

    GlanceWAM runs a video-based robot policy by generating a single future visual frame asynchronously, while the action head operates continuously in latent space at a 48 ms control interval. On RoboCasa it reaches 72.2% success versus 67.1% for synchronous imagination and 64.4% without imagination, while running 24× faster; it also reports 99.0% on LIBERO.

    The asynchronous, stale-tolerant use of sparse visual imagination directly addresses a major latency–performance trade-off, with substantial benchmark and speed gains, but the evidence is limited to simulated benchmarks and an abstract-level report.

  116. maybe AI / ML score 4.2

    Evidence Blindness in Direct Corpus Interaction: Persistent Navigation with AtlasNav

    Hongyu Guo, Zhiyu Zheng, Zhao Cao

    The paper argues that direct corpus access is not enough for agentic search: under finite interaction budgets, agents can miss the relevant document, fail to open it, or miss the decisive passage. AtlasNav builds a persistent, query-navigable multi-view “Corpus Atlas” once, rather than reconstructing a workspace for every query, and reports 92.05% strict accuracy with 30.21% lower online inference cost than a dynamic-workspace baseline, plus results across scaling and enterprise settings.

    The persistent corpus-structure formulation and explicit stage-wise account of evidence loss are potentially important, with substantial reported accuracy and cost gains, but the abstract does not establish how much of the result comes from representation design versus benchmark-specific indexing and engineering.

  117. maybe Robotics score 4.2

    LM-X: Explainable Action Modeling with Progress, Event, and Uncertainty Prediction for Generalist Robot Manipulation

    Jin Lou, Zhiyuan Jing, Andong Chen et al.

    LM-X augments a vision-language-action robot policy with three online predictive signals: task progress, the next semantic event, and uncertainty in the action itself. These signals are trained explicitly and used to condition control, yielding large reported gains over an action-only model and GR00T across randomized simulation tasks and seven real-robot tasks, while uncertainty correlates with hesitation and oscillation.

    The explicit, jointly conditioned progress/event/uncertainty formulation is a meaningful alternative to opaque short-horizon action prediction, and the reported gains span many tasks and real robots, though the abstract does not establish how much comes from extra supervision, scale, or the specific architecture.

  118. maybe AI / ML ▲ 172 score 4.2

    VoiceMem: Streaming Dual-Brain Memory for Real-Time Interaction

    Zhifei Xie, Jiaqi Lang, Ze An et al.

    VoiceMem proposes a streaming memory system for duplex speech models, separating factual/user information from emotional and persona-related state, with parallel retrieval and interchangeable backends. It reports better retrieval and persona-benchmark scores than prior systems and 134 ms retrieval latency, but the abstract gives limited detail about the architecture, baselines, and real-world evaluation.

    The combination of low-latency streaming memory with affective/persona memory is relevant to real-time voice agents, but the “dual-brain” framing appears largely architectural packaging and the reported gains are not yet clearly enough supported to warrant a strong recommendation.

  119. maybe AI / ML score 4.2

    Trajectory-Level Speculative Decoding for Diffusion Language Models

    Tianxiang Pan, Baitao Gong, Mo Guang et al.

    The paper extends speculative decoding to diffusion language models by predicting and verifying entire denoising trajectories, including token positions and unmasking order, rather than a left-to-right token sequence. It reports 30–40% fewer denoising iterations, 7–14× speedup over vanilla diffusion LMs, and 1.3× over Fast-dLLM, with under 1% accuracy change on reasoning and code benchmarks.

    This is a genuinely tailored decoding framework for a key diffusion-LM bottleneck with substantial reported speedups, but the evidence is limited to abstract-level claims and the gains over the strongest stated baseline are comparatively modest.

  120. maybe AI / ML score 4.2

    Enforcing Dirichlet Boundary Conditions in Operator Learning

    Andrew M. Stuart, Margaret Trautner

    The paper introduces a neural-operator architecture whose outputs satisfy homogeneous Dirichlet boundary conditions by construction, using finite spans of Laplacian eigenfunctions at every layer. Unlike prior approaches, it supports arbitrary bounded Lipschitz domains and discretizations rather than requiring smooth boundaries, regular grids, or box-shaped geometries; the authors prove universal approximation and test it on Darcy flow and Helmholtz problems.

    The combination of exact boundary enforcement, general geometries and meshes, and a universality proof is a substantive advance for PDE operator learning, though the abstract provides limited evidence of a large practical performance or efficiency gain.

  121. maybe Robotics score 4.2

    CLAP: Cross-Embodiment Video World Models are Zero-Shot Physical Simulators

    Kechen Liu, Ola Shorinwa

    CLAP trains action-conditioned video world models across humans and multiple robot embodiments, using latent actions and later grounding them in end-effector actions. The authors claim that pretraining on heterogeneous unlabeled videos matches or exceeds single-robot models and enables zero-shot deployment, with further gains from few-shot adaptation across several robot platforms.

    Cross-embodiment world-model pretraining from internet-scale video is a genuinely important direction, but the abstract gives no quantitative results or details sufficient to validate the strong zero-shot and physical-simulation claims.

  122. maybe AI / ML score 4.2

    Report Supervision

    Pedro R. A. S. Bassia, Wenxuan Li, Jakob Wasserthal et al.

    R-Super uses radiology reports as weak supervision for 3D tumor segmentation, designing losses that encourage predicted tumors to match the reported counts, sizes, and locations. Across kidney and pancreatic CT data, including up to 41,418 report-linked scans and 3,488 masked scans, it reportedly improves external detection F1 and segmentation Dice by as much as 15% over mask-only training, especially when masks are scarce.

    This is a concrete and potentially scalable way to turn routinely produced reports into supervision for an otherwise mask-limited task, with substantial gains shown across data regimes, though the contribution is confined to medical tumor segmentation and needs scrutiny of report quality and external validation details.

  123. maybe AI / ML score 4.2

    Nested Byte-Level Vocabularies Are Cheap to Deploy and Expensive to Share: A Pre-Registered Negative Result

    Christos Koutsiaris

    This preregistered study tests whether one byte-level BPE language model can support multiple vocabulary sizes simply by slicing its embedding and output layers. Slicing is exactly equivalent and removes 66% of deployed weights, but the shared model is modestly worse than separately trained specialists; multi-vocabulary training nevertheless substantially improves robustness to typographical noise. The experiments suggest that the robustness comes from exposure to multiple granularities, not from the conditioning token, and that sharing incurs a penalty related to how often each vocabulary size is trained.

    The preregistered negative result cleanly separates deployment savings, quality loss, and robustness benefits in nested-vocabulary models, with unusually strong evidence for a practical tradeoff rather than another claimed win.

  124. maybe Robotics score 4.2

    GeoFF3D: Coordinate-Anchored Feed-Forward Reconstruction for Large-Scale UAV Mapping

    Xiang Yang, Yongli Wang, Yunsheng Zhang et al.

    GeoFF3D reconstructs large UAV image collections in a shared, gravity-aligned metric coordinate frame rather than independently normalized local frames. It combines georeferenced camera translations and optional geometric priors with overlapping spatial chunks and hierarchical aggregation, reporting substantially better reconstruction quality than prior feed-forward and SLAM-style baselines while processing 2,000 images in about five minutes.

    The coordinate-anchored formulation and scalable aggregation address a real limitation of feed-forward 3D reconstruction, with sizable results across multiple aerial sequences, but the contribution is specialized to UAV mapping and the abstract does not establish broad generality beyond this setting.

  125. maybe Robotics score 4.2

    DriftingVLA: Native One-Step Vision-Language-Action Generation via Per-Dimension Temporal Drifting

    Yuxuan Gao, Shiqi Zhang, Yedong Shen et al.

    DriftingVLA trains a vision-language-action model to generate an entire robot action chunk in one forward pass instead of iteratively refining it with a flow-based policy. Its main idea is to apply the training objective separately to each action dimension while still producing the joint action together, yielding reported gains on LIBERO, RoboTwin, and six real-world single- and dual-arm tasks, along with a 3.36× generation speedup.

    Native one-step action generation with substantial measured latency reduction and real-robot evaluation is a meaningful advance, but the abstract does not establish whether the method generalizes beyond the tested tasks or whether the comparison baselines are equally optimized.

  126. maybe Robotics score 4.2

    Learning Agile Perceptive Traversal of Sparse 3D Structures for Humanoids

    Efe Ongan, Chong Zhang, Boyang Sun et al.

    The authors train a humanoid to use head-mounted lidar and recurrent reinforcement learning to jump onto, traverse, and jump off sparse monkey-bar structures. A phase-based teacher–student system and hardware-aware simulation enabled successful full-sequence traversal in 14 of 15 hardware trials, with speeds up to 0.5 m/s; the same lidar encoder also supports ducking under very thin overhead obstacles.

    This demonstrates a substantial real-world humanoid capability—perceptive, agile traversal of sparse structures—but the approach is a specialized combination of lidar encoding, privileged RL, and hardware engineering rather than an obviously general new paradigm.

  127. maybe AI / ML score 4.2

    FoundYou: A Unified Model for Personalized Segmentation and Retrieval

    Gabriele Trivigno, Marcos Alfaro, Claudia Cuttano et al.

    FoundYou uses a frozen SAM 2-small backbone to treat personalized segmentation and image retrieval as the same cross-image object-matching problem. It adds only 5.9M trainable parameters yet reports large gains—+18.4 mIoU on PerMIS and +17.8 mAP on ILIAS—while being over 75× faster and 20× smaller than the prior unified model; it also supports few-shot retrieval and flexible segmentation prompts.

    The unified formulation and strong efficiency/results are genuinely interesting, but the contribution remains a specialized vision framework and the abstract does not establish whether the gains generalize beyond the reported benchmarks.

  128. maybe BCI ▲ 3 score 4.2

    LibriBrain100: One Hundred Hours of Broad and Deep MEG Data for Neural Speech Decoding at Scale

    Francesco Mantegna, Dulhan Jayalath, Gereon Elvers et al.

    The paper introduces a large MEG speech-decoding dataset with over 100 hours of recordings, including an unusually deep ~80-hour single-subject corpus and ~40 minutes from each of 32 additional subjects. Using an existing model, the authors report state-of-the-art word classification and show that multi-subject pretraining can partly offset limited data for new subjects. The dataset, standardized splits, loaders, and competition are intended to make noninvasive neural speech-decoding comparisons more reproducible.

    The exceptional within-subject scale and explicit depth-versus-breadth comparison could materially affect MEG speech-decoding research, but this is primarily a dataset and benchmark release using an existing model, with no actual brain-to-text capability demonstrated.

  129. maybe AI / ML ▲ 9 score 4.2

    CritICL: Inference-Time Weak-to-Strong Generalization from Small Language Model Failure Modes

    Yufan Wu, Yinghui He, Zhengyi Hu et al.

    CritICL uses recurring mistakes from weaker models as critique examples for stronger models at inference time. It offers a dynamic version that predicts which failure mode applies to each input and a static version using a general failure profile, claiming better reasoning than ordinary in-context learning with fewer generations and tokens than test-time scaling methods.

    The failure-mode-as-guidance framing could be a useful and relatively novel alternative to repeated sampling, but the abstract gives no quantitative results, task breadth, or details sufficient to judge how substantial the efficiency and accuracy gains are.

  130. maybe AI / ML ▲ 14 score 4.1

    The Handoff Tax: Continuing Non-Native Trajectories in LLM Agents

    Roy Ganz, Mor Shpigel Nacson, Adi Kalyanpur et al.

    This paper studies what happens when a coding agent switches between weaker and stronger models during a long task. Across Claude and GPT model pairs, handing a stronger model the full weaker model’s trajectory recovers less than half of the potential quality improvement and costs substantially more; downshifting is more favorable. The surprising result is directional asymmetry: escalation works better when the weaker model’s trajectory is omitted or compressed, while downshifting benefits from retaining the stronger model’s trajectory.

    The directional “handoff tax” and opposite effects of trajectory visibility are practically relevant and non-obvious for multi-model coding agents, though the abstract lacks enough experimental detail to establish how broad or robust the effect is.

  131. maybe AI / ML picked▲ 17 score 4.1

    Sliding-window beats linear attention

    Alexia Jolicoeur-Martineau, Rhea Sanjay Sukthanker, Pashmina Cameron et al.

    This paper compares post-trained linear-attention replacements with the much simpler combination of sliding-window attention and attention sinks. Across several LLMs and tasks, the authors report that sliding-window attention matches or outperforms linear attention, with a 2–10× advantage on long-context retrieval and reasoning tests, while requiring no post-training and less inference memory. The main contribution is an empirical warning that linear attention should be compared against strong, simple local-attention baselines.

    The methods are established, but the reported large advantage over post-trained linear attention could challenge a popular efficiency direction; the abstract provides broad but not sufficiently detailed evidence for a stronger verdict.

  132. maybe AI / ML ▲ 1 score 4.1

    MemUse: Moving Memory Evaluation from Direct QA to Natural Integration in Long-Term Human-AI Conversation

    Ryuichi Sumida, Koji Inoue, Tatsuya Kawahara

    This paper argues that conversational memory should be evaluated not by whether an LLM can answer explicit questions about past dialogue, but by whether it notices when a memory is relevant and incorporates it naturally. In a four-month deployment with 40 users, Direct QA accuracy varied widely without affecting satisfaction, while a new MemUse evaluation of natural integration tracked satisfaction better; the same system recalled 78.8% of facts when asked but used only 7.9% naturally in conversation.

    The distinction between retrieval-on-demand and contextually appropriate memory use is a useful, non-obvious reframing backed by real deployment data, but the evidence is limited by the small user population and judgment-based evaluation.

  133. maybe Robotics ▲ 4 score 4.1

    Generative Semantic Scene Completion

    Shi Chen, Weifeng Ge

    The paper reframes outdoor LiDAR semantic scene completion as discrete generative modeling. It introduces a synthetic paired sparse/dense-scene corpus, a diffusion model conditioned on sparse scans, and a one-step refinement method that improves both its own predictions and those of external completion models. On SemanticKITTI, the best causal single-sweep single-sample result reaches 38.8% mIoU, 2.1 points above the prior published result under the same restriction.

    The unified generative formulation and model-agnostic one-step refinement are substantive ideas, but the demonstrated gain is still a moderate benchmark improvement rather than a clear capability or efficiency breakthrough.

  134. maybe AI / ML ▲ 12 score 4.1

    Are Android GUI Agents Robust Against Runtime Anomalies? AnTrap: Evaluating Agents in Dynamic Adversarial Environments

    Guo Gan, Yilun Zhao, Cong Chen et al.

    The paper introduces AnTrap, a benchmark that tests Android GUI agents under injected runtime anomalies such as pop-ups, state changes, action misuse, and deadlocks. Across 16 GUI models, it reports broad performance degradation and finds that adversarial reinforcement learning helps with simpler traps, while deeper contextual failures remain difficult to fix through trap-focused training alone.

    This is primarily a benchmark, but its layered diagnosis of which runtime failures are learnable versus reasoning-limited—and the reported persistent vulnerability of strong agents—offers a potentially useful non-obvious finding.

  135. maybe AI / ML score 4.1

    Does Rank Still Matter? Position Bias When AI Agents Shop on Our Behalf

    Davood Wadi, Yu Ma

    The study randomizes the order of 100 hotel listings across 5,000 shopping sessions involving four LLM-based agents, and compares their behavior with human data. Agents inspect deeper than humans and show much weaker, non-monotonic position bias: middle-ranked items are least likely to be inspected, while displayed attributes matter more than placement; however, models differ in whether position affects final choice and converge on the same undominated option.

    The large randomized comparison suggests a genuinely non-obvious pattern in how AI agents process ranked results, though the finding is based on a narrow hotel-search setting and its broader generality is not yet established.

  136. maybe AI / ML score 4.1

    The Compaction Cliff in Long-Running AI Agent Memory

    Saber Zerhoudi, Jelena Mitrovic, Michael Granitzer

    The paper shows that ordinary context compaction can rapidly destroy exact safety rules in long-running agents: only 10% remain after five rounds in one tested production configuration. It proposes Knowledge Triage, which assigns different retention, decomposition, and retrieval policies to different knowledge types, reporting substantially better rule preservation and downstream task compliance across several corpora and domains.

    The compaction cliff is a concrete and non-obvious failure mode, and type-specific memory operations are a meaningful alternative to uniform summarization, but the evidence is limited to agent-memory benchmarks and the practical generality of the reported gains is not yet clear.

  137. maybe AI / ML score 4.1

    XTC: Head-Aware Sampling by Excluding Top Choices

    Philipp Emanuel Weidmann, Allen Roush, Judah Goldfeder et al.

    XTC is a decoding method for open-ended text generation that identifies several sufficiently plausible next tokens, probabilistically removes the most likely ones, and samples from the weakest remaining plausible option. Across several quantized language models, it substantially reduces repetition and increases lexical diversity, while human raters preferred its outputs and instruction-following accuracy stayed close to baseline; the method has also been integrated into several inference tools.

    The head-aware strategy is a genuinely nonstandard alternative to temperature and truncation, with unusually large reported diversity gains and human-evaluation support, but its benefits appear concentrated in creative generation rather than representing a broad capability advance.

  138. maybe AI / ML score 4.1

    Unsupervised Speech Recognition at the Syllable Level

    Liming Wang, Kai-Wei Chang, Kunio Kashino et al.

    The paper proposes unsupervised speech recognition using syllable-level units and masked language modeling, avoiding grapheme-to-phoneme converters and GAN-based training. It reports up to a 40% relative reduction in character error rate on LibriSpeech and improved generalization to difficult low-resource languages, though the abstract gives few experimental details.

    Syllable-level modeling could be a useful new direction for genuinely unsupervised, low-resource ASR, and the reported error reduction is substantial, but the abstract does not provide enough baseline, language, or scale details to justify a strong recommendation.

  139. maybe AI / ML score 4.1

    The Illusion of Control: Why Bare Classifier Inversion Silently Fails in Concept-Bottleneck Text Generation

    Qi Bing, Xiaowei Shao

    This paper studies how to construct inference-time concept codes for controllable text generation when the desired attribute combination was not directly observed. Across three model families (124M–8B), it finds that classifier inversion—including several distributional regularizers—consistently loses to a simple post-hoc prior over encoder-produced concept codes; bare inversion collapses to chance because it produces off-manifold codes. The diagnosis is also tested on real-world benchmarks and with external evaluators.

    The cross-model negative result and direct off-manifold explanation challenge the assumption that classifier inversion provides controllable concept codes, though the finding is specialized to concept-bottleneck generation and does not yet imply a broad change to generative modeling.

  140. maybe AI / ML score 4.1

    Budget-Constrained Embodied Perception: Four Resource Walls and a Pre-Registered Evaluation of Access-Structured Perception on Open Models at less than 31B

    Defu Lin, Wenhui Chen, Ziyao Lin et al.

    The paper studies how embodied multimodal agents should spend a fixed token budget when answering questions over long observation streams. Its training-free ASP wrapper uses query-conditioned retrieval and iterative access, and on a synthetic long-horizon benchmark it substantially beats query-independent frame sampling; however, its learned/compressive state component hurts performance, and the full proposed system does not beat a verbatim retrieval baseline. The strongest conclusion is that adaptive access matters more than simply increasing context or model size, though the evidence is limited to synthetic scenes.

    The resource-wall framing and preregistered ablations yield a non-obvious result about adaptive observation access, but the core retrieval idea is familiar and the claims are supported only on a synthetic benchmark, with the proposed full architecture largely falsified.

  141. maybe AI / ML score 4.1

    Signal or Noise? A Benchmark Study of Agent Skills in Web Development

    Ziyue Yang, Fan Ding

    This paper benchmarks 31 reusable coding-agent Skills across 50 web projects, 1,000 tasks, and four models, using length-matched irrelevant controls and component ablations. Contrary to the usual assumption that more task-specific guidance helps, injecting Skills generally reduced Pass@2, sometimes substantially increased token cost, and helped only a minority of Skill–project pairs; the losses arose either from prompt-length distraction or misleading content. It argues that Skills should be selectively routed for a particular model, project, and deployment rather than treated as generally reusable assets.

    The controlled finding that supposedly relevant agent Skills often hurt performance—and that the failure can be separated into length distraction versus content misdirection—is a non-obvious and practically important result, though it remains a benchmark study rather than a demonstrated general solution.

  142. maybe AI / ML score 4.1

    Feedback That Backfires: Why Small Language Model Agents Repeat the Call They Just Watched Fail

    Esmail Gumaan

    This study finds that small instruction-tuned language models often become more likely to repeat a tool call after seeing that exact call fail, rather than treating the error as corrective feedback. Across six models and two environments, the failed call’s surface form caused most of the effect; replacing it with a generated description of the failure largely removed the problem, while “do not repeat” instructions and clean-context retries did not. The result suggests that transcript design, not merely error understanding, is a major cause of repetitive agent failures.

    The consistent, counterintuitive finding that verbatim failed actions can reinforce repetition—and the targeted counterfactuals identifying a practical harness-level remedy—makes this worth a closer look, though the evidence is limited to small models and two environments.

  143. maybe AI / ML score 4.1

    When a neural surrogate cannot accelerate a solver: runtime share, closed-loop drift, and the economics of uncertainty gating in a stiff coupled simulation

    L. Thümmler, T. Kuroda

    This paper evaluates neural surrogates inside a stiff, coupled radiation-hydrodynamics solver and finds that apparent per-call speedups do not translate into end-to-end acceleration. Runtime profiling, fourteen surrogate models, and long closed-loop runs show three barriers: the replaced block is only 16.9% of runtime, offline error does not predict deployment survival, and distribution-shift gating rejects nearly all surrogate calls while the accepted predictions accumulate a substantial directional bias. The paper also derives a break-even rule for gating and distinguishes stability from fidelity.

    The controlled negative result offers unusually concrete, potentially reusable lessons about evaluating learned solver surrogates—especially runtime-share limits, misleading offline metrics, and bias accumulation—but its evidence comes from one specialized multiphysics testbed.

  144. maybe AI / ML score 4.1

    Definitional Sensitivity in Media Bias Detection: A Multi-Definition Dataset and Benchmark

    Martin Wessel, Timo Spinde, Jürgen Pfeffer et al.

    The paper tests how changing the definition of media bias affects ratings by 354 people and four LLMs, using 8,496 human and 28,800 model judgments across six articles and four bias categories. Changing the underlying conceptual target substantially changes annotations, while adding detail without changing the concept has little effect; LLMs are even more definition-sensitive than humans. The authors release a multi-definition dataset and discuss consequences for annotation protocols and prompt-based evaluation.

    It provides a reasonably well-supported, non-obvious finding that definitions—not merely annotation detail—can change the construct being measured, especially for LLMs, though the study remains centered on a narrow media-bias setting.

  145. maybe AI / ML score 4.1

    How Much Regularization Survives Averaging? Update Masking in Federated Learning

    Wenhao Yan, Fu Kuroda, Yucheng Jin et al.

    This paper analyzes why update-masking noise, which can act as an implicit regularizer in centralized training, largely loses that effect under federated averaging. It proves that independently masked client updates dilute the regularization by the cohort size, while using a shared mask can recover it according to client gradient diversity; experiments on CIFAR-10 find only a 1.19× recovery in the default setting, rising to 8.96× without minibatch noise, but the configurations that preserve regularization train poorly.

    The exact analysis of how federated averaging destroys update-masking regularization is a non-obvious and potentially useful clarification, but the empirical setting is narrow and the proposed configurations appear impractical.

  146. maybe AI / ML score 4.1

    SRPO: Self-Reflective Policy Optimization for Long-Horizon Reasoning

    Jialong Liu, Yuling Shi, Ning Yang et al.

    SRPO has the model inspect its own completed reasoning trajectories, summarize mistakes as short reflection patches, and use those reflections to produce dense token-level training signals for on-policy rollouts. The authors report strong results from an 8B Qwen3 model, including 73.3% on AIME’24 with 8% of the FLOPs of scaled supervised fine-tuning, plus gains on WebShop, ALFWorld, and SWE-Bench-Lite, without external critics or reward models.

    The combination of self-generated error summaries with dense policy optimization and the claimed compute reduction is a substantive idea, but the abstract does not establish how comparisons were controlled or whether the broad benchmark gains are robust.

  147. maybe AI / ML score 4.1

    Tight Majorizations and Convergence Rates of Nuclear Norm Minimization IRLS

    Christian Kümmerle, Tomas Masak, Dominik Stöger

    This paper analyzes IRLS algorithms for recovering low-rank matrices via nuclear-norm minimization. It proves that a harmonic-mean weight operator gives a global quadratic majorizer, is optimal among power-mean choices, and yields dimension-independent local linear convergence; one-sided weighting schemes cannot generally achieve the same rate. The theory is supported by counterexamples and numerical experiments across several recovery settings.

    The harmonic-mean reweighting analysis and dimension-independent rate provide a genuinely non-obvious theoretical clarification of why common IRLS variants differ, but the contribution is specialized to low-rank recovery and does not demonstrate a broad practical capability jump.

  148. maybe AI / ML score 4.1

    A Theory of Speciation in Generative Diffusion Models on Compact Riemannian Manifolds

    Alessio Marta, Paola Causin

    This paper develops a geometric theory of how diffusion-model denoising trajectories split into distinct, class-like branches on compact Riemannian manifolds. It argues that generic transitions are usually fold (saddle-node) bifurcations rather than the symmetric pitchforks often assumed, derives topology- and geometry-dependent constraints and timing estimates, and studies robustness to score-estimation error. Sphere experiments with von Mises–Fisher mixtures and chart-based neural score learning illustrate pitchforks, saddle-nodes, and hierarchical speciation.

    The generic-fold characterization and topological treatment offer a genuinely nonstandard lens on diffusion dynamics, with substantial mathematical support, but the abstract does not establish a major practical consequence for current diffusion models.

  149. maybe AI / ML score 4.1

    Wontopos Tablet 2: Measuring Multilingual and Multimodal Memory Retrieval Without Lexical Matching

    Sunwoo Kim

    The paper evaluates a production long-term memory retrieval system that uses neither lexical matching nor an internal language model. It reports strong multilingual text retrieval and unusually robust cross-lingual retrieval of captionless photographs, while also showing that reader choice and re-asking budget can change benchmark scores substantially. The most interesting findings are that dense visual retrieval is not automatically language-independent, and that adding captions can hurt cross-lingual retrieval in this system.

    The combination of language-free memory retrieval, multilingual text queries, and captionless cross-lingual image retrieval is a genuinely interesting systems direction, supported by reasonably broad controls, but the evidence is still centered on one engine and benchmark methodology rather than a clearly general breakthrough.

  150. maybe AI / ML score 4.1

    ALPHABET: A Laplace-Pole History Aggregator with Banked Exponential Transport

    Daehwa Ko, JaeHyeon Kim, Oh Seong Kwon et al.

    ALPHABET is a very small linear-time sequence model that represents history using banks of stable complex exponential modes, then predicts from modal energies and lag statistics rather than reconstructing the full sequence. The authors give a spectral interpretation of these descriptors and report competitive results across 82 tasks, with substantially lower parameter count and faster training and inference than nine baselines.

    The combination of an auditable pole-based history representation, spectral-identification analysis, and strong efficiency claims is meaningfully distinctive, but the abstract gives rankings rather than task-level gains and does not establish a major capability advance.