Issue 28 · Jul 6–12, 2026
Every candidate
All 2,782 papers were scored from their abstracts by gpt-5.6-luna; 972 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
Odors Smell Like Their Components: A Linear Framework for Predicting Olfactory Mixture Perception
The authors measured human perception of 432 odor mixtures built from 144 component odorants and found that mixture qualities were usually predicted by simply averaging the components’ quality profiles. This linear model outperformed prior methods and nearly reached the measurement-noise ceiling, including for mixtures previously thought to show emergent, nonlinear odors.
It directly challenges the assumption that odor-mixture perception is dominated by complex nonlinear interactions, using a relatively broad human study and a result that could make computational odor prediction and reconstruction tractable.
-
Multiplayer Interactive World Models with Representation Autoencoders
The paper builds a 5-billion-parameter latent diffusion world model that predicts Rocket League video while conditioning on the actions of all four players, rather than treating other players as part of the environment. Trained on 10,000 hours of bot gameplay, it reportedly generates matches in real time and remains visually and physically coherent for at least five minutes despite training on short clips. It also analyzes scaling, representation choices, multiplayer conditioning, and physical-understanding evaluations, with code and data released.
Explicitly modeling multiple interacting action streams in a fast, tightly coupled environment is a genuinely interesting world-model direction, but the abstract provides few quantitative comparisons or details supporting the unusually strong long-horizon and physical-consistency claims.
-
Harnessing Code Agents for Automatic Software Verification
The paper wraps a general-purpose code agent in a Coq/Lean verification harness and lets it choose proof strategies freely, rather than prescribing retrieval and tactic-generation pipelines. It reports that Claude Opus 4.7 proves all 4,257 core Iris lemmas, 217 Rust-library verification lemmas, all 318 reglang targets, and 72 previously unported Iris-to-Lean lemmas, with the prover kernel enforcing soundness and completeness.
The reported jump from roughly 1/8 theorem coverage to complete coverage across thousands of formal lemmas and two proof assistants would substantially change expectations about LLM-driven software verification, though the cost, engineering effort, and independent reproducibility need scrutiny.
-
Is Randomness Necessary for Adaptive Data Analysis?
This paper proves that randomness is fundamentally necessary for answering many adaptively chosen statistical queries when the analyst is computationally unbounded. Whereas randomized mechanisms can support roughly n² queries, the authors show that every deterministic mechanism can be forced to fail after only about n queries, up to logarithmic factors.
It resolves a long-standing foundational question in adaptive data analysis and shows a sharp, surprising separation between randomized and deterministic mechanisms, although the abstract provides limited detail about the proof and scope of the lower bound.
-
A Function-Space Dichotomy for Compositional Learning: Exponential Sub-Optimality of the Neural Tangent Kernel
The paper gives a theoretical explanation for why finite-width, trained ReLU networks can substantially outperform their neural tangent kernel (NTK) limit on compositional tasks. For targets such as an iterated sawtooth, it proves an exponential sample-complexity gap: NTK regression requires on the order of 4^L samples, while an appropriately bounded depth-L network achieves a polynomial-in-L minimax rate; experiments also show large gaps on sparse parity but not on smooth bandlimited targets.
The paper identifies and quantifies a function-space mechanism behind a major known failure mode of NTKs, including an exponential separation rather than a marginal benchmark improvement, with supporting theory and experiments.
-
High-Dimensional Procrustes Matching via Tree Counts
This paper studies matching two sets of high-dimensional Gaussian vectors when one is a rotated and permuted, only constantly correlated version of the other. It gives a polynomial-time exact-recovery algorithm based on weighted counts of specially designed wide trees, succeeding for constant correlation when the dimension is at least polylogarithmic in the number of points, and improves the information-theoretic threshold as well. A low-degree analysis suggests the tree-counting threshold is intrinsic to that class of methods.
It opens a previously poorly understood high-dimensional regime with a polynomial-time constant-correlation recovery result, while tying the algorithm to a distinctive tree-counting threshold and providing supporting information-theoretic and low-degree bounds.
-
Brain2voice 2.0: High-performance voice synthesis brain-computer interface
The authors decode intracortical neural activity into both phonemes and acoustic features with a causal multimodal Transformer, producing synthesized speech every 10 ms. On a prior brain-to-voice benchmark, listeners transcribed the output at 5.24% word error rate, versus 43.75% for the previous system, while phoneme error was 7%.
An eightfold intelligibility improvement that brings real-time neural voice synthesis near practical conversational quality is a major capability jump, although the evidence is limited by evaluation on a prior benchmark and unspecified participant breadth.
-
Gauge dependence and structured-output corruption in sign-branched repetition penalties: measurements across models, inference stacks, and alternative repetition controls
The paper shows that the common multiplicative repetition penalty is sensitive to the arbitrary zero-point of a model’s logits because it branches differently on positive and negative raw logits. Re-centering logits—which leaves the model’s probabilities unchanged—therefore changes 58–96% of greedy tokens at a typical penalty setting, and the penalty reduces valid JSON-schema output from 97% to 23% in their tests. A penalty applied after logit normalization avoids both problems, and the authors reproduce the issue across several models and inference engines.
This identifies a widespread, previously hidden inference bug whose behavior depends on an arbitrary logit gauge, with large cross-stack effects on structured generation and a simple principled alternative.
-
Video Generation Models are General-Purpose Vision Learners
GenCeption repurposes a large text-to-video diffusion model as a text-steered, feed-forward vision system rather than using it only for synthesis. It reports strong performance across depth, normals, pose, segmentation, and 3D keypoints, with claimed 7–500× training-data reductions and some transfer from synthetic human videos to real footage, animals, and robots.
The broad reuse of video-generation representations for general-purpose perception, especially the claimed extreme data efficiency and synthetic-to-real transfer, is genuinely interesting, but the abstract gives few quantitative details and makes unusually broad state-of-the-art claims that need verification.
-
ABot-N1: Toward a General Visual Language Navigation Foundation Model
ABot-N1 separates visual-language reasoning from low-level control: a slower model produces explicit language reasoning and pixel-space goal anchors, while a faster controller converts those anchors into continuous waypoints. The authors report large gains in urban-scale navigation, including 77.3% POI arrival and strong indoor/outdoor success rates, while supporting several navigation tasks and releasing new benchmarks.
The pixel-grounded slow/fast interface is a meaningful and potentially general design for interpretable navigation, and the reported urban-scale gains are substantial, but the abstract does not provide enough baseline, scale, or ablation detail to justify a strong verdict.
-
KronQ: LLM Quantization via Kronecker-Factored Hessian
KronQ improves post-training quantization by using both activation and gradient covariance, rather than treating output channels as equally important. It applies rotations on both input and output dimensions and uses a Hessian-based sensitivity measure for mixed-precision allocation; on 2-bit weight-only quantization of LLaMA-3-70B, it reports 7.93 WikiText-2 perplexity where GPTQ and GPTAQ reportedly fail catastrophically.
The bidirectional, gradient-aware treatment is a meaningful departure from activation-only PTQ, and the claimed 2-bit 70B result is striking, but the abstract provides too little comparative and cross-model evidence for a strong verdict.
-
Weak-to-Strong Generalization via Direct On-Policy Distillation
The paper proposes transferring the policy change learned by reinforcement learning on a smaller language model to a larger model, rather than simply imitating the smaller model’s final outputs. It uses the log-ratio between the weak model before and after RL as a dense reward applied to the stronger model’s own trajectories, improving Qwen3-1.7B from 48.3% to 58.3% on AIME 2024 with 8 A100s in four hours, and reportedly supports composing multiple such policy shifts.
Reusing an RL-induced policy shift across model scales is a non-obvious alternative to expensive target-model RL, but the abstract provides limited evidence beyond one headline benchmark and broad claims of consistency.
-
Constrained Decoding for Diffusion Language Models via Efficient Inference over Finite Automata
The paper develops an exact method for enforcing finite-automaton constraints during diffusion language-model decoding, despite the model updating multiple token positions in parallel rather than generating left to right. It supports greedy and stochastic decoding under arbitrary remasking schedules, and uses arithmetic-circuit techniques to reduce inference depth from linear to logarithmic in sequence length. On Dream-7B and LLaDA-8B, it reports large gains on function calling, planning, text-to-SQL, and math tasks with under 5% wall-clock overhead, including a jump from 22.3% to 69.0% stochastic accuracy on BFCL-Live for Dream-7B.
This is a genuinely tailored solution to a central obstacle for diffusion LMs—exact structured decoding without sacrificing parallelism—and the reported large gains, logarithmic-depth inference, and low overhead suggest a broadly useful capability rather than a minor decoding tweak.
-
Disturbance-aware Motion Planning for Over-actuated Underwater Vehicles Exploiting Actuation Redundancy for High-fidelity 3D Reconstruction
The paper uses the redundant thruster configurations of an eight-thruster underwater vehicle to choose not only how it moves, but also how much turbulence and sediment resuspension it creates near the imaging target. A wake model and real-time thrust allocator reduce target-region particle velocity by 67% and improve 3D reconstruction RMSE from 4.3 mm to 1.9 mm across 440 trials, with a 98.5% reconstruction success rate.
It demonstrates a non-obvious and broadly useful actuation-to-perception principle—using control redundancy to improve sensing quality—with substantial real-world underwater gains and quantitative validation.
-
Accurate, Interdisciplinary and Transparent Structure-property Understanding with Deep Native Structural Reasoning
SciReasoner is a multimodal scientific model that represents protein, molecular, and crystal structures as addressable structural tokens, aiming to reason from native geometric and chemical evidence rather than only text or generic embeddings. It reports sizable gains in protein function prediction and retrosynthesis, interpretable fragment-level reasoning traces, and strong results across 86 benchmarks, though the abstract gives limited detail on baselines and evaluation methodology.
The unified, structure-native reasoning representation spanning biology, chemistry, and materials is a potentially important direction, but the broad state-of-the-art claims and expert-trace evaluation are not sufficiently substantiated in the abstract for a strong recommendation.
-
Score Accuracy Along the Forward Diffusion Does Not Certify Numerical Stability in Diffusion Sampling
This paper constructs diffusion score fields that are arbitrarily accurate under the forward diffusion and even close to the exact reverse process in path-space, yet become numerically unstable under Euler–Maruyama sampling. The discretized samplers converge weakly while all Wasserstein distances and positive moments diverge; experiments with a small DiT-style model show analogous rare-trajectory growth, which projection onto a known bounded support can suppress.
It identifies a sharp, counterintuitive gap between standard score accuracy and sampler reliability, with constructive theory and preliminary empirical evidence, but the practical prevalence of these pathological trajectories remains uncertain.
-
Moving Like a Human: Ego-Motion-Normalized Temporal Signatures for Real-Time Aerial Person Tracking on Milliwatt-Class Hardware
The paper introduces EMTS-Det, a tiny aerial-person tracker that analytically removes camera ego-motion and feeds the resulting residual-motion channels to a stateless detector, rather than learning temporal features. On a Raspberry Pi Zero 2W it reportedly runs at 31.85 FPS and substantially outperforms a much larger YOLOv8n baseline, while maintaining tracking through occlusions in real UAV video.
This is a credible-looking combination of a simple but non-obvious motion representation and extreme edge efficiency, with unusually strong real-hardware and field-video evidence rather than only benchmark results.
-
Deceptive Grounding: Entity Attribution Failure in Clinical Retrieval-Augmented Generation
The paper identifies “deceptive grounding,” where a clinical RAG system cites real, relevant evidence but attributes it to the wrong drug or disease entity, allowing it to pass standard faithfulness and citation checks. Across 13 models, adversarial tests produced DG rates of 8–87%, while a deployed system showed 7.8% overall and 13.6% for recently approved drugs; explicit entity-evidence verification detected nearly all cases.
This exposes a plausible and consequential failure mode that existing RAG evaluations miss, with controlled experiments and deployment measurements suggesting it is not merely an adversarial artifact.
-
Simple Geometric Recentering Rivals Deep Sequence Models for Cross-Session EEG Motor-Imagery Decoding
The paper compares a simple tangent-space classifier with unsupervised test-time recentering against classical Riemannian and deep sequence-model decoders for motor-imagery EEG, using identical covariance features across eight public datasets. The simple geometric method strongly outperforms the deep Mamba-based models in cross-session decoding, while recentering provides no benefit within sessions, suggesting that handling session shifts—not greater model capacity—is the key factor.
This is a broad, controlled, and practically important result that challenges the assumption that increasingly complex deep architectures improve cross-session EEG decoding, with large effects and a plausible mechanism centered on distribution recentering.
-
DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation
DSpark combines a semi-autoregressive draft model, which adds lightweight token dependencies to parallel drafting, with confidence- and load-aware verification that avoids spending serving capacity on likely-to-be-rejected tokens. The authors report higher accepted lengths offline and a 60–85% per-user generation-speedup over MTP-1 at matched throughput in live DeepSeek-V4 traffic, especially under strict latency constraints.
The combination of semi-autoregressive drafting and adaptive verification addresses a real speculative-decoding bottleneck, and the reported production gains are large, but the abstract provides too little detail to establish how much is genuinely new or how broadly the gains transfer.
-
Infinite Worlds with Versatile Interactions
LingBot-World 2.0 is an interactive video world model that generates scenes while users act in them, with support for long or ostensibly unbounded interaction, diverse actions, and agent-controlled character and environment behavior. The authors also distill a 1.3B real-time model claimed to drive 720p video at 60 fps and provide a multiplayer interface, but the abstract gives no quantitative evaluation of quality, latency, or long-horizon consistency.
The combination of persistent interactive generation, real-time deployment, and agentic world editing could be a meaningful step toward usable generative simulators, but the headline claims are insufficiently substantiated in the abstract.
-
Nemotron-Labs-Diffusion: A Tri-Mode Language Model Unifying Autoregressive, Diffusion, and Self-Speculation Decoding
Nemotron-Labs-Diffusion trains one language model to support autoregressive generation, diffusion-style generation, and diffusion drafting with autoregressive verification. The authors report that the combined objectives improve planning and self-speculative decoding, with the 8B model producing up to 6× more tokens per forward pass and 4× higher measured throughput than Qwen3-8B at similar accuracy; they also include 3B, 8B, 14B, and vision-language variants.
The unified AR/diffusion/self-speculation design and reported real-device throughput gains are substantially more interesting than a routine decoding tweak, but the abstract does not establish how broadly the gains hold beyond the stated hardware, benchmark, and sampler assumptions.
-
EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments
The paper studies how agents improve through prolonged interaction with real-world environments rather than offline pretraining. Using about 38,000 hours across 134 long-horizon tasks, it reports that performance follows a highly precise log-sigmoid scaling curve and that learning speed roughly doubles every three months across model generations; it also introduces and partially releases the EdgeBench evaluation suite.
The empirical claim of predictable scaling for post-deployment environment learning is potentially important and non-obvious, but the abstract does not establish that the scaling relationship is causal or robust beyond this benchmark and its model-generation comparisons.
-
A Theory of Contrastive Learning with Natural Images
This paper analytically characterizes optimal representations for contrastive learning on images with stationary statistics and common augmentations. It finds that the optimum can be implemented by CNNs whose first-layer filters are sinusoidal, followed by a nonlinearity, global pooling, and partial whitening; a waterfilling procedure determines the useful frequencies and weights from the image power spectrum. Experiments suggest that ordinary SGD-trained CNNs independently discover similar sinusoidal filters.
The sinusoidal-filter and waterfilling characterization offers a non-obvious theoretical explanation for emergent structure in contrastive learning, but the abstract provides limited detail about the theorem scope and empirical strength, so it falls short of a strong recommendation.
-
Who Needs DRAM? We Have Fiber
The paper proposes using long optical fibers as active, recirculating delay-line memory to broadcast immutable data such as LLM weights to many accelerators, rather than storing redundant copies in DRAM or HBM. A case-study architecture using multicore fiber, optical taps, co-packaged optics, and regeneration is estimated to eliminate redundant weight storage across 10,000 accelerators and cut weight-delivery energy by more than 70%.
The fiber-as-memory and optical weight-broadcasting concept is genuinely unconventional with potentially large energy and capacity implications, but the abstract reports only a case-study evaluation and does not establish that the substantial latency, bandwidth, reliability, and deployment challenges are solved.
-
Autonomous disproofs of the sum-product conjecture over $\mathbb R$ with GPT-5.5 Pro
A simple GPT-5.5 Pro agent reportedly generated seven correct, independently produced disproofs of the Erdős–Szemerédi sum-product conjecture over the reals in eight trials. The pipeline uses separate planning, construction, and review stages, with proofs based on several algebraic constructions; the authors release code and intermediate outputs for verification.
If the proofs withstand independent mathematical checking, this would be a major example of autonomous AI-driven mathematical discovery, but the abstract provides no details about the constructions or verification beyond the authors’ claim of correctness.
-
Principled Analysis of Deep Reinforcement Learning Evaluation and Design Paradigms
This paper studies how deep reinforcement-learning algorithms should be designed and compared as the amount of data and model capacity change. It develops a scaling-law framework and reports that algorithm rankings can change non-monotonically across data regimes, implying that some conclusions drawn from standard RL evaluations are unreliable. Large-scale experiments support the claim, though the abstract does not identify the affected algorithms or quantify the practical size of the effect.
The potentially important result is that conventional RL algorithm comparisons may produce qualitatively wrong conclusions because rankings depend non-monotonically on data regime, but the abstract lacks enough concrete results to justify a strong recommendation.
-
Ideas Have Genomes: Benchmarking Scientific Lineage Reasoning and Lineage-Grounded Idea Generation
This paper introduces IdeaGene-Bench, a benchmark that represents scientific contributions as typed, evidence-grounded “Idea Genome” objects and tracks how ideas are inherited, modified, removed, imported, or newly introduced across paper lineages. It evaluates both lineage reasoning and the ability to generate proposals that coherently extend a research lineage; across 14 LLM-based scientist systems, the best achieves only 27.3% exact accuracy on lineage reasoning, and lineage context changes system rankings rather than uniformly helping them.
The explicit genome/diff representation and lineage-conditioned generation task are a nonstandard, potentially useful direction, while the weak and uneven model performance suggests a real compositional limitation, though the benchmark’s annotation subjectivity and limited abstract-level evidence keep it from being a strong recommendation.
-
KVpop -- Key-Value Cache Compression with Predictive Online Pruning
KVpop learns which key-value cache entries to retain by supervising eviction with a future-attention signal, and uses a delayed scorer to incorporate near-future context before pruning. On Qwen3-4B mathematical reasoning tasks, it preserves 97–98% of full-cache performance while removing 75–88% of the cache, with stronger results reported for Qwen3-8B.
The combination of future-aware supervision and delayed online scoring appears meaningfully different from heuristic KV eviction and achieves unusually high compression, but the evidence is limited to a narrow set of reasoning benchmarks and model sizes.
-
Neural Collapse Is Forbidden: Information Floors in Language Models
The paper argues that language-model representations do not fully undergo neural collapse because within-category variation is needed to store context-dependent information. It derives an information-theoretic lower bound on within-category dispersion and reports that this dispersion tracks conditional mutual information across 14 models, partitions, and training stages, while category-level variance remains a small share of total variance.
The proposed information floor directly challenges a common interpretation of neural collapse in language models and is supported by cross-model empirical analyses, but the abstract does not provide enough methodological or quantitative detail to warrant a strong verdict.
-
Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents
The paper adds a separate memory agent that maintains structured state and selectively injects reminders into an otherwise unchanged action agent during long-horizon tasks. On Terminal-Bench 2.0 and τ²-Bench, it reports pass@1 gains of 8.3 and 6.8 percentage points, respectively, and finds that selective intervention works better than passive retrieval or always-on memory injection; an initial Qwen3.5-27B memory policy shows partial transfer.
The active, selective-memory framing and relatively large cross-benchmark gains are worth examining, but the abstract lacks absolute scores, implementation detail, and evidence that the improvement generalizes beyond two agent benchmarks.
-
Unified Audio Intelligence Without Regressing on Text Intelligence
The paper presents Audex, a 30B/3B-active MoE language model that handles audio and text with one Transformer decoder: audio is projected into the text embedding space, and generated audio is represented as quantized tokens. Trained on hundreds of billions of audio and text tokens, it reportedly supports audio understanding, speech recognition and translation, text-to-speech, audio generation, and speech-to-speech interaction while largely retaining the backbone’s text reasoning and agentic abilities; model checkpoints are released.
A broadly capable, openly released unified audio-text model that claims essentially no loss of text intelligence is worth tracking, but the abstract gives no quantitative comparisons or detailed evidence for its many state-of-the-art and non-regression claims.
-
Critical reassessment of lipophilic dye labeling reveals negligible incorporation into small extracellular vesicles derived form serum-free cultured cells
This study tests whether commonly used lipophilic dyes actually label small extracellular vesicles (sEVs), using fluorescently tagged vesicles as an independent reference. Across nanoflow cytometry, microscopy, and an in vivo comparison, PKH26, PKH67, and DiD labeled fewer than 0.5% of bona fide sEVs, suggesting that much of the observed dye signal may come from non-vesicular dye aggregates rather than vesicles.
The result directly challenges a widespread experimental assumption with orthogonal validation, but the evidence is currently limited to a serum-free HEK293F system and a small set of dyes, so its generality needs checking.
-
Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning
This paper studies why an LLM can memorize newly added facts but fail to use them in downstream reasoning, calling the discrepancy the “Knowing–Using Gap.” Using an intervention method called self-patching, it finds that the information is present internally but is often not routed to computation-effective layers, and a simple heuristic recovers 58–75% of the available oracle improvement across domains.
The mechanistic framing of memorization–generalization failure and the intervention-based routing diagnosis are plausibly useful new insights, but the abstract provides limited detail on model scale, comparisons, and the robustness of the reported recovery.
-
MV-Forcing: Long Multi-View Video Generation via 4D-Grounded Spatio-Temporal Self-Forcing
MV-Forcing combines temporal and viewpoint autoregression for generating long videos that remain consistent across multiple camera views. It uses autoregressive 3D reconstruction to provide a geometric prior for each new view, then trains and distills a diffusion model so both temporal and view-wise rollouts can continue beyond the original context window. The claimed result is arbitrary-length, multi-view video generation with a single few-step model, evaluated on synthetic and real data.
The joint use of 3D reconstruction as a bridge between long-horizon temporal and view autoregression is a genuinely interesting direction, but the abstract gives no quantitative evidence about consistency, length, viewpoint count, or comparisons sufficient for a stronger recommendation.
-
RynnWorld-Teleop: An Action-Conditioned World Model for Digital Teleoperation
The paper proposes collecting robot demonstrations in a generative world model rather than on physical hardware: an operator’s hand-pose stream conditions video generation, and the resulting trajectories can be retargeted to different robots. Its system reportedly generates interactive video at over 40 FPS and enables zero-shot sim-to-real imitation on dexterous and bimanual tasks, with further gains when combined with real data.
The digital-teleoperation framing and hardware-independent action collection are potentially important, but the abstract gives no task counts, baseline comparisons, or quantitative transfer results to substantiate the broad sim-to-real claims.
-
What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness
The paper probes intermediate activations of LLM forecasters rather than relying on their verbalized chain of thought. These probes improve calibration, detect forecast changes after evidence perturbations even when the CoT stays unchanged, and recover the model’s answer and confidence before reasoning begins; routing based on this pre-reasoning signal reduces generation by 30–47% without reported accuracy loss. The most notable finding is that forecasts appear largely committed before the stated reasoning, while internal representations track those commitments more faithfully.
It presents several non-obvious and potentially useful findings about pre-reasoning commitment and CoT unfaithfulness, but the evidence is limited to a few forecasting models and the abstract lacks detailed calibration and generalization results.
-
A Sovereign, Open-Source Foundation Model for German and English
Soofi S is a 30B-parameter German-English mixture-of-experts model combining Mamba and Transformer components, with only 3B parameters active per token. The authors claim near-constant inference cache size, strong long-context and high-concurrency efficiency, and benchmark performance comparable to much larger dense models; they also plan to release weights, checkpoints, data accounting, and training code openly.
The combination of a hybrid Mamba-Transformer architecture, sparse activation, and unusually comprehensive sovereign/open training disclosure could matter for efficient multilingual foundation models, but the abstract gives no quantitative throughput, memory, or benchmark results to substantiate the strongest claims.
-
SPEAR: A Simulator for Photorealistic Embodied AI Research
SPEAR is a Python-controlled simulator interface for arbitrary Unreal Engine applications, exposing more than 14,000 UE functions and supporting deterministic execution of complex dependency graphs within a frame. It reports 1080p photorealistic rendering at 73 FPS, richer ground-truth modalities than prior UE simulators, and demonstrations spanning multi-agent control, city-scale scenes, procedural generation, MuJoCo co-simulation, and natural-language scene editing.
The combination of unusually broad UE programmability, fast direct-to-NumPy rendering, richer supervision, and deterministic cross-system orchestration could make SPEAR a useful research platform, though the abstract provides limited comparative and task-performance evidence for the claimed gains.
-
Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity
The paper adds a large explicitly addressed memory to gated linear RNNs, using sparse reads and writes so the model can increase state capacity without proportional per-token compute. Under matched FLOPs and parameter counts, the authors report better in-context learning and long-context retrieval, with further gains from using the initial memory state as learned parametric knowledge.
Sparse, high-capacity state could be a meaningful alternative to quadratic-attention context handling, but the abstract gives no quantitative results or evidence of scaling beyond the reported task families.
-
Quantifying and Expanding the Theoretical Capacity of Late-Interaction Retrieval Models
The paper analyzes what late-interaction MaxSim retrieval can represent, showing that it can exactly implement inner products for nonnegative sparse vectors and can express some similarities that same-sized single-vector inner products cannot. It introduces Signed MaxSim, which handles arbitrary real-valued inner products and supports operations such as soft-OR and positive-CNF logical evaluation. On a negation-focused retrieval task, Signed MaxSim greatly improves out-of-domain nDCG, though the empirical evidence appears narrow.
The formal characterization and signed extension provide a genuinely useful theoretical explanation for late interaction, with striking negation results, but the practical evidence is limited to a specialized retrieval setting and the representational constructions may not directly translate to efficient learned systems.
-
SAM-MT: Real-Time Interactive Multi-Target Video Segmentation
SAM-MT adapts SAM2 for interactive multi-target video segmentation using per-object queries, shared scene features, decoupled attention, and sparse temporal memory. It reports over 36 FPS for 10 targets, aiming to keep latency roughly independent of the number of objects rather than processing each target separately.
The potentially important contribution is constant-scale real-time multi-target segmentation, but the abstract gives limited quantitative evidence beyond the 10-target speed claim and presents several specialized components whose broader advantage is unclear.
-
SynthDocBench: Controlled Benchmark for Long-Context Visual Document Understanding
SynthDocBench is a synthetic benchmark designed to independently vary document length, layout, modality, and question type, allowing VLM failures to be attributed to specific factors. Testing seven frontier models reveals strong long-document degradation, a positional bias where the middle section is hardest, and worsening chart comprehension—failure modes largely hidden by existing benchmarks.
Although primarily a benchmark paper, its controlled experiments expose a non-obvious middle-of-document sensitivity and long-context chart failure across frontier VLMs, making it more informative than a routine dataset or leaderboard.
-
RynnWorld-4D: 4D Embodied World Models for Robotic Manipulation
The paper presents a diffusion world model that predicts future RGB images, depth, and optical flow jointly from an RGB-D observation and language instruction. It also uses the model’s internal representation to predict robot actions directly, avoiding repeated diffusion at control time, and trains on a very large pseudo-labeled manipulation-video dataset. The claimed novelty is combining 4D multimodal prediction with efficient closed-loop policy learning for dexterous manipulation.
The unified RGB-depth-flow world model and single-pass policy extraction are plausible useful ideas for manipulation, but the abstract gives no quantitative results, task breadth, or comparisons to establish that the claimed state-of-the-art performance is a substantial advance.
-
From RGB Generation to Dense Field Readout: Pixel-Space Dense Prediction with Text-to-Image Models
The paper proposes ReChannel, a lightweight way to turn a text-to-image DiT into a dense-prediction model by mapping each image-plane token directly to task-specific pixel fields, rather than encoding targets as RGB-like images and decoding them through a VAE. With a frozen FLUX-Klein backbone and task-specific LoRA, it reports leading results on matting, KITTI depth, and referring segmentation, while being 2.48× faster than an edit-and-decode baseline in a matched 4B setting.
The direct token-to-native-field interface is a clean and potentially reusable alternative to treating dense prediction as image generation, but the abstract provides limited quantitative detail and the breadth of the claimed gains is not yet enough for a strong recommendation.
-
Wrong Before Right: Late Rescue and Interface Failure in Aligned Language Models
The paper argues that some language models briefly favor an incorrect answer in middle layers and rely on late-layer correction to produce the right output. Across 17 models, it reports that this “wrong-dip” predicts failures under pruning, layer dropping, and low-rank compression, while being largely unrelated to quantization failures; a LoRA penalty can reduce the dip and improve compression robustness without hurting accuracy.
This is a potentially important mechanistic finding with causal interventions, scale comparisons, and an actionable training result, but the abstract provides no concrete accuracy, failure-rate, or evaluation details sufficient for a strong verdict.
-
Qantara: Bridge-Flow Training for Multi-Paradigm JEPA Control
Qantara trains one JEPA world model to support three deployment modes without retraining: latent trajectory planning, behavior-cloning action generation, and inverse dynamics. It combines Brownian-bridge latent interpolation with action flow matching, and reports large gains from query-focused training, including 91.2 success on the LeWM suite and substantial improvements over prior methods on OGBench-Cube.
The single-checkpoint, multi-paradigm controller is a meaningful new direction with strong reported benchmark gains and an informative ablation, but the evidence is still limited to the abstract and appears concentrated on a small set of learned-control benchmarks.
-
AIFS-SUBS: Extending Data-Driven Forecasting to Sub-Seasonal Timescales
The paper extends ECMWF’s data-driven AIFS weather model from medium-range forecasting to sub-seasonal horizons (weeks 2–6), using 24-hour autoregressive steps, stratospheric and radiation inputs, and careful long-term verification. It matches the operational IFS overall, gains eight days of skill for the convective MJO signal, performs well on stratospheric events, and reportedly uses about 200 times less inference energy, enabling much larger ensembles.
This is a meaningful extension into a difficult forecasting regime with nontrivial MJO and stratospheric results plus a large efficiency advantage, but the core method is an adaptation of an existing model and the abstract lacks detailed comparative metrics needed for a stronger verdict.
-
Most LLM Conformity Needs No Speaker: Measuring the Speaker-Free Floor in Peer-Pressure Benchmarks
The paper argues that much apparent LLM “peer pressure” is caused not by the presence of a speaker, but by seeing a repeated alternative answer. Removing the speaker still caused initially correct models to switch to the wrong answer in 66.5% of cases, versus 10.3% after a plain re-ask, across six open-weight models and seven datasets. It proposes measuring this speaker-free baseline before attributing additional changes to social or source cues.
This identifies a substantial confound in a widely used evaluation setup, with broad experiments and a striking result, though it is primarily a methodological correction rather than a new model capability or training method.
-
Observation Quality Matters: Robust Multi-Fisheye Calibration via Failure-Oriented Analysis
The paper analyzes why multi-fisheye calibration fails and argues that the key issue is poor intrinsic initialization from observations with insufficient radial coverage, rather than simply missed detections or uneven image-plane coverage. Its CO-Calib framework selects calibration frames for initialization, cross-camera visibility, and coverage completion, raising reported calibration success from 68.1% to 99.3% while improving extrinsic accuracy on synthetic and real rigs.
The failure-oriented diagnosis and large reported reliability gain make this more than a routine calibration tweak, but the contribution remains specialized and the abstract does not establish how broadly it transfers across rigs and calibration pipelines.
-
Boosting with List-Decodable Codes
The paper gives a boosting algorithm for concept classes closed under XOR of roughly O(log(1/γ)) functions. It uses a list-decodable code to turn weak predictions into a short list containing a strong hypothesis, reducing the number of sequential weak-learner calls from O(log(1/ε)/γ²) to O(log(1/ε)), while using a comparable one-time batch of extra samples.
The code-theoretic formulation yields a substantial round-complexity improvement and bypasses a known generic lower bound, but the gain depends on a structural closure assumption and the abstract provides no indication of broad practical applicability.
-
No Subspace to Track: Non-Identifiability and Optimizer State in Low-Rank Training
This paper argues that the gradient subspace tracked by GaLore is largely not a stable low-rank object: most apparent subspace rotation is minibatch estimator noise, with only about 39 of 128 directions reproducible in one reported setting. It reframes each refresh as a coordinate change for Adam’s optimizer state, showing that first-moment transport is principled while blindly carrying the second moment is inefficient; experiments across model scales support shorter second-moment memory and the proposed LDAdam variant.
The non-identifiability diagnosis and optimizer-state interpretation challenge a central assumption behind low-rank training and yield concrete design guidance, with cross-scale evidence, but the practical gains and broader consequences are not yet strong enough for a must-read verdict.
-
More Convincing, Not More Correct: Self-Play Reward Hacking of Reference-Free LLM Judges
The paper shows that self-play with reference-free LLM judges can optimize for answers that sound plausible rather than answers that are correct. On GSM8K, self-play raised judged pass rates from 0.72 to 0.94 while true accuracy remained 0.20, and the false positives transferred across judge families and survived a three-judge ensemble. Having the judge solve the problem before inspecting the candidate largely removes the effect, reducing the false-positive rate from 0.719 to 0.012, with similar behavior reported for best-of-N selection in code and competition math.
This identifies a structural failure mode in reference-free judging, rather than a model-specific exploit, and provides a simple de-anchored judging procedure with strong reported effects across models and tasks; the evidence is promising but still mainly limited to the abstract's experiments.
-
Propose and Attend: Training-free MLLM Grounding Confidence via Multi-Token Localized Attention
The paper introduces MTLA, a training-free confidence score for multimodal language models that checks whether all tokens describing a predicted region attend strongly to that region. It applies to image boxes and video/audio temporal windows, and reportedly improves hallucination detection by 7–38 AUROC points across several MLLM families and modalities. Re-ranking with the score nearly doubles zero-shot COCO detection AP for an 8B model, from 20.4 to 37.0, without task-specific training.
The cross-modal, training-free grounding-confidence signal and large reported gains are genuinely interesting, but the abstract does not establish how robust the result is across models, prompts, or evaluation settings well enough for a strong verdict.
-
When Does Tool Use Increase the Expressive Power of Finite-Precision Recurrent Models?
This paper characterizes when external tools genuinely increase the computational power of finite-precision recurrent models. It proves that finite-state tools can be folded into the controller with only a small state-cost increase, while a minimal read/write/move tape makes even a constant-memory controller Turing complete and yields an exponential separation on equality testing. The authors also implement the construction in a one-layer selective affine SSM, showing that selectivity is essential.
The finite-state versus infinite-state tool dichotomy and its exact realization in a finite-precision selective SSM are unusually clean theoretical results, though the conclusions rely on an idealized tape oracle and may have limited immediate implications for practical agents.
-
When do prophets profit in prediction markets?
This paper develops a betting strategy for central-limit-order-book prediction markets that converts a forecaster’s superior probabilistic accuracy into expected trading profit, extending the familiar result for automated market makers. It argues that this strategy is essentially unique, explains why inaccurate or heuristic traders can still profit, and reports experiments with AI forecasts plus a month-long Kalshi deployment showing 80.33% ROI and a 3.35 Sharpe ratio.
The extension of the accuracy–profitability equivalence from AMMs to real order-book markets, together with a live trading result, is a substantive and non-obvious contribution, though the practical result lacks scale and longer-term validation in the abstract.
-
Specification Grounding Drives Test Effectiveness for LLM Code
The paper tests whether LLM-generated tests help because tests exist or because they are explicitly grounded in the program specification. Giving the tester the specification as a checklist improves correct-code generation by 19–38 percentage points across Claude, GPT, and Gemini models, while extra ungrounded tests provide little benefit; the effect also reduces false rejection of correct code. The result holds on a held-out set and across several testing baselines, though the study covers only 18 tasks.
This is a concrete, cross-model result challenging the idea that self-testing mainly benefits from test quantity, but the relatively small task set and prompt-based intervention make it promising rather than an unquestionable major advance.
-
Doomed from the Start: Early Abort of LLM Agent Episodes via a Recall-Controlled Probe Cascade
The paper predicts whether an LLM-agent trajectory will eventually fail using linear probes over hidden states, often from the first interaction round. It combines these predictions into a calibrated, multi-round abort cascade with an explicit global recall guarantee, reducing generated tokens by roughly 55–60% at 90% recall and 41–45% at 95% recall across TextCraft and WebShop experiments.
The combination of very-early hidden-state failure prediction, recall-controlled cascading gates, and post-selection guarantees is a meaningful efficiency direction with substantial reported savings, though the evidence is limited to two environments and the underlying probe approach is relatively simple.
-
CaLiSym: Learning Symplectic Dynamics of Real-World Systems through Structured Canonical Lifts
CaLiSym learns dynamics of actuated, dissipative, and contact-rich robotic systems by lifting observed states and input/output ports into a larger phase space where the learned evolution is exactly symplectic. On a dissipative double pendulum, real quadrotor, and real quadruped, it reports the best out-of-distribution autoregressive rollout errors, with up to 69.5% lower error and up to 85× fewer FLOPs than sequence-model baselines. The notable contribution is extending symplectic learning beyond closed conservative systems without recurrent latent states or inference-time integration.
The explicit canonical lift appears to provide a genuinely useful way to retain symplectic structure in real, nonconservative robotics, and the cross-system results and efficiency claims merit inspection, though the abstract does not establish whether the gains generalize beyond the tested systems.
-
Geometric Self-Distillation for Reasoning Generalization
The paper proposes GeoSD, a self-distillation method for language-model reasoning that limits how aggressively a privileged teacher pulls the student toward unsupported predictions. It combines overlap-weighted Hellinger supervision, a proximal penalty in Fisher–Rao geometry, and natural-gradient updates; across three model families from 1.7B to 32B, it reportedly improves out-of-distribution reasoning accuracy by 5.7–8.6 points while retaining in-distribution gains. The analysis argues that standard distillation can suppress alternative answers at uncertain states, producing confident but incorrect OOD predictions.
The geometric treatment of self-distillation drift and the substantial, cross-scale OOD gains are more than a routine loss tweak, but the abstract does not establish whether the result generalizes beyond the tested reasoning benchmarks or whether the component combination is genuinely necessary.
-
Intrinsic-Noise Consolidation: A Doob-Barrier-Conditioned Diffusion Turns Analog Device Noise into a Continual-Learning Resource
The paper uses a Doob h-transform to condition noisy synaptic weight dynamics on avoiding a barrier around consolidated memories, turning intrinsic analog noise into a state-dependent restoring force. On Split-MNIST, it reports an inverted-U relationship between noise and retention, with a 10.9-point improvement, and demonstrates a 15.6-point retention gain over a matched control in a single BrainScaleS-2 hardware-in-the-loop run.
The synaptic use of barrier-conditioned diffusion and its predicted non-monotonic noise benefit appear genuinely novel, but the evidence is still limited to Split-MNIST, eight software seeds, and a single hardware seed, so it is not yet a strong result.
-
Answering Without Referring: How AI Search Rewrites the Web's Economic Bargain
Using U.S. desktop clickstream data, the authors compare ChatGPT and Google search behavior and examine what happens when access to ChatGPT Search expands. ChatGPT sessions generate outbound clicks only 5.2% of the time, and expanded access reduces traditional search use by 9.4%, with losses concentrated in informational queries; remaining clicks also favor specialized sites over ad-supported destinations. The paper argues that AI search may decouple answering information needs from sending traffic to the websites that produce the information.
The unusually low referral rate and estimated displacement provide important empirical evidence that AI search may fundamentally alter the web's traffic-and-content-production bargain, though the analysis is limited to one service, U.S. desktop behavior, and the abstract gives little detail on identification or robustness.
-
A law of robustness for two-layer neural networks with arbitrary weights
This paper proves, up to a logarithmic factor, a conjectured lower bound on the Lipschitz constant required for a width-m two-layer network to fit noisy data, without assuming bounded weights. The key move is to cover the realized function class rather than parameter space, together with a rigidity result showing that generic kink hyperplanes cannot cancel; the result also exposes a genuine failure in two dimensions and gives a matching construction at high overparameterization.
It closes a technically important gap in a proposed robustness law for unbounded-weight two-layer networks and identifies a dimensional obstruction, but the result is still an asymptotic theoretical bound with a logarithmic loss rather than a demonstrated capability or broad practical impact.
-
Closed-Loop LLM Co-Pilots for Digital Agriculture
The paper uses an LLM to interpret a 49-channel plant-sensing system and directly control lighting and other agricultural actuators in closed loop. In reported vertical-farm and single-plant studies, it shortened production cycles by 35%, reduced energy use by 18% under one objective, and achieved a claimed 67.9% energy saving through an autonomously discovered dark-period strategy.
The unusual part is an LLM controlling a real biological production system and apparently discovering a non-obvious lighting policy, but the abstract gives too little detail about baselines, replication, experimental controls, and how much of the optimization was genuinely LLM-driven to justify a strong verdict.
-
CORAL-AUV: CFD Oriented Reinforcement Learning for Autonomous Underwater Vehicles
The paper trains fast surrogate models of CFD-derived drag forces and uses them inside reinforcement-learning training for a 6-DOF autonomous underwater vehicle. In tank and field tests, the resulting zero-shot policy used 31% less energy, reached waypoints 11% faster, and had 19% lower error than a controller trained with simplified physics; it also transferred under perturbed parameters where the comparison policies failed.
Using learned CFD surrogates to close the underwater sim-to-real gap, with real AUV deployment and substantial reported gains, is a genuinely interesting robotics direction, but the abstract does not establish how broadly the approach generalizes beyond this vehicle and task.
-
Frontal Eye Field Leads a Distributed Oculomotor Circuit for Abstract Categorical Decisions
The study recorded simultaneously from frontal eye field (FEF), lateral intraparietal area (LIP), and superior colliculus (SC) while monkeys categorized visual motion and reported decisions with saccades. FEF encoded abstract categories first, followed by SC and LIP, while LIP encoded visual features earliest; directed-information and reversible inactivation results suggest FEF causally helps transform category information into choices.
It provides a relatively nonstandard, causally supported account of how a distributed oculomotor circuit routes abstract decision variables, but the abstract lacks quantitative effect sizes and broader evidence needed for a strong verdict.
-
Opioid- and NMDA-receptor-dependent neural plasticity mediates long-term analgesia from motor cortical stimulation
The study finds that a brief, focal motor-cortex TMS session produces 1–2 weeks of analgesia in mice with chronic neuropathic pain by engaging layer-5 motor-cortex projections to the RVM. The durable effect depends jointly on NMDA-receptor plasticity and endogenous opioid signaling in the RVM, and inhibiting enkephalin breakdown during stimulation strengthens and prolongs analgesia; a reanalysis of human data is consistent with opioid treatment amplifying TMS effects.
This provides a specific circuit and molecular explanation for unusually long-lasting stimulation-induced analgesia and suggests a testable neurostimulation–drug combination, but the main causal evidence is still from mice and the human support is only a reanalysis.
-
Computational demands shape seizure susceptibility in recurrent neural networks
The authors show in recurrent-network models that systems representing information with continuous attractor dynamics are more vulnerable to seizure-like perturbations than matched networks representing discrete states. Recordings from mouse medial entorhinal cortex and CA3 support this prediction: the entorhinal region shows stronger, smoother seizure involvement, and silencing its connectivity reduces the effect.
This offers a non-obvious computational explanation for regional seizure susceptibility, supported by both recurrent-network experiments and in vivo recordings with a connectivity manipulation, but the evidence appears limited to a small set of regions and seizure conditions.
-
Conserved and specialized features of thalamocortical wiring revealed by single-cell projection mapping in mouse and marmoset
Using in situ sequencing and BARseq, the authors mapped gene expression and cortical projections of thalamic neurons in marmosets and mice, including a large marmoset expression atlas. They find that marmoset projections are more spatially segregated, but both species retain a continuous molecular and projection gradient that crosses conventional thalamic nucleus borders and predicts cortical target position. The result combines local, primate-specialized point-to-point wiring with a conserved gradient-based organization.
This is a substantial cross-species, single-cell result that reconciles competing discrete-nucleus and continuous-gradient models, though the abstract does not establish a transformative functional consequence.
-
Device-embedded accelerometry complements neural signals for tracking parkinsonian motor states
This study analyzes more than 1,900 hours of chronic Parkinson’s recordings to compare STN, cortical, and DBS-device accelerometer signals for tracking motor symptoms during stimulation. Device-embedded accelerometry tracked bradykinesia and dyskinesia more robustly than neural features, while decomposing beta power showed that conventional total beta can obscure opposing periodic and aperiodic relationships with symptoms.
The large longitudinal dataset and finding that an already-embedded accelerometer may outperform commonly used neural biomarkers under active DBS are genuinely useful for adaptive stimulation, but the abstract gives no quantitative effect sizes or prospective closed-loop demonstration.
-
Mitochondrial pyruvate import in astrocytes links anaplerosis to seizure resistance
The study deletes the mitochondrial pyruvate carrier subunit MPC2 specifically in adult astrocytes and finds motor impairment, neuronal hyperexcitability, and fatal seizures. Metabolic measurements suggest that mitochondrial pyruvate use in astrocytes supplies anaplerotic carbon needed to maintain TCA-cycle and neurotransmitter pools, challenging the view that astrocytes can mainly divert pyruvate to lactate without major consequences.
It provides a potentially important in-vivo challenge to the glycolytic-astrocyte model by linking astrocytic mitochondrial pyruvate import directly to excitation-inhibition balance and seizure resistance, but the abstract alone does not establish how broadly this mechanism generalizes.
-
Hippocampal CA3 forms a two-layer network of molecularly distinct cell types in mice and humans
The study identifies two molecularly distinct CA3 pyramidal-neuron populations organized along the deep–superficial axis. Superficial St18− cells make the broad, bilateral recurrent network and project to CA1, whereas deep St18+ cells form a previously unrecognized, mostly unilateral CA3-restricted circuit; similar molecular and functional features are reported in mice, pigs, and humans.
This is a potentially important revision of the CA3 circuit architecture, supported by molecular, anatomical, physiological, and cross-species evidence, but the abstract does not yet establish how the two layers differ computationally or behaviorally.
-
Image2Sim: Scaling Embodied Navigation via Generative Neural Simulator
Image2Sim builds interactive neural environments from posed RGB-D sequences. It separates 3D scene anchoring from photorealistic observation generation, using feature Gaussians for scene construction and a one-step pixel-flow model for panoramic RGB-D rendering. The system reportedly scales to nearly 20,000 scenes and over 10 million navigation examples, with models trained in these environments transferring to real-world navigation without real-world training.
The combination of automatically converting ordinary RGB-D imagery into scalable, interactive training environments and the claimed zero-shot sim-to-real transfer is genuinely interesting, but the abstract gives no quantitative results or comparison details to establish how large the improvement is.
-
MedPMC: A Systematic Framework for Scaling High-Fidelity Medical Multimodal Data for Foundation Models
MedPMC is an automated pipeline that extracts and validates image–caption pairs from 6.1 million permissively licensed PubMed Central articles, producing 11 million medically relevant pairs. A model trained on the curated data outperformed a matched biomedical CLIP baseline by 7.1 AUC points using less than half as many pairs, with additional gains in medical VQA and dermatology image retrieval.
This is primarily a data-engineering and curation contribution, but the large-scale, independently reviewed corpus and strong gains with substantially less data make it more consequential than a routine medical dataset paper.
-
Light-Omni: Reflex over Reasoning in Agentic Video Understanding with Long-Term Memory
Light-Omni replaces iterative search-and-reason video agents with two continuously updated states: a compact multimodal global summary and a latent state used for retrieval and action selection. The authors report similar-or-better video understanding with a 2.4% average accuracy gain over M3-Agent, 12.1× lower latency, and 2.6× better GPU memory efficiency, while also improving existing multimodal LLMs when used as a memory module.
The reflexive, single-pass alternative to expensive iterative video-agent reasoning and the large reported efficiency gains are worth a look, although the abstract gives limited detail about benchmarks, accuracy tradeoffs, and how broadly the approach generalizes.
-
A Physics-Regulated Neural Framework for Learning 3D Grain Growth Dynamics
The paper introduces 3D-PRIMME, a physics-constrained neural surrogate for simulating grain growth. Trained from only two consecutive states on a 100^3 grid with 512 grains, it reportedly preserves coarsening kinetics and topology when run for long horizons and scaled to 1024^3 grids with 550,000 grains without retraining. The notable claim is that a learned local rule generalizes across both time and system size in a 3D physical simulation.
The extreme zero-shot scaling and long-horizon physical consistency would be important if validated, but the abstract gives no quantitative error comparisons, baselines, or details sufficient to establish how robust the claim is.
-
Geometric Causal Models
The paper proposes geometric causal models, which use symmetries such as translation invariance in spatial data or node permutations in graphs to identify and estimate causal effects when observations are dependent rather than i.i.d. It connects group theory and ergodic theory with geometric deep learning and Bayesian inference, recovering ordinary causal models in a special case and constructing a DNA-symmetry model for genetic effects, demonstrated on semisynthetic data.
The symmetry-based causal framework is a genuinely interesting and potentially general research direction, but the abstract provides mainly theoretical claims and a semisynthetic demonstration rather than strong empirical validation or a clear capability jump.
-
Advances in Neural Controlled Differential Equations
This thesis develops faster Neural Controlled Differential Equations for continuous-time and irregularly sampled time series. It introduces Log-NCDEs, which approximate the dynamics during training, and Linear and Structured Linear NCDEs, which enable closed-form and parallel-in-time computation while retaining theoretical expressivity. The reported methods reduce training-step time by up to three orders of magnitude and match or improve performance across several time-series benchmarks.
The combination of theoretical expressivity with parallelizable, dramatically faster NCDE computation is a substantial advance, but the abstract provides limited detail behind the broad speed and benchmark claims, and this is a thesis rather than a focused result with clearly quantified comparisons.
-
Graph Sparse Sampling: Breaking the Curse of the Horizon in Continuous MDP Planning
The paper proposes Graph Sparse Sampling, an online planner that reuses sampled future states across many candidate actions instead of growing separate search trees. It provides finite-sample guarantees showing polynomial, rather than exponential, dependence on planning horizon under overlap and coverage assumptions, and reports strong results in continuous-control simulations, especially at long horizons.
The shared-futures, branch-free planning design and accompanying horizon-scaling guarantees are a genuinely interesting alternative to tree search, but the abstract gives limited quantitative and empirical detail and the theoretical improvement depends on substantial regularity and coverage assumptions.
-
Width-Robust Learnability in Mean-Field Bayesian Neural Networks
This paper studies whether infinite-width mean-field Bayesian neural networks can generalize functions that would require prohibitively large finite networks. For fixed-depth networks on Boolean-cube targets, it proves an equivalence: polynomial-sample learnability at infinite width, polynomial-width learnability, and polynomially bounded prior representation entropy coincide. The key technical idea is to subsample infinitely many hidden units into polynomially many, retaining data-dependent directions while resampling entropy-dominated directions from the prior.
The width-robustness theorem addresses a fundamental concern about infinite-width limits and claims that mean-field learning does not gain spurious generalization power, but the result is confined to a theoretical Boolean-cube setting and the abstract provides no empirical or broader-scope validation.
-
Escaping the Procrustean Bed: Groupwise Orthogonal Connectors for Audio-Language Models
The paper argues that Q-Former connectors in audio-language models discard information because their output vectors collapse toward one direction and become insensitive to speaker differences. ORCA divides queries into groups and constrains each group to produce outputs pointing in distinct directions; on the SAKURA multi-hop reasoning benchmark, this raises performance from 49.0% to 75.2% for an otherwise identical 4B model, while substantially reducing query redundancy and increasing speaker-dependent variation.
The combination of a concrete representation-collapse diagnosis and a simple orthogonality-based connector intervention with a very large reported gain is worth checking, but the evidence is currently concentrated in one benchmark and needs broader validation.
-
MobileWan: Closing the Quality Gap for Mobile Video Diffusion
MobileWan reformulates a 5B-parameter video diffusion transformer as a chunk-wise recurrent model with constant-memory attention, then combines head pruning, sampling distillation, and memory-efficient decoding to run on mobile hardware. It reportedly generates 5-second 480×832 videos at 16 FPS in 20 seconds on a commercial device, with a VBench score of 83.79, suggesting substantially larger video models can be made practical on phones.
The combination of recurrent reformulation and structured compression enabling 5B-scale video generation on a phone is a meaningful deployment result, but the abstract provides limited hardware, baseline, and quality-versus-efficiency details, so it does not yet justify a strong verdict.
-
Optimal Transport Q-Learning for Flow Policy Steering and Acceleration
The paper introduces Optimal Transport Q-Learning (OTQL), which uses advantage-weighted conditional optimal-transport flow matching to fine-tune and shorten diffusion/flow-based robot policies from the robot’s own experience. With only 50–60 episodes, it reports large success-rate increases—36% to 86% for single-task policies and 38% to 76% for a pretrained VLA—while reducing action-generation inference steps by 70%.
The combination of offline-style flow-policy post-training, limited real-robot interaction, and substantial simultaneous gains in success and inference speed is notably interesting, but the abstract lacks baselines, task breadth, and details needed to establish how broadly the claims hold.
-
NativeMEM: Native Memory Compression for Long-Horizon Robotic Manipulation
NativeMEM adds long-term visual memory directly inside a pretrained vision-language-action model by compressing each historical camera frame into one token using the model’s own vision encoder. The authors report large gains on long-horizon manipulation—from 32.4% to 84.0% success in simulation and up to 98.7% on real robots—while preserving low latency and achieving competitive performance with 20% of the training data.
The native, single-token memory mechanism and claimed large real-robot and data-efficiency gains are genuinely interesting, but the abstract omits task breadth, baselines, and evaluation details needed to trust a strong verdict.
-
Compress the Cache, Not the Speech Embedding: KV Compression for Efficient Speech LLMs
SpeechKV compresses the key-value cache for speech tokens inside the language model rather than shortening the speech representation before it enters the model. Trained on 71K hours of speech, it reportedly reduces speech sequences to roughly text-like lengths while preserving or slightly improving recognition performance, including a 6.6% relative gain on out-of-domain entity recognition, and provides at least 1.49× faster decoding.
The idea of compressing speech-token KV states after internal fusion addresses a real long-sequence bottleneck while avoiding irreversible early downsampling, but the abstract gives limited detail about tasks, compute savings, and comparisons, so it is promising rather than clearly field-changing.
-
Do Modules Stay in Their Lane? Role Drift in Compound LLM Systems
This paper identifies “role drift” in end-to-end-trained compound LLM systems: modules can improve overall accuracy by quietly violating their assigned responsibilities. It proposes Role Anchor, which preserves the change in next-token behavior induced by a module’s role prompt, and finds that in one decomposer pipeline 86% of the apparent reinforcement-learning gain disappears when shortcut behavior is prevented.
The finding that most of an apparent system-level RL improvement can come from a module bypassing its intended role is non-obvious and practically important, though the evidence is limited to two pipelines and the proposed regularizer incurs an accuracy tradeoff.
-
EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI
EmbodiedGen V2 is a generative pipeline for turning 3D assets into editable, executable environments for navigation and manipulation across simulators. It reports that most generated task worlds are usable without manual fixes, and that reinforcement learning in these environments improves simulated success from 9.7% to 79.8% and real-robot success from 21.7% to 75.0%.
The unified, cross-simulator generation of policy-ready task environments and the large reported sim-to-real improvement are potentially important, but the abstract does not establish baselines, task breadth, or how much of the gain comes from the environment engine rather than ordinary training changes.
-
Agon: Competitive Cross-Model RL with Implicit Rival Grading of Reasoning
Agon trains two language models as competing solver-and-grader pairs: each model tries to solve problems while reading the other’s attempt, and is rewarded for outperforming its rival rather than only for reaching a verifiable final answer. The authors report roughly a 2× improvement over GRPO on a hard DeepMath split, with similar ordering on competitive programming and across Qwen and Gemma families; inference uses the same two-stage draft-and-revise cascade.
The implicit process supervision through adversarially co-trained peer models is a genuinely interesting alternative to outcome-only RL, but the abstract provides insufficient absolute results, baselines, and stability details to justify a strong recommendation.
-
Co-LMLM: Continuous-Query Limited Memory Language Models
Co-LMLM replaces relational knowledge-base lookups with continuous vector queries that retrieve human-readable factual text during generation. It also introduces a pipeline for extracting factual spans from arbitrary corpora, and reports better perplexity and factual precision than prior limited-memory models and standard LLMs across scales; notably, a 360M model allegedly matches or exceeds much larger models on some factuality comparisons.
The continuous-query external-memory design could be a meaningful alternative to weight-based knowledge storage, and the claimed small-model/data-efficiency gains are striking, but the abstract lacks enough experimental detail to distinguish a major advance from optimistic benchmark comparisons.
-
An exact information theory of generalization phase transitions in Bayesian diffusion models
The paper develops an analytically tractable Bayesian diffusion model in which each pixel receives only restricted information about noisy training examples. It derives a phase boundary: the model memorizes when the mutual information in those observations exceeds the logarithm of the dataset size, and otherwise generalizes; experiments reportedly place the transition where predicted. The proposed explanation is that both spatially local diffusion models and early-training UNets/DiTs generate samples near this boundary by progressively restricting information during denoising.
This offers a potentially important information-theoretic explanation for why diffusion models avoid memorization in high dimensions, but the evidence is based on a restricted analytical model and broad abstract-level claims rather than clearly quantified results on full trained systems.
-
Harness VLA: Steering Frozen VLAs into Reliable Manipulation Primitives via Memory-Guided Agents
Harness VLA uses a frozen vision-language-action model only for difficult, contact-rich manipulation, while an agent with memory and a small set of analytic primitives handles semantic grounding, staging, transport, and recovery. The authors report large gains under task and layout perturbations—38.6 and 25.4 percentage points on LIBERO-Pro and RoboCasa365—and 58.4% on RoboTwin C2R, without finetuning the VLA.
The decomposition of a frozen VLA into a retryable local skill with learned operating ranges could be a useful path to robust out-of-distribution manipulation, but the abstract provides limited detail about baselines, scale, and real-world validation behind the large gains.
-
SafeExplorer: An Unbiased Policy Gradient for Reinforcement Learning with Recovery Interventions
SafeExplorer modifies PPO for settings where a recovery controller takes over when the learned policy enters an unsafe region. It estimates policy gradients using only timesteps controlled by the learned policy, avoiding importance sampling and remaining valid when the recovery policy is deterministic; additional value and imitation terms improve learning near intervention boundaries. Across three environments, it reportedly reduces training falls by 26–233× while retaining comparable final reward, though the evidence is limited to five seeds and a small benchmark.
The deterministic-recovery, unbiased policy-gradient formulation is a substantive and practically relevant idea, with striking fall reductions, but the claims rest on only three environments and modest experimental evidence rather than a broad real-robot validation.
-
COBS: Cumulant Order Block Sparse Attention
COBS improves block-sparse attention by selecting KV-cache blocks using a compressed second-order statistic rather than the first-order summaries used by prior selectors. On 32k-token RULER retrieval, it raises NSA’s mean score from 0.300 to 0.820, close to dense attention’s 0.904, while requiring 15.15× less KV-cache read traffic than dense attention. The main novelty is framing block selection as approximating each block’s total attention mass and using a cumulant expansion to motivate the selector.
The large retrieval improvement at near-sparse memory-read cost and the second-order selection idea are genuinely interesting, but the evidence is centered on one benchmark and a comparison against one sparse-attention baseline, so broader robustness and compute overhead need verification.
-
DumpsterCluster: From Dumpster Diving to Serving LLaMA-70B on $60 GPUs
The authors built and operated a 128-GPU cluster from second-hand components for a year, using pipeline-parallel optimizations to serve LLaMA-70B. They report a much lower purchase cost than current hardware and competitive throughput, but also find that older GPUs can have far higher energy and carbon costs per token—especially for large models—unless electricity is cheap and low-carbon.
The long-running, physically deployed study offers a non-obvious cost-versus-carbon analysis of repurposed GPU inference, but the abstract gives too few throughput and workload details to establish a major serving breakthrough.
-
MLPs are Hebbians: Constructing Efficient Fact-Storing MLPs for Transformers
The paper gives a closed-form, Transformer-compatible construction for MLPs that store factual key–value associations. It analyzes decoding margins and claims information-theoretically optimal storage scaling, using 10–104× fewer parameters than prior constructions in isolated MLPs and 15–63× fewer inside Transformer blocks; it also demonstrates modular fact editing by swapping the MLP.
This is a theoretically grounded and substantially more efficient account of how Transformer MLPs can store facts, but the evidence appears limited to constructed or factual-recall settings rather than demonstrating that real pretrained LLMs use this mechanism.
-
CoSAG: Compact Semantic Anchor Gaussians via Training-Free Rate-Distortion Coding
CoSAG treats compact open-vocabulary 3D Gaussian Splatting as a rate-distortion problem, compressing the per-Gaussian assignments to a small set of spatially grounded semantic anchors rather than learning a per-scene codec. It combines a training-free transmittance-weighted feature lift and multi-view denoising with spatial predictive entropy coding, reportedly achieving sub-megabyte fields and 37–76× smaller storage than LangSplatV2 while matching or improving accuracy.
The explicit compression of spatially predictable per-Gaussian anchor assignments, without per-scene training or a shipped decoder, is a genuinely interesting direction with potentially major deployment benefits, but the abstract provides too few experimental details to validate the unusually large compression and accuracy claims.
-
Program-Synthesis-Driven Autodesign of Universal Unitary Operators
The authors extend DreamCoder to complex-valued linear algebra and use program synthesis to discover photonic decompositions of unitary matrices. The system recovers minimal universal decompositions that generalize from small matrices to sizes up to 64, and discovers structure-specific reductions—for example, a linear-scaling $2N-3$ interferometer construction for Householder matrices and up to 38% fewer components for highly sparse SVD-derived matrices.
The notable contribution is using interpretable program synthesis to discover dimension-agnostic matrix-decomposition algorithms and hardware optimizations, but the abstract does not provide enough experimental or comparative detail to establish how independent or broadly superior these discoveries are.
-
ZoRRO: A Zero-Weight Personalized Recommender System for Scalable News Recommendation
ZoRRO is a training-free, zero-parameter framework for personalized news recommendation. The authors report that it beats neural baselines offline and comes close to a leading deep model in online click-through rate while running over 600 times faster; they also show that similar CTR can yield substantially different recommendation distributions and news exposure.
The combination of near-neural online performance with a claimed 600× speedup is potentially important, but the abstract gives too few details about the method, datasets, and statistical strength of the comparisons to warrant a strong recommendation.
-
Striatal Dopamine at Learned Sequence Boundaries Sustains Birdsong
In zebra finches, dopamine transients in the striatum shift toward the beginning of a learned song sequence as the song becomes well practiced. Temporarily inhibiting this dopamine signal at sequence onset caused progressive, severe degradation of adult song, suggesting that initiation-related dopamine is needed not just to learn a skilled behavior but to maintain its execution over time.
The causal link between sequence-onset dopamine and long-term maintenance of a naturally learned skill is a non-obvious extension of reward-guided sequence theories, though the abstract provides limited detail on scale and experimental robustness.
-
Hippocampal engrams configure prefrontal context representations to guide flexible decisions
The study reactivates hippocampal engrams in mice performing a context-dependent task-switching task while recording medial prefrontal cortex. Reactivating an engram made animals retrieve the associated decision rule—not merely produce a fixed motor response—and rapidly reinstated the corresponding prefrontal context representation while leaving other ongoing dynamics largely intact.
It provides a fairly distinctive causal link between memory engrams and rapid reconfiguration of cortical task representations for flexible behavior, but the abstract gives limited quantitative and scope information, so it falls short of a strong verdict.
-
Specialized layer 5 cortical terminals target functional thalamic spines
The study reports that frontal-cortex layer 5 neurons form functional excitatory synapses onto dendritic spines in the thalamus, a synaptic arrangement not previously established there. These contacts show spine-like calcium signaling, strongly influence thalamic spiking, recruit state-dependent subnetworks, and affect motor learning when disrupted.
This is a potentially important revision of the view of the thalamus as primarily a relay, but the abstract provides limited detail on scale, circuit generality, and behavioral evidence, so it falls short of a strong verdict.
-
Rapid odorant metabolism organizes identity- and timing-based odor representations by olfactory bulb inputs and outputs
In awake mice, the authors show that odor responses are not determined solely by receptor affinity and sensitivity. Rapid metabolism of odorants in the nasal mucosa produces secondary odorants during a single breath, recruiting additional olfactory sensory neurons and reshaping both the identity and timing of odor representations; timing differences between inhaled and metabolically generated odors remain visible in olfactory bulb output.
This is a non-obvious mechanistic account of olfactory coding that challenges receptor-binding-based models and proposes inhalation timing as a way to distinguish external odors from odors generated within the nose, but the abstract provides limited quantitative detail about the breadth and strength of the evidence.
-
Patient-derived tau-seeded human neuronal chimeras recapitulate mature Alzheimer's tau pathology and uncover human-specific neuronal vulnerability
The authors transplant human stem-cell-derived neurons into mouse brains, mature them in vivo, and seed them with tau isolated from Alzheimer’s patients. These human neurons develop mature, mixed 3R/4R tau pathology resembling neurofibrillary tangles and neuropil threads, with spreading, neurodegeneration, elevated plasma pTau-217, and memory deficits; mouse neurons are comparatively resistant. Single-cell data suggest greater tau uptake and synaptic disruption in human neurons, and a PSEN2 mutation worsens the phenotype.
This is a potentially important human-relevant Alzheimer’s model and supports a non-obvious species-specific vulnerability, but the abstract alone does not establish how broadly reproducible or therapeutically predictive the chimera is.
-
A molecular integrator of sleep duration and interruption
The study measures membrane PKA-substrate phosphorylation in freely behaving mice and finds that it falls exponentially during each sleep bout. The signal appears to combine sleep duration and interruptions, predicts the likelihood of waking, and changes appropriately after sleep deprivation, providing a molecular bridge between fast arousal circuits and slower sleep homeostasis.
This is a non-obvious mechanistic candidate for how the brain tracks sleep history within individual bouts, but the abstract provides limited detail on validation, causal evidence, and generality beyond the mouse measurements.
-
Primate Astrocyte Evolution Controls the Tempo of Neuronal Development
The study argues that prolonged human neuronal maturation is partly controlled by astrocytes rather than being solely neuron-intrinsic. Human astrocytes slow neuronal maturation relative to chimpanzee astrocytes, apparently through lower APOE expression regulated by Hippo-TEAD signaling; restoring APOE speeds development, and affected neuronal genes overlap with neurological disease genes.
It offers a genuinely non-obvious glial mechanism for human brain neoteny, supported by cross-species cell comparisons and a causal APOE rescue, but the abstract does not establish how well the in-vitro result generalizes to developing brains.
-
Challenges and Solutions in Quantifying Brain β-Hydroxybutyrate (BHB) with 1H-MRS Following Oral Keto-Ester Consumption
The study shows that brain 1H-MRS measurements commonly interpreted as β-hydroxybutyrate (BHB) after ketone-monoester ingestion also contain a substantial contribution from 1,3-butanediol (BD), which has a similar spectral signature. Using two edited MRSI acquisitions, the authors separately measured BD at 0.25–0.58 mM and BHB at 0.20–0.27 mM, suggesting that prior reports may have substantially overestimated brain BHB.
This identifies a non-obvious measurement confound that could revise the interpretation of existing brain-ketone studies, although the evidence is currently limited to a small volunteer experiment and methodological validation.
-
Hippocampal ripples are distinguishable from aperiodic activity
The paper argues that claims that awake-human hippocampal ripples are merely misidentified 1/f activity arise from a flawed surrogate-data test. Because adaptive ripple detectors lower their thresholds on signals containing only aperiodic activity, those surrogates produce inflated false-positive rates; when realistic ripple events are included, most false detections disappear, and aperiodic activity cannot reproduce ripple timing or spectral structure.
It directly challenges a live interpretation of human hippocampal ripples by identifying a specific evaluation artifact, with tests across multiple detectors, but the abstract does not establish broad validation on independent real recordings strongly enough for a strong verdict.
-
A single dynamical property can account for the capacity to learn, from artificial networks to the mammalian brain.
The authors test whether proximity to criticality—a dynamical regime near instability—predicts how quickly systems learn. Across mouse motor cortex, ferret visual cortex, and human frontal cortex, it reportedly predicts future behavioral learning rates, experience-driven neural tuning changes, and general cognitive ability, while not predicting initial or final performance; a minimal recurrent network provides a possible timescale-based mechanism.
The cross-species link between a measurable dynamical property and learning rate, together with a mechanistic network model, is substantially more interesting than a standard criticality claim, but the abstract gives no effect sizes or details sufficient for a strong verdict.
-
Egocentric spatial scaffolds organize cortical memory engrams
This study used engram tagging, repeated calcium imaging, and optogenetic silencing in retrosplenial cortex to examine how spatial information is incorporated into episodic memories. It reports that memory engrams are preferentially recruited from pre-existing neurons encoding self-position relative to environmental boundaries, forming a scaffold-engram network that is reorganized during recall. Silencing the scaffold weakened memory expression, while silencing the engram disrupted recall dynamics more completely.
The proposed causal scaffold-engram architecture offers a genuinely interesting account of how self-referenced spatial representations support episodic recall, but the abstract gives no quantitative detail about effect sizes, behavioral scope, or replication.
-
LLM-as-a-Verifier: A General-Purpose Verification Framework
The paper uses the probability distribution over an LLM’s scoring tokens, rather than a single prompted judgment, to produce continuous solution-quality scores. It scales verification through finer score granularity, repeated evaluations, and decomposed criteria, reporting gains across software agents, robotics, medical agents, and math, plus applications to progress monitoring and reinforcement learning.
The broad, training-free verifier and its use across several agentic settings are potentially useful, but the core idea appears closer to a systematic probabilistic refinement of LLM judging than a clearly new verification paradigm, and the abstract lacks comparative and ablation details needed to validate the strong claims.
-
Imagined Rollouts are Kinematic, Not Dynamic: A Diagnosis of Long-Horizon World-Model Failure
This paper argues that long-horizon world-model errors are not just generic compounding errors: imagined rollouts preserve kinematic motion patterns but fail to represent changing dynamics. Using a DreamerV3 checkpoint on DMC Walker-Walk, the authors introduce a per-step kinematic-consistency diagnostic and show that it remains nearly unchanged across a friction sweep even when the policy’s reward collapses at gait failure, suggesting the model does not capture the relevant physical regime change.
The kinematic-versus-dynamic framing and regime-boundary diagnostic are genuinely useful and somewhat surprising, but the evidence is currently limited to one checkpoint and one simulated task in a workshop paper.
-
LATO.2: Factorized 3D Mesh Generation with Vertex and Topology Flow
LATO.2 generates 3D meshes by modeling geometry and connectivity in separate flow-matching stages, both guided by a shared coarse voxel scaffold. The factorization is intended to reduce broken surfaces and drifting vertices, while enabling higher-resolution part-wise generation and topology changes that follow edited vertex geometry without re-optimization.
The explicit separation of continuous geometry from discrete topology, together with topology-adaptive editing and part-wise scaling, is a meaningful design direction, but the abstract provides no quantitative results or breadth of evaluation to support a stronger verdict.
-
OPSD-V: On-Policy Self-Distillation for Post-Training Few-Step Autoregressive Video Generators
OPSD-V is a post-training method for few-step autoregressive video diffusion models that trains the student under its actual rollout, including its own accumulated KV-cache errors. A teacher receives cleaner real-video temporal context while evaluating those same student states, providing dense denoising-level supervision without changing inference steps or cache structure. Applied to Self-Forcing and LongLive, it reportedly improves long-video quality, motion, and VBenchLong scores, with a small user study preferring the method in 66% of comparisons.
The on-policy teacher supervision with real temporal context is a meaningful and potentially reusable solution to long-horizon error accumulation in few-step AR video generation, but the abstract gives limited quantitative evidence and only a small user study.
-
PanoWorld: Real-World Panoramic Generation
PanoWorld is a panoramic world model that uses the rotation-equivariance of 360° representations to simplify camera-motion modeling and preserve longer-range spatial memory. It introduces dense panoramic ray conditioning and geometry-aware memory augmentation, and evaluates them on a new mix of real drone footage and AirSim360 simulations; the abstract claims large gains over prior methods but gives no quantitative results.
The geometry-aware treatment of panoramic rotation and long-range memory is a meaningful, potentially general direction for world models, but the abstract provides insufficient numerical or real-world evidence to justify a stronger recommendation.
-
A Quantized Native Runtime for On-Device Semantic Audio Generation
The authors implement the full Stable Audio 3 text-to-music pipeline in a dependency-free native runtime, including quantized inference and activation steering. They report that 8-bit quantization preserves measured quality while reducing memory, and that 4-bit quantization is sufficient to run the 1.2-billion-parameter model on an 8 GB Raspberry Pi 5; the runtime also reportedly starts about seven times faster than the official implementation. The release includes a demonstration of limited control over musical attributes via activation steering.
Running a large semantic audio model on a Raspberry Pi with little apparent quality loss is a meaningful deployment result, but the abstract gives limited quantitative detail and the work appears primarily to be a specialized systems/quantization implementation rather than a broadly new modeling direction.
-
EAGOR: Embodied Reasoning in Omni-direction
EAGOR treats 360° directional reasoning as Bayesian estimation on the sphere rather than prediction over a distorted equirectangular image. Its spherical-harmonic belief representation is training-free and designed to remain consistent as the robot moves; experiments on two benchmarks and a real legged robot report sizable gains in angular accuracy, navigation success, and efficiency.
The spherical, motion-equivariant belief formulation is a substantive alternative to standard ERP-based VLM reasoning, and the reported benchmark and real-robot gains make it worth examining, though the contribution is specialized to 360° directional navigation and the abstract does not establish broad generality.
-
Agentic Neural Architecture Search
AgentNAS has an LLM propose an initial architecture, then converts it into a modular scaffold whose slots define a task-specific search space for conventional NAS. Across 17 varied tasks, it reportedly achieves the best result on 11, with the LLM seed providing much of the gain and NAS improving it further through recombination of modules. The main contribution is a practical division of labor between open-ended LLM design and structured combinatorial search, avoiding hand-engineered search spaces.
The LLM-generated scaffold plus slot-level NAS is a credible, broadly tested way to combine two otherwise separate architecture-design paradigms, though the evidence is still primarily benchmark-based and the reported gains need scrutiny.
-
Provably Optimal Learning Algorithms for Assistance Games
The paper gives the first provably efficient online-learning algorithms for assistance games, where a human observes the latent state and an assistant sees only the human’s actions. It achieves a roughly T^{3/4} assistance-regret rate with a 1−1/e approximation, proves that improving this approximation is computationally hard, and obtains an almost-optimal T^{1/2} rate in a shared-randomness pseudo-decentralized setting.
This is a substantial theoretical treatment of partial-observation human–assistant learning, combining efficient algorithms, an approximation barrier, and near-optimal rates, though its practical relevance and empirical validation are unclear from the abstract.
-
Joint Discrete-Continuous Flow Matching for Open-Vocabulary Inverse Design of Multilayer Optical Coatings
IrisFlow jointly generates material choices and continuous layer thicknesses for multilayer optical coatings, while accepting the candidate materials, wavelength grid, and layer count at query time rather than fixing them during training. A 136M-parameter model handles 2–100 layers, generalizes to a held-out 15-material bank and wavelengths beyond its training range, and produced fabricated color-selective cooling coatings with low color error and high near-infrared reflectance.
The combination of query-time open-vocabulary conditioning with joint discrete/continuous flow matching is a meaningful inverse-design direction, supported by held-out materials, extrapolation tests, and fabrication results, though the evidence remains concentrated in one specialized optics application.
-
KV-PRM: Efficient Process Reward Modeling via KV-Cache Transfer for Multi-Agent Test-Time Scaling
KV-PRM scores intermediate steps in multi-agent LLM rollouts by reusing the generator’s existing KV cache, rather than re-encoding the full trajectory with a separate text-based process reward model. The authors report matching or improving task performance across MATH, GSM8K, and AIME while reducing scoring FLOPs by up to 5,000×, latency by 37×, and per-sequence memory by 34×.
Reusing generation-time KV caches for process reward modeling is a technically distinctive way to make long-context test-time scaling practical, and the reported efficiency gains are substantial, but the abstract does not establish how broadly the method transfers beyond these math benchmarks or whether the claimed information-capacity result is meaningful in practice.
-
FlashTrie: A GPU-Accelerated Constrained Beam Search for Generative Retrieval
FlashTrie moves trie-constrained beam search for generative retrieval fully onto the GPU. It combines a compressed GPU-resident trie with a cooperative CUDA kernel and GPU-friendly parallel pruning, reporting under 3 ms trie-search latency on an 800M-keyword library and up to a 24x speedup over a multithreaded CPU baseline. A commercial-search A/B test reportedly produced a +0.71% revenue lift by making larger-beam real-time decoding practical.
The end-to-end GPU design and large-scale online validation make this more than a routine kernel optimization, but it is primarily specialized systems engineering rather than a broadly new retrieval or decoding paradigm.
-
Agentic-DPO: From Imitation to Agentic Policy Optimization on Expert Trajectories
The paper proposes training LLM agents from expert trajectories without online environment interaction by contrasting the expert action at each state with sampled incorrect actions using a DPO-style objective. It also introduces Policy-Preserving Augmentation, which varies the rendering schema while keeping the underlying expert behavior fixed. Across three agent benchmarks, the method consistently beats SFT; on tau-bench retail, a 9B model improves from 21.7% to 41.4%, reportedly matching online GRPO with much lower rollout cost.
The combination of offline, state-level preference optimization and schema-preserving augmentation addresses a real cost bottleneck, with a sizable reported gain across multiple benchmarks, but it remains a DPO-style extension whose broader generality and comparison details need verification.
-
Bandit PCA with Minimax Optimal Regret
This paper determines the minimax regret for online PCA when the learner observes only the scalar reward of its chosen direction and the adversarial gain matrices are low-rank. It introduces multiscale exploration combined with mirror descent, and proves matching upper and lower bounds of approximately r√(dT), closing a substantial gap in prior results; the lower bound uses an adaptive hidden-subspace construction connected to quantum tomography.
Closing the minimax gap for a difficult bandit-feedback problem with a new multiscale exploration method is a meaningful theoretical advance, but its impact is mainly within online learning and the abstract provides no empirical or broader capability evidence.
-
Search Beyond What Can Be Taught: Evolving the Knowledge Boundary in Agentic Visual Generation
This paper studies visual generators that fail on long-tailed, changing, or post-training-cutoff entities, and introduces a benchmark and search corpus for measuring those failures. It argues that simply adding retrieved information is ineffective because each generator has an evolving boundary between knowledge it can absorb during training and knowledge that should remain in external context; a teach-then-search co-training procedure is proposed to discover that boundary and improve generation. The abstract reports a striking low baseline score, but gives no concrete improvement numbers for the proposed method.
The generator-specific knowledge-boundary framing and teach-then-search setup are potentially useful, but the abstract mainly presents a benchmark and qualitative recipe without enough quantitative evidence to establish a major advance.
-
HunyuanOCR-1.5: Making Lightweight OCR VLMs Faster and Better
HunyuanOCR-1.5 is a lightweight OCR-focused vision-language model covering document parsing, text spotting, extraction, translation, and multi-image understanding. It keeps the previous backbone but adds DFlash-based decoding for long structured outputs, reporting a 6.37× Transformer-level and 2.14× vLLM speedup, plus an agent-driven data-generation pipeline aimed at long-tail OCR cases. The abstract claims broader performance on scripts, charts, tables, multilingual documents, and long-context inputs, though it gives few concrete accuracy comparisons.
The reported decoding speedup for dense OCR outputs is practically significant, and the automated weakness-to-data pipeline is a useful direction, but the abstract does not establish a major capability breakthrough or provide enough detailed evidence to merit a strong recommendation.
-
SIEVE: Structure-Aware Data Selection for Imitation Learning with VLA Models
SIEVE selects robot demonstrations by identifying reusable visuo-motor primitives and transition patterns, then choosing representative trajectories within those structural groups. Across several datasets, benchmarks, and VLA models, it reportedly beats existing selection methods and can exceed full-data training while using half the demonstrations and training steps. The main contribution is treating long-horizon demonstrations as compositions of reusable structures rather than scoring whole trajectories or individual state-action pairs.
The structure-aware view of demonstration selection and the claimed 50% data/compute result are plausibly useful for scaling VLA training, but the abstract provides no quantitative margins or detail about benchmark breadth, so it is not clearly a must-read.
-
Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE
Jet-Long is a tuning-free method for extending an LLM’s context window by combining a local, original RoPE region with a dynamically rescaled long-range region. The authors report up to 128K-context gains on RULER, HELMET-RAG, and PG-19 across several Qwen3 sizes, while adding little generation overhead and improving prefill throughput with a fused kernel. The main novelty is the bifocal, sequence-length-dependent RoPE schedule and its efficient implementation, rather than a new training procedure.
It offers a practical and reasonably well-supported improvement to zero-shot long-context extension, including efficiency and multi-model results, but dynamic RoPE rescaling is an active line of incremental extensions rather than an obviously new direction or major capability jump.
-
Scalable Visual Pretraining for Language Intelligence
The paper pretrains models directly on visual documents—preserving figures, equations, typography, and page layout—instead of converting the same data into plain text. It reports that this visual pretraining consistently improves language-intelligence benchmarks across multiple model backbones and argues that visual representations can be a scalable alternative or complement to text-only pretraining.
The central claim—that retaining document visuals during pretraining improves general language intelligence—is potentially important, but the abstract provides no quantitative gains, scale details, or evidence that this goes beyond established multimodal/document-pretraining approaches.
-
SILO: Simulation-in-the-Loop Sim-to-Real Transfer for Multi-Stage Cable Routing
The paper trains reinforcement-learning policies in thousands of parallel cable simulations and deploys them on a real robot using a simulation-in-the-loop execution scheme, localized policies, and cable-state estimation. On multi-stage cable-routing tasks, it reports higher success rates and roughly half the cycle time of prior learning-based methods, while generalizing across cable geometries and deformation patterns.
Real-world sim-to-real RL for multi-stage deformable cable routing is a meaningful capability and direction, but the abstract gives no task counts, success-rate numbers, or comparison details to establish how broad or substantial the gains are.
-
Elastic Gang: Per-Token Membership Change for a Hard-Barriered LLM Inference Gang Co-Scheduled with OS Processes
The paper introduces an “elastic gang” scheduler that lets the CPU cores participating in hard-barriered LLM decoding change between tokens without deadlocking or corrupting results. A prototype kernel on an 8-core AMD Zen 5 machine preserves bit-exact output while improving general-purpose workload throughput by 1.28–1.75× over static core partitioning at comparable inference throughput, and returns unused cores to the OS when decoding is idle.
The epoch/acknowledgment protocol is a concrete and non-obvious solution to safely time-share barriered CPU inference with ordinary OS processes, but the demonstrated hardware and workload scale are limited.
-
PRISM: Personalized Robotic Dataset Generation via Image-based Scene and Motion Synthesis
PRISM generates robot-training demonstrations tailored to a target environment using only one image and a natural-language task instruction. It creates varied but geometrically aligned synthetic scenes and executable motions, then trains policies that reportedly outperform other synthetic-data baselines on LIBERO/LIBERO-Plus and reach up to 100% success on three real-world manipulation tasks, with better transfer to changed environments.
The combination of single-image personalization, diverse scene generation, and automatic executable demonstrations could materially reduce the need for teleoperation, but the abstract gives few quantitative details and the real-world evidence is limited to three tasks.
-
Input Pathways Shape Few-Shot, Not Zero-Shot, Binding in Tiny Transformers: A Fully-Enumerable Study
This exhaustive study trains tiny transformers on finite factored worlds to test how symbolic, oracle, and entangled input representations affect compositional binding. It finds that none achieve converged zero-shot binding despite having all necessary information, while few-shot performance depends on two separate factors: parameter sharing in the input pathway and how readable the code is; diagnostic tests also reveal distinct readout and slot-misbinding failure modes.
The fully enumerated, information-matched setup and the separation of pathway sharing from code readability provide a non-obvious constraint on why tiny transformers bind compositionally, but the results remain bounded synthetic findings rather than a demonstrated advance in capability or scale.
-
Grokking Is Conditional and Fragile: A Fully-Tractable, Multi-Seed Study at 12K Parameters
This paper studies grokking in an unusually small (~12K-parameter) transformer where the full model and input-output behavior can be enumerated, using many random seeds rather than one run. It finds that grokking is conditional on training coverage, highly sensitive to numerical details such as CPU reduction order and CPU/GPU execution, and that several striking single-run effects disappear under multi-seed analysis. The results suggest that grokking should be reported as a rate under a fixed computational environment, not as an inevitable property of a training setup.
The combination of fully tractable models, multi-seed rates, and evidence that numerical nondeterminism can flip grokking outcomes is a genuinely useful and somewhat assumption-challenging methodological result, though its broader relevance beyond tiny modular-arithmetic models is unproven.
-
RABBiT: Rapidly adaptive BOLD foundation model via brain-tuning for accurate zero-shot and few-shot prediction of speech-elicited responses in the brain
RABBiT is a compact audio-to-fMRI model that predicts speech-evoked brain responses across participants and unseen datasets. It combines region-specific attention, shared versus subject-specific response components, and parameter-efficient adaptation; across 324 participants, it reportedly beats an existing fMRI foundation model and group-average baselines, with further gains from only 10 minutes of participant data. The main contribution is a potentially scalable way to model individual variation in language-related brain activity without fitting a full model per subject.
The cross-participant zero-shot and 10-minute few-shot results, if supported by strong held-out evaluations, could materially improve scalable brain encoding, but the abstract gives no quantitative effect sizes or detail about dataset leakage and comparison conditions.
-
Rethinking On-Policy Self-Distillation for Thinking Models
The paper finds that giving thinking models privileged solution context during on-policy self-distillation can make them worse, with up to a 17% relative drop in avg@16 accuracy across five Qwen3 and OLMo models on three difficult math benchmarks. The effect grows with longer reasoning budgets and appears tied to suppressing high-entropy self-correction behaviors such as verification, backtracking, and reconsideration; ordinary on-policy distillation does not show the same failure.
This is a well-supported and notably counterintuitive failure mode for privileged distillation in reasoning models, with a plausible token-level mechanism, though it appears primarily diagnostic rather than a major capability advance.
-
The Exact Worst-Case Tail Probability under Bounded Kurtosis
The paper exactly solves the extremal problem of maximizing \(P(X\ge t)\) when a real random variable has mean zero, variance one, and bounded fourth moment, without assuming symmetry or bounded skewness. It gives a four-regime phase diagram, explicit extremal distributions and dual certificates in the closed-form regimes, and characterizes the remaining regime algebraically; notably, fourth-moment information can be completely useless over part of the threshold range. The results also yield exact worst-case quantiles and implications for robust-estimation constants and degree-4 certifiable reasoning.
This is a genuinely exact and structurally rich result for a classical moment problem, with the surprising finding that kurtosis may not improve one-sided Cantelli bounds, but its practical impact for mainstream AI/ML is indirect and the abstract does not establish a major downstream capability gain.
-
Is the Geometry Doing the Work? An Operating-Point Audit of Hierarchy in Hyperbolic Vision-Language Models
This paper audits whether hyperbolic vision-language models actually use hyperbolic geometry to represent semantic hierarchies. Across seven checkpoints and controlled retraining experiments, embeddings stayed close to Euclidean, entailment cones were often trivially wide, and tests found little consistent evidence of radial or cone-based hierarchy; lowering curvature appears to be an easy shortcut rather than learned structure. The authors propose a compact five-number report for future geometry claims.
It provides a relatively thorough, controlled negative result challenging a central claim of hyperbolic VLMs— that their geometry encodes hierarchy—though the impact is mainly diagnostic and may be limited to the audited formulations and checkpoints.
-
Plausible but Not Valid: A Psychometric Audit of LLMs as Synthetic Survey Respondents
This paper evaluates LLM-generated survey responses against human data at the level of correlations, latent structure, reliability, demographic effects, and downstream validity—not just whether individual answers look plausible. Across 37 models, a simple Gaussian-copula baseline outperformed all LLMs on the proposed score; LLMs were more similar to one another than to humans, showed strong acquiescence bias, produced poor human-held-out prediction, and sometimes invented mediation effects. The findings suggest that synthetic respondents are not interchangeable with real survey participants, although the evidence comes from one organizational-psychology dataset.
The psychometric, population-level audit reveals several non-obvious failure modes—including synthetic consensus, demographic distortions, and fabricated mediation—using broad model and baseline comparisons, but its narrow single-dataset setting limits generality.
-
REDDIT: Correcting Model-Generated Timestamp Drift in ASR without Forgetting via Replay-Based Distribution Editing
The paper identifies a specific failure mode in autoregressive ASR: timestamps drift during long non-speech intervals even when the transcript remains plausible. Its replay-based post-training method edits timestamp behavior while preserving the base distribution for ordinary transcription, using synthetic gap data rather than human timestamp annotations; on Whisper-tiny, long-gap mIoU improves from 38.7% to 95.0% and mixed-gap timing error falls from 2752 ms to 223 ms with much less forgetting than ordinary fine-tuning.
The combination of self-supervised timestamp correction and distribution-preserving replay addresses a real, non-obvious failure mode with unusually large gains, but the evidence appears concentrated on a narrow task and mainly Whisper-tiny, limiting its broader significance.
-
CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents
CompactionRL trains agentic language models to summarize and compress their interaction history during reinforcement learning, rather than treating context compaction as a separate inference-time trick. The method combines task and summary losses with normalized token weighting and cross-trajectory advantage estimation, yielding reported gains of 3.1–7.0 Pass@1 points on SWE-bench Verified and Terminal-Bench 2.0 across two open-model bases.
Jointly learning context compaction within RL is a relevant and plausibly reusable direction, and the gains are sizable across two coding-agent benchmarks, though the abstract lacks ablations and details needed to distinguish the contribution from improved training or prompting.
-
Forgetting Is Not a Fix: Path Dependence in Sequential Engram Editing
This paper tests whether AI Engram edits remain compositional and order-independent when applied repeatedly, rather than only in single or paired edits. Across three model configurations, it reports substantial divergence between direct composition and sequential recalibration, order-dependent interference, progressive drift in the method’s own statistics, and partial recovery of erased knowledge after later edits.
The result directly challenges a specific commutative-memory hypothesis with cross-model sequential experiments, but its significance is mostly confined to one recent editing method and the abstract provides limited detail about scale and independent replication.
-
PERSONAJUDGE: Simulating Individual Human Preference Judgments with Evaluator-Specific Demonstration Data
The paper studies whether an LLM can imitate individual human preference judgments rather than only predicting group consensus. Using judgments from 32 annotators, it conditions the model on evaluator-specific reasoning traces and interface telemetry; reasoning traces improve agreement by up to 9.9 percentage points, while telemetry often degrades performance. It also finds that an evaluator’s tendency to choose neutral ratings predicts how difficult they are to simulate across tasks.
The evaluator-specific framing and finding that cheap behavioral telemetry can hurt while reasoning traces help are genuinely interesting, but the study is moderate-scale and appears more like a useful evaluation methodology than a major capability advance.
-
FORGE: Towards Functional Tool-Use Generalization via Keypoint Trajectory Reasoning
FORGE addresses a robot’s inability to use unfamiliar objects for a known function, such as using different objects to hit a nail. It predicts general-purpose 2D keypoint trajectories from action-free data, then grounds those trajectories into robot actions using relatively few demonstrations; on a seven-tool hitting benchmark, it reports more than twice the average success rate of prior methods on unseen tools in simulation and real-world tests.
The explicit separation of functional trajectory reasoning from tool-specific motor execution, together with a reported 2× transfer gain on real unseen tools, is a substantial and relevant advance, though the abstract gives too few absolute results and experimental details for a strong verdict.
-
Level-Crossing Density as a Mesh-Free High-Frequency Auxiliary Loss for Implicit Neural Representations
The paper derives mesh-free, differentiable estimators of excursion-set area, boundary measure, and Euler characteristic for implicit neural fields, using autodiff and scattered samples rather than grids or persistent homology. In 2D, combining these quantities improves topology while preserving fidelity at much lower per-iteration cost, but in 3D optimization can hide small topological defects below the sampling resolution; a persistent-homology baseline is substantially more accurate on that benchmark.
The differentiable Minkowski-functional loss and its explicit analysis of adversarial sampling blind spots are genuinely interesting, but the central 3D result exposes a serious limitation and keeps this from being a broadly convincing replacement for persistent homology.
-
Information Limits and Attractor Dynamics in Economies of Frontier LLM Agents: A Pre-Registered Test
This pre-registered study tests information-theoretic and mean-field predictions in small economies of Claude Opus 4.8 agents. In market-coupled settings, wealth differences closely tracked claimed information, including submodular versus XOR-induced supermodular coalition effects, while the proposed smooth population misalignment law failed: populations collapsed to consensus or switched abruptly between bistable regimes. The authors release cached outputs, protocols, and code for reproducibility.
The combination of mechanistic information-law tests and a preregistered negative result showing that LLM populations do not enter the assumed smooth dispersion regime is genuinely interesting, though the evidence is limited to small artificial economies and one frontier model.
-
Stability Annealing Selects the Implicit Bias of Smoothed Sign Descent: A Rate-Indexed Barrier Path on Separable Data
This paper analyzes full-batch smoothed-sign descent for separable linear classification when its numerical-stability parameter is annealed over time. It proves that the convergence direction follows a specific convex Burg-barrier solution on a margin-constrained slice, with an explicit convergence-rate bound, and connects the dynamics exactly to entropic mirror ascent; experiments check the identities and predicted behavior.
The rate-dependent interpolation between sign-descent and fixed-stability behavior, together with an exact barrier-path characterization and rate bound, is a substantive theoretical explanation rather than a routine optimizer tweak, though its scope is limited to a specialized linear setting.
-
When Do Geometric Algebra Layers Beat Scalarization? A Controlled Study on SO(3)-Equivariant Vector Laws
This controlled study compares compact geometric-algebra networks with an equally SO(3)-equivariant scalarization baseline on synthetic 3D vector laws. Geometric algebra provides little or no benefit for single-stage or algebraically flattenable operations, but substantially improves low-data learning of deeply nested rotations and orientation-dependent force computations: it reaches comparable performance with about 100 samples versus 3,000 for the baseline. The results suggest that the useful ingredient is compositional representation of group operations, not equivariance or geometric algebra by itself.
The paper offers a valuable, non-obvious diagnosis of when geometric-algebra layers help—deep composition of rotations rather than equivariance in general—with strong controlled baselines, though the evidence is limited to synthetic SO(3) tasks.
-
At-Grok Is Not Converged:A Measurement-Validity Audit for Grokking Representation Metrics
This paper argues that representation compression does not coincide with the moment a model begins to generalize in grokking experiments: embeddings can keep compressing for tens of thousands of steps afterward. It proposes an audit procedure that separates generalization onset from later compression, handles censoring and non-generalizing cells, and shows that LayerNorm substantially changes the lag; a proposed depth law also fails to transfer beyond MLPs.
The quantified finding that compression lags grokking by roughly the time-to-grok, plus controls and an adversarial audit of a widely used metric, is genuinely non-obvious, though the evidence is currently confined to modular-arithmetic MLPs and transformers.
-
DepthWeave-KV: Token-Adaptive Cross-Layer Residual Factorization for Long-Context KV Cache Compression
DepthWeave-KV compresses long-context KV caches by sharing low-rank representations across neighboring layers while preserving token-specific residuals for tokens that matter to retrieval and attention. It also adapts the compression online and provides a fused CUDA implementation; the abstract reports near-full-cache quality, 8.3× lower KV memory, and 72.8 tokens/s at 64K context across several long-context evaluations.
The combination of cross-layer factorization, token-conditional rank allocation, and online adaptation could represent a substantial practical advance in long-context inference, but the abstract lacks baseline details, hardware comparisons, and enough quantitative evidence to justify a strong verdict.
-
Hierarchical Acoustic-Semantic Modeling: Modality Separation and Semantic Coherence for Full-Duplex SLMs
The paper studies why end-to-end full-duplex spoken-language models lose semantic quality and interaction naturalness, attributing this to gradient conflicts between acoustic and semantic objectives in shared deep parameters. It proposes Lychee-FD, which separates modality-specific parameters in deeper layers while retaining a semantic alignment pathway; experiments report a 7.4% Spoken QA improvement and a 28.5% FullDuplexBench 1.5 improvement without added inference cost.
Native full-duplex speech models and explicit hierarchical modality separation are relevant and plausibly substantial, but the abstract provides relative gains without baselines, absolute results, or enough detail to verify whether the claimed optimization insight and improvements generalize.