Paper Feed

Issue 33 · Aug 10–16, 2026

Every candidate

All 2,925 papers were scored from their abstracts by gpt-5.6-luna; 1,027 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.

  1. strong Tech picked score 5.9

    Exponential quantum advantage for learning signals with a single qubit

    Ishaan Kannan, Sridhar Prabhu, Saeed A. Khan et al.

    The paper develops “quantum feature sensing,” using a single controllable qubit coupled to a conventional sensor to learn classical signals with dramatically fewer measurements. It gives formal advantages for Fourier coefficients, temporal correlations, and observable transformations, and reports superconducting-cavity experiments claiming up to 10^7-fold measurement reductions, with demonstrations motivated by dark-matter sensing and wireless communications.

    The combination of a single-qubit architecture, claimed exponential measurement advantage, and an experimental 10^7-fold reduction is an unusually large and assumption-challenging result, though the precise resource accounting and baseline assumptions merit close scrutiny.

  2. strong AI / ML score 5.9

    A Constant-Competitive Algorithm for Dynamic Mixture-of-Experts Serving

    Ian D'Ambrosio

    This paper claims that dynamic Mixture-of-Experts serving can be handled with a randomized algorithm whose competitive ratio is constant, independent of the number of experts, overturning the previously known O(sqrt(log k)) bound and its apparent lower-barrier intuition. The proof reduces the problem to chasing positive bodies, uses a finite approximation and resource-augmentation removal, then applies lazy threshold rounding; the main guarantee and reduction are machine-checked in Lean 4, though they depend on formal interfaces to cited theorems.

    A credible constant-competitive result for arbitrary-scale dynamic MoE serving would resolve an open order-of-growth question and overturn the expectation suggested by the prior sqrt(log k) bound, with unusually strong formal verification evidence.

  3. strong Neuroscience picked score 5.8

    Magnetoencephalography Without a Shielded Room

    Bezsudnova, Y., Alexander, N. A., Mellor, S. J. et al.

    The authors build a lightweight, multichannel MEG system that operates without a magnetically shielded room, using active shielding and software spatial filtering to separate extremely weak brain signals from environmental noise. In vivo recordings reportedly achieve signal-to-noise ratios comparable to conventional cryogenic MEG in a shielded room, while the system weighs under 75 kg.

    Removing the shielded-room requirement would substantially reduce MEG’s cost and infrastructure barrier and enable deployment in settings previously inaccessible to the technology, although the abstract provides limited quantitative validation beyond headline claims.

  4. strong AI / ML picked score 5.6

    Bagging Robustly Learns VC Classes with Linear Sample Complexity

    Omar Montasser

    This paper proves that any VC class can be learned robustly against test-time adversarial examples using only a number of samples linear in its VC dimension, improving exponentially over the prior upper bound. The method is notably simple: train robust ERM predictors on independent bootstrap samples and take their majority vote; a matching lower bound shows that the dependence on the dual VC dimension is necessary in the stated oracle model.

    An exponential improvement for a fundamental robust-learning problem, obtained through an unexpectedly simple bagging construction and accompanied by a lower bound, is a genuinely notable theoretical result.

  5. strong Neuroscience picked score 5.5

    Activity-dependent structural plasticity of myelinated axons tunes their excitability

    Bennett, D. J., Rancic, V., Zhao, Z. et al.

    The authors report that brief, subthreshold stimulation causes myelinated mouse axons to rapidly shrink near stimulated nodes to about 70% of their original diameter, with neighboring glia swelling into the space. This structural change increases axonal excitability for hours, apparently through altered geometry and a mechanism involving Kv1 channels, Na/K pumps, bicarbonate, and pH rather than transient active currents; GABAA activation produces a similar effect.

    This challenges the view of myelinated axons as fixed cables by proposing fast, lasting activity-dependent structural tuning of conduction, although the evidence is currently from a preprint and appears limited to mouse axons and specialized preparations.

  6. strong AI / ML score 5.4

    Optimistic Rates for Multiclass PAC Learning

    Xiaoyu Li, Andi Han, Jiaojiao Jiang et al.

    This paper derives the optimal intermediate-noise (optimistic) excess-risk rate for multiclass PAC learning: approximately \(\sqrt{L^\star d_N/n}+d_{DS}/n\), interpolating between realizable and agnostic learning and remaining uniform in the number of labels. It introduces a comparator-relative compression result for the upper bound, proves matching lower bounds, and extends the result to list learning; the main theorems are formally verified in Lean 4.

    It claims to close an explicitly open gap in multiclass learning theory with matching rates, new proof machinery that avoids binary-specific geometry, and machine-checked main theorems.

  7. strong AI / ML picked score 5.4

    Algebraic Decomposition Theory for Transformer Length Generalization

    Andy Yang, Blerta Veseli, Corentin Barloy et al.

    This paper gives a complete algebraic characterization of which regular languages transformers can recognize while generalizing to sequence lengths beyond training, together with a polynomial-time decision procedure based on decompositions involving the infinite additive group of integers. It argues that classical finite-semigroup theory cannot capture this behavior and reports broad experiments showing better agreement with observed transformer behavior than prior classifications.

    It appears to provide the first general, efficiently decidable theory of transformer length generalization for regular languages, replacing inadequate classical finite-semigroup tools with a new decomposition framework tied to unbounded counting.

  8. strong AI / ML picked score 5.4

    Equivariant learning of a transferable three-dimensional classical density functional

    Bingqing Cheng

    The paper learns a three-dimensional classical density-functional free-energy functional directly from equilibrium liquid density fields, while enforcing spatial symmetries and variational consistency without free-energy labels. One learned functional transfers across temperatures, system sizes, ensembles, and complex geometries, predicting structure factors, equations of state, phase coexistence, interfacial behavior, and colloidal and porous-media phenomena that were not training targets.

    This is a genuinely broad, label-efficient approach to learning a reusable thermodynamic functional, with unusually strong transfer and emergent-property claims beyond the training observables.

  9. strong AI / ML picked score 5.3

    Sampling Luck Masquerades as Allocation Gain: Auditing Test-Time Budget Allocation for Neural Combinatorial Optimization

    Jinhyung Bae

    The paper shows that apparent gains from allocating more test-time samples to some combinatorial-optimization instances can be an artifact of choosing and evaluating allocations on the same sampled solutions. On in-distribution TSP, the reported 2.2–2.6% gains disappear out of sample, but under distribution shift, held-out allocation statistics produce genuine 11.5–12.0% gains at equal evaluation budget, with smaller gains after charging for probing.

    This is a well-controlled, preregistered result overturning a seemingly plausible test-time allocation improvement while also identifying a setting where allocation genuinely works, backed by null calibration, replication, negative controls, and released artifacts.

  10. maybe AI / ML picked▲ 170 score 5.3

    Self-Supervised Visual On-Policy Distillation

    Yijiang Li, Yijun Liang, Yunjie Tian et al.

    The paper proposes creating the teacher–student asymmetry needed for on-policy distillation by weakening the student’s visual input rather than giving the teacher privileged information. The teacher sees the original image while the student learns from a strongly augmented version, and the method improves Qwen3.5-4B from 70.7% to 77.4% across six fine-grained perception benchmarks, reportedly recovering most of the gain from privileged-information methods without extra labels, rewards, or a stronger teacher.

    The inversion of the usual distillation asymmetry is a clear and potentially general idea, with a substantial reported gain, but the abstract provides limited detail and the broad superiority claims need careful verification.

  11. maybe AI / ML ▲ 445 score 5.1

    StateM: Reaching 95.3% Raw Accuracy, or a \$15 Frontier Run, on Terminal-Bench 2.1 via Harness Scaling

    Ziheng Qin, Yaxin Lu, Zhangyang Atlas Wang et al.

    StateM improves long-horizon coding agents by adding an external runtime for durable state, phase-specific context, checked transitions, recoverable runbooks, and versioned procedures, without changing model weights. On Terminal-Bench 2.1 it reports large gains across several models—up to 95.3% raw accuracy—and substantially lower API cost, while held-out BusinessBench results suggest some runbook rules transfer when task structure is shared.

    The reported cross-model accuracy and cost gains from a reusable execution harness are unusually large, but the approach combines familiar agent scaffolding ideas and the abstract does not establish how much is benchmark-specific or independently reproducible.

  12. maybe AI / ML ▲ 30 score 5.1

    V-RAE: Rethinking Video Latent Spaces for Generation

    Minghui Guo, Shengqiong Wu, Hao Fei

    V-RAE replaces the usual pixel-reconstruction-oriented video VAE latent with a compressed representation built from frozen vision foundation models, using temporal pooling and a decoder to recover motion. Across Kinetics and UCF101, it reports better reconstruction and semantic content, up to 6× faster generation convergence, and introduces tFVD as a temporal-coherence metric that supposedly predicts generation quality better than reconstruction metrics.

    The central claim—that semantically organized frozen representations are better generative latents than reconstruction-optimized video tokenizers—is a meaningful and somewhat assumption-challenging direction, supported by multiple encoders, tasks, and benchmarks, but the abstract does not establish whether the gains generalize beyond these settings.

  13. maybe AI / ML picked▲ 49 score 5.1

    MOSS-VL Technical Report

    Pengyu Wang, Chenkun Tan, Shaojun Zhou et al.

    MOSS-VL is an open vision-language model designed to perceive incoming video while it is speaking, using gated visual cross-attention, synthetic supervision for turn-taking and revision, and a final real-time training stage. It reports substantially better proactive behavior than open baselines on streaming benchmarks, including 66.0 versus 37.5 on OmniMMI Proactive Alerting, while also improving time-to-first-token as visual context grows by keeping visual tokens outside the decoded sequence.

    The combination of genuinely interactive streaming generation, proactive speech behavior, and a favorable latency design is notable, but the abstract provides mostly benchmark claims against open-source baselines rather than broad evidence of a field-changing capability.

  14. maybe AI / ML ▲ 35 score 5.0

    SimpleOPD: Simple Tokenizer-Agnostic On-Policy Distillation for Long-Context Reasoning

    Haonan He, Haodi Lei, Yun Luo et al.

    The paper adapts on-policy distillation to transfer long-context reasoning from a teacher to shorter-context students despite different tokenizers. It distills in shared text spans, adds a reference KL constraint, and suppresses learning signals on termination tokens to control response-length growth and training instability. Across several student model families, it reports substantial gains in mathematical proof reasoning, including a 21.2-point improvement on ProofBench, plus improvements on science benchmarks.

    The combination of tokenizer-agnostic span-level OPD and length-control mechanisms addresses a real practical obstacle, and the reported cross-family, cross-domain gains are unusually large, but the abstract provides too little detail to establish whether the headline comparison is robust.

  15. maybe AI / ML ▲ 165 score 5.0

    Alaya-EVOKE: From Linear-Scaling Supervision to Endless World

    Yuanyang Yin, Gongxuan Wang, Yifan Zhan et al.

    Alaya-EVOKE is an interactive video/world model that keeps persistent scene geometry in an external, camera-indexed memory rather than growing the denoiser context. It trains a three-step student from a long-horizon teacher using recurrent self-forced rollouts and claims 30-second generation, controllable prompt changes, reduced drift, and 1.5-second chunks generated in 2.11 seconds on one H200.

    The combination of external world-state memory and long-horizon teacher distillation is a substantive direction for interactive video generation, but the abstract gives few quantitative comparisons beyond broad state-of-the-art claims, so it does not yet warrant a strong recommendation.

  16. maybe AI / ML ▲ 113 score 4.9

    AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

    Cheng Qian, Wenting Zhao, Liangwei Yang et al.

    The paper studies whether a strong model can improve a weaker model at inference time by constructing an external “harness,” without changing the weaker model’s parameters. On four Theory-of-Mind benchmarks, these harnesses reportedly raise average performance from 0.49 to 0.91, mainly by moving brittle reasoning into deterministic code, routing by benchmark/task type, and enforcing answer formats; however, the evaluation appears closely tied to those benchmarks and a validation set.

    Test-time transfer through automatically designed scaffolding is a genuinely interesting alternative to distillation, and the large reported gain is non-obvious, but the narrow four-benchmark evaluation and likely benchmark-specific engineering limit how broadly convincing the result is.

  17. maybe Robotics picked▲ 14 score 4.9

    RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance

    Dongchi Huang, Hongyin Zhang, Bohan Hou et al.

    RynnValue trains a general-purpose robotic value model using temporal distance—the estimated time from an observation to a language-specified goal—instead of preference or normalized-progress labels. Using timestamp-derived supervision from about 7,000 hours and 3 million clips, it reports better out-of-distribution ranking than preference-supervised and progress-based baselines, plus substantial real-world policy-success gains when used for reward shaping.

    The combination of timestamp-only supervision, large-scale cross-embodiment training, and strong reported generalization is a genuinely interesting direction for scalable robot reward modeling, but the abstract does not establish whether the gains hold across sufficiently broad tasks and independent evaluations.

  18. maybe AI / ML ▲ 23 score 4.9

    LiveAnimate: Stable Long-Form Streaming Human Animation in Real-Time

    Yuxuan Zhang, Haozhong Xiong, Yubo Huang et al.

    LiveAnimate turns a large bidirectional video diffusion transformer into a block-causal streaming generator for pose-driven human animation. It combines teacher-forcing adaptation, three-step self-forcing distillation, and a bounded pose-retrieval KV cache to maintain identity and visual quality over long streams; the system reportedly reaches 19.63 FPS on two H100s and remains stable over three minutes.

    Real-time, long-duration diffusion-based human animation with constant-memory streaming would be a meaningful capability advance, but the abstract provides limited comparative detail and only a three-minute benchmark, so it merits a look rather than a strong recommendation.

  19. maybe AI / ML ▲ 33 score 4.9

    Intern-S2-Mobius: Foundation Model with Decoupled Knowledge and Reasoning

    Kai Chen, Jifeng Ding, Ning Ding et al.

    The paper separates stored knowledge from computation: a shared FFN-like Memory holds knowledge vectors, while multiple self-attention Reasoners repeatedly retrieve and use them for compositional reasoning. A 7B model reportedly matches a Transformer baseline using 62.6% as much training data, while a 35B Qwen-based model reportedly reaches similar scores with nearly 4× faster end-to-end inference.

    The explicit separation of knowledge storage from reasoning, together with substantial claimed data-efficiency and inference-speed gains, is potentially important, but the abstract provides too few experimental details to justify a stronger recommendation.

  20. maybe Robotics picked score 4.9

    Flex-$π$: A Multi-Stream World-Action Model with Compute Flexibility

    Ge Yan, Jinghao Liu, Yuzhi Fan et al.

    Flex-π trains a 6B-parameter world-action model to jointly predict RGB, 3D pointmaps, object-centric semantics, and actions in a shared frozen video-VAE latent space. Stream dropout allows the same model to operate in faster action-only mode or use any combination of visual streams; the authors report 2–7× better performance than strong baselines on precise real-world bimanual manipulation, with no added inference latency and faster execution than π0.5.

    The claimed lossless reuse of a video VAE for 3D geometry, combined with flexible multimodal inference and large real-robot gains, is genuinely unusual, but the abstract gives too few experimental details to justify a strong verdict.

  21. maybe AI / ML picked▲ 764 score 4.8

    BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

    Björn Engdahl, Adrian Kosowski, Jan Chorowski et al.

    BDH-CQ combines recurrent latent computation with in-context learning: demonstrations update a persistent memory, and the model iteratively solves the query without emitting verbal chain-of-thought. A 150M-parameter model achieves 29.5% pass@2 on ARC-AGI-1 at a reported cost of $0.0007 per task, beating the prior cost–accuracy frontier; controlled ARC-like tests examine what transformations it can infer and apply.

    The recurrent latent-memory formulation and unusually low-cost ARC performance are worth checking, but the evidence is centered on one narrow benchmark and the abstract does not establish broader reasoning gains.

  22. maybe AI / ML ▲ 187 score 4.8

    Beyond Pixels: From Video Priors to 4D Worlds

    Zihao Liu, Xiaolong Shen, Zhenglin Zhou et al.

    The paper proposes converting the final latent representation of a video diffusion model directly into a dynamic 3D scene, rather than first generating RGB video and reconstructing it. A lightweight model trained on about 1,000 reconstruction clips transfers across several video generators sharing a VAE, and improves projection-based DINO-F1 over matched RGB-to-4D cascades by 2.88–5.81 points, with better human ratings for geometry and temporal stability.

    The reusable latent-to-4D interface and cross-generator transfer are a genuinely interesting direction, but the evidence is limited to small benchmarks and a narrow same-VAE setting, so it is not yet a clear field-level breakthrough.

  23. maybe AI / ML picked score 4.8

    Twin: Playing an Unknown Game with a Test-Time Digital Twin

    Alexy Skoutnev, Kirill Acharya, Gaston Longhitano et al.

    Twin gives a coding agent a test-time loop for writing and repeatedly repairing an executable world model of an unknown game. On ARC-AGI-3, it clears 179/183 levels and raises the same base model from 7.8% to 93.3% completion, with most solutions more efficient than first-time humans. The notable idea is enforcing replay agreement with all observed transitions before allowing new actions, while using mismatches as automatic counterexamples; goal inference remains the main difficulty.

    This is a genuinely distinctive test-time program-synthesis/world-model approach with a large reported gain, but the evidence is concentrated on one benchmark and the abstract does not establish how broadly the method transfers beyond these games.

  24. maybe AI / ML picked▲ 8 score 4.7

    Beyond Visual CoT: Internalized Visual Thinking for Proactive Video Reasoning

    Xiaoyu Zhu, Xinke Deng, Suresh Taddewadikar et al.

    The paper trains a multimodal model to predict latent representations of future video frames while learning to produce textual answers, so it can acquire visual foresight without generating images at inference time. This internalized visual thinking reportedly improves over direct-answer fine-tuning across six settings and matches or exceeds explicit visual chain-of-thought while reducing end-to-end latency by over 5×.

    The combination of latent future-state prediction and direct-answer inference is a meaningful alternative to pixel-level visual CoT, with a potentially important efficiency gain, but the abstract gives no absolute results or detail about evaluation breadth.

  25. maybe AI / ML ▲ 33 score 4.7

    Marionette: Predicting World States, Rendering Geometry, Painting Appearance

    Zian Meng, Zhen Li, Chuanhao Li et al.

    Marionette separates game-world modeling into three parts: an autoregressive model predicts an explicit 3D state for articulated characters, a fixed renderer computes geometry and occlusion, and a diffusion model renders the final RGB video. This makes the state directly controllable and allows simple rules such as terrain collision and character separation to repair long-horizon failures without retraining the video model; reported visual fidelity is close to rendering from recorded poses.

    The explicit-state-plus-fixed-renderer decomposition is a meaningful and somewhat surprising alternative to pixel- or latent-space game world models, but the evidence is limited to a narrow articulated-character setting and 48 held-out segments.

  26. maybe AI / ML ▲ 91 score 4.7

    OmniScientist: An Omni-Modal Omni-Discipline AI Scientist

    Bobo Li, Hao Fei, Tianjie Ju et al.

    OmniScientist is an end-to-end system that uses multimodal raw scientific evidence—such as images, audio, video, signals, trajectories, and 3-D structures—to generate ideas, run experiments, and write manuscripts. Across 36 cases in five discipline families, it completed the workflow and beat a scalar-feature-only variant in 85% of paired comparisons, suggesting that direct perception improves evidence-grounded research decisions.

    The lifecycle-wide use of raw multimodal evidence in an autonomous research workflow is a meaningful direction, but the small heterogeneous evaluation and aggregate paper scores do not yet establish broadly reliable scientific discovery.

  27. maybe AI / ML ▲ 169 score 4.7

    Demystifying Agent Skills: Why They Work-Until They Don't

    Zhiyuan Jiang, Fangrui Huang, Hanwen Xing et al.

    This paper studies when inference-time “skills” help LLM agents, separating procedural guidance from factual knowledge and analyzing retrieval failures. Across 8,135 controlled trials and paired trajectory analyses, it finds that skills mainly stabilize execution by turning noisy experience into procedural anchors, while their usefulness degrades sharply as retrieval pools grow; exact skill selection is not required for success.

    The procedural-anchoring explanation and the finding that retrieval precision collapses without necessarily hurting downstream success are non-obvious, but the work appears primarily to be a careful empirical taxonomy rather than a major new method or capability.

  28. maybe AI / ML ▲ 11 score 4.7

    DistilVDR: A Compact End-to-End Visual Document Retriever via Dual-Student Distillation

    Zhuchenyang Liu, Ziyi Wang, Yao Zhang et al.

    DistilVDR compresses an 8B visual document retriever into a 524M asymmetric student using direct embedding distillation on both query and document encoders. It reports 61.74 average NDCG@5 across three ViDoRe versions, while using a single-vector index that is 15.6× smaller than the strongest sub-1B multi-vector baseline and indexing roughly 10× faster; a lower-resolution variant trades some accuracy for lower visual-token cost.

    The combination of bilateral embedding distillation, asymmetric capacity allocation, and single-vector indexing addresses a real deployment bottleneck with substantial reported efficiency gains, though the evidence is limited to benchmark results and claims relative to reproduced baselines.

  29. maybe AI / ML ▲ 6 score 4.7

    Persistent Recursive Worlds Enable Autonomous Software Evolution

    Beichen Huang, Zhenyu Liang, Bowen Zheng et al.

    Genesis treats the software project—not any individual coding agent—as the persistent entity. Finite-lived agents make changes within versioned repository “worlds,” and recursive delegation coordinates work across them; in reported runs it built a roughly 250k-line Rust C compiler from scratch, maintained development across agent replacement, and reimplemented 13 MESA modules with substantial speedups. The potentially new contribution is an organizational architecture for long-horizon autonomous software development, supported by unusually large end-to-end demonstrations.

    The persistent-project framing and large autonomous compiler/software-engineering runs are genuinely notable, but the abstract provides limited detail about baselines, agent costs, human intervention, and test validity, so the strong claims need verification.

  30. maybe AI / ML ▲ 64 score 4.6

    Apodex Discovery: Reality Benchmarks and Environments for Evaluating and Building Discoverative Artificial Intelligence

    Brian Wang, Bin Feng, Xiaoman Pan et al.

    Apodex Discovery proposes evaluating AI systems as extended, stateful investigations rather than isolated answers. It combines a foundation model with tools, control policies, recorded trajectories, intermediate verification, and metrics for repair, alternatives, coherence, evidence, and scope, covering 20 real-world problems; experiments report gains in AAV capsid design and drug repurposing over closed-book models. The potentially new contribution is an environment and evaluation framework for verifiable scientific and industrial discovery, though the abstract gives limited detail about task breadth and independent validation.

    Worth a closer look because it frames discovery as verifiable, tool-using, long-horizon investigation with component-level evaluation, but the claimed real-world gains and benchmark validity are not yet supported strongly enough in the abstract for a strong recommendation.

  31. maybe AI / ML ▲ 2 score 4.6

    From Atomic Evidence to Logical Composition: Structured Compositional Reasoning over Compound Answer Options

    Obed Junias, Maria Leonor Pacheco

    The paper addresses a specific failure mode where language models can judge individual statements correctly but fail to combine them under AND, OR, or NEITHER/NOR answer options. It decomposes each compound option into atomic judgments, calibrates contrastive scores, and uses an operator-constrained integer program to recombine them; this raises Macro-F1 from 48.3 to 77.0 on LOGICAL-COMMONSENSEQA and from 47.0 to 75.6 on LOGICAL-SATA, with especially large gains for NEITHER/NOR.

    The large, cross-benchmark gains suggest a genuinely useful way to separate atomic reasoning from logical composition, but the contribution is evaluated on specialized derived benchmarks and the abstract does not establish broad generality beyond this task family.

  32. maybe AI / ML ▲ 59 score 4.6

    VibeWorlding: Can Multimodal Agents Construct 3D Open Worlds End-to-End?

    Yansong Ning, Jingwen Ye, Zhongkai Wu et al.

    The paper introduces a benchmark and training framework for multimodal agents that turn natural-language requests into editable, interactive 3D worlds using asset retrieval, scene editing, rendering, and iterative feedback. It reports that current multimodal models perform poorly, especially on precise 3D editing, while reinforcement-learning post-training substantially improves performance; a 30B-A3B model reportedly achieves the best Pass@1 among tested systems.

    This targets a relatively new, practically important agent capability and combines tool use, multimodal verification, and RL, but the abstract gives limited detail about task realism, baselines, and the credibility of the reported frontier-model comparisons.

  33. maybe AI / ML ▲ 55 score 4.6

    AutoDesign: Meta-Harness Optimization for Long-Horizon Agentic Design

    Yaxin Luo, Haobin Jiang, Jialv Zou et al.

    AutoDesign uses a meta-level optimizer to have a code agent iteratively improve the harness that generates academic posters from papers, using rollout feedback and reusable design procedures. On its newly introduced PosterBench, it reports a 7.45-point advantage over Claude Design, and transferring the learned harness across seven agent/model setups raises average performance by 12.4%; a 40-minute autonomous run produces a poster judged roughly conference quality by humans.

    The recursive optimization of an agent harness, with reported transfer across model configurations, is more interesting than a normal prompt or workflow tweak, but the evidence is centered on a newly defined and narrow poster-generation benchmark and comes from a tech report.

  34. maybe AI / ML picked score 4.6

    The sharp SAT/UNSAT phase transition in random ellipsoid fitting

    Theodor Misiakiewicz, Garrett G. Wen

    The paper proves that fitting a centered ellipsoid exactly through Gaussian points has a sharp feasibility transition at approximately n=d^2/4 constraints. Below the threshold, a well-conditioned positive semidefinite fit exists with high probability; above it, no fit exists at all. The proof resolves prior technical gaps and links the threshold to the statistical dimension of the PSD cone.

    This is a genuinely sharp resolution of a high-dimensional semidefinite-feasibility conjecture, with techniques potentially relevant to random optimization and learning theory, but its direct impact on mainstream AI/ML is uncertain.

  35. maybe AI / ML score 4.6

    Similarity Gates Approve Reversals: A Validity Audit of Embedding-Cosine Thresholds in Agent Systems

    Scott E. Frias

    This paper tests whether embedding-cosine thresholds actually detect meaning changes in agent-system gates such as drift guards, deduplication, and answer checks. It finds that small polarity reversals can receive very high similarity— including 0.9608 for changing “withhold the study drug” to “administer the study drug”—while naive evaluation corpora can produce completely inverted conclusions; common repairs and an NLI model largely fail, although matched-pair designs work better. The authors release an audit corpus, harness, and frozen results.

    The paper presents a concrete, broadly relevant measurement failure in a widely used agent-system pattern, with unusually striking reversal and AUROC results, but the evidence remains an audit of a relatively narrow class of configurations rather than a demonstrated general solution.

  36. maybe AI / ML ▲ 18 score 4.6

    Training Leaves Traces: Centered Residual Signatures for Language Model Lineage Verification

    Aman Singh Thakur, Rayan Khoury

    The paper proposes verifying whether two compatible language-model checkpoints share ancestry using only their weights, without data or inference. It removes shared residual-training structure and compares checkpoint-specific patterns across residual blocks; on several model families, it reportedly separates fine-tuned, merged, pruned, and quantized descendants from independent or distilled models, including under checkpoint-laundering tests, while being much faster than a baseline.

    The potentially important contribution is a data-free weight-space lineage signal that claims to distinguish true ancestry from behavioral similarity and survive function-preserving laundering, but the evidence is largely summarized by benchmark claims and may not establish robustness beyond compatible checkpoints and the tested transformations.

  37. maybe AI / ML ▲ 16 score 4.6

    Second Thought: Reasoning in Parallel as LLM Agents Act and Observe

    Zhensu Sun, Chengran Yang, Yunbo Lyu et al.

    Second Thought uses the time an LLM agent spends waiting for an action or observation to run several speculative reasoning branches in parallel. Merging those thoughts when the observation arrives reduces sequential decoding and often lowers the number of interaction turns, while maintaining or improving Pass@1 across three benchmarks and three models.

    The idea of exploiting environment-waiting time for parallel future reasoning is a non-obvious systems improvement, supported by consistent but moderately scoped benchmark results and a compute-matched comparison.

  38. maybe AI / ML ▲ 9 score 4.6

    Forecast Collapse in Time-Series Foundation Models

    Shu Wan, Miles Ma, Hank Zhu et al.

    The paper identifies “forecast collapse”: time-series foundation models and conventional forecasters produce nearly flat hourly stock-return predictions, despite appearing calibrated under standard per-series losses. Across 97 configurations, the authors link this to weak target predictability and objectives that ignore cross-series structure; their CalibRank loss substantially improves cross-sectional ranking while preserving forecast amplitude. The main contribution is highlighting a calibration-versus-ranking failure mode that standard time-series evaluation can miss.

    The broad empirical diagnosis of flat forecasts and the explicit calibration–ranking tradeoff are genuinely useful and somewhat surprising, but the evidence is centered on hourly equity returns and the proposed fix is a relatively simple objective change.

  39. maybe AI / ML ▲ 42 score 4.6

    AVA-Encoder: Towards Agent-Native Video Representation Learning

    Chuyue Li, Jinpeng Yu, Haozhe Wang et al.

    AVA-Encoder represents films as structured Film Knowledge Graphs containing entities, events, assets, and multimodal relationships, then reconstructs the video from that representation. It uses reconstruction residuals and a dual-loop textual optimization process to improve both the graph and the video encoder, reporting large gains over an external baseline and lower prompt-token use in a policy experiment.

    The agent-native, manipulable video representation is a genuinely interesting direction, but the abstract does not clarify the evaluation metrics, baseline quality, or whether the large gains generalize beyond its new benchmark and self-defined policy setting.

  40. maybe AI / ML picked▲ 14 score 4.5

    Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

    Haodong Li, Shaoteng Liu, Tianyu Wang et al.

    The paper proposes Latent Dynamics Reasoning, which explicitly integrates lower-order motion in a structured latent space and asks a model to predict only higher-order residual dynamics. On a five-task synthetic physics benchmark, it reportedly extrapolates much better to out-of-distribution motions than a video diffusion baseline, while using 26× fewer parameters and running 143× faster; it also handles changes in object color, shape, and direction.

    The explicit kinematic latent rollout and strong OOD extrapolation are genuinely interesting, but the evidence is confined to a controlled synthetic benchmark, so it is not yet clear whether this represents a broadly useful video-world-modeling advance.

  41. maybe AI / ML ▲ 8 score 4.5

    QuoteBench: How Matched Scores Can Hide Command-Path Failures

    Shangao Li, Yao Zhang, Volker Tresp et al.

    QuoteBench evaluates coding agents separately from the shell transport layer that serializes and reparses their commands. On 56 incident-derived tasks, inserting one unescaped parser reduced replayed-command success by 55.4–73.2 points, while explicitly disclosing the boundary let models recover much—but not all—of the loss; this shows that matched execution scores can conceal both transport damage and model adaptation. The paper argues that agent evaluations should report the generation contract, execution path, operating point, and exact final-state validator.

    The decomposition of coding-agent failures into generation versus command-path effects, including model reordering and large hidden compensation, is a genuinely useful and non-obvious evaluation insight, though the evidence is limited to 56 tasks and one deliberately inserted parser.

  42. maybe AI / ML picked score 4.5

    Thinking With Tools, Not With Pixels: Tool Calls as Text Scaffolds for Visual Reasoning

    Jiahao Shao, Yuanbo Yang, Yiyi Liao et al.

    The paper argues that in current tool-augmented vision-language training, the useful information comes mainly from the text surrounding a tool call—such as the tool name, coordinates, target, and intent—rather than from the returned image. Its TextCall method removes image outputs while preserving this textual scaffold, reportedly matching or exceeding image-returning systems across LoRA, full fine-tuning, and RL, while reducing latency by 29–46% and eliminating API calls.

    This is a genuinely non-obvious challenge to the assumption that visual tool calls help because models inspect returned pixels, but the abstract provides insufficient benchmark and scale details to justify a strong verdict.

  43. maybe AI / ML score 4.5

    Intrinsic Structure: Spectral Identifiability for Mechanistic Interpretability

    Ashim Dhor, Pin-Yu Chen

    The paper gives a formal identifiability theory for mechanistic-interpretability features. By modeling a network’s depth evolution with a Koopman operator, it argues that the operator’s spectrum is an intrinsic, coordinate-independent property recoverable from activations at a provable M^-1/2 rate, while also showing that the directions carrying variance need not be the directions carrying information. Experiments on GPT-2, Gemma-2, and Qwen models support the convergence predictions, but the resulting Koopman modes are less useful for one interpretability task than principal components.

    This is a genuinely new theoretical framing for separating model-intrinsic structure from artifacts of interpretability methods, with matching lower bounds and multi-model validation, but the identified spectrum is not itself a clearly useful or legible explanation and the practical gains remain limited.

  44. maybe AI / ML picked score 4.5

    Cracks in the Foundation: Seemingly Minor Architectural Choices Impact Long Context Extension

    Amanda Bertsch, Luca Soldaini, Matthew R. Gormley et al.

    This study shows that several individually minor Transformer design choices—normalization, grouped-query attention, pretraining context length, and sliding-window attention—interact badly when extending models to long contexts. In controlled experiments across 26 comparable 7B models and more than 170,000 GPU hours, combining three or more choices reduced downstream long-context performance by up to 47%, despite leaving short-context loss largely unchanged; the authors also analyze associated attention-sink behavior and release the models as OlmPool.

    The controlled evidence that seemingly small architectural choices compound into large, otherwise hidden failures in long-context extension is a genuinely useful and somewhat assumption-challenging finding, though its impact is concentrated on long-context training and extension.

  45. maybe AI / ML score 4.5

    Actions Speak Louder than Words: Measuring Cross-Lingual Policy Retention in Tool-Using Agents

    Sourabrata Mukherjee, Kalika Bali, Sunayana Sitaram

    This paper evaluates whether tool-using agents follow the same action policy when tasks are presented in different languages, rather than comparing only their final answers. Across 2.38 million rollouts in 41 languages, it develops corrections for several misleading trace-similarity confounds and finds that large frontier models retain about 71–73% of their policy across languages under greedy decoding, while smaller models degrade and often route tasks through English. It also shows that evaluation artifacts, including a trace-extraction regex, can create apparent multilingual failures.

    The combination of large-scale cross-lingual action-policy measurement, reproducibility-normalized results, and evidence that English pivoting and extraction artifacts materially shape conclusions is substantially more informative than a standard multilingual benchmark, though it remains primarily an evaluation study rather than a new capability or method.

  46. maybe AI / ML score 4.5

    A Probe Direction Is a Property of Its Prompt

    Valentin Noël

    The paper argues that activation probes for detecting whether a model recognizes an evaluation are not measuring a stable model property. By varying only the wording of the evaluation-announcing prompt, the authors reproduce opposite model-scaling trends reported in prior work, and find that prompt choice explains much more variance than model identity. They also show that surface-form artifacts account for a substantial fraction of probe performance and propose varying prompts, rather than merely adding evaluation examples, for defensible comparisons.

    This is a potentially important measurement critique because it suggests published conclusions about evaluation awareness and scaling may be prompt-design artifacts, but the abstract provides few quantitative details and the topic is adjacent to the reader’s lower-priority safety/evaluation interests.

  47. maybe AI / ML ▲ 9 score 4.5

    TinyCast: Probabilistic Zero-Shot Forecasting with Computed Periodicity

    Armin Steinhauser

    TinyCast is a very small zero-shot probabilistic time-series forecaster that computes dominant periodicities at inference time instead of learning them, then uses convolutions and a quantile decoder to produce forecasts. With only 146,505 parameters, it reportedly matches or exceeds larger zero-shot models on probabilistic benchmarks and runs with static INT8 inference on embedded hardware without per-series fitting.

    The combination of explicit, computed periodicity with probabilistic zero-shot forecasting appears to deliver an unusually strong size–accuracy tradeoff, but the abstract provides leaderboard claims rather than detailed numerical comparisons or broad evidence of generality.

  48. maybe AI / ML ▲ 25 score 4.5

    Latent On-Policy Self-Distillation

    Guibin Zhang, Jiayang Lyu, Ran Sun et al.

    The paper proposes making the privileged information used in on-policy self-distillation learnable rather than hand-designed. A teacher retrieves relevant past experiences, compresses them into continuous latent tokens, and provides dense supervision on the student’s own trajectories; experiments on tool use and code generation reportedly outperform several RL and self-distillation baselines while using under 30% of their rollout budget.

    The learnable latent privileged context is a meaningful departure from hand-crafted self-distillation, and the claimed rollout-efficiency gains are potentially important, but the abstract gives no quantitative results or details about scale and generalization.

  49. maybe AI / ML ▲ 4 score 4.4

    Gathered, Not Admitted: How Attention Brings a Latent Variable into Verbalizable Form

    Parsa Mazaheri

    This paper studies how a latent variable in a language model becomes readable and reusable at a queried position. Across several open-weight models, it finds evidence against a discrete admission gate: attention-mediated gathering in a mid-depth layer window increases the variable’s verbalizable representation, while MLP outputs do not contribute positively there; the effect appears at a similar relative depth across two model families. It also shows that probe visibility is not a reliable measure of functional use, since components with similar readout effects differ substantially in their impact on answers.

    The attention-mediated, demand-specific account and the explicit separation of representational visibility from functional use are genuinely interesting mechanistic findings, but they remain a narrow analysis of one latent-variable behavior rather than a broadly demonstrated advance.

  50. maybe AI / ML score 4.4

    Logarithmic-Free Moment and Generalization Bounds for Uniformly Stable Algorithms

    Thanh Nguyen-Cung, Binh T. Nguyen

    The paper proves that a logarithmic factor in a moment bound underlying uniform-stability generalization guarantees is unnecessary. It establishes a bound for sums of weakly interacting functions that removes the prior \(\log n\) term and matches the known lower bound up to constants, using a Rademacher-cube argument followed by a transfer to general product distributions.

    It resolves an explicitly posed technical open question and reaches the correct scaling up to constants, though the result is theoretical and its direct practical impact on learning algorithms is not demonstrated in the abstract.

  51. maybe AI / ML score 4.4

    Decodable But Not Detachable: Training Data Granularity Determines Parametric Modularity in Large Language Models

    Marcus Armstrong, Navid Ayoobi, Arjun Mukherjee

    This paper tests whether LLMs store knowledge in domain-specific groups of neurons, comparing broad academic subjects with token-level domains such as languages and modalities. Although subject identity is highly decodable, causal ablations find no selective neuron groups for subjects; in contrast, small, largely disjoint neuron sets for languages and code show strong selective effects, with code-neuron ablation also harming mathematical reasoning. The authors argue that parametric modularity arises mainly when the training data is modular at the token level, though the reported interleaving makes selective quantization difficult.

    The decodability-versus-causal-modularity distinction and the claim that token-level training structure, rather than semantic domains, determines modularity are genuinely interesting, with unusually broad ablation evidence, but the conclusions need scrutiny of the causal methodology and domain construction.

  52. maybe AI / ML score 4.4

    Frozen Brain-MRI Foundation Models Are Site Fingerprints

    Saman Rahbar

    The paper shows that frozen 3-D brain-MRI encoders strongly encode acquisition site, with site classification reaching about 0.9 balanced accuracy across two cohorts, several encoder types, and even random networks. Because site is similarly decodable from raw downsampled images, the effect appears to come from low-level scanner and acquisition differences rather than pretraining; removing it post hoc also damages segmentation performance, suggesting entanglement with anatomical information.

    This is a well-supported and somewhat assumption-challenging audit showing that apparent foundation-model representations may be dominated by scanner fingerprints, though the underlying acquisition confound is familiar and the work is primarily diagnostic rather than a new modeling direction.

  53. maybe AI / ML score 4.4

    A second-order theory of texture for depth from focus

    Sreekar Ranganathan, Ioannis Gkioulekas

    The paper develops a wave-optics account of “second-order texture”: rough surfaces that appear textureless in ordinary computer vision can still produce focus-dependent subjective speckle from their microscopic structure. It shows that a narrowband spectral filter can increase this texture under ambient light, enabling substantially better passive depth-from-focus reconstruction, supported by theory, simulations, and real-world experiments.

    This offers a non-obvious physical source of depth cues for textureless scenes and a simple optical intervention, but the abstract gives no quantitative scale or comparison strong enough to warrant a strong verdict.

  54. maybe Neuroscience score 4.4

    Modeling and Interpreting Correlations, Null Distributions and Significance Levels in Neural Tracking of Natural Stimuli

    Simon Geirnaert, Alexander Bertrand, Tom Francart et al.

    This paper argues that raw stimulus–neural correlations are misleading because they depend strongly on the statistical properties and reconstructability of the signals, not just on neural tracking strength. It compares permutation-based nulls, recommends stimulus–response misalignment, and introduces a Fisher-transformed semi-parametric null model plus a null-normalized tracking score; in EEG from 121 speech-listening participants, this normalization reverses some conclusions based on raw correlations.

    The proposed null-normalized measure addresses a widespread interpretational problem and reportedly changes empirical conclusions at substantial scale, but the abstract does not establish how broadly the correction generalizes beyond this analysis setting.

  55. maybe Neuroscience score 4.4

    Evaluation Resolution Confounds Learning-Rule Comparisons in Model-Brain RSA of Early Visual Cortex

    Nils Leutenegger

    This paper shows that conclusions about which learning rules produce brain-like early visual representations can change dramatically with the image resolution used for evaluation. In their experiments, the apparent advantage of untrained or locally trained networks at V1 disappears or reverses at higher resolutions, and the authors localize the effect to image content rather than several alternative explanations; a scalar luminance feature performs nearly as well as the untrained model in one comparison.

    It identifies a concrete, substantial confound that can undermine a line of model-brain learning-rule comparisons, with multi-resolution experiments and several controls, though the result is primarily methodological and its broader generality remains to be established.

  56. maybe AI / ML score 4.4

    PAIR: Pairwise-Aware Inclusion Reweighting for Adaptive Rollout Allocation in RLVR

    Pixel Nomand, Elena Voss, Marcus Hale et al.

    PAIR treats RLVR rollout allocation as a pairwise sampling problem rather than assigning importance independently to each rollout. It uses prefix predictions to decide which continuations to generate, then corrects the resulting pairwise gradient estimate with joint inclusion probabilities; experiments on Qwen3-1.7B/4B report roughly 51–52% fewer generated tokens than full-group GRPO while improving accuracy over a pointwise allocator by 1.2–1.4 points. The notable contribution is the statistical correction for adaptive selection when the learning signal depends on rollout contrasts.

    The U-statistic/inclusion-probability framing is a genuinely nonstandard way to make adaptive RLVR allocation unbiased, and the reported compute reduction is substantial, but the evidence is limited to two model scales and a specialized RLVR setting.

  57. maybe Robotics picked score 4.4

    Keep the Future, Drop the Rollout: RIFT for World Action Models

    Chushan Zhang, Jinguang Tong, Xuesong Li et al.

    The paper finds that some world action models mainly need the final future representation, not the full sequence of intermediate rollout states. It introduces RIFT, which uses learned anticipation tokens to build that future cache in one pass, achieving near-identical or better success on LIBERO and RoboTwin while reducing action-chunk latency by 68–89%.

    The separation of future-cache consumption from rollout production and the resulting large latency reduction are genuinely interesting, but the evidence is limited to a few models and simulation-style benchmarks rather than broad real-robot validation.

  58. maybe AI / ML score 4.4

    XBridge: Entity-Grounded Latent Bridge for Heterogeneous LLM Communication

    Wooseong Yang, Wei-Chieh Huang, Weizhi Zhang et al.

    XBRIDGE enables LLM agents from different model families to communicate through hidden representations without decoding them into text. It combines receiver-vocabulary entity anchors with a learned latent bridge, reportedly preserving entity identity while adding contextual information; across Llama, Qwen, and Mistral pairs, it beats text communication on seven benchmarks with 11× lower latency and beats KV sharing in most same-architecture tests.

    Cross-architecture latent communication that avoids entity-loss in continuous bottlenecks is a substantive idea with broad comparative results, but the abstract does not establish how robust the gains are beyond the tested models and tasks.

  59. maybe AI / ML score 4.4

    Reading the Gate, Not the Interference: Output-Side Interference Measurement Does Not Track Merge Collapse

    Chencheng Zhu

    This paper tests whether layerwise activation cross-terms—the usual measure of interference in task arithmetic—actually explain merge collapse. It finds that the cross-term is largely propagated rather than created within layers, is strongly affected by output-side geometry and normalization, and can be removed without repairing collapse; instead, collapse appears to reside in the marginal parameter displacements, which state-space measures detect. The authors support this with a preregistered set of 81 predictions across two scales and tasks including GSM8K and HumanEval.

    If the results generalize, they overturn a common diagnostic for model-merging failure and provide a useful carrier-versus-bystander framing, but the evidence is still limited to two collapse parameterizations and a narrow set of merging experiments.

  60. maybe AI / ML ▲ 289 score 4.4

    Spark-to-Paper: End-to-End Research Paper Generation as a Composable Skill

    Zhuoyang Qian, Biao Wu, Yiran Wang et al.

    Spark-to-Paper is a workflow embedded in a coding assistant that turns a research idea into experiments, analysis, figures, and a paper using thirteen composable skills. It separates planning from reporting, uses deterministic integrity checks alongside model critique, and revises or abandons claims when experiments fail. On eight controlled topics, it reports high citation validity and figure editability, substantially better fabrication detection than single-pass drafting, at roughly $8 and 3.2 hours per manuscript.

    The planning/reporting separation and explicit handling of self-refuting research trajectories are useful workflow ideas, but the evidence is limited to eight controlled topics and does not establish that the system can conduct genuinely novel, reliable research beyond structured demonstrations.

  61. maybe AI / ML score 4.4

    An Efficient Minimax-Optimal Algorithm for Adversarial $m$-Set Bandits

    Francesco Bacchiocchi, Tommaso Cesari, Roberto Colomboni

    The paper gives a polynomial-time algorithm for adversarial combinatorial bandits where the learner chooses an m-item subset but observes only its total loss. It avoids explicitly handling the exponentially many subsets, achieves the optimal regret rate up to constants, and proves a matching lower bound; the m=1 case also shows that a log K factor is unavoidable for adaptive adversaries.

    This resolves an open computational problem and closes a theoretical regret-gap debate, but its impact is mainly within adversarial bandit theory rather than a broadly influential ML capability advance.

  62. maybe AI / ML score 4.4

    One Frozen Simulator Is Not Enough: Simulator Collapse in Multi-Agent RL

    Simon Yu, Nicholas Tomlin, Marwa Abdulhai et al.

    The paper argues that training an LLM policy against one frozen user simulator causes “simulator collapse”: the policy learns to exploit that simulator’s dominant behavior and transfers poorly to other simulators and real people. It proposes broadening simulator behavior at inference time through Verbalized Sampling, or training against a population of jointly optimized simulators via Co-Training; across three multi-turn benchmarks, these improve held-out performance by up to 9% and 14%, respectively, with similar gains in a human study.

    The simulator-collapse framing and population-based training target a real but underemphasized source of poor human-interaction generalization, supported by multiple benchmarks and a human evaluation, though the reported gains do not yet establish a broadly transformative result.

  63. maybe AI / ML score 4.4

    A Contract-Grade Verifier for LLM-Generated GPU Kernels, and a Native Blackwell Backward for the Gated-Linear-Recurrence Family

    Rishi Shah, Rishav Shrestha

    The paper proposes a verifier with twelve adversarial, often tolerance-free checks for LLM-generated GPU kernels, covering issues such as NaN/Inf behavior, nondeterminism, shape generalization, and accumulation precision. Auditing 2,638 kernels that had passed an existing harness, it rejects 39.5% outright and finds violations in 62.1%; it also demonstrates a native Blackwell training backward for gated-linear recurrences, including a previously fallback-based reverse-state stage.

    The large, independently cross-checked audit suggests that standard kernel-generation correctness metrics are substantially misleading, while the verifier and native Blackwell backward offer practically relevant tools, though the abstract alone does not establish field-wide adoption or generality.

  64. maybe AI / ML score 4.4

    Agent Behavioral Contracts II: Certifying Compositional Reliability Without Assuming Independence

    Varun Pratap Bhardwaj, Garima Singh, Arun Pratap Bhardwaj

    This paper shows that reliability calculations for multi-agent handoffs can be badly overoptimistic when agents share a model: in 18,000 preregistered missions, two copies of the same model co-failed on 90% of missions where either failed. It proves limitations of fitted dependence models and proposes a finite-sample linear-programming certificate using measured joint moments, substantially tightening the certified reliability interval when more moments are available.

    The empirical demonstration that shared-model redundancy can systematically overstate reliability, together with an assumption-free certification method, is genuinely useful and somewhat assumption-breaking, but the practical scope appears limited by the reported two-agent setting and task-specific experiments.

  65. maybe AI / ML score 4.4

    Exponential Convex Calibration Dimension for the Multi-Label Jaccard Measure

    Mingyuan Zhang

    This paper studies how many prediction coordinates a convex surrogate needs to achieve exact calibration for multi-label Jaccard/IoU loss. It proves an exponential lower bound—at least 2^(s-1) dimensions—and gives polynomial-dimensional MinHash and F1-based surrogates when a fixed nonzero regret is acceptable, revealing a sharp tradeoff between exactness and efficiency.

    The exponential-versus-polynomial dimension tradeoff for calibrated Jaccard optimization is a substantive theoretical result, but the work is specialized and the abstract does not establish an exact dimension or broad practical impact.

  66. maybe AI / ML score 4.4

    Doomed to Re-Annotate, Forever: The ImageNet Story

    Illia Volkov, Nikita Kisel, Tetiana Mishkina et al.

    The authors comprehensively reannotate ImageNet-1k validation images, adding multilabel annotations, localization, revised class definitions, and semantic attributes. They find substantial problems in the original benchmark—about 12% incorrect labels, 33.3% multilabel images, and 3.8% with no matching class—and show that corrected labels raise reported accuracy especially for multimodal LLMs, by 5–6%.

    This is more than another dataset release: it provides credible evidence that ImageNet evaluation is structurally mismeasuring modern models and offers a broadly useful replacement, though its main contribution is dataset infrastructure rather than a new modeling capability.

  67. maybe AI / ML score 4.4

    Attributing Preprocessing Invariance in Spectral Foundation Models

    Dongjun Wei, Hongyi Wu, Yinuo Zou

    The paper argues that apparent preprocessing invariance in spectral foundation models may come entirely from input normalization rather than from learned representations. For a Raman model and two other released systems, normalization alone matches the model’s cross-preprocessing performance across the reported evaluations, while learning only helps when the transformation survives normalization. It provides a simple numerical test for determining which transformations a normalization removes, and shows that several systems may be attributing normalization effects to training.

    This is a useful and potentially broadly applicable attribution correction: it shows that a commonly claimed learned invariance can be explained by a parameter-free preprocessing step, though the evidence is still centered on spectral models and limited replications rather than a demonstrated broad failure across foundation models.

  68. maybe AI / ML score 4.4

    AI-Assisted Discovery and Construction of a Counterexample to the Convergence of Three-Block ADMM with the Identity Matrix as its Third Constraint Block

    Kenan Xu, Xiangfeng Wang

    The paper gives an explicit rational counterexample showing that direct three-block ADMM can fail to converge even when the third constraint block is the identity and the first two blocks are strongly convex quadratics. Exact analysis reportedly produces bounded periodic orbits (period 66, and independently period 23), while experiments with relaxation identify conditions that restore convergence for a particular problem but not uniformly for the whole problem class. AI coding systems were used to discover and verify the counterexamples, but the main contribution is the mathematical construction and certificate.

    It resolves a specific longstanding gap in three-block ADMM theory with an exact, machine-assisted counterexample, though the result is mathematically niche and its practical impact is less clear.

  69. maybe AI / ML score 4.4

    From LLM Inference to Agentic Workloads: Characterization and Implications for Serving Systems

    Chaokun Chang, Yukun Zhou, Kaihua Fu et al.

    The paper measures agentic LLM workloads as long-running, stateful systems rather than isolated model calls, using ten applications, production traces, and unified instrumentation. It finds that sandboxes, retrieval, idle state, control-plane overhead, shifting bottlenecks, and redundant tool calls can dominate performance, and shows sizable gains from task-aware scheduling, placement, state offloading, and caching.

    This is a useful systems-level reframing of agent serving backed by concrete production observations and large optimization opportunities, though it is primarily a characterization and design-exploration paper rather than a fundamentally new serving method.

  70. maybe AI / ML score 4.4

    KV-Rescue: Recovering Reasoning Language Model KV Eviction Loss via Stepwise Interleaving

    Minsoo Cheong, Woosang Lim, Vincent-Daniel Yun et al.

    KV-Rescue addresses quality collapse from aggressive KV-cache eviction by using a smaller full-context helper model to interleave and repair reasoning steps from the evicted large model. Across five math benchmarks, it reportedly recovers 87% of the accuracy lost at a very small eviction budget (B=64), while early termination of repetitive generations reduces base-model decoding by 43%.

    The combination of a lightweight full-context helper, stepwise trajectory interleaving, and degeneration detection is a meaningful inference-time approach with strong reported recovery, but the evidence is limited to math benchmarks and the overall compute/memory tradeoff is not yet clear enough for a strong verdict.

  71. maybe Tech picked score 4.4

    A cuffed CRISPR guide RNA for microRNA activity-dependent genome editing

    Adel, A., Shuto, Y., Kawasaki, S. et al.

    The authors redesign the CRISPR guide RNA so that it remains inactive until a target microRNA cleaves it, enabling genome editing only in cells where that microRNA is active. They demonstrate this in human and mouse cell lines, use a genome-wide mouse microRNA-targeting library to map activity, and show that the system can permanently record molecular-state information in DNA.

    This is a genuinely new form of endogenous-signal-gated genome editing with demonstrations beyond a single proof of concept, but its broader utility, specificity, editing efficiency, and performance in complex organisms are not established in the abstract.

  72. maybe AI / ML ▲ 19 score 4.4

    Context-Matched Distillation: Teacher Causality for Autoregressive Video Distillation

    Hmrishav Bandyopadhyay, Xuanchi Ren, Zijian Huang et al.

    The paper proposes Context-Matched Distillation, which trains and distills video generators with a causal teacher that only sees the frames, controls, and prefixes available at generation time. It adds prefix scoring using the student’s actual rollout context and prefix corruption for unstable early training, reporting better aggregate autoregressive video results and substantially improved time-varying camera-control adherence on short- and long-video benchmarks.

    The causal teacher and rollout-context matching address a real mismatch in existing video distillation pipelines and appear broadly useful for interactive video generation, but the abstract gives no quantitative results or evidence that the gains are more than a well-motivated training refinement.

  73. maybe AI / ML ▲ 18 score 4.4

    Evo-Bench: Can Language Models Improve Agent Harness?

    Lisheng Huang, Chen Yang, Hao Zhou et al.

    Evo-Bench evaluates whether language models can improve the tools, prompts, and control logic surrounding an agent, rather than merely solve tasks with a fixed harness. Across nine models and three domains, autonomous evolution reportedly yields gains up to 16.6 points and produces harnesses that transfer across policy models, but performance depends strongly on the task: it works well for search, less so for workflow-specific office tasks, and often saturates early.

    The benchmark targets a relatively under-measured capability—agents improving their own operating harness—and reports potentially useful transfer and task-sensitivity findings, but the abstract does not establish how robust the gains are or whether the benchmark avoids substantial design-specific effects.

  74. maybe AI / ML ▲ 10 score 4.4

    Specification-first convergence with an AI coding agent: a case study of dismantling a core architectural invariant across 189 files in a 717k-line codebase with no test oracle and no human code review

    Joel Abenhaim

    This paper documents an AI coding agent carrying out a large refactor in a 717k-line TypeScript production codebase: removing the rule that a UI panel must remain open during an AI request, while preserving and reattaching live streams. The agent used a specification-first process with 31 audit passes and reportedly changed 189 files without human code review; the author observed no failures over subsequent sessions, but the result is a single self-reported case study without an independent behavioral oracle.

    The scale and specification-driven workflow for an otherwise rewrite-like architectural change are genuinely interesting, but the evidence is limited by one case, author-observed success, and the lack of independent tests or replication.

  75. maybe AI / ML ▲ 3 score 4.4

    Amplified Does Not Mean Predictive: Reasoning Behaviors in Thinking Models

    Jean de Dieu Nyandwi, Leena Mathur, Yonatan Bisk et al.

    This paper compares the reasoning behaviors that models produce with the behaviors that actually predict correct answers. Across 15 text and vision-language models and six benchmarks, it finds that training amplifies visible behaviors such as self-correction and uncertainty acknowledgment, while more predictive behaviors—especially confidence calibration, knowledge alignment, and self-awareness—are not amplified as much. It introduces Behavioral Lift to quantify this mismatch and argues for process objectives based on calibrated, grounded reasoning rather than deliberative-looking traces.

    The cross-model, cross-modal finding that commonly amplified reasoning behaviors are not the ones most associated with correctness is a non-obvious and potentially useful critique of current process training, though the behavioral annotations and correlational analysis limit how strongly it supports new training methods.

  76. maybe AI / ML ▲ 27 score 4.4

    Agentic Transaction: Towards ACID-Compliant Agent Systems

    Zhaoyan Sun, Xiaoxiao Wang, Guoliang Li

    The paper reframes reliable long-horizon LLM-agent execution as a database transaction problem, defining semantic versions of atomicity, consistency, isolation, and durability. It implements this idea in a data agent using exploration/execution/validation cycles, dependency-aware isolation, and transaction-aware state management, reporting a 10.6% improvement over prior agents, including Claude Code.

    The ACID-based framing and semantic transaction mechanisms are a potentially useful new direction for agent systems, but the abstract gives too little detail about benchmarks, baselines, and what the reported 10.6% improvement measures to justify a stronger verdict.

  77. maybe AI / ML ▲ 8 score 4.3

    Intent Speaks Louder: Controllable User Simulation Beyond Response Imitation

    Bo Wang, Ruixing Zhang, Yunqi Liu et al.

    The paper makes user simulators explicitly choose a local dialogue intent—such as accepting, repairing, or disagreeing—before generating the wording of the next turn. Its UserIDA method combines intent-conditioned fine-tuning with intent-aware reinforcement learning, and reports large gains in intent accuracy and targeted interventions while also improving semantic and stylistic similarity on LMSYS-USP.

    The explicit separation of interaction intent from linguistic realization is a meaningful idea for controllable dialogue simulation, supported by large reported gains, though the contribution remains focused on a specialized simulator benchmark rather than a broad capability jump.

  78. maybe AI / ML ▲ 8 score 4.3

    MathForm: Scaling Mathematical Autoformalization with Knowledge Retrieval and Verification-Guided Refinement

    Lushi Pu, Weiming Zhang, Xinheng Xie et al.

    MathForm combines retrieval from Lean’s Mathlib library with compiler- and semantics-based iterative revision to generate more faithful formalizations of natural-language mathematics. The authors use it to build a 367K-example verified dataset and train an 8B model, which reportedly outperforms specialized 32B baselines across six benchmarks, including difficult consistency-check subsets.

    The combination of library retrieval and verification-guided refinement is a plausible but meaningful advance, and the reported 8B-versus-32B results are notable, though the abstract does not establish how strong or comparable the baselines and benchmarks are.

  79. maybe AI / ML ▲ 2 score 4.3

    CardioState-JEPA: Delay-Aware Cross-Modal Learning of a Shared Cardiac Representation

    Hamza Shafiq, Hung Manh Pham, Bin Zhu et al.

    CardioState-JEPA pretrains one shared cardiac representation from ECG, PPG, and PCG rather than modeling each sensor separately. It combines masked latent-state prediction with a learned temporal-delay aligner, and first learns from abundant single-modality data before aligning the modalities using paired recordings. Across 25 downstream tasks, the authors report sizable gains over self-supervised signal baselines, especially for PPG and PCG.

    The shared physiology representation and explicit alignment of delayed cross-modal cardiac events are a meaningful direction with apparently large multi-task gains, but the abstract lacks dataset, scale, ablation, and statistical details needed to establish that the improvements are robust rather than benchmark- or pretraining-specific.

  80. maybe AI / ML score 4.3

    A Tight Lower Bound for Smooth Nonconvex Stochastic Optimization with Bounded Gradient Noise

    Jikai Jin

    This paper proves a query-complexity lower bound for finding an approximately stationary point in smooth nonconvex stochastic optimization when gradient noise is uniformly bounded. The bound, Ω(ΔL/ε² + ΔLσ²/ε⁴), matches standard upper bounds, suggesting that almost-surely bounded noise does not improve the asymptotic rate over bounded-variance noise in the fresh-sample setting. The proof was reportedly generated with GPT-5.6 Sol and then checked and edited by the author.

    It appears to resolve a specific open question with a tight lower bound and an assumption-relevant negative result, but the abstract provides limited proof or experimental detail and the unusual AI-generated-proof claim warrants closer verification.

  81. maybe AI / ML score 4.3

    Omni2LoRA: Coherence-Preserving Parametric Memory for Efficient Omni Language Models

    Puneet Mathur, Manan Suri, Dinesh Manocha

    Omni2LoRA compresses a long audio-visual context into a reusable LoRA-style parameter state rather than retaining multimodal tokens. A Perceiver hypernetwork produces the adapter, while a learned rank-allocation policy tries to preserve cross-modal information; the authors report 8–12% gains over token-compression baselines and up to 12× lower time to first token on four audiovisual QA benchmarks.

    The idea of turning an entire multimodal context into a fixed, query-reusable parameter memory is substantially more interesting than ordinary token pruning, but the abstract provides limited experimental detail and makes unusually strong claims that need careful verification.

  82. maybe AI / ML score 4.3

    Listen, See and Track: Spatio-Temporal Audio-Visual Sound Event Reasoning for Omni-Modal Language Models

    Zhi Zeng, Cheng Zhang, Zesheng Yang et al.

    The paper introduces ST-OmniQA, a large benchmark for answering questions about what sound sources are doing, where they are, how far away they are, and how they move, using panoramic video with synchronized spatial audio. Its proposed ST-Omni-R1 combines audio-derived spatial and trajectory features with visual context and reinforcement learning, reaching 77.83% versus 37.28% for the best reported baseline, with some transfer to three existing benchmarks.

    Joint audio-visual source localization and motion reasoning is a meaningful capability beyond global sound recognition, and the large reported gain is notable, but the evidence is primarily on a newly introduced benchmark and the abstract does not establish how strong or fair the baselines are.

  83. maybe AI / ML score 4.3

    Towards Expert-level Medical AI for Real-time Video Consultations

    Mahvish Nagda, Jihyeon Lee, Matthew Thompson et al.

    The paper evaluates a real-time, audio-visual version of AMIE, a Gemini-based multi-agent medical consultation system, against text-only AMIE and primary-care physicians in a randomized OSCE with 30 physicians, 15 patient actors, and 100 scenarios. Evaluators rated the video system at least comparable to physicians across several clinical tasks, while patients preferred it for assessment and explanation but preferred physicians for rapport; the study also identifies remaining weaknesses in fine anatomical and affective perception.

    The combination of real-time multimodal clinical interaction and a physician-controlled OSCE comparison is genuinely notable, but the evidence is still simulation-based, the cohort is modest, and the abstract provides few quantitative details to support an expert-level claim.

  84. maybe AI / ML score 4.3

    CHORUS: Complementary Experts for High-Coverage Testbench Stimulus Generation

    Hejia Zhang, Sheng Lu, Zhongming Yu et al.

    CHORUS trains several specialized 4B language-model checkpoints for hardware-verification stimulus generation, then combines their complementary abilities through model merging or additional post-training. The resulting model reaches 88.0% Pass@1 on CVDP-ECov, reportedly 13.5 points above the much larger 671B DeepSeek-R1, suggesting that diverse specialists can outperform a single standard SFT-to-RL pipeline.

    The combination of complementary post-trained experts and a large efficiency gap versus a 671B model is notable, but the evidence is currently limited to one specialized hardware-verification benchmark and builds on established model-merging and RL ideas.

  85. maybe AI / ML score 4.3

    Human versus Computer Vision

    Elena Sirotkina

    This study evaluates visual-saliency models against 11.4 million webcam gaze observations from 3,023 demographically quota-matched US adults viewing news photographs. It reports that a simple center-bias baseline beats all tested neural models, and that the models’ residual accuracy is uneven across demographic groups, especially underperforming for older, Black, and ideologically extreme viewers. The paper proposes evaluating whether models can learn the gaze patterns of each group rather than reporting only aggregate accuracy.

    The large population-scale comparison and apparent failure of modern saliency models against a trivial baseline could challenge common assumptions, but the abstract provides no quantitative margins or details establishing how broadly the result generalizes beyond news imagery and this sample.

  86. maybe AI / ML score 4.3

    Black-Box Knowledge Transfer across Distinct Feature Sets

    Oh-Ran Kwon, Daeyoung Ham

    The paper presents a way to use a pretrained black-box predictor whose inputs differ from the features available for a new task. Using many unlabeled paired examples to connect the feature spaces and a small labeled set, it separates the target into a component predictable from the black box and a residual specific to the new features; theory shows gains when that residual is small or smooth, including better minimax rates under conditions. It also combines multiple black boxes and reports simulations and real-data experiments.

    The distinct-feature, black-box transfer setting and its claim of strictly better polynomial risk rates are genuinely interesting, but the abstract gives no quantitative empirical results and the gains depend on potentially strong assumptions.

  87. maybe Tech score 4.3

    Media-over-Multipath-QUIC for Realtime Video Applications

    Tanya Shreedhar, Zuji Zhou, Nitinder Mohan et al.

    MoMQ extends Media over QUIC so multipath schedulers can use media-object metadata—such as frame importance—to choose among WiFi, cellular, and satellite paths. In a Starlink/WiFi testbed, four declarative edge rules reduced P99.9 frame-completion latency from 384.7 ms to 114.1 ms and cut the needed playback buffer by 61.5%, meeting a 150 ms interactive target where transport-only scheduling did not.

    The interesting contribution is a clean cross-layer mechanism that exposes already-available media semantics to multipath scheduling, with a large reported latency improvement, but the abstract provides limited detail on baselines, workloads, and generality beyond the testbed.

  88. maybe AI / ML score 4.3

    Weightless Fine-Tuning: Personalizing LLMs via Logit-Space Transport

    Bohan Zhang, Anqi Ni, Yixin Wang et al.

    The paper proposes a training-free personalization method that estimates how supervised fine-tuning would change a model’s logits, then applies those corrections at decoding time without modifying or storing model weights. On three LaMP benchmarks, it reportedly matches or sometimes exceeds SFT while using under 7% of the effective computation, and its logit shifts closely resemble those from SFT.

    The logit-space transport formulation could substantially reduce the per-user cost of personalization, but the evidence is limited to three benchmarks and the abstract does not establish robustness across models, sequence lengths, or more demanding adaptation settings.

  89. maybe Robotics score 4.3

    Adaptation of Generalist Robot Policies with Minimal Data

    Shreyas Kowshik, Sreyas Venkataraman, Leo Wang et al.

    MiDAS adapts a pretrained vision-language-action robot policy to a new task using only one or a few demonstrations, followed by autonomous online reinforcement learning. The method uses behavior cloning to anchor the policy and value-based RL on a residual policy; it reportedly works across LIBERO and RoboCasa and improves a fragile bimanual YAM policy over about six hours of real-robot interaction.

    Single-demonstration adaptation with reliable autonomous improvement is a meaningful step toward self-improving robots, but the abstract gives no quantitative results and the core recipe appears to combine known offline-to-online RL ingredients.

  90. maybe Tech score 4.3

    Extremely Large Beyond-Diagonal RIS: Low-Rank Modal Optimization for Near-Field Communications

    Giovanni Iacovelli, Chandan Kumar Sheemar, Eva Lagunas et al.

    The paper proposes scaling fully connected beyond-diagonal reconfigurable intelligent surfaces by optimizing in a geometry-derived low-dimensional spherical-wave modal subspace. It claims that near-field propagation makes the effective aperture response low rank, allowing a 24×24 panel to match the fully connected optimum with roughly 200 tunable entries rather than 300,000, while retaining panel-size-independent optimization cost. The main novelty is using terminal localization and near-field structure to avoid the usual quadratic/cubic hardware and computation burden.

    This is a non-obvious, potentially important scaling result for near-field wireless surfaces, but the abstract provides mainly a theoretical claim and one numerical panel-size example, so it merits a look rather than a strong recommendation.

  91. maybe Robotics score 4.3

    StellaVLA: In-Context Structured Demonstration for Generalizable Vision-Language-Action Models

    Siyu Xu, Yunke Wang, Zijian Wang et al.

    StellaVLA adapts a vision-language-action policy at test time using one retrieved demonstration that has been automatically converted into a structured task plan, subgoals, and verbalized 3D motion. The model is trained jointly on language reasoning and actions but runs only the action component at inference, preserving real-time control; it reports large gains on VLA-Arena, LIBERO, and LIBERO-Plus, plus transfer from human, robot, and XR demonstrations to real robots.

    The combination of automatically structured demonstrations, single-example OOD adaptation, and cross-embodiment transfer is a genuinely interesting direction, but the abstract gives limited detail about evaluation breadth and baseline comparability, so the strong leaderboard claims need verification.

  92. maybe AI / ML score 4.3

    Ripple-Pivot Search: Active Parallel Decoding for Diffusion Large Language Models

    Yushi Ye, Xu Chen, Haoyun Jiang et al.

    The paper proposes a training-free decoding scheduler for diffusion language models that deliberately commits a mid-uncertainty token as a “pivot,” because fixing it can reduce uncertainty in many other masked positions. Its lookahead search chooses both the pivot location and token, and reportedly yields 4–10× wall-clock speedups while preserving quality, with up to 18× when combined with KV caching and accuracy gains over a prior lookahead method.

    The proposed ripple-effect framing and large claimed inference gains are genuinely interesting for dLLM decoding, but the evidence is limited to three models and four benchmarks, so the broad practical significance is not yet established.

  93. maybe AI / ML score 4.3

    MiDashengLM-Gen: Unified Audio Scene Generation via LLM-Driven Autoregressive Flow Matching

    Xingwei Sun, Heinrich Dinkel, Gang Li et al.

    MiDashengLM-Gen presents an end-to-end model for generating mixed audio scenes containing speech, music, and sound effects from text. It couples an LLM with per-token conditional flow matching and reports a large reduction in speech WER, from 12.15% to 2.79% on Seed-TTS, while retaining competitive mixed-audio quality and supporting multiple languages. The notable claim is that one jointly trained model can handle general text-to-audio generation without the usual separate text-encoder and audio-decoder pipeline.

    The substantial speech-intelligibility improvement in a unified mixed-audio generator is genuinely interesting, but the abstract provides limited evidence beyond benchmark results and does not establish how broad or robust the gains are.

  94. maybe AI / ML score 4.3

    Small-Scale Experiments: Are We There Yet?

    Nicholas Lourie, Kyunghyun Cho, Karen Ullrich et al.

    This paper argues that small models can support useful scaling-law experiments, but only after extensive hyperparameter tuning: sensitivity is high at small scale, while the best hyperparameter region becomes simpler as models grow. It shows that tuning quality matters more than other parts of the usual scaling-law recipe, explains why extrapolation is still statistically limited, and demonstrates the approach by recovering the large-scale preference for transformer pre-normalization from small experiments.

    The potentially important result is that careful small-scale tuning—not larger models—is the key prerequisite for reliable scaling-law conclusions, but the abstract provides no quantitative breadth or cost comparison strong enough for a strong verdict.

  95. maybe AI / ML score 4.3

    Information Abundance Paradox: Long-Context Training Undermines Parametric Knowledge

    Arda Uzunoglu, Benjamin Van Durme, Daniel Khashabi

    The paper argues that giving language models more relevant information during training can make them less likely to store that information in their parameters, causing them to rely more heavily on context at inference time. Across long-document pretraining and supervised fine-tuning, it reports an intermediate context-length optimum: beyond it, language modeling, understanding, and closed-book QA decline, while robustness without or against misleading context worsens. Gradient analyses and causal interventions link this behavior to a shift from feed-forward components toward attention-based processing.

    The proposed tradeoff between context scaling and parametric knowledge is a non-obvious, practically important challenge to the assumption that longer training contexts are uniformly beneficial, but the abstract provides no quantitative details about scale, reproducibility, or breadth of models and tasks.

  96. maybe AI / ML score 4.3

    Jagged Judges: Epistemic Stability Under Perturbation, Pressure, and Persistence

    Justin Zhao, Himaghna Bhattacharjee, Hannah Korevaar et al.

    The paper proposes the “Wiggle Framework” for testing whether LLM judges change their verdicts when re-prompted, challenged once, or subjected to sustained persuasion. Across 9 frontier models and 14 evaluation tasks, judges frequently flip decisions—25–71% under static pushback and 62–91% with an adversarial persuader—and successful persuasion is usually wrong relative to ground truth. It also reports that the initial majority strength of a jury is the best single-shot predictor of which judgments will change.

    This is a useful and potentially important characterization of LLM-judge reliability, with broad cross-model evidence and striking instability results, but the abstract does not establish whether the stress tests and ground-truth methodology are robust enough to merit a stronger recommendation.

  97. maybe AI / ML score 4.3

    Fast A/B/n Testing: Exact Multi-Policy Comparison via Tree-Coupled Feedback Sharing

    Yuxiao Wen

    The paper proposes coupling multiple adaptive contextual-bandit policies so that they can share reward observations while each still has exactly the same trajectory distribution it would have in an independent run. A dynamically chosen tree links similar policy histories, and the number of reward queries is exactly one per round plus the number of edge mismatches; under low regret, this becomes T+o(T) rather than JT, with variance bounds for policy comparisons and experiments across several evaluation settings.

    The exact distribution-preserving coupling of arbitrary history-dependent policies, together with near-single-trajectory evaluation cost, is a genuinely non-obvious direction, though the abstract gives limited quantitative detail and the strong savings rely on low-regret and unique-optimum assumptions.

  98. maybe AI / ML score 4.3

    From Local Mismatch to Global Impact: Optimizing Cache Reuse Policy for Efficient Diffusion

    Xichen Ye, Yifan Wu, Zhikang Xie et al.

    The paper argues that cache-reuse decisions for diffusion inference should be based on their eventual effect on final generation quality, rather than local similarity between denoising states. GCache combines an error-propagation analysis with bilevel optimization to learn which computations can be reused and which deserve fresh computation; on image and video models, it reports better quality than prior caching methods, including 2.17× speedup on Wan2.1 while lowering LPIPS from 0.1095 to 0.0316.

    The global, trajectory-aware optimization of cache policies is a meaningful departure from local heuristics, and the reported speedup-quality tradeoff is unusually strong, but the abstract does not establish how broadly the gains hold or what baselines and evaluation protocol produce the large LPIPS improvement.

  99. maybe AI / ML score 4.3

    ChartProbe: A Diagnostic Study on Visual Reasoning through Perception, Grounding, and Simple Reasoning

    Mahsa Khoshnoodi, Sarah Adel Bargal

    The paper separates chart-question failures in vision-language models into perception, grounding, and simple one-step reasoning, using automatically generated probes with exact answers. It finds that training only these basic skills—without complex questions or reasoning traces—substantially improves performance on unseen complex chart questions, and reports transfer to pie charts, ChartQA, and CLEVR.

    The potentially important result is that much of apparent visual reasoning failure may be caused by weak perception and grounding, but the abstract gives no quantitative gains or experimental detail, so it merits a look rather than a strong recommendation.

  100. maybe AI / ML score 4.3

    From Passive Delegates to Strategic Negotiators: Reinforcing Social Reasoning in Small Language Models with SocialRL

    Wenyue Hua, Zachary Huang, Tyler Payne et al.

    The paper trains a 4B language model with reinforcement learning specifically for negotiation and delegation, rather than relying on generic helpfulness. Across six simulated social environments, domain-specific policies reportedly match or exceed much larger frontier models, while transfer-aware training and on-policy distillation produce one unified model competitive with GPT-4.1–5.2; distilling explicit theory-of-mind reasoning helps generalization, especially through next-action prediction.

    The combination of social-reasoning RL, structural cross-domain transfer, and a small model matching frontier negotiation performance is genuinely interesting, but the evidence is confined to simulated negotiation benchmarks and the abstract does not establish real-world agent capability or robustness.

  101. maybe AI / ML score 4.3

    PPAPlace: Differentiable Cross-Stage Objectives for Chip Placement Optimization

    Ruogu Chen, Jie Han

    PPAPlace trains a differentiable surrogate to predict post-route timing from both the netlist and full macro/cell placement, using post-global-routing results as training labels rather than HPWL or pre-route timing. Its gradients are used either inside an analytical placer or for macro-position refinement; on five held-out ChiPBench circuits, it reports 22% better WNS and 51% better TNS than a hierarchical baseline without retraining, while preserving power and routability.

    The combination of stage-selection evidence, graph-and-spatial differentiable modeling, and end-to-end timing optimization is a meaningful direction, but the evidence is limited to five test circuits and the gains come from a specialized EDA setting.

  102. maybe Robotics score 4.3

    Imagining Recovery: Inference-Time Counterfactual Realignment for Vision-Language-Action Models

    Yanyan Zhang, Disheng Liu, Kai Ye et al.

    CoRe is a training-free recovery method for frozen vision-language-action policies. When execution deviates, it simulates how the policy would proceed from a recent valid state, then physically realigns the robot and scene to that imagined trajectory instead of learning from failures or trying arbitrary corrective actions. Across simulators and real robots, it reportedly raises success by up to 85 percentage points and cuts physical restorations by 42.2%.

    Inference-time counterfactual recovery that unifies instruction changes and physical perturbations without retraining is a genuinely interesting direction, but the abstract does not establish how broadly the large gains generalize or how demanding the realignment procedure is.

  103. maybe Tech score 4.3

    Generative data assimilation highlights fronts as key regulators of ocean energy cascade

    Scott A. Martin, Georgy E. Manucharyan, Patrice Klein

    The authors combine satellite observations with a generative deep-learning data-assimilation method to reconstruct gap-free, kilometer-scale surface currents in the Agulhas Current. They find that submescale fronts concentrate both upscale energy transfer that strengthens mesoscale eddies and downscale transfer toward dissipation, making fronts disproportionately important despite their small area.

    The result offers a non-obvious, physically meaningful picture of ocean energy transfer and a potentially useful route to improved eddy parameterizations, but the evidence is limited to a reconstructed field in one energetic region and depends on the validity of the learned assimilation model.

  104. maybe Robotics score 4.3

    Max-Q Selective Imitation for Human-in-the-Loop Online Robot Learning

    Zihang Wang, Yishan Wang

    The paper proposes a human-in-the-loop robot learning method that uses Monte Carlo returns for chunk-level value estimation and imitates whichever is judged better by the critic: the human-demonstrated action or the robot’s own action. On a real USB pick-and-insertion task, it reportedly reaches 99% success in 30 minutes with 20 demonstrations, versus about 5 hours for HIL-SERL, with similar advantages in simulated insertion tasks.

    The hard max-Q switch between human interventions and autonomous actions is a meaningful approach to avoiding policy–demonstration distribution shift, and the reported order-of-magnitude training-speedup on a real robot is notable, but the evidence is limited to narrow tasks and abstract-level results.

  105. maybe AI / ML score 4.3

    MAPLE: MoE Adaptive Plug-and-play Layer-wise Expert allocation

    Lie Li, Wen Li, Junxiao Shen et al.

    MAPLE reallocates a fixed routed-expert budget unevenly across layers in pretrained MoE language models, without changing weights or retraining. It uses layer-sensitivity measurements to derive an initial allocation and then refines it with a guided genetic search; across four MoE models, a 75% expert budget reportedly beats uniform full-budget routing on several tasks and improves serving latency and throughput.

    The plug-and-play, post-training allocation of MoE capacity is a meaningful efficiency direction, and the reported gains over full uniform routing are surprising, but the abstract provides limited detail about evaluation breadth and whether the allocation/search overhead is practical.

  106. maybe AI / ML score 4.3

    RoofGS: Roofline-Guided End-to-End Acceleration of 3D Gaussian Splatting

    Yang Luo, Yan Gong, Yongsheng Gao et al.

    RoofGS profiles 3D Gaussian Splatting with the Roofline model and optimizes each rendering stage according to its bottleneck. It uses a 32-bit adaptive depth-sorting key for the memory-bound front end and a bounded-range fast exponential approximation for the compute-bound rasterizer, plus kernel fusion and compact storage. On an RTX 4090 at 4K, it reports 10.1× end-to-end speedup—from 61 to 616 FPS—with only a 0.028 dB PSNR decrease.

    The reported 10× high-resolution speedup is substantial and the stage-specific hardware analysis is more than a routine kernel tweak, but evidence is limited in the abstract to one GPU and one primary workload, so it falls short of a strong verdict.

  107. maybe Robotics score 4.3

    Tactile Sim2Real without Tactile Simulation via Bottlenecked Latent Reconstruction

    Fan Yang, Youngsun Wi, Jinhao Yu et al.

    The paper proposes training tactile-control policies using an easy-to-generate simulator “oracle” representation, then mapping real tactile observations into the oracle’s latent space at deployment. It uses bottlenecked reconstruction and unpaired simulation/real random-play data, avoiding sensor-specific tactile simulation and calibration; on two real tasks with GelSight Mini and DIGIT sensors, it reports 85–97.5% zero-shot success and a 7.5–15% gain over physics-based tactile simulation.

    This is a genuinely useful sim-to-real framing that could make rapidly changing tactile sensors much easier to deploy, with credible hardware results, but the evidence is limited to a few tasks and sensors and the abstract does not establish broad generality.

  108. maybe Neuroscience score 4.3

    LRRC57/RABIN is a presynaptic inhibitor of Rab GTPases in glutamatergic neurons

    Garic, D., Murphy, J. G., Teubner, B. J. W. et al.

    The study identifies LRRC57/RABIN as a brain-enriched presynaptic protein that binds active Rab GTPases and blocks their interaction with trafficking effectors. Removing it increases vesicle cycling, glutamate release, and network excitability, while overexpression suppresses release and reduces seizure-like activity, suggesting a previously unrecognized brake on excitatory synapses.

    The proposed decoy-effector mechanism is a genuinely interesting addition to the control logic of synaptic vesicle trafficking, with bidirectional genetic manipulation and seizure-related phenotypes, but the abstract lacks quantitative and independent evidence needed for a strong verdict.

  109. maybe Neuroscience score 4.3

    Arc capsid signaling between serotonergic and dopaminergic neurons sets sleep depth in Drosophila

    Butts, A. R., DeNiro, K., Bervoets, S. et al.

    This study reports that the retrotransposon-derived protein dArc forms capsids that mediate signaling between serotonergic and PAM dopaminergic neurons in flies. Removing dArc, or disrupting associated Sas–Ptp10D components, makes sleep deeper and harder to interrupt, while restoring capsid-forming dArc reverses the effect; the circadian timing of sleep is unchanged.

    It presents unusually direct evidence that Arc capsid-based intercellular communication can regulate an ongoing behavior, but the evidence is confined to a Drosophila circuit and the abstract does not establish the transferred RNA or the molecular signaling process in detail.

  110. maybe Neuroscience picked score 4.3

    Neural prediction decorrelation reveals that adversarial robustness substantially improves DNN prediction accuracy across the entire human auditory cortex

    Skrill, D., Feather, J., Norman-Haignere, S. V.

    The authors introduce neural prediction decorrelation (NPD), a stimulus-synthesis method designed to make competing neural encoding models produce different predictions, then use fMRI to compare standard and adversarially robust auditory DNNs. Robust models predicted responses across all tested auditory-cortex regions better for these diagnostic sounds, despite the difference being invisible with ordinary natural sounds; their low-dimensional organization also matched population responses.

    This combines a potentially general method for resolving indistinguishable encoding models with a non-obvious result that adversarial robustness improves neural predictivity, but the abstract provides no sample sizes or quantitative effect sizes, so the strength and generality of the claim need verification.

  111. maybe Neuroscience score 4.3

    Retinal degeneration in a mouse model of CRB1 disease rescued by the photoreceptor-specific CRB1-B isoform

    Dembla, E. M., Valdez-Lopez, J. C., Kozlowski, C. et al.

    This study identifies CRB1-B as the retinal isoform expressed specifically by photoreceptors and tests it in a mouse model of CRB1-associated degeneration. Restoring CRB1-B in photoreceptors prevented loss of photoreceptor–glia adherens junctions and reportedly completely rescued degeneration and visual function, suggesting that this isoform alone may be sufficient for therapy.

    The photoreceptor-specific isoform both explains a disease mechanism and produces an apparently complete functional rescue in mice, but the evidence is still preclinical and the abstract gives too little detail to justify a strong verdict.

  112. maybe Neuroscience score 4.3

    Reversible m6Am methylation of snRNA by FTO controls morphine reward and tolerance without altering analgesia

    Liu, S., Malik, A. F., Chien, R. et al.

    The study identifies the RNA demethylase FTO as a regulator of morphine reward and opioid tolerance in mice. Genetic or pharmacological FTO inhibition reduced reward and tolerance to morphine and fentanyl while preserving analgesia; experiments point specifically to increased m6Am modification on small nuclear RNAs rather than the m6A marks typically associated with FTO. The m6Am dependence and separation of analgesia from reward/tolerance are the main new findings.

    This is a non-obvious mechanistic link between snRNA m6Am regulation and opioid reward/tolerance, with a potentially important analgesia-sparing effect, but the evidence is preclinical and the abstract does not establish translational efficacy or safety.

  113. maybe Neuroscience score 4.3

    Selective GRAB sensors reveal distinct endocannabinoid dynamics in vivo

    Cai, R., Yang, Y., Cai, S. et al.

    The authors developed two genetically encoded fluorescent sensors that separately measure the endocannabinoids 2-AG and anandamide in living tissue and behaving mice. The sensors revealed different ligand dynamics across stimuli, including a sustained 2-AG response to Δ9-THC in the nucleus accumbens; deleting the 2-AG-synthesizing enzyme reduced both this signal and THC-induced hypolocomotion.

    Selective real-time measurement of the two major endocannabinoids addresses a longstanding technical limitation and supports a non-obvious endogenous 2-AG contribution to THC behavior, though the abstract provides limited validation and scope for the broader claims.

  114. maybe Neuroscience score 4.3

    Combination spinal cord stimulation at different frequencies produces sustained pain relief with immune activation

    Huh, Y., Song, S., Chen, T. et al.

    In mouse and rat nerve-injury models, combining 60 Hz and 1 kHz spinal cord stimulation produced pain relief that continued for days to weeks after stimulation stopped, rather than only during active stimulation. The authors link this persistence to spinal immune changes, especially neutrophils and S100A8–CD69/SOCS3 signaling; neutrophil depletion weakened the effect and S100A8 reproduced it.

    The potentially important result is that a frequency-combination SCS protocol induces prolonged analgesia through a neuroimmune mechanism, but the evidence is limited to preclinical injury models and the abstract gives little detail on effect sizes, durability, or translational feasibility.

  115. maybe AI / ML ▲ 19 score 4.3

    MegaParts: Scaling Part-Aware 3D Object Generation to 300 Parts via Token-Efficient Autoregressive Modeling

    Manwen Liao, Xinyu Lian, Jian Mao et al.

    MegaParts presents an autoregressive system for generating part-structured 3D objects with up to 300 parts. It uses adaptive-length discrete tokens for each part’s geometry, jointly generates object and part layouts, and applies long-context training to reach sequences up to 256k tokens; the authors report better mesh quality than autoregressive and diffusion baselines. The main claimed contribution is showing that token compression can make large-scale, controllable part generation feasible without abandoning autoregressive modeling.

    The combination of adaptive geometry tokenization and very long structured sequences is a meaningful scaling direction, but the abstract provides no quantitative results, dataset breadth, or convincing evidence for the claimed advantage over diffusion.

  116. maybe AI / ML ▲ 10 score 4.3

    Multimodal Model Diffing for Feature Discovery and Control

    Hunar Batra, Lachin Naghashyar, Ashkan Khakzar et al.

    The paper introduces MMDiff, which compares sparse autoencoder features in a base language model and its multimodal version to identify features changed by visual training. It also uses contrastive activation analysis to find task-specific features and then removes or steers them; across three MLLM families, this selectively affects spatial reasoning and OCR, improves steering over a single-layer baseline, and reduces multimodal attack success without hurting VQA.

    The base-versus-multimodal SAE diffing framework and feature-level causal control are a meaningful extension of interpretability methods, with experiments across several model families, but the reported gains are moderate and the abstract does not establish broad capability or robustness beyond a few tasks.

  117. maybe AI / ML ▲ 28 score 4.3

    PACE-Bench: Benchmarking Physics Adaptation via Code Evolution in Dynamic Environments

    Yuhao Zhan, Bingxiang He, Zecong Tang et al.

    PACE-Bench tests whether code-generating agents can adapt designs when the underlying physics changes, using 144 source-to-target simulator pairs across six domains. Agents must diagnose failures and modify their code under a limited budget; performance is low, and exposing the exact physical changes does not remove the difficulty, suggesting that redesigning mechanisms—not merely estimating new parameters—is the main challenge.

    The benchmark itself is a relatively narrow evaluation contribution, but the evidence that even knowing the physics change does not solve adaptation—and that reflection beats unverified revision—offers a non-obvious insight into the limits of self-evolving agents.

  118. maybe AI / ML ▲ 46 score 4.3

    PlayWorld: Benchmarking World Models with Agent Players over Long-Horizon Objectives

    Kaixin Ding, Xi Chen, Minghong Cai et al.

    PlayWorld evaluates interactive video world models by having multimodal agents pursue specified long-horizon goals, rather than comparing models on identical action sequences. Across 171 scenarios and nine current models, it finds that models often fail to preserve spatial consistency and persistent state changes, even when short-term video quality and action controllability look strong. The benchmark separates geometry, interaction, out-of-sight, and state-evolution behavior.

    The agent-driven evaluation protocol targets a real weakness of fixed-action benchmarks and reports broad evidence that long-horizon world-model reliability remains substantially worse than surface video quality suggests, though it is still primarily a benchmark rather than a new modeling advance.

  119. maybe AI / ML ▲ 16 score 4.3

    Ex-Omni-2D: Expressive Omni-Modal Dialogue Models with Native Visual Presence

    Haoyu Zhang, Zhipeng Li, Xiaoying Tang et al.

    Ex-Omni-2D is an omni-modal dialogue system that produces text, personalized speech, and a synchronized talking-avatar video response. It introduces a visual planning representation and uses shared speech units to align audio with video, then distills a full-sequence video model into a four-step streaming model; the reported pipeline runs at 1.293× real-time factor on 400×720/720×400 video.

    The shared acoustic-temporal interface and streaming distillation for coordinated speech-video dialogue are plausible new system ideas relevant to interactive omni-modal models, but the abstract gives no quality comparisons or evidence that the claimed coordination is substantially better than existing avatar-generation pipelines.

  120. maybe AI / ML ▲ 15 score 4.3

    The Embedder's Dilemma: LLMs Are Better, but at What Cost?

    Adnan El Assadi, Niklas Muennighoff, Jinhyuk Lee

    This paper compares 10 LLMs and 26 embedding models across 37 tasks, finding that their best overall quality is nearly tied, but their strengths differ: LLMs help most on reasoning-heavy retrieval, while embedding models are better for classification and far cheaper. LLM inference costs up to 1,431 times more and is substantially slower, with reasoning tokens responsible for much of the cost; reducing reasoning budgets often preserves retrieval quality.

    The broad, cost-controlled comparison yields a useful and somewhat surprising result—that LLMs do not generally outperform dedicated embedders despite their greater capability—but it is primarily a benchmark and synthesis rather than a new method.

  121. maybe AI / ML ▲ 8 score 4.2

    LittleLearner: Language Models Under Pedagogically Controlled Knowledge Exposure

    Fanfei Li, Jana Zeller, Manuel Prada-Corral et al.

    The paper introduces an 88B-token corpus drawn from U.S. elementary-school material, deliberately excluding concepts and vocabulary above Grade 5, and trains a 5B-parameter model on it. The resulting model has ordinary language competence but interpretable limits on knowledge and skills, providing a controlled testbed for studying acquisition, post-training, and in-context learning; initial experiments suggest these methods improve use of existing knowledge without creating genuinely out-of-scope capabilities.

    The carefully bounded pretraining environment is a genuinely useful research direction for disentangling learned knowledge from prior exposure, but the abstract gives limited quantitative evidence and the contribution is primarily a released sandbox rather than a demonstrated major scientific finding.

  122. maybe AI / ML ▲ 57 score 4.2

    Beyond Final Scores: A Systematic Evaluation of Agents for Long-Horizon AI Research and Development

    Yiwei Li, Wanli Yang, Hexiang Tan et al.

    The paper evaluates seven frontier models on 36 long-horizon AI research and engineering tasks using process-level metrics, rather than judging only final performance. It finds that agents can implement and optimize practical solutions, but are inconsistent across runs, rarely produce genuinely new methods, and reuse prior experience in ways that can either help or mislead them.

    It offers a relevant process-oriented evaluation of autonomous R&D agents and a useful diagnosis of bottlenecks, but it is primarily a benchmark/evaluation study and the abstract gives too little quantitative detail to establish a major finding.

  123. maybe Robotics score 4.2

    Topological Feasibility Guarantees for Differentiable Predictive Control

    Guangyu Wu, Ján Drgoňa

    The paper analyzes differentiable predictive-control policies through the topology and geometry of their reachable safe sets, and proposes training with a Control Barrier Function proxy loss. It claims that, under its structural assumptions, a finite number of offline training samples can yield deterministic feasibility guarantees without an online safety filter; simulations show constraint violations decreasing to zero as sample count grows.

    Deterministic, offline feasibility guarantees for learned MPC policies would be important and the topological framing appears novel, but the abstract provides no theorem conditions, quantitative scope, or real-system evidence, so the unusually strong claims need close inspection.

  124. maybe AI / ML score 4.2

    Multi-Agent Closed-Loop Reasoning for Organic Structure Elucidation from Multimodal Spectra

    Bingsen Xue, Zhuojun Jiang, Jianhao Zhang et al.

    MACROS is a multi-agent, closed-loop system for identifying molecular structures from arbitrary combinations of spectroscopic data. Trained on large simulated and experimental spectra datasets, it reportedly generalizes to unseen real compounds—including large natural products and metabolites—and recovers familiar spectroscopic reasoning patterns without assigned inputs. The abstract also claims sixfold faster and 40% more accurate collaboration with chemists, but gives few details about benchmarks or evaluation design.

    The combination of multi-agent hypothesis testing, multimodal spectra, and claimed zero-shot elucidation could represent a meaningful new AI capability, but the unusually broad claims and lack of concrete comparative results make the evidence too weak for a strong recommendation.

  125. maybe AI / ML score 4.2

    SoftWater: Class-Aware Rate Allocation for Softmax Quantization

    Joao V. Cavalcanti, Ashia C. Wilson

    SoftWater treats quantizing an LLM’s softmax head as a rate–distortion problem using output-distribution KL divergence, rather than ordinary weight error. It allocates finer quantization to frequent, sensitive vocabulary classes and coarser quantization to rare ones, achieving 6.5–8.3× lower head-induced KL than a prior quantizer at 2 bits across five models, while reducing storage substantially. On Llama-3.2-1B, a 2-bit head causes a 2.9–3.7% perplexity increase, and a 4-bit head is nearly lossless in a tied model.

    This is a technically distinct, practically relevant approach to a neglected source of quantization cost, backed by broad matched-rate experiments, but it remains a specialized compression advance rather than a clearly field-changing result.

  126. maybe AI / ML picked score 4.2

    CAKE: Compiler-Agent Co-Design for Frontier Kernel Evolution

    Zihao Ye, Yingyi Huang, Hongyi Jin et al.

    CAKE jointly designs a GPU-kernel agent workflow and an intermediate representation that exposes hardware scheduling, synchronization, memory movement, and pipelines while giving the agent verification and cost feedback. On NVIDIA GPUs, it reportedly beats a tuned FlashML baseline on Flash-KMeans, gets a 2.05x geometric-mean speedup over FlashKDA, and improves KNN/KMeans across more than 400 shapes, with several changes prepared as upstream PRs.

    The compiler-agent co-design and evolving verifier/IR are a substantive direction with real kernel and multi-shape results, but the headline gains are reported by the authors on a limited set of workloads and do not yet establish broad superiority over expert optimization.

  127. maybe AI / ML score 4.2

    vToken: Token-Level Virtualization for Reclaimable KV Caches

    Yuanhang Gao, Xiangrui Yang, Yuanfeng Chen et al.

    vToken adds a token-level indirection layer on top of block-based KV-cache allocation, allowing dead or evicted tokens to be reclaimed even when they share blocks with live tokens. It asynchronously repacks live tokens while retaining existing PagedAttention kernels and CUDA Graph support. In a vLLM implementation, it reduces retained KV blocks by 27.2%–72.3%, raises SLA-constrained throughput by up to 1.37×, and can double feasible concurrency under a fixed KV-memory budget.

    This addresses a real systems mismatch between token-level eviction and block-level allocation, with substantial reported memory and throughput gains and low integration cost, though it is an optimization layer rather than a major new serving paradigm.

  128. maybe AI / ML score 4.2

    AdsWorldEngine: A Self-Evolving Conversational Advertising Agent through Orchestrator and Tool Coevolution

    Simiao Zuo, Chenhui Xu, Yimeng Jia et al.

    AdsWorldEngine is a conversational advertising system that decides when an ad is appropriate, infers commercial intent from dialogue, selects ads through tools, and trains both the orchestrator and those tools in an iterative reinforcement/preference-learning loop. The authors report large offline gains in diversity and relevance, plus a 22% RPM increase and 74% higher ad coverage in an online A/B test, although the abstract gives little detail about baselines, scale, or experimental controls.

    The actor-and-tool coevolution setup and reported production A/B gains are worth a look, but the contribution is tightly tied to advertising and the abstract does not establish how much is a general agent-training advance versus system engineering.

  129. maybe AI / ML score 4.2

    QUASAR: Lowering the Loss Floor of Quantization-Aware Training with Loss-Aware Reconstruction

    Vincent Counathe, Ben Athiwaratkun, Christopher De Sa et al.

    QUASAR adds lightweight, loss-aware weight reconstruction directly inside quantization-aware training, using gradient-based saliency, clipping-range search, and affine dequantization fitting. Across Qwen3 and Llama-3.1 at 2–4 bits, it reports the best held-out KL divergence among tested QAT methods, with especially large gains at 2 bits and 3.5–4.3 percentage-point average accuracy improvements across eight tasks.

    The online reconstruction objective is a substantive and principled improvement over standard QAT, with unusually strong reported gains at very low precision, but it remains a specialized quantization-training method and the abstract does not establish broad superiority beyond the tested models and baselines.

  130. maybe AI / ML score 4.2

    Quantum Multi-Armed Bandits and Linear Bandits: Lower Bounds and Algorithms

    Maoli Liu, Zhuohua Li, John C. S. Lui

    This paper proves near-matching lower bounds for quantum multi-armed and finite-action linear bandits: regret must grow as Ω(K log(T/K)) and Ω(d log(T/d)), so quantum access cannot eliminate the usual logarithmic-in-horizon cost. It also gives a design-based elimination algorithm for quantum linear bandits whose regret is linear in dimension for action sets of polynomial size, improving the previous quadratic dependence up to polylogarithmic factors.

    The combination of a first horizon-dependent quantum bandit lower bound and a substantial d²-to-d algorithmic improvement is genuinely notable, but the results are confined to an idealized quantum-oracle model and have mainly theoretical rather than immediate practical impact.

  131. maybe Robotics score 4.2

    Grouping Auction-Consensus Algorithm for Decentralized Task Allocation in Multi-Robot Systems

    Jose Rodriguez, Sven Koenig, Wenjie Dong et al.

    The paper modifies CBBA for multi-robot task allocation by having robots bid on spatially coherent groups of tasks rather than individual tasks. In experiments across thousands of simulated worlds and swarm sizes up to 20, the method reached about 97% of MILP-optimal allocations versus 81–84% for CBBA, with no more consensus iterations. The main contribution is a structured group-level bidding and consensus mechanism aimed at better matching the min-sum travel objective.

    The large and broadly evaluated improvement over a standard decentralized baseline is notable, but the core grouping heuristic is relatively plausible and the abstract provides no real-world validation or detailed computational-cost analysis to justify a strong verdict.

  132. maybe AI / ML ▲ 11 score 4.2

    InSight-doc: Agentic Visual Perception for Long-Document Understanding

    Kaican Li, Weiyan Xie, Lewei Yao et al.

    InSight-doc trains an 8B document-understanding agent to begin with low-resolution page views and selectively zoom into regions that need closer inspection, using supervised trajectories and reinforcement learning. The authors report 4.3–16.4 point gains on document VQA, over 40% lower hallucination on long documents, and 41–68% lower latency, though the abstract gives limited detail about benchmarks and baselines.

    Adaptive visual computation for long-document agents is a useful and reasonably novel direction, with potentially large latency gains, but the abstract does not establish how broad or robust the results are beyond benchmark claims.

  133. maybe AI / ML ▲ 11 score 4.2

    Hybrid-Policy Self-Editing for Composable Unstructured Knowledge Editing

    Tianci Liu, Zihan Dong, Tianchun Li et al.

    This paper treats unstructured knowledge editing as a self-distillation problem rather than simply inserting a passage into the model. Its HPSE method mixes the model’s own rollout with targeted insertion of missing facts, aiming to make edited knowledge usable for atomic questions and multi-hop composition without external supervision; experiments cover four LLM backbones and two editors, though no quantitative results are given in the abstract.

    The hybrid on-policy/off-policy formulation addresses a real limitation of unstructured editing and appears technically distinct, but the abstract provides insufficient quantitative evidence to establish a substantial practical improvement.

  134. maybe AI / ML ▲ 6 score 4.2

    Simplex Relaxation for Discrete Diffusion

    Jinya Sakurai, Patrick Pynadath, Satoshi Hayakawa et al.

    The paper augments uniform categorical diffusion with an auxiliary Dirichlet-simplex variable, while keeping the original categorical corruption process unchanged. This enables a Rao–Blackwellized reverse-training objective and stochastic sampler; experiments report better perplexity–entropy tradeoffs on OpenWebText and stronger Sudoku accuracy across clue densities, including 17-clue puzzles, when trained only on 30-clue instances.

    The exact simplex augmentation and tractable reverse-bridge formulation are a substantive new direction for discrete diffusion, and the reported Sudoku cross-density generalization is non-obvious, though the abstract gives limited quantitative evidence and unclear breadth beyond two tasks.

  135. maybe AI / ML ▲ 47 score 4.2

    UI-Mate: Advancing Open-Weight Foundation GUI Agents with In-Context Demonstrations

    Zihan Ding, Longxu Dou, Qi Gao et al.

    UI-Mate is an open-weight GUI agent trained with a large automated environment-and-verification pipeline, plus a mechanism for using multimodal demonstrations as subtask workflows while replanning from the live interface. Its 27B model reports 77.0% on OSWorld-Verified and 66.2% on WindowsAgentArena, while a single demonstration roughly doubles strict success on its 33-task self-demo subset, from 17.2% to 35.4%.

    The potentially useful result is that demonstrations substantially improve long-horizon GUI execution, but the core contribution appears to be a fairly broad training-and-agent system integration, with evidence largely limited to reported benchmark and technical-report results.

  136. maybe AI / ML ▲ 9 score 4.2

    Dion3: Full-Stack Orthogonal Updates

    Noah Amsel, Jack Zhang, Kwangjun Ahn et al.

    Dion3 is a faster implementation and variant of the Muon optimizer. It reduces orthogonalization and distributed-communication costs through a cheaper Gram-based Newton–Schulz procedure, optimized kernels, megabatching, and orthogonalizing only some momentum rows per update. The authors report matching or improving Muon’s loss while reducing optimizer step time by up to 6×, with a drop-in implementation.

    The reported up-to-6× optimizer speedup and full-stack treatment could matter in large-scale training, but the abstract gives limited detail about model scales, workloads, and how broadly the performance and loss claims hold.

  137. maybe AI / ML ▲ 26 score 4.2

    StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization

    Yuyang Yin, Zixiang Li, Longxuan Deng et al.

    StateFlow treats generative previsualization as editing a persistent, structured 3D world rather than regenerating each image or video from prompts. It constructs a 3D state from generated views, applies user edits as structured state transitions while preserving scene memory, and uses render feedback to refine camera trajectories; the abstract reports qualitative success for video creation and game-like prototyping.

    The persistent 3D working-state framing and explicit separation of construction, evolution, and camera access are a plausible step beyond one-shot generation, but the abstract provides no quantitative comparisons or details establishing a substantial capability gain.

  138. maybe AI / ML ▲ 5 score 4.2

    Gaze Target Estimation Anywhere with Concepts

    Xu Cao, Houze Yang, Vipin Gunda et al.

    The paper introduces promptable gaze target estimation: a single model uses text or visual prompts to identify which person to analyze, then predicts that person’s gaze target and whether it lies outside the image. It provides a 120K-example Gaze-Co benchmark and GazeAnywhere, which jointly performs subject localization and gaze estimation rather than relying on a brittle multi-stage pipeline; the abstract reports strong results, including on an out-of-domain clinical dataset.

    The concept-driven, natural-language interface to gaze estimation is a meaningful new task framing with an end-to-end model, but the abstract gives no quantitative gains and the practical impact beyond a new benchmark and baseline is not yet clear.

  139. maybe AI / ML ▲ 15 score 4.2

    Scaling Domain Data Repetition in LLM Pretraining

    Jingwei Li, Xinran Gu, Rui Dai et al.

    The paper studies how often scarce, high-quality domain data should be repeated during LLM pretraining as models and token budgets grow. It finds that optimal repetition rises slightly with model size at fixed tokens-per-parameter, and is better predicted by a domain’s validation loss than by the amount of unique data; small-model experiments may therefore help tune repetition for larger models.

    The counterintuitive result that easier-to-model domains benefit from more repetition, and that proxy-model tuning transfers across scale, is practically relevant, but the abstract gives no quantitative effect sizes or evidence of broad robustness.

  140. maybe AI / ML ▲ 8 score 4.1

    Don't Scroll Back: Missing-Evidence Memory for Streaming Dialogue Summarization

    Hyangsuk Min, Hwanjun Song

    This paper frames streaming dialogue summarization as recovering the historical evidence that the current context implicitly depends on, rather than simply retrieving more past text. It introduces a benchmark separating evidence recall from whether summaries use that evidence, plus ReMEMBER, which retrieves and compresses history around unresolved dependencies; experiments up to 160K tokens report better memory recall and gap resolution at a fixed budget, though no numerical gains are given in the abstract.

    The missing-evidence formulation and separate evaluation of retrieval versus summary use are a useful conceptual contribution, but the abstract provides insufficient quantitative evidence to establish a major advance.

  141. maybe AI / ML ▲ 4 score 4.1

    UniProbe: A Learnable Token-Level Hallucination Detector for Large VLMs using Multi-Structural Internal Representations

    Dvir Samuel, Guy Bar-Shalom, Fabrizio Frasca et al.

    UniProbe detects hallucinated tokens in vision-language model outputs using the model’s frozen internal activations and attention. It represents image patches, prompt tokens, and generated tokens as a structured trace, combining graph, visual, and sequential processing, and can resample suspicious tokens during decoding. Across several LVLMs, it reportedly reduces object hallucinations by up to 55% with only 1.06× generation latency, though the abstract gives few details about baselines and evaluation breadth.

    The unified use of structured internal traces for token-level detection plus near-real-time corrective decoding is a meaningful idea, but the abstract’s evidence is limited to broad state-of-the-art claims and a best-case improvement number.

  142. maybe AI / ML ▲ 12 score 4.1

    SPIEval: Evaluating Large Language Models as Mobile Assistants over Scattered Personal Information

    Junjie Ye, Zhuohui Sheng, Shaofan Liu et al.

    SPIEval is a benchmark for testing whether LLM-based mobile assistants can retrieve and combine personal information scattered across 10 apps while handling reasoning, ambiguity, preferences, and multi-intent requests. Across 250 tasks and nine models, performance is weak: the best model reaches 57.3% accuracy, and 79% of failures come from stopping at plausible but incorrect information rather than verifying it; advanced search is used in under 2% of retrieval actions.

    Although primarily a benchmark paper, its realistic cross-app setting and specific finding that premature, incorrect information localization dominates failures make it useful for understanding a poorly measured limitation of LLM agents.

  143. maybe AI / ML score 4.1

    Not All Visual Tokens Are Equally Safe to Remove:Consequence-Sensitive Visual Token Compression

    Jingbo Wen, Liang He, Mingyu Cao et al.

    The paper proposes allocating visual-token budgets based on the cost of making an error, rather than treating every VLM question equally. It calibrates error-versus-budget curves offline and uses task or question information to give more computation to high-consequence queries; on a controlled document benchmark, high-stakes errors fall from 0.300 to 0.133 at the same total token budget, with reported generalization across datasets, VLMs, and compression mechanisms.

    The consequence-sensitive allocation framing and within-image control are genuinely non-obvious, but the advance is primarily a risk-aware inference policy rather than a major new VLM capability, and the abstract does not establish broad real-world cost-model validity.

  144. maybe AI / ML score 4.1

    Privileged Likelihood Is Not Automatically Value: Three Checks for Token Credit in On-Policy Self-Distillation

    Xuan-Phi Nguyen, Zeyu Leo Liu, Yang Li et al.

    The paper tests whether token-level likelihood scores from privileged feedback are actually useful proxies for token credit in on-policy self-distillation. On AIME 2025, the scores were nearly unrelated to correctness (AUC 0.505), concentrated training pressure on high-entropy tokens, and produced much worse results than outcome-only GRPO; on Biology, where score alignment was strong (AUC 0.81–0.92), the method improved held-out performance by 28%.

    It provides a useful, empirically supported warning that privileged likelihood is not automatically token value, with a striking failure on mathematical reasoning and a concrete diagnostic for when the method may work.

  145. maybe AI / ML score 4.1

    Temporal Misgrounding in Legal RAG: A Versioned-Corpus Benchmark for French Tax Law

    Rose Cymbler, Daniel Guez, Laurent Fabre

    The paper studies a failure mode in legal RAG where systems retrieve the currently valid law even when a question requires a past or future version. On a versioned French tax-law benchmark, eleven models achieve about 3% accuracy with parametric knowledge and 2.7% with current-version RAG, while a version-aware retriever reaches 98.3%; the remaining errors are mostly article-retrieval failures. It also releases the versioned corpus, benchmark, and citation-link data.

    The near-total failure of static RAG on temporally indexed questions and the large gain from explicit version-aware retrieval are genuinely non-obvious, though the evidence comes from a narrow French tax-law benchmark and may not generalize broadly.

  146. maybe AI / ML score 4.1

    Evaluating Generative Time-Series Models on Data with Point Masses

    Jian Xu

    This paper shows that standard rolling-origin evaluation can badly misrepresent generative time-series models when the data has a large point mass, such as zero demand or no rainfall, because evaluation windows may have very different occurrence rates from the full dataset. It proposes matched evaluation and a CRPS-preserving control that destroys temporal dependence, finding that an autoregressive hurdle model outperforms a conditional flow by up to 153x on five of six datasets, while model rankings depend strongly on the occurrence statistic used.

    The paper identifies a concrete but likely underappreciated evaluation failure, demonstrates that it can reverse conclusions, and provides useful controls and multi-dataset evidence, though the impact is mainly on a specialized time-series benchmarking practice rather than a broad modeling advance.

  147. maybe AI / ML score 4.1

    Decodable but Not Accessible: Auditing Distance-Based Reliability Estimation on Disentangled Skin-Lesion Representations

    Duc-Vinh Tran

    The paper tests whether changes in representation geometry actually make distance-based reliability or OOD-confidence scores more useful. Across representations trained with increasing orthogonality pressure, geometry changed substantially, but seven of eight reliability scorers remained poor (Mahalanobis AUROC about 0.40), even though a supervised probe could still decode domain membership from the same embeddings at 0.72–0.81 AUROC. The main result is that information can remain present and decodable while being inaccessible to simple downstream reliability estimators.

    The controlled dose-response audit and decodable-versus-accessible distinction provide a non-obvious warning about representation geometry and reliability estimation, though the evidence is limited to one skin-lesion setting and a relatively small representation family.

  148. maybe AI / ML score 4.1

    GENCO - A Unified Neural Solver Embedded in a Development Framework for Steady-State Grid Analysis

    Alban Puech, Matteo Mazzonelli, Tamara R. Govindasamy et al.

    GENCO is a shared neural architecture for power flow, optimal power flow, and state estimation, paired with a framework for generating data and training grid solvers. The authors report substantial speedups over Newton–Raphson and IPOPT while maintaining or improving feasibility, and better robustness than weighted least squares under noisy or incorrect grid parameters, including tests on Hydro-Québec SCADA data.

    The unified solver and reported large speedups across several physically constrained grid-analysis tasks are potentially important, but the abstract does not establish how well the neural approach generalizes beyond its training distributions or whether the headline comparisons are fully apples-to-apples.

  149. maybe AI / ML score 4.1

    Off-Axis, On Purpose: Where a Transformer Computes Concepts and Why it Does So

    Mark Oskin

    The paper argues that a transformer’s intermediate representations deliberately compute in a subspace nearly orthogonal to the direction used to read out token probabilities. In experiments with a 12-layer model, forcing computation onto the read-out direction damages cross-token composition, while a late additive step brings the answer onto that direction; the authors also show that this off-axis workspace can be imposed with rotations without hurting benchmark performance. The results suggest that internal concept representations have substantial basis freedom, though the evidence is limited to a small model and a narrow set of tasks.

    The specific geometric account of a protected, rotating concept-computation subspace—and the demonstration that its basis can be arbitrarily prescribed—is genuinely interesting, but the conclusions rely mainly on one small transformer and do not yet establish broad architectural or practical significance.

  150. maybe Robotics score 4.1

    Hip Energized Monopedal Hopping

    Shane Rozen-Levy, Griffon McMahon, Daniel Koditschek

    The paper shows that a monoped can use the reaction torques generated by pitch stabilization as an energy source, rather than treating them only as a corrective cost. A controller that shifts the center of mass and reallocates energy between radial and angular motion produces stable hopping, with analytical predictions matching simulations and Penn Jerboa experiments at 1.02–1.77 m/s.

    It presents a genuinely interesting control/robotics insight—turning stabilization torques into propulsion—and supports it with hybrid analysis, simulation, and hardware experiments, but the demonstrated scope is limited to a specialized planar hopping robot.