Issue 29 · Jul 13–19, 2026
Every candidate
All 2,668 papers were scored from their abstracts by gpt-5.6-luna; 944 were not skipped. Shown here: the top 150 of those, in score order. Picks are marked.
-
RoboTTT: Context Scaling for Robot Policies
RoboTTT uses test-time training: gradient-updated fast weights act as a recurrent memory, allowing robot policies to condition on up to 8K visuomotor timesteps without increasing inference latency. On real-robot manipulation, it enables one-shot imitation from human video, online improvement, perturbation recovery, and completion of a five-minute, ten-stage assembly task; 8K-context pretraining improves performance by 62% over 1K context and by 87% over a single-step baseline.
This presents a genuinely new context-scaling mechanism for robot foundation models and reports substantial long-horizon capability gains, including a task no baseline completed, though the abstract does not establish how broadly the results generalize.
-
The Benjamini--Hochberg Procedure Can Fail to Control the FDR for Correlated Two-Sided Gaussian Tests
This paper gives a correlated Gaussian factor model in which the standard Benjamini–Hochberg multiple-testing procedure exceeds its nominal false-discovery-rate level: at α=0.01, the authors rigorously certify FDR > 0.0104 for sufficiently many tests. The result overturns a long-standing conjecture about two-sided Gaussian tests under dependence, with simulations agreeing; the proof was generated with GPT-5.6 Pro and checked by the author.
It presents a rigorous, assumption-revising counterexample to a widely believed foundational claim about FDR control under dependence, rather than an incremental statistical improvement.
-
Agora: Collective and Permissionless Internet-Scale Pretraining of Large Language Models
Agora presents a system for training large language models across heterogeneous, unreliable, individually owned GPUs connected only over the internet. Its 40-day demonstration trained an 8.6B-parameter model on 500B tokens using 330 changing contributor nodes, reportedly reaching 63% of a centralized H100 baseline and converging similarly to a centralized run, while no participant held the full model weights.
The combination of permissionless collective ownership, internet-scale pipeline parallelism, and a substantial 8B pretraining run would establish a genuinely new direction if the reported efficiency and convergence claims hold, though the abstract alone provides limited independent validation.
-
TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs
TimeLens2 trains multimodal video models to return a variable number of evidence intervals rather than a single temporal segment. It introduces interval-set supervision and a matching-free Wasserstein reward, supported by a 93K-example dataset built with proposal generation, consensus, verification, and boundary refinement; the authors report large gains over Qwen3-VL backbones and strong results across seven benchmarks, including smaller models beating much larger open models.
The set-valued formulation and matching-free temporal reward are a substantive, plausibly general solution to a real weakness in video grounding, but the unusually strong benchmark claims are presented in a technical report with limited methodological and experimental detail in the abstract.
-
Auditing the Risk Claims of Distributional Reinforcement Learning
This paper tests whether the return distributions learned by distributional RL agents actually support the risk-sensitive claims made from them. Across QR-DQN, C51, and IQN, it finds that most of the strongest apparent risk trade-offs are statistically false, appear early and idiosyncratically across seeds, and persist even at Atari scale; positive controls suggest the audit can detect real effects. The result is that distributional heads may be encoding training artifacts rather than useful information about environmental uncertainty or risk.
It directly challenges a widespread interpretation of distributional RL, with strong statistical controls and replication across algorithms and scales suggesting that learned risk signals may be fundamentally uninformative rather than merely miscalibrated.
-
Innocuous-Seeming Data, Latent Ideology: Ideological Generalisation in Finetuned LLMs
The paper shows that fine-tuning an LLM on small, seemingly innocuous datasets can induce ideological or epistemic shifts on unrelated topics, while leaving general task performance largely unchanged. Across GPT-4.1 and Gemma-3, these shifts generalize beyond the training domain and become more extreme than the effects of few-shot prompting, including endorsement of highly out-of-distribution views.
The broad, replicated finding that narrow fine-tuning can amplify latent ideological tendencies far outside the training distribution—and do so without obvious capability loss—challenges the assumption that domain adaptation is behaviorally localized.
-
An Exam for Active Observers
The paper introduces ActiveVision, a 17-task benchmark intended to test whether multimodal language models actively choose and revise visual observations rather than answer from a single image. Reported frontier models perform extremely poorly—GPT-5.5 solves 10.6% of items and Claude Fable 5 only 3.5%, versus 96.1% for three humans—even when allowed to write and execute vision code.
The striking result is that strong MLLMs may lack the closed-loop visual information-gathering ability assumed by many interactive reasoning systems, though this is still primarily a benchmark paper with limited evidence about generality beyond its 17 tasks.
-
Loop the Loopies!
The paper introduces Loopie, two sparse Mixture-of-Experts models that repeatedly apply a relatively small Transformer while using only a fraction of their total parameters per token. It claims that, unlike earlier looped Transformers, this design beats conventional models at equal pretraining compute and reaches frontier-level reasoning after post-training, supported by ablations against a 30B-A3B baseline.
The potentially important result is that looping may outperform simply scaling parameter count under a fixed compute budget, but the abstract gives no quantitative results or details sufficient to justify a strong recommendation.
-
Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model
Xiaomi-Robotics-U0 is a 38B multimodal autoregressive model that unifies image/video generation with robot-specific tasks such as multi-view scene synthesis, embodiment transfer, and embodied video generation. The authors report better human-rated embodied generation than GPT-Image-2.0 and a large improvement in out-of-distribution manipulation success for pi_0.5, from 36.9% to 63.2%, suggesting the model can also act as a scalable source of robot-training data.
The unified world-model/data-engine approach and reported real-robot generalization gain are genuinely notable, but the abstract gives few experimental details and relies heavily on broad state-of-the-art claims, so it merits inspection rather than a strong recommendation.
-
Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models
The paper uses dependency-graph analysis to mask whole functions during fill-in-the-middle mid-training, treating missing function outputs as analogous to tool returns in coding-agent trajectories. Training on 2.6B Python tokens improves SWE-Bench by roughly 2.8–5.4 points across three models and two agent-training setups, while also reducing losses on ordinary coding and non-coding tool-use tasks after agentic post-training.
The function-call framing and targeted FIM objective are a meaningful intervention, and the cross-model, cross-pipeline gains plus preservation of general tool use are non-obvious, but the reported improvements remain moderate and may reflect a specialized training refinement rather than a major new direction.
-
StructAgent: Harness Long-horizon Digital Agents with Unified Causal Structure
StructAgent represents long-running computer-use tasks with an explicit, compact state and a verifier-controlled workflow instead of relying on raw interaction history. It adds checkpointing, evidence-based completion, and targeted recovery, with reported gains from 27.0% to 46.9% and 31.6% to 62.2% on OSWorld-Verified for two Qwen models, plus 78.9% with MiniMax-M3 and transfer to Minecraft.
The combination of explicit verifiable progress state and structured recovery appears to produce unusually large long-horizon agent gains across backbones and environments, though the causal framing and benchmark setup warrant verification.
-
Answer-Conditioned Chains of Thought Degrade Verifiable-Reasoning Distillation in Large Language Models
The paper studies a common reasoning-distillation practice: showing a language model the correct answer and asking it to generate a chain of thought that reaches that answer. Across multiple reasoning models and problem difficulties, these answer-conditioned traces appear to rationalize backward rather than genuinely derive the solution, causing large downstream accuracy losses—up to roughly 27 points on the hardest competition problems—even when incorrect traces are filtered out. The authors show that the degradation can be detected from unlabeled traces and is caused by the instruction to rationalize toward the answer, not merely by revealing the answer.
This identifies a non-obvious failure mode in a widely used reasoning-distillation recipe, with unusually large reported effects, cross-model transfer, controlled ablations, and a potentially actionable data-generation rule.
-
Low-latency neuromorphic closed-loop control of hippocampal ripples in vivo
The authors deploy a tiny spiking neural network on SpiNNaker to detect hippocampal sharp-wave ripples in awake mice and trigger optogenetic inhibition in real time. Across 23 recording sessions, the system uses up to 200-fold less energy than deep-learning alternatives, achieves about 50 ms end-to-end latency, and can stimulate during up to 80% of ripples; the intervention measurably changes ripple dynamics.
This is a rare in-vivo demonstration of an integrated neuromorphic sensing-to-stimulation loop operating fast enough to manipulate transient hippocampal events, combining substantial energy savings with causal neural intervention rather than offline detection alone.
-
Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation
The paper uses a pretrained multimodal language model as a text-to-image RL reward model by scoring how likely it is to reconstruct the original prompt from the generated image, requiring no preference-data training. It also proposes a self-rewarding variant in which a unified model’s understanding branch evaluates its own generations; experiments across diffusion models, RL algorithms, nine MLLMs, and five out-of-distribution benchmarks reportedly show consistent gains and that policy–reward alignment can matter more than reward-model size.
The prompt-recovery reward and self-evaluation loop are a non-obvious, potentially useful alternative to separately trained preference models, supported by unusually broad experiments, but the abstract gives no quantitative gains or evidence that the approach generalizes beyond this setting.
-
MonkeyOCRv2: A Visual-Text Foundation Model for Document AI
MonkeyOCRv2 pretrains a document-focused vision encoder on 113 million images in 17 languages using both image-to-text generation and pixel-level document reconstruction. It improves several document-analysis tasks and, when frozen, enables a 0.7B document parser that reportedly beats a 3B competitor by 2.8 percentage points while using an encoder about 11 times smaller.
The combination of large-scale document-specific pretraining with pixel reconstruction and the reported accuracy-versus-model-size advantage is substantially more than an incremental OCR tweak, but the abstract provides limited detail beyond benchmark and SOTA claims.
-
Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories
Xiaomi-Robotics-1 is a large vision-language-action model trained on more than 100,000 hours of real-world manipulation trajectories, with automatically generated language descriptions of scene and action transitions. The authors report that performance scales with both data and model size, transfers to better zero-shot real-robot behavior, and improves fine-tuning efficiency; it reaches 57.4% on RoboCasa365 and 20.07 on RoboDojo, above prior reported results.
The combination of unusually large real-robot data, scalable auto-labeling, and evidence that pretraining scale transfers to downstream robot performance is worth examining, but the abstract gives limited quantitative detail about real-world gains and much of the result may be a scale-and-engineering advance rather than a fundamentally new method.
-
xHC: Expanded Hyper-Connections
xHC extends Hyper-Connections to 16 parallel residual streams by enriching write-back information and updating only 4 streams per layer while still reading the full state. In experiments on 18B and 28B MoE models, it reports a 4-point average downstream improvement over mHC, lower compute to reach the same loss, and an xHC-Flash implementation that substantially reduces memory traffic.
This is a potentially important new scaling axis for Transformer residual capacity, supported by relatively large-model and compute-efficiency results, but the abstract provides limited detail for judging how broadly the gains transfer or how much depends on the specific MoE setups.
-
Towards Human-level Dexterous Teleoperation
TeleDexter is a teleoperation controller that jointly tracks the operator’s hand and the manipulated object, using learned low-level contact behaviors rather than imitating every reference frame. It reports zero-shot transfer to two real dexterous hands across seven object-reorientation and long-horizon tool-use tasks, with 75% average success while the baselines reportedly fail; the demonstrations also train autonomous policies through behavioral cloning.
This presents a potentially important real-world capability jump in dexterous robot teleoperation—dynamic grasp changes, in-hand manipulation, and tool use—with unusually broad task and hardware evaluation, though the abstract provides limited detail on baselines and experimental scale.
-
LongStraw: Long-Context RL Beyond 2M Tokens under a Fixed GPU Budget
LongStraw is a systems method for running GRPO-style long-context RL on prompts exceeding 2M tokens without keeping the full prompt and all response computation graphs resident on GPU. It virtualizes architecture-specific recurrent/KV/attention/MoE state, replays responses one at a time, and delays distributed gradient finalization; the design is implemented for two substantially different model architectures. The provided abstract describes the mechanism but gives no quantitative memory, throughput, context-length, or training-quality results.
The architecture-aware state virtualization and response-replay design could materially expand feasible long-context RL, but the abstract provides no results to establish that it actually works efficiently or improves training.
-
NexForge: Scaling Agent Capabilities through Requirement-Driven Task Synthesis for LLMs
NexForge generates executable agent-training tasks from high-level capability requirements rather than relying on predefined tools, repositories, or skill graphs. It automatically constructs the needed environments and collects expert trajectories; training on its data reportedly raises Terminal-Bench performance from 22.5% to 52.0% with 3.6K tasks, and larger scaling reaches 58.4% and 75.3% in later models, alongside strong GDPval gains.
The requirement-driven, substrate-independent task-generation approach and large reported agent gains are genuinely notable, but the abstract provides too little detail to validate the ambitious state-of-the-art and frontier-model comparison claims.
-
MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators
MeanFlowNFT adapts forward-process reinforcement learning, previously developed for diffusion models, to MeanFlow generators that predict average rather than instantaneous velocities. The method uses an induced instantaneous-velocity predictor for the RL objective while retaining MeanFlow’s few-step sampling, and reports improvements on image and video generation; notably, 4-step generation on Wan 2.1 reportedly exceeds a 50-step RL-tuned baseline on VBench.
The technically nontrivial extension of forward-process RL to average-velocity models, plus the claimed 4-step-versus-50-step video result, is worth checking, but the abstract provides limited experimental detail and the cross-model benchmark comparison may not establish a genuinely broad capability jump.
-
Qwen-Music Technical Report
Qwen-Music is an end-to-end music generation system for text/lyrics-to-song and cover-song generation, combining semantic audio tokens, an autoregressive language model, melody-first planning, and a separate waveform renderer. It is trained on more than 5 million hours of multilingual music and reportedly leads on 13 of 16 objective metrics, with human preference over proprietary systems, though the abstract gives no numerical margins or evaluation details.
The melody-planning mechanism and reported combination of coherent singing, melody-preserving covers, and high-fidelity rendering are plausibly meaningful advances, but the evidence is mainly unsupported state-of-the-art claims without quantitative margins or reproducible comparisons.
-
FlashDecoder: Real-Time Latent-to-Pixel Streaming Decoder with Transformers
FlashDecoder replaces the usual 3D-convolutional video latent decoder with a causal Transformer that generates pixels frame by frame using a bounded rolling KV cache. On Wan2.1 and Wan2.2 latent spaces, it reportedly preserves reconstruction quality at 1080p while decoding 3.6–4.7× faster, using up to 11× less memory, and reaching 12× speedup with inference optimizations.
This addresses a major but sometimes neglected bottleneck in real-time video generation with a genuinely streaming, memory-bounded decoder and large reported speed and memory gains at high resolution, though the evidence is limited to two latent spaces and one GPU.
-
Hierarchical Denoising For Multi-Step Visual Reasoning
The paper proposes HDR, a video-generation framework that uses hierarchical latent representations and sparse attention to plan coarse visual hypotheses before refining them into streaming outputs. On a new benchmark of multi-step visual tasks, it reports substantially higher success than streaming diffusion baselines, much lower inference cost than bidirectional diffusion, and better data efficiency; robot experiments provide a limited demonstration of physical-world use.
The hierarchical coarse-to-fine reasoning design and reported gains in consistency, latency, and low-data training are genuinely notable, but the evidence relies heavily on a new benchmark and the abstract gives too little detail about baseline fairness, scale, and the robot results for a strong recommendation.
-
Globally Consistent Coloring Schemes for Language Identification
This paper studies whether adversarial language identification needs a full color trace attached to every example. It proves that for any countable collection of infinite languages, a single terminal bit per string is sufficient—and that one globally chosen two-color assignment works for every countable subcollection. The construction is necessarily nonconstructive in a precise sense: no Borel finite-color global assignment can have this property, while constructive trace-coloring approaches require infinitely many colors.
The striking result is the collapse from an entire per-example color trace to one globally fixed bit, together with a sharp constructive/nonconstructive boundary, but its impact is mainly foundational and confined to an abstract learning model.
-
Verbalizable Representations Form a Global Workspace in Language Models
The paper proposes a Jacobian-based interpretability method for finding representations that a language model is ready to verbalize, calling this subset the “J-space.” Across several tests, these representations appear to support reportable content, deliberate reasoning, memory, and broadcasting across the network, while routine processing occurs outside them. The authors argue this provides an operational analogue of a global workspace and a way to inspect otherwise unspoken model reasoning.
The global-workspace framing and Jacobian lens could be a genuinely important direction for mechanistic interpretability, but the abstract gives few quantitative details and makes broad cognitive claims that need careful validation.
-
Understanding Reasoning from Pretraining to Post-Training
The paper studies how pretraining affects later reinforcement learning for reasoning, using small language models trained on chess games, synthetic reasoning traces, and verifiable-reward RL. It reports that pretraining loss predicts post-RL performance and that more pretraining tokens make RL improve faster; mechanistically, RL reinforces already-likely correct moves on easy problems but can uncover nearly absent correct moves on hard ones. A similar pattern appears in a 1B-parameter model trained on mathematical text.
The controlled pretraining-to-RL analysis and claim that RL can retrieve weakly represented correct solutions—not merely sharpen the SFT policy—are genuinely interesting, though the evidence is based mainly on synthetic chess/math settings and the abstract gives limited quantitative detail.
-
Do Video-LLMs Actually Watch? Diagnosing Character-Tracking Failures in Long-Form Video
The paper tests whether Video-LLMs genuinely track a named person through a long video, rather than exploiting easier cues in multiple-choice questions. Across three open-source models and Gemini, changing the queried character often leaves the answer unchanged, with models responding mainly to gender differences; open-ended questioning and extra frames do not fix the problem. The results suggest that current benchmark scores substantially overstate identity-aware character tracking, and the authors release diagnostic tests for detecting this.
This is a useful, well-supported diagnosis of a non-obvious failure mode in video understanding: strong benchmark scores can come from ignoring the person named in the question and relying on coarse visual or answer-option cues.
-
RAGU: A Multi-Step GraphRAG Engine with a Compact Domain-Adapted LLM
RAGU is an open-source GraphRAG pipeline that separates graph extraction from consolidation, using typed extraction, entity deduplication, summarization, and community detection. Its accompanying 7B Meno-Lite model is trained for extraction and contextual reasoning rather than broad factual knowledge, and reportedly beats a 32B model on graph construction while matching it on GraphRAG tasks; on a medical benchmark, it achieves higher evidence recall than competing systems. The main contribution is a compact, single-GPU GraphRAG system, though the evidence is largely benchmark-specific.
The claim that a compact model specialized for language and extraction skills can outperform a much larger general model is non-obvious, but the results appear limited to GraphRAG construction and a medical benchmark, so this is more worth checking than clearly field-changing.
-
Auditing Data Leakage in Whole-Slide Image Multimodal Benchmarks
This paper audits whole-slide image VQA benchmarks and finds severe leakage at both the patient and institution levels: reported train–test case overlap reaches 92.3–100%, with near-complete overlap in staining and scanner sources. The authors show that these confounds are decodable from model features, create an accuracy gap on clean versus leaked cases, and appear to explain high results across several published WSI VLMs; they propose procedures for contamination-free evaluation.
The finding that headline WSI VQA performance may largely reflect patient- and institution-specific retrieval rather than multimodal reasoning is highly consequential and assumption-challenging, though this is an evaluation audit rather than a new modeling capability.
-
Detector Confidence Signals Presence Rather Than Occlusion in Cluttered Manipulation
The paper tests whether open-vocabulary detector confidence indicates that a particular object is visible, rather than merely that an object of that category exists somewhere in the scene. Across three detectors, simulated and real scenes, and severe occlusion, confidence remains almost unchanged as visibility drops and often fires on same-category distractors, causing confidence-based gating to activate when the target is hidden. The authors show this can make active-perception metrics dramatically underestimate the value of resolving occlusion and release a controlled benchmark for evaluating the issue.
This is a broad, assumption-challenging audit with strong cross-detector and real-video evidence, but it diagnoses a failure mode and recommends target-grounded signals rather than introducing a demonstrated new solution.
-
LPM: Industrial-Scale Generative Video Restoration
LPM is a diffusion-based video restoration system designed for the varied, long-form degradations found in user-generated video. It combines large-scale data and foundation-model training with progressive training and temporal-pyramid inference, and is reportedly deployed at Kuaishou on videos representing about 45% of viewing time. The claimed production impact is substantial: similar perceptual quality at 20% lower bitrate, with hundreds of millions in annual bandwidth savings, though the abstract gives few technical or independently verifiable details.
The industrial-scale deployment and claimed system-level bitrate and cost gains are unusually important, but the abstract does not provide enough quantitative evaluation or architectural detail to justify a strong recommendation.
-
There Is No Neutral Harness: Modern LLM Leaderboards Are Manufactured by Config-Fragile Items
The paper tests 12 open-weight LLMs on 3,679 multiple-choice items under 26 equally defensible evaluation-harness configurations, varying prompt wording, option order, and answer-scoring method. It finds that scores and rankings can change dramatically: one model ranges from 31% to 89%, four models become the winner under different configurations, and most pairwise score gaps come from items that are harness-fragile. The authors release item-level results and a reproducible analysis tool.
This is more than a benchmark variant: if the reported item-level analysis holds up, it shows that leaderboard rankings can be artifacts of scoring configuration, though the evidence is limited to 12 open models and four benchmarks and needs independent verification.
-
On-Policy Delta Distillation
The paper proposes using the token-level difference between an instruction-tuned reasoning teacher and its pre-tuning base model as the distillation signal, rather than directly matching the teacher’s output distribution. Across math, science, and code-reasoning tasks, this on-policy delta distillation reportedly transfers reasoning ability more effectively than conventional on-policy distillation with short post-training.
The teacher-minus-base “delta” is a meaningful and potentially general way to isolate capabilities introduced by reasoning tuning, but the abstract gives no quantitative gains or evidence that the method generalizes beyond the reported benchmarks.
-
Generative Compilation: On-the-Fly Compiler Feedback as AI Generates Code
The paper introduces “generative compilation,” which turns an incomplete program produced during autoregressive decoding into a compilable surrogate so a standard compiler can provide feedback before generation finishes. Its “sealor” transformation is formally shown not to reject partial programs that could still be completed, and an implementation for real Rust improves compilation and functional correctness on repository-level coding tasks across black-box and open-weight models.
This is a genuinely new way to use ordinary compiler semantics during black-box or open-weight code generation, with formal guarantees and multi-model Rust experiments, but the abstract gives no quantitative gains and the demonstrated scope is still relatively narrow.
-
ShortOPD: Recovering Pruned LLMs with Short-to-Long On-Policy Distillation
The paper studies why structured-pruned LLMs often retain useful solutions under sampling but fail under greedy free-form generation, finding that compression mainly demotes good continuations and causes repetitive suffixes. It proposes ShortOPD, which distills from the original model on the pruned model’s own rollouts while gradually extending rollout lengths based on how much useful prefix the policy can currently exploit. On math, code, and open-ended generation, it reports roughly 9x recovery over the unrecovered model and comparable quality to long-rollout OPD with 71% fewer rollout tokens and about one-quarter of the training time.
This is a plausible but meaningful training recipe for a real deployment bottleneck, with an interesting diagnosis of pruning-induced generation failure and substantial reported compute savings, though the abstract does not establish broad gains beyond the tested pruning and model settings.
-
The generator is the tracker: Multi-object tracking by painting persistent identity colours
The paper fine-tunes a large video diffusion model to convert videos into identity-colored clips, using persistent colors as the model’s internal tracking state. On DanceTrack it achieves only 40.3 HOTA—well below specialist trackers—but its association quality is unusually strong, and ablations suggest the generated colors preserve identities through occlusion and long temporal gaps better than conventional post-hoc association.
This is a genuinely unusual reframing of tracking as persistent pixel generation, supported by controlled comparisons and occlusion analyses, but the large gap to state of the art makes its immediate practical significance uncertain.
-
RecGPT-V3 Technical Report
RecGPT-V3 is a production recommender system that maintains persistent structured user memories, jointly reasons over natural-language descriptors and item Semantic IDs, and compresses explicit reasoning into latent tokens. In Taobao’s feed, it reportedly reduces user-modeling computation by 55.8% and overall serving resources by 52.4%, while improving online metrics including GMV by 3.97% and CTR by 1.00%.
The combination of stateful user memory, direct item-level semantic grounding, and latent reasoning addresses real scaling bottlenecks, and the large production A/B test plus resource reductions make it worth a look, though the underlying components are extensions of established ideas and the abstract gives limited experimental detail.
-
Evidence-Backed Video Question Answering
The paper proposes video QA that must provide both an answer and precise spatio-temporal visual evidence, including tracked segmentation masks rather than only text rationales or sparse boxes. A human-verified benchmark and a 160k-example synthetic training set reveal that QA accuracy can be largely decoupled from genuine visual grounding, while fine-tuning on the new data substantially improves grounding on 7B models.
The combination of dense tracked evidence with video-language QA and the reported accuracy–perception decoupling is genuinely interesting, but the contribution is still primarily a benchmark/data-and-fine-tuning package whose broad robustness is not established in the abstract.
-
Partition, Prompt, Aggregate: Statistical Self-Consistency in Language Models
The paper tests whether language models’ answers obey the law of total probability: estimates for subpopulations should combine into the corresponding estimate for the whole population. Across frontier models and tasks, they find frequent inconsistencies, and a surprising “macro fallacy”: aggregating answers about finer-grained personas often matches human data better than asking directly about the overall population. The authors propose statistical self-consistency as a reference-free way to evaluate whether models use their latent subpopulation knowledge coherently.
The partition-and-aggregation diagnostic and the finding that indirect, fine-grained estimates can outperform direct population estimates are genuinely non-obvious, though the abstract gives limited quantitative detail about effect sizes and breadth.
-
SUFLECA: Scaling Up Feature Learning for CAD-to-image Alignment
SUFLECA learns geometry-aware visual features from Normalized Object Coordinates across 674K real and synthetic images, then uses geometrically consistent one-to-one matching for zero-shot 9D CAD-to-image alignment. It reports 33.4% category and 42.3% instance accuracy on ScanNet25k—10.3 and 12.2 points above the strongest zero-shot baseline—while running in under a second and reportedly exceeding fully supervised methods on this benchmark.
The combination of large-scale geometry-grounded feature learning and non-iterative matching produces a substantial, computationally efficient gain and an unexpected result against supervised methods, though the evidence is centered on one benchmark and the method extends familiar correspondence-learning ideas.
-
Demystifying On-Policy Distillation: Roles, Pathologies, and Regulations
This paper studies why on-policy distillation (OPD) helps language-model post-training, arguing that it mainly improves exploration by steering students toward useful reasoning paths rather than raising their ultimate capability. It identifies teacher–student distribution mismatch and length-based gaming as failure modes, then shows that advantage clipping and log-scale signal compression reduce these problems across seven benchmarks and outperform several OPD and RLVR baselines.
The explicit diagnosis of OPD as an exploration mechanism, together with concrete length-exploitation and distribution-mismatch pathologies and lightweight fixes, is more informative than a routine distillation variant, though the abstract provides no quantitative gains or theoretical validation.
-
Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes
The paper proposes a masked-diffusion sampler that jointly updates image and text while allowing each modality’s current confidence to influence the other, including remasking decisions that can undo inconsistent commitments. Its training-free CO₂Jump sampler is evaluated on image editing, multimodal understanding, and visual reasoning using three newly introduced corpora, where the authors report the best joint performance and improvements with more denoising steps.
The within-step cross-modal coupling and explicit self-correction are a meaningful sampler-level idea, but the abstract gives no quantitative comparisons or evidence beyond broad best-performance claims on largely new benchmarks.
-
Can a Language Model Learn Facts Continually in Its Weights?
The paper studies whether language models can accumulate invented facts through sequential weight updates, rather than relying on in-context prompts. It finds that broad, varied training examples make facts more usable and durable than bare statements, but later writes still make earlier facts unreachable even when their weight changes remain detectable; context can recover some of this knowledge, while composition and long-term accessibility remain unreliable.
This offers a useful empirical framing of continual learning as a problem of question-keyed accessibility rather than simple storage, with substantial degradation after 20–100 writes, but the evidence is limited to invented facts and Qwen3 experiments.
-
Fundamental Limitations of Fixed-Budget Best-Arm Identification
This paper proves that no adaptive algorithm for fixed-budget best-arm identification can achieve the static oracle’s optimal error-exponent uniformly over all problem instances when there are at least three arms. For any algorithm and any one-parameter natural exponential family, there is an instance where its error exponent is bounded by a constant factor below the oracle’s, resolving an open question about whether oracle-level performance is attainable.
It establishes a broad impossibility result with an explicit nonvanishing gap from the oracle, overturning the plausible assumption that adaptive sampling can uniformly match oracle error rates, though the impact is mainly within sequential decision theory.
-
Robust bipedal locomotion on flowable slopes via foot-driven terrain manipulation
The paper studies how a biped’s foot design can control deformation of granular terrain rather than merely compensating for it through body motion. Experiments show that intermediate cleat spacing avoids both excessive yielding and excessive resistance, enabling a small biped to walk on granular slopes up to 30 degrees; an actively adjustable foot also transfers to rigid terrain and a larger autonomous biped.
The limb-centric idea of deliberately regulating terrain deformation is a genuinely interesting alternative to conventional body-centric locomotion control, supported by systematic robophysical experiments and a larger-robot demonstration, though the reported evidence is not yet enough to make it a must-read.
-
Requential Coding: Pushing the Limits of Model Compression with Self-Generated Training Data
The paper proposes requential coding, a model-compression scheme in which a teacher generates samples from the student’s distribution and the code records only where the student and teacher disagree. Unlike parameter quantization or standard prequential coding, its code length is intended to avoid scaling with parameter count or raw data entropy. Experiments reportedly show much shorter codes for large models and ensembles, tighter PAC-Bayes bounds for billion-parameter LLMs, increasing relative compressibility with scale, and a separation between learnable structure and unpredictable data content.
This is a genuinely nonstandard approach to measuring and exploiting model compressibility, with potentially important implications for scaling laws and generalization, but the abstract gives few concrete quantitative details to justify a strong verdict.
-
Thompson Sampling Is 2-Competitive for Mistakes
The paper proves that Thompson sampling incurs at most twice the expected number of suboptimal-arm selections of any competing policy in a broad class of Bayesian bandits. This applies to independent latent arm processes that evolve only when played, covers common horizon and discounting schemes, and settles a 2014 conjecture for stochastic bandits; the factor of two is tight.
It resolves a longstanding conjecture and gives a tight, broadly applicable optimality guarantee for Thompson sampling, but it is a theoretical characterization rather than a major new capability or practical algorithmic advance.
-
MESH: Scaling Up Retrieval with Heterogeneous Content Unification
MESH is a unified retrieval architecture intended to replace separate models for different content tiers, especially popular versus fresh or long-tail items. It separates feature domains and uses gated bias correction to reduce interference from high-frequency engagement signals; Pinterest reports a 14× improvement in the scaling exponent for fresh items, a 5.5% increase in fresh-item repins, higher funnel efficiency, retention gains, and 2.87× serving throughput.
The combination of a concrete explanation for heterogeneous-retrieval scaling failure with substantial billion-scale online results is worth examining, although the abstract does not establish how much of the gain comes from the architectural idea versus system or baseline changes.
-
Who Grades the Grader? Co-Evolving Evaluation Metrics and Skills for Self-Improving LLM Agents
The paper proposes evolving the evaluator alongside an LLM agent’s skills, rather than assuming a fixed rubric or judge. It searches interpretable compositions of defect detectors, anchors selection to a small labeled reference set, and uses unlabeled-output consensus; on code generation, the learned metric improves agreement with hidden ground truth by 0.21 and the full loop retains 88–110% of the improvement obtained using ground-truth or hand-written evaluation. The authors also show that removing the anchors produces a degenerate always-pass evaluator, demonstrating that task performance alone cannot validate a self-evolved metric.
Co-evolving the evaluator with the agent, together with evidence of evaluator collapse without anchored validation, is a meaningful and somewhat assumption-challenging direction, but the evidence is still limited to a few tasks and a small ten-item anchor set.
-
The Seriality Gap in Video Diffusion Models
The paper studies video diffusion on controlled multi-ball collision dynamics, where later events depend on the consequences of earlier ones. It finds that prediction quality worsens as the causal chain grows, even with more denoising steps, while autoregressive or deeper models improve more reliably; a theoretical analysis argues that standard denoising iterations do not provide additional serial computation beyond the backbone for deterministic prediction.
The controlled causal-chain experiments and computational explanation suggest a potentially important limitation of video diffusion, but the abstract does not establish how broadly the result transfers beyond synthetic dynamics or whether the theoretical claim applies to practical stochastic generation.
-
Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution
The paper studies whether coding agents can estimate how much work a task actually needs instead of immediately inspecting the whole available context. Its E3 policy first estimates scope, follows a minimum viable execution path, and expands only after failed verification; on a 121-edit simulator benchmark it reportedly preserves 100% success while reducing cost, tokens, and files inspected by 85–92%, with smaller but similar effects in a live GPT-4o code-editing harness.
The execution-scope framing and large efficiency gains are genuinely interesting, but the strongest evidence comes from a controlled simulator and a small real-model validation, so it is not yet strong enough to establish broad agentic impact.
-
Accuracy Without Grounding: Diagnosing Visual Dependency Dissociation in Video LLM Benchmarks
This paper tests whether video-LLM benchmark accuracy actually depends on seeing and understanding the video. Across 20 open-weight models and four API models, black-screen controls and frame-order ablations suggest that much of the measured gain comes from frame diversity, while temporal order contributes little; it proposes the Visual Dependency Gap as a diagnostic for this problem. It also finds that aggregate accuracy can hide substantial question-level answer changes under video perturbations.
The broad, controlled audit and finding that video-model accuracy may rely on unordered visual snapshots rather than temporal understanding are genuinely interesting, but the evidence is centered on MVBench and may reflect benchmark limitations rather than a general property of video LLMs.
-
Learned Pairwise Deep Dual-Optimal Inequalities for Stabilizing Column Generation
The paper learns pairwise ordering constraints on dual variables and adds them to column-generation formulations to reduce dual instability, avoiding problem-specific hand-designed exchange arguments. On capacitated vehicle routing and VRP with time windows, it reports 89.7–93.9% reductions in root column-generation time with small bound losses; a recovery procedure restores the original bound while retaining 54.8–83.1% of the speedup.
The combination of learned dual restrictions, graph-based conflict filtering, and certified relaxation produces unusually large optimization speedups, but the evidence is limited to two related routing problem families and may not yet establish broad generality.
-
GFlowRL: Scaling Distribution-Matching RL to Large Language Models
GFlowRL adapts GFlowNet-style reinforcement learning to large language-model post-training without a separately learned, prompt-conditioned partition function. It estimates the normalization from rollout groups already used for training, then adds importance-sampling correction and asymmetric clipping to handle rollout drift and outliers. The method reportedly trains stably across dense and MoE models up to 235B parameters, outperforming prior GFlowNet-style methods on math, code, and adversarial-generation tasks, including a 14B model reaching a 2048 Codeforces rating.
Removing the partition network while retaining distribution matching, with reported stability at scales where the prior method diverges, is a substantive and somewhat surprising algorithmic advance, but the abstract provides insufficient comparative detail to merit a strong verdict.
-
Non-Expansive Two-Time-Scale Stochastic Approximation: A Fixed-Schedule One-Quarter Barrier and Bias-Corrected Acceleration
This paper studies stochastic two-time-scale fixed-point methods when the slow map is non-expansive rather than contractive. It proves a worst-case one-quarter-rate barrier for standard Krasnosel’skii–Mann updates, then uses a residual-based bias correction to improve nested methods to roughly a $T^{-1/3}$ rate and a smooth derivative-oracle variant to roughly $T^{-1/2}$ with constant samples per iteration.
The combination of a schedule-independent lower bound, bias cancellation, and a single-loop square-root-rate method is a substantial theoretical advance, but its relevance depends on fairly specialized oracle and smoothness assumptions.
-
AI advice suppresses people's willingness to say "I don't know", even when the advice is wrong and accuracy is incentivized
Across five experiments with 3,132 participants, the authors found that merely having access to AI advice sharply reduced people’s willingness to answer “I don’t know,” even when the advice was deliberately wrong. Participants answered more questions but became much less accurate and substantially more confident; accuracy incentives improved behavior but did not restore the baseline willingness to defer judgment. The paper’s main contribution is linking AI assistance to a shift in people’s metacognitive threshold, not just to errors caused by bad advice.
The preregistered experiments and replication support a notably non-obvious human-AI effect—AI availability may suppress uncertainty acknowledgment even when it reduces accuracy—but the artificial question/incorrect-advice setup limits how broadly the result can yet be generalized.
-
Branching Policy Optimization: Sandbox-Native Language Agent Reinforcement Learning
The paper proposes Branching Policy Optimization, which exploits deterministic, resumable agent sandboxes by branching multiple rollouts from shared intermediate states instead of starting independent trajectories. It uses sibling outcomes to estimate per-step advantages, and reports lower gradient variance, 3.6–6.1 point success gains, and 38% fewer policy updates than GRPO/RLOO across WebShop, ALFWorld, and SWE-bench Verified.
The sandbox-native rollout topology and claimed variance reduction are a genuinely non-obvious alternative to standard LLM-agent RL, with results across several tasks, but the abstract alone does not establish whether the gains survive careful compute, branching-cost, and estimator-validity scrutiny.
-
Posterior Variance Is a Constraint Map, Not an Error Map: Closed-Form Uncertainty for Radiative Gaussian Splatting in Sparse-View CT
The paper derives an exact, clamp-aware posterior-variance computation for radiative Gaussian splatting in sparse-view CT, avoiding the sampling estimator and reducing uncertainty-map cost by about 8x. More importantly, it shows that posterior variance mostly identifies where the reconstruction is constrained by observations, not where the reconstruction is wrong: apparent full-volume correlations are largely caused by object/background contrast, while in-object error ranking fails across 15 scenes and several posterior constructions. It also proposes a posterior reparameterization that makes uncertainty scale more transferable across scenes and photon counts, along with stricter evaluation protocols.
The central empirical finding—that uncertainty maps can look well calibrated only because of object/background structure while missing systematic in-object bias—is a useful, non-obvious warning supported by multiple posterior and estimator variants, though its impact is mainly methodological and domain-specific rather than a major new capability.
-
Ask Twice, Look Twice: Prompt Echoing Resolves the Question-First Paradox in Vision-Language Models
The paper shows that putting the question before an image helps a vision-language model focus on relevant visual content, but causes the answer token to lose access to the question after processing many image tokens. It proposes repeating the question before and after the image, so the first copy guides perception and the second remains available for answering; across five open VLMs, this reportedly fixes the ordering problem and improves several VQA benchmarks, with gains as large as 19 Winoground group-accuracy points.
This combines a non-obvious mechanistic explanation with a simple, training-free prompting method and unusually large reported gains, though the abstract does not establish how robust the results are across models, prompt formats, or evaluation leakage concerns.
-
Think at 5 Hz, Act at 20 Hz: Asynchronous Fast-Slow Vision-Language-Action Inference for Closed-Loop Driving
The paper separates slow scene and instruction reasoning from fast control: a frozen 7B vision-language model maintains a cached scene representation, while a small action module uses that cache plus the latest camera frame to predict waypoints every 50 ms. In CARLA, training the action module to tolerate stale caches and producing fresh per-tick actions raises route completion from 37.0 to 94.0, reduces red-light violations, and transfers reasonably to unseen towns; the system runs in 32 ms per tick on one consumer GPU.
The asynchronous fast-slow design and explicit training for representation staleness are genuinely useful ideas, with unusually large simulated-driving gains and transfer evidence, but the results remain limited to CARLA and a narrow benchmark, so they do not yet justify a strong verdict.
-
A Geometry-Limited Identification Floor and Its Consequences for Voice-Clone Attribution in Professional Voice Actors
The paper studies speaker identification and voice-clone attribution using 1,168 Japanese professional voice actors and finds a persistent error floor that calibration, normalization, and reranking cannot eliminate. Embedding-space crowding and style variation cause both misidentification of real actors and false accusations from clones; a voice-actor-specific encoder reduces errors but does not remove them, motivating spoof-aware open-set systems with abstention rather than automatic attribution.
The combination of a large, difficult professional-voice cohort and evidence for a geometry-limited attribution floor is a non-obvious result with practical consequences, though the claims remain based on particular encoders, languages, and cloning systems rather than a broadly established law.
-
Data and Learning Where it Matters for Contact-Rich Manipulation
This paper separates contact-rich manipulation from easy free-space motion: it uses conventional planning for the latter and collects dense autonomous data only around the critical contact segment, then trains an offline RL policy. On four real-world tasks, 2–2.5 hours of data collection reportedly yields 96% average success, versus 55% for the strongest baseline, while retaining performance in out-of-distribution settings.
The targeted-data principle, autonomous collection, and large real-world success gap are genuinely interesting, but the abstract does not establish how broad or difficult the tasks are enough to warrant a strong verdict.
-
Learning Standard Model structure from LHC data with Riemannian flow matching
The paper proposes ShellFlow, a transformer-based conditional flow model that generates collider particles directly on their on-shell manifolds, using only kinematic constraints and recorded ATLAS events. Trained on about 10^9 events, it reportedly recovers known resonances, particle masses, the Weinberg angle, and correlations that were not explicit training targets, suggesting that substantial Standard Model structure can emerge from data alone.
The claimed unsupervised recovery of broad particle-physics structure from one generative model is highly non-obvious and potentially important, but the abstract gives no quantitative errors, baselines, ablations, or evidence that the recovered quantities are not inherited from dataset and preprocessing artifacts.
-
Poor Man's Agentic Modeling: Simulating Large LLM-Agent Societies on a Laptop
The paper proposes replacing expensive LLM agents in large-population simulations with simple low-parameter behavioral surrogates fitted from a few hundred to few thousand LLM queries. It introduces an interaction-order-by-memory taxonomy that predicts how surrogate error changes with population size, and reports validation across EconAgent and seven other LLM-based simulations, including quantitative explanations of two apparent prediction failures without fitted parameters.
This is a potentially useful new way to study emergent behavior at large agent counts, with unusually strong claims of theory-backed, cross-simulation validation, but the abstract lacks enough quantitative detail to establish that the surrogates preserve important macro-level phenomena.
-
Kernelized Linear Attention: Breaking the Capacity Wall with Symmetric Cones
The paper reframes associative recall in linear attention as a spherical-packing problem and proposes KATA, using rank-one PSD feature maps derived from cone geometry to increase memory capacity without adding parameters. It reports efficient implementations—up to roughly 11× FlashAttention-2 forward throughput in a long-context chunked setting—and strong recall on MQAR, including 0.985 accuracy at 16× the training length, while using about one quarter of the KV-cache entries. Experiments also examine how the feature map interacts with positional embeddings, delta rules, and decay gates in 340M-parameter language models.
The combination of a principled geometric capacity analysis, parameter-free state expansion, and unusually strong long-context efficiency/recall results is genuinely interesting, but the abstract does not establish broad language-model gains or independent validation beyond specialized associative-recall tests.
-
Myosin 5A mediates membrane-associated periodic skeleton reassembly during axon regeneration in response to ROCK-2 inhibition
Using STED microscopy and perturbation experiments, the authors show that axons rebuild their 190-nm membrane-associated periodic skeleton only gradually after injury, and that inhibiting ROCK-2 accelerates this recovery about fivefold. They identify myosin 5A as a downstream effector: it binds more strongly to spectrin after injury, is required for ROCK-2-inhibition-enhanced regrowth, and helps maintain the periodic skeleton even in uninjured axons.
This is a non-obvious mechanistic link between ROCK-2, myosin 5A, nanoscale axonal lattice repair, and regeneration, supported across mouse and human neuron cultures, but its importance remains uncertain without in vivo regeneration or functional recovery evidence.
-
Convergent Inhibitory Cortical Circuit Disruption Drives Genetically Distinct Absence Seizures
This study compares two genetically different mouse models of absence epilepsy and finds the same early disruption in a layer-6 corticothalamic to Tac1+ Pvalb interneuron circuit. Disrupting this circuit in otherwise normal mice was sufficient to produce later Sst interneuron dysfunction and spike-wave seizures, suggesting a causal cascade rather than unrelated downstream abnormalities.
The convergent, experimentally induced circuit cascade across distinct epilepsy genes is a substantial mechanistic result, but the abstract provides limited detail on effect sizes, controls, and generality beyond two mouse models.
-
Distributed resonant coupling enables high efficiency power transfer to mm-scale bioelectronics
The paper introduces distributed resonant coupling, a three-coil wireless-power architecture in which a millimeter-scale implant receiver participates in the resonant circuit rather than acting as a passive harvester. It reports 56% measured transfer efficiency, 42% end-to-end DC efficiency, and over 420 mW delivered from 1 W transmitted power, with tolerance to orientation and tissue; demonstrations include implant localization, battery charging, and drug delivery.
The reported power and efficiency for mm-scale, orientation-tolerant implant receivers would substantially expand wireless bioelectronics if validated across realistic implant sizes and depths, but the abstract does not provide enough experimental detail to make this a strong recommendation.
-
Early Adversity Selectively Reshapes the Somato-Cognitive Action Network in the Developing Brain
Using longitudinal imaging data from 4,525 youths, the study examined how different forms of early adversity relate to individualized cortical network organization. Threat exposure, unlike deprivation or unpredictability, was associated with a dose-dependent expansion of the somato-cognitive action network toward sensorimotor cortex, altered connectivity, and poorer cognition; the effect replicated across waves, within participants, and out of sample.
The selective, large-scale remodeling of a single brain–body integrative network is a non-obvious developmental finding supported by a large longitudinal sample and multiple replication analyses, though the observational design limits causal interpretation.
-
UniVR: Thinking in Visual Space for Unified Visual Reasoning
UniVR trains a vision-based model to perform reasoning, physical-dynamics prediction, and long-horizon planning from visual demonstrations alone. It introduces a reinforcement-learning method with both whole-task and step-level rewards, plus the VR-X evaluation suite, and reports improvements of up to 25%, though the abstract does not specify baselines, tasks, or absolute results.
The unified pure-visual training setup spanning reasoning, dynamics, and planning is potentially a meaningful direction, but the evidence is too underspecified and the contribution is partly a new benchmark, so it does not yet warrant strong attention.
-
Generalizable VLA Finetuning via Representation Anchoring and Language-Action Alignment
Anchor-Align fine-tunes vision-language-action policies while trying to preserve the pretrained VLM’s visual-semantic representations and explicitly aligning language with discretized motion directions. The method improves physical xArm7 success from 28% to 54% and 37% to 60% across two VLA architectures, with additional gains under simulated distribution shifts, perceptual perturbations, and long-horizon tasks.
The combination of representation distillation and same-observation language-action supervision is fairly incremental, but the large gains on a real robot across two architectures and broad robustness evaluations make it worth a closer look.
-
Latent-Identity Tuning in Text-to-Image Personalization Models
The paper edits a person’s identity directly in the latent tokens of a frozen text-to-image personalization encoder, rather than editing each generated image or retraining the model. By identifying token- and subspace-level semantic directions, it enables localized facial changes that remain consistent across newly generated images, with experiments supporting identity preservation and edit coherence.
The training-free discovery and manipulation of localized identity directions in an existing personalization latent space is a meaningful idea, but the abstract gives no quantitative results or comparisons showing a major capability jump.
-
VideoRAE: Taming Video Foundation Models for Generative Modeling via Representation Autoencoders
VideoRAE turns features from a frozen video foundation model into compact video latents that can be used by either diffusion-transformer or autoregressive generators. It uses a lightweight projector, multi-scale features, and representation-alignment decoding; on UCF-101 it reports better gFVDs and roughly 5× faster convergence than prior autoencoder baselines, with faster convergence also observed in a controlled 2B-parameter text-to-video setup.
Using frozen video-understanding representations as generation-ready latents is a meaningful architectural direction, but the evidence is mainly UCF-101 plus a convergence comparison, so it does not yet establish a major general video-generation advance.
-
DiffGI: Differentiable Geometry Images for High-Fidelity Thin-Shell 3D Generation
The paper represents 3D surfaces as continuous 2D truncated signed-distance fields rather than binary geometry images, then makes contour extraction differentiable so surface losses can train the 2D representation directly. A VAE and flow-matching diffusion model use this compact representation to generate thin-shell shapes such as garments, with reported improvements in reconstruction and boundary precision at lower computational cost.
The combination of subpixel TSDF geometry images and differentiable Marching Squares is a meaningful representation and optimization idea for thin-shell 3D generation, but the abstract gives no quantitative results or evidence that the gains generalize beyond the reported datasets.
-
DSWorld: A Data Science World Model for Efficient Autonomous Agents
DSWorld treats a data-science workflow as an environment whose state changes can be predicted before running expensive operations. It combines structured state representations, cost-aware decisions, lightweight execution, and an LLM simulator, reporting 14× faster RL training and 3–6× faster search with similar performance; its transition-prediction model beats an LLM baseline by 35.6%.
The world-model framing for data-science agents and explicit cost-aware simulation are meaningful, potentially reusable ideas, but the evidence is limited to abstract-level claims and a specialized 8K trajectory dataset rather than a demonstrated broad capability jump.
-
Learning to Navigate Efficiently with Only 0.58M Trainable Parameters
The paper decomposes visual navigation into analytic geometric operations plus three small learned modules, training only 0.58M parameters rather than a large end-to-end policy. On 6,060 point-goal episodes across 60 environments, it reportedly approaches state-of-the-art performance, achieves the lowest collision rate, runs above 10 Hz on a Jetson Xavier NX, and adapts to exploration by retraining only a 123k-parameter head.
The combination of analytic structure and extremely low trainable parameter count, with apparently competitive navigation and edge-device performance, is genuinely notable, but the abstract provides too few comparative numbers and robustness details to justify a strong verdict.
-
Omni-Decision: A Progressive Evidence-State Agent System for Omni-Modal QA
Omni-Decision is a training-free agent framework that represents multimodal QA as maintaining and closing a structured evidence state, rather than relying on scratchpads or tool histories. It tracks confirmed facts, conflicts, dependencies, and missing evidence while coordinating media, web, computation, and verification tools. The system reports large gains on OmniGAIA and WorldSense, although the abstract gives limited detail about baseline strength and evaluation breadth.
The explicit, query-scoped evidence state and deterministic evidence updates are a substantive alternative to trajectory-based multimodal agents, and the reported 27–30 point gains are notable, but the claims need closer inspection of benchmarks, baselines, and ablations.
-
HyperGS: Fast and Generalizable Gaussian Video Representation
HyperGS replaces per-video Gaussian Splatting optimization with a feedforward Transformer that predicts frame-level 2D Gaussian representations in one pass. A dynamically weighted rank-based geometric regularizer prevents degenerate Gaussian shapes, and the method reportedly matches reconstruction quality while encoding 10^4–10^5 times faster, improves PSNR by about 3 dB, and transfers zero-shot to 720p videos.
The combination of amortized Gaussian video encoding, extreme claimed speedups, and zero-shot resolution generalization is a genuinely interesting direction, but the abstract does not establish enough detail about compute, baselines, or the breadth of the evidence for a strong verdict.
-
Tropical Circuits with Scalar Multiplication Gates
The paper proves exponential lower bounds for tropical circuits that use max, addition, and positive scalar multiplication when computing maximum-weight directed spanning trees and bipartite perfect matchings. It also translates these results into an exponential size gap between monotone and unrestricted maxout networks, showing that convexity-constrained models such as ICNNs can require exponentially more units to represent some functions.
The concrete exponential lower bounds and their connection to the expressive cost of convexity constraints are genuinely interesting, but the abstract provides no experimental or broader theoretical context beyond the stated theorems, so this is worth a look rather than a must-read.
-
Markov Chain Monte Carlo with Diffusion Paths
The paper replaces conventional tempering with the marginals of a diffusion noising process, which preserves relative weights between modes while moving the target toward a Gaussian. It introduces MAD-Path, a Metropolis-adjusted sampler that uses learned intermediate scores but remains exactly unbiased despite score and discretization errors, and provides spectral-gap and acceptance-probability analyses plus Bayesian-posterior experiments showing better mode exploration than tempering.
The diffusion-path formulation with an exact augmented-space Metropolis correction is a substantive new direction for multimodal MCMC, but the abstract gives no quantitative scale or especially decisive empirical comparison warranting a strong verdict.
-
Think Through a Bottleneck: Hourglass Reasoning for Rigorous Induction
The paper proposes “Hourglass reasoning,” a prompting procedure that strictly separates induction, deduction, implementation, and refinement, allowing only a compressed symbolic schema and rule to pass between stages. Across visual reasoning, Verilog synthesis, and linguistic-rule tasks, it reports sizable gains over iterative refinement, including 14 points on ARC-AGI-2 and 31% to 58% on ChipBench, while reducing the harm from explicit verbalization on linguistics puzzles.
The enforced information bottleneck and cross-domain gains are a genuinely interesting alternative to unrestricted chain-of-thought refinement, but the abstract provides limited detail about scale, variance, and baseline comparisons, so the strong claims need verification.
-
Paradoxes of Game Theoretic Equilibria and Price of Anarchy
This theoretical paper argues that standard equilibrium and Price-of-Anarchy analyses can miss important instability in multi-agent learning dynamics. It presents results claiming that Nash and correlated-equilibrium formulations may permit dominated or non-rationalizable behavior, that low regret can coexist with chaotic trajectories, and that learning dynamics can make apparently stable congestion-game equilibria highly inefficient over time.
The paper offers a potentially important dynamical critique of equilibrium-based multi-agent analysis, including claims of chaos and exponentially worsening inefficiency, but the abstract is theorem-heavy and provides too little detail to judge the scope and robustness of these striking results.
-
Encoder-Side Neuron Identification and Amplification for Acoustic Perception in Large Audio-Language Models
The paper introduces IAAN, a training-free inference-time method that identifies individual neurons in an audio encoder as especially responsive to real speech rather than noise, then amplifies those neurons. On ten non-semantic speech attributes, it reports large accuracy gains across three large audio-language models, including improvements even for a model already fine-tuned to use acoustic information; ablations suggest that intervening inside the encoder and selecting specific neurons are both important.
The encoder-level, neuron-selective intervention and the reported 9.7–25.7 point gains are genuinely notable, but the abstract lacks enough detail about baselines, datasets, and robustness to justify a strong recommendation.
-
From Geometric Recovery to Causal Validation: A Reproducible Audit of Sparse Autoencoder Features, from Superposition Geometry to Causal Inertness
This paper audits sparse autoencoder features using causal tests rather than relying only on decoder–ground-truth cosine similarity. It reports that many apparently well-recovered features are causally inert: up to 77% in a degraded SAE and 9% in a well-trained one, with distinct structural and TopK-related causes; the authors release a deterministic auditing pipeline and observe similar, smaller effects in a production SAE.
The separation of geometric recovery from causal usefulness, especially the finding that near-perfect cosine matches can be inert, is a potentially important challenge to standard SAE evaluation, but the evidence appears mostly limited to controlled reproductions and a small production-scale audit rather than broad validation.
-
Flatness-Preserving Residual Learning for Real-Time Tight Quadrotor Formation Flight
The paper learns aerodynamic interaction effects in tightly flying quadrotors as a residual model while preserving the overall system’s differential flatness. This enables a lightweight feedback-linearization controller with feedforward disturbance compensation; hardware experiments report 31% lower tracking error than nominal control and performance comparable to NMPC at roughly one-tenth the computation, using less than 30 seconds of training data and a 5 ms control loop.
The combination of learned downwash compensation, preserved flatness, very limited training data, and real-time hardware performance is a genuinely useful direction, but the abstract does not establish the formation size, flight conditions, or breadth of comparisons well enough for a strong verdict.
-
A Bearing-Strength Method for Motion Estimation of Unknown Energy Emitters
The paper combines bearing measurements with received signal strength to estimate the motion of unknown moving light, acoustic, or radio emitters using passive sensors. Its observability analysis claims that, unlike bearing-only estimation, the sensor no longer needs lateral motion, and real-world experiments support this result without requiring additional hardware.
Removing the standard lateral-motion observability constraint using an already available signal measurement is a non-obvious and potentially useful advance, but the abstract gives limited detail about the experimental breadth and quantitative gains.
-
Less Experts, Faster Decoding: Cost-Aware Speculative Decoding for Mixture-of-Experts
The paper argues that speculative decoding for sparse MoE models can become slower when likely draft tokens route to many different experts, increasing memory traffic during verification. EcoSpec selects draft paths using both token acceptance probability and predicted marginal expert cost, with an expert predictor and buffer; across three very large MoE models, it reports up to 1.62× end-to-end decoding speedup while reducing active expert footprints.
The expert-scattering failure mode and cost-aware draft-tree selection are a meaningful, practically relevant extension of speculative decoding, with unusually large-model experiments and a substantial reported speedup, but the abstract lacks enough baseline and systems-detail evidence for a stronger verdict.
-
MixCIT: A Kernel Based Local-Polynomial Debiased Test for Conditional Independence on Mixed-Type Data
MixCIT proposes a conditional-independence test for mixtures of discrete and continuous variables. It combines exact matching for discrete conditioning variables with nearest-neighbor neighborhoods for continuous ones, and uses local-polynomial debiasing to remove smoothing bias; the authors claim valid asymptotic calibration, a dimension-free n^-1/4 detection threshold, and near-quadratic computation.
The unified mixed-type treatment and claimed dimension-free detection rate would be meaningful advances for causal-discovery tooling, but the abstract offers theoretical claims without empirical comparisons or evidence that the method works broadly in practice.
-
TerraZero: Procedural Driving Simulation for Zero-Demonstration Self-Play at Scale
TerraZero is a high-throughput procedural driving simulator and self-play training system that uses real maps but generates traffic, dynamics, and scenarios procedurally. It trains driving policies entirely with reinforcement learning—without demonstrations, logged trajectories, or fallback planners—and reports 1.3 million agent-steps per second, zero-shot transfer across cities and datasets, emergent left-hand driving, and strong results on several simulation benchmarks.
The combination of million-step simulation, procedural long-tail generation, and fully demonstration-free self-play with cross-city transfer is unusually ambitious, but the abstract gives too few quantitative comparisons and relies heavily on broad state-of-the-art claims to justify a strong verdict.
-
A vision foundation model for single-cell biology via spatial gene cartography
scVision represents each cell’s gene-expression profile as an image: optimal transport places genes on a shared layout so co-expressed genes become nearby, and a vision transformer is pretrained with masked-image modeling on 72 million cells. Without fine-tuning, it reportedly improves zero-shot cell-type annotation and gene-program recovery across six held-out studies, while matching the best token-based model for integration. A notable ablation suggests the biologically informed gene layout contributes more than the particular vision architecture.
The image-based representation and evidence that gene placement matters more than the network are genuinely non-obvious, but the abstract provides no quantitative margins and the broad superiority claims need careful validation.
-
Ego-Dynamics-Augmented World Model for Autonomous Driving with Zero-Shot Cross-Chassis Adaptation
The paper adds an explicit, physics-informed model of the ego vehicle’s motion to a Dreamer-style latent world model, so the learned model can focus on changes in the surrounding scene rather than reconstructing the vehicle’s own motion. It also uses chassis parameters to adapt the dynamics context to previously unseen vehicles without retraining. Reported task-success gains are 28% in urban driving, 61% on highways, and up to 73% for unseen chassis, though the abstract does not specify absolute performance or experimental scale.
Explicitly separating ego dynamics from scene dynamics and demonstrating claimed zero-shot cross-chassis transfer is a meaningful, potentially broadly useful direction, but the unusually large gains are difficult to assess without benchmark details, baselines, and absolute results.
-
Consensus as Privileged Context for Label-Free Self-Distillation
CANON uses majority agreement among multiple model-generated solutions to create a teacher trace, then provides token-level self-distillation targets for new rollouts without labeled answers. On math and science reasoning tasks, it reportedly improves pass@1 by up to 12 points, beats label-free reinforcement learning with far less compute, transfers across benchmarks, and helps the model solve problems outside its previous 32-sample capabilities.
The consensus-anchored, dense token-level supervision is a meaningful extension beyond filtering, preference learning, or scalar rewards, and the claimed gains and transfer are notable, but the abstract lacks enough experimental detail to justify a strong verdict.
-
ECG-LLM: Foundation Model for ECG-Based Cardiac Reasoning
The paper trains an ECG-conditioned language model on 679,112 studies from 186,409 patients, using question-answer supervision grounded in ECGs, clinical context, CMR, and echocardiography. It answers free-form cardiovascular questions from a 12-lead ECG and predicts both conventional measurements and less directly observable phenotypes such as chamber volumes, ventricular function, wall thickening, aortic stenosis, and right-ventricular dysfunction.
The notable contribution is moving ECG models from fixed labels and report generation toward question-driven multimodal clinical reasoning, with unusually large-scale supervision and predictions of imaging-derived phenotypes; however, the abstract gives no quantitative results, external validation, or evidence that the apparent reasoning is reliable rather than learned phenotype correlations.
-
Cluster with Auctions for Vector Search
CwA jointly learns how database vectors are partitioned and how queries probe those partitions, instead of forcing both to use the same assignment rule. It alternates training a probing network with a balanced auction-based reassignment of database vectors, and reports up to 4.7× higher throughput at equal recall when query and database distributions differ; even a linear probe performs well in the in-distribution setting.
The separation and joint optimization of database partitioning from query probing is a meaningful departure from standard vector-index design, with a potentially large throughput gain, but the abstract gives insufficient detail to establish how broad or reproducible the 4.7× result is.
-
PiVoT: A Variational Solution for Real-time Large-scale Multi-object Detection and Tracking under Heavy Clutter
PiVoT is a training-free variational Bayesian method that jointly detects and tracks many radar objects directly from positional and Doppler point clouds, without a separate detector or clustering stage. The authors report real-time operation on full-resolution automotive radar data, robustness to severe clutter, and scaling to roughly 1,000 objects while approaching a deep-learning detection benchmark.
The combination of joint detection/tracking, Doppler modeling, clutter resilience, and claimed thousand-object real-time scaling is unusually ambitious, but the abstract gives no quantitative comparisons or experimental detail sufficient for a stronger recommendation.
-
TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents
TRACE assigns dense rewards at tool-call boundaries by measuring how each action changes a frozen reference model’s log-probability of the gold answer, using temporal-difference differences instead of a separately trained critic. On BrowseComp-Plus, pure RL reportedly improves Qwen3-4B from 7.2 to 35.6 and Qwen3-30B-A3B from 8.4 to 42.6, with faster learning and some transfer to open-web tasks.
The unusually large gains from a simple, critic-free turn-level credit signal make this worth checking, but the abstract provides limited evidence beyond one main benchmark and does not establish how much comes from the particular reward construction versus other training details.
-
Seeing the End at Step Zero: Accelerating Diffusion MLLMs via MLP Sparsity-Aware Truncation
The paper finds that diffusion multimodal language models reveal the useful end of their output sequence during the first denoising step, via a change in MLP activation sparsity. Seer uses this signal to truncate redundant [EOS] suffixes without retraining, with a batching strategy for variable lengths; it reports up to 31× higher throughput across nine benchmarks while preserving or slightly improving accuracy.
The early-step sparsity signal enabling training-free one-shot sequence truncation is a genuinely interesting efficiency idea, but the abstract gives limited detail about model scale, baselines, and how often the headline 31× speedup applies.
-
SoftNav: Injecting 3D Scene Tokens into VLMs for Embodied Navigation
SoftNav feeds object- and frontier-level 3D scene representations directly into a frozen vision-language model as learned soft tokens, rather than serializing the scene as text. With only about 1,200 training examples and 17M trainable parameters, it reports higher HM3D-OVON navigation success and SPL than prior methods, plus zero-shot transfer to several benchmarks and a real robot.
The direct embedding-level interface between 3D scene encoders and VLMs, combined with low-data training and claimed cross-domain and real-world transfer, is a meaningful result, though the abstract provides limited detail to establish how broad or reproducible the gains are.
-
Looped Latent Attention: Cross-Loop KV Compression for Looped Transformers
The paper compresses the key/value cache across recurrence steps in weight-tied (looped) Transformers, exploiting the observation that each token’s K/V vectors follow a low-rank trajectory across loops. Its post-training latent codec reportedly achieves near-lossless compression up to 32×, improves batch capacity from 32 to 768 sequences at 4k context in one H200 test, and transfers across models; on-policy refinement also recovers reasoning accuracy at higher compression.
The cross-loop compression axis is a genuinely useful and non-obvious alternative to ordinary head/layer KV sharing, with unusually large reported memory gains, but the evidence is limited to looped Transformers and the abstract does not establish broad reproducibility or end-to-end decoding gains.
-
Retraining Seeks Stable Signals
This paper studies repeated retraining when deployed predictions change the future data distribution. It argues that even very strong model-induced feedback can converge geometrically toward a model-independent “stable signal,” provided such a signal exists and retraining is suitably regularized, and extends the analysis to time-varying, nonlinear, and feature-changing feedback processes. The perspective also offers a theoretical explanation for why training on model-generated language data can remain stable.
The claim that a small invariant signal can control retraining despite arbitrarily strong performative feedback is a genuinely non-obvious theoretical framing, but the abstract gives no empirical validation and the result may depend heavily on its modeling and regularization assumptions.
-
Toward a mechanistic understanding of inference in visual cortex and diffusion models
The paper builds a recurrent sparse-coding model of V1 whose dynamics are mathematically equivalent to a minimal diffusion model, trained with denoising score matching. On natural images, the learned latent interactions resemble V1 horizontal connections and support strong denoising of ambiguous structures; analyzing the Jacobian exposes how the circuit represents coherent image deformations and includes latent variables that enforce global consistency without directly receiving visual input.
The explicit mechanistic bridge between recurrent V1-like circuits and diffusion inference, including testable connectivity predictions and an interpretable account of global consistency, is genuinely interesting, but the abstract provides limited quantitative or comparative evidence for the broad claims.
-
Implicit Virtual Leader: Decentralized Vision-Only Relative Pose Estimation for Multi-Robot Formations
The paper proposes a decentralized Transformer/GNN estimator that uses each robot’s monocular image and exchanged messages to infer relative poses through a learned, non-physical team-centroid reference frame. It reports robustness to robot and communication-link removal, simulation-to-real transfer, generalization to larger teams and unseen scenes, and closed-loop formation control on heterogeneous physical robots, with 0.22 m and 1.6° pose errors.
The combination of vision-only decentralized estimation, an implicitly learned leaderless reference frame, and demonstrated sim-to-real formation control is a meaningful direction, but the abstract does not establish enough detail or comparative evidence to justify a strong recommendation.
-
When Can Safe Controllers Adapt? Information before Commitment
This paper studies when a controller can safely learn which environment it is operating in while adapting online. It argues that safety may force the controller to decide before taking actions that would reveal the true model, so limited information available before this “commitment” can make a constant fraction of the oracle gap—and even linear regret—unavoidable. The paper formalizes this with a precommitment KL-information measure, proves the obstruction for constrained linear systems, and gives recovery results and semidefinite certificates for special cases.
The commitment/precommitment-information framing identifies a non-obvious information-theoretic barrier to safe adaptation, but the abstract only reports theoretical results in constrained linear systems without enough evidence to merit a strong recommendation.
-
Afadin Loss Uncovers an Ectopic Neurogenic Niche and Reorganizes the Adult Ventricular-Subventricular Zone
The study shows that deleting the adhesion protein Afadin during cortical development creates a stable, ectopic neurogenic niche in the neocortex, independent of the ventricles, that contains self-renewing multipotent neural stem cells into adulthood. Afadin loss also disrupts the normal ventricular-subventricular zone, keeping stem cells activated and shifting gene expression toward proliferation and neurogenesis; mosaic deletions suggest both cell-intrinsic and surrounding-tissue effects.
The formation of a persistent, ventricular-independent adult neural stem-cell niche through developmental loss of an adhesion protein is a genuinely non-obvious result, though the abstract provides limited quantitative and functional evidence for its broader significance.
-
Teneurins Are SPARCL1 Receptors
This study identifies teneurins as binding partners for SPARCL1/Hevin, an astrocyte-secreted protein that promotes synapse formation. In cultured mouse neurons, SPARCL1’s follistatin-like domain binds teneurins, while synaptogenic activity requires that domain together with a separate calcium-binding domain, suggesting recruitment and activation are distinct molecular steps.
It offers a specific mechanism that could reconcile conflicting accounts of how SPARCL1 acts at synapses, but the evidence is mainly from cultured neurons and binding/domain experiments rather than in vivo circuit or behavioral validation.
-
CRISPR-Associated Transposases Enable Programmable DNA Integration in Plants
The authors adapt and evolve a CRISPR-associated transposase (CAST) system for RNA-guided DNA insertion in plant cells without double-strand breaks. Their evoCAST variant reaches 2.7% chromosomal integration at one locus, 8% insertion into a synthetic landing pad, and can target six endogenous loci; the plant chromatin factor AtHMGB2 further boosts activity.
This is an important first demonstration that CAST systems can mediate programmable chromosomal integration in plants, but the efficiencies and early-stage cell-based validation fall short of a strong practical genome-engineering advance.
-
Bifurcation Structure and Cross Nuclei Universality Govern Frequency-Selective Deep Brain Stimulation
The paper proposes an excitatory-inhibitory attractor model in which a boundary-equilibrium bifurcation, rather than a change in the oscillatory Hopf condition, explains why DBS effects switch sharply near 90 Hz. It also claims that response curves from four DBS targets collapse onto two common rescaled master curves, suggesting a cross-nucleus universality that could yield testable predictions.
The bifurcation-based separation of stimulation threshold from endogenous beta dynamics and the claimed cross-nucleus universality are genuinely interesting, but the evidence is mainly model-derived and based on limited intraoperative recordings, so it is not yet strong enough for a must-read recommendation.
-
A claustro-cortical loop times state transitions for flexible behavior
In freely behaving mice, the authors show that the claustrum helps time transitions between active and silent states in anterior cingulate cortex during a cue-guided switching task. Claustrum population synchrony peaked just before cortical UP-to-DOWN transitions, and silencing claustral inputs prolonged cortical UP states and selectively impaired flexible switching while sparing exploration.
This is a mechanistically specific and causal link between claustral coordination, cortical metastable-state timing, and behavioral flexibility, but the evidence is currently limited to one circuit and task in mice and lacks quantitative detail in the abstract.
-
Selectivity for high-level language processing is highly localized in individual brains
Using precision neuroimaging in individual participants, the authors report small, sharply bounded cortical patches that respond selectively and reliably to language. Although these language regions vary in anatomical position across people, their distinct functional profiles are consistent within each person, suggesting that high-level language is organized into true cortical areas rather than only a diffuse network.
The potentially important contribution is evidence that receptive language has individually localized, sharply bounded cortical areas, challenging the common view that higher-level language functions are primarily distributed; however, the abstract gives few quantitative details about participants, methods, or effect sizes.
-
Projectomic Organization of the Serotonin System of the Mouse Brain
The authors combine whole-brain axon tracing, spatial transcriptomics, and projection-specific depletion to organize mouse serotonin neurons into five groups based on their target territories. These groups are associated with distinct molecular and anatomical features, and selectively removing them produces different behavioral effects, suggesting that serotonin signaling is structured into functionally specialized channels rather than being globally diffuse.
This is a potentially important systems-level reframing of serotonin organization, linking projections, molecular identity, and behavior, but the abstract gives limited detail about scale, causal specificity, and the behavioral evidence.
-
Amplitude-Modulated Kilohertz Stimulation Targeting Beta-Band Activity Disrupts Motor Learning
The study uses amplitude-modulated kilohertz transcranial magnetic perturbation to deliver continuous, frequency-specific electric fields to human motor cortex during a force-learning task. In a double-blind study of 40 participants, beta-frequency stimulation reduced learning-related performance gains compared with delta stimulation and sham, while delta stimulation had no detectable effect. This provides preliminary causal human evidence that elevated beta activity can interfere with motor learning, while introducing kTMP as a tool for manipulating cortical rhythms.
The combination of a potentially useful new noninvasive frequency-targeting method and causal evidence linking beta activity to impaired human motor learning is genuinely interesting, but the evidence is still limited to one task and a modest sample in a preprint.
-
Modulation of retroviral capsid assembly halts ARC-mediated TDP-43 intercellular spreading
The study argues that ARC/Arg3.1, a protein with retroviral-like capsid properties, helps package and spread pathological TDP-43 between cells. In fly and human-cell models, reducing ARC or treating with the HIV capsid drug lenacapavir reduced TDP-43 transmission and improved disease-related outcomes, suggesting a possible therapeutic route for ALS and FTD.
The proposed ARC-mediated capsid mechanism and repurposing of lenacapavir are unusually interesting, but the evidence is still mainly from cellular systems and a Drosophila model, so translation to mammalian disease and humans remains uncertain.
-
From Human-Centric to Agentic Code Review: The Impact of Different Generations of Generative AI Technology on Review Quality
This study analyzes 1.02 million pull requests across 207 GitHub projects as code review moves from human-only review to LLM-assisted and agentic review. It finds that AI-agent involvement—especially agent-initiated reviews or reviews involving multiple agents—is associated with faster decisions, but not better review quality, and that human–AI interaction patterns become the main predictor of efficiency in later adoption stages.
The large-scale longitudinal comparison and finding that agentic review improves speed without improving quality are useful and somewhat counterintuitive, though the observational design and lack of quantitative effect sizes in the abstract limit confidence.
-
Partially Correlated Verifier Cascades in LLM Harnesses: Concave Log-Odds, Polynomial Reliability, and Blind-Spot Ceilings
This paper analyzes cascades of LLM verifiers when their errors are correlated across instances rather than independent. It shows that surviving errors become increasingly concentrated among “hard” blind-spot cases, so reliability may improve only polynomially, plateau because of persistent blind spots, or even worsen when true-accept rates vary; synthetic experiments show large errors from independence-based extrapolation. The main practical implication is that diversifying verifiers may matter more than simply adding more sequential checks.
The polynomial-reliability and blind-spot-ceiling framing gives a useful, non-obvious correction to the common exponential-gains assumption, but the evidence is mainly theoretical plus synthetic experiments rather than validation on real LLM harnesses.
-
Motion4Motion: Motion Transfer Across Subjects at Inference
Motion4Motion transfers movement from a source video to a different character without relying on a predefined skeleton or training on labeled skeleton-specific data. It instead models the character’s motion flow, aiming to support transfer across humans, animals, and other diverse body structures at inference time; the abstract claims strong results and applications but gives no quantitative details.
Skeleton-free, training-free motion transfer across species is a meaningful departure from standard pipelines, but the abstract provides no numbers or detail sufficient to judge whether the claimed gains are genuinely substantial.
-
OvisOCR2 Technical Report
OvisOCR2 is a 0.8B end-to-end vision-language model that converts document page images into Markdown containing ordered text, formulas, tables, and visual regions. It combines real and HTML-derived synthetic data with reinforcement learning, distillation, and model fusion, and reports top results on OmniDocBench v1.6 (96.58) and PureDocBench (75.06), outperforming pipeline-based systems on these evaluations.
The notable result is a small end-to-end parser reaching the top of benchmarks previously dominated by pipelines, but the abstract does not establish how broad or durable the advantage is beyond leaderboard scores and a vague in-house evaluation.
-
Causal Graphs, Markov Properties and Do-calculus for Stochastic Differential Equations
This paper develops a formal causal graphical-model framework for stochastic differential equations, including conditions ensuring that observational and intervention distributions remain well defined under marginalization. It proves Markov properties and do-calculus for sample-path interventions, obtains stronger results for some cyclic additive-noise systems, and connects the framework to subsampled time series, Granger non-causality, and causal-discovery algorithms.
The unified extension of graphical causal reasoning and do-calculus to continuous-time stochastic systems, including cyclic SDEs and subsampling, is a substantial foundational contribution, but its impact is primarily theoretical and the abstract gives no empirical demonstration of practical causal discovery.
-
Worlds in One Demo: A Synthetic Data Engine for Learning Open-World Mobile Manipulation
WANDA turns a single real RGB-D mobile-manipulation demonstration into many synthetic training trajectories. It reconstructs the scene and interaction segments, recombines them across spatial configurations and generated environments, expands states for corrective robustness, and renders photorealistic observations; the authors report improved long-horizon, spatial, and cross-environment generalization, including zero-shot deployment on a different robot morphology.
The one-demonstration-to-open-world-data pipeline and claimed cross-embodiment transfer are genuinely interesting, but the abstract gives no quantitative comparisons or task scale, so the strength of the reported gains is difficult to assess.
-
G$^2$SR: Geometric Methods for Fast and Memory-Efficient Gaussian-based Surface Reconstruction
G2SR reconstructs 3D Gaussian splats from only 2–3 posed RGB views by detecting cross-view 2D splat correspondences and analytically triangulating them, rather than using a large end-to-end network. It reportedly matches or exceeds the geometric accuracy of heavier methods on ScanNet, Replica, and DTU while achieving 69–89 reconstructions per second using just 203 MB of GPU memory.
The explicit separation of lightweight correspondence detection from analytic multi-view geometry yields an unusually large speed and memory improvement for few-view Gaussian reconstruction, though the abstract does not establish a major capability or accuracy breakthrough beyond existing methods.
-
DriftWorld: Fast World Modeling through Drifting
DriftWorld replaces iterative diffusion sampling with a learned action-conditioned drift, generating predicted robot futures in one forward pass at over 30 fps. It reports roughly 17× faster rollouts than diffusion-based world models across several manipulation benchmarks, while retaining strong control performance and achieving up to 0.99 correlation when ranking policies offline.
The combination of a substantially faster world-model sampling mechanism and broad manipulation evaluation is promising for practical model-predictive control, but the abstract does not provide enough detail about absolute performance or comparisons beyond diffusion baselines for a strong recommendation.
-
MegaSlide-DiT: Memory-Centric Adaptation and Deformable Local Attention for Efficient Video Diffusion
MegaSlide-DiT proposes adapting a 105B-parameter video diffusion Transformer using host RAM for persistent weights and optimizer state, streaming only temporary shards to a single H200 GPU. It also replaces global spatiotemporal attention with motion-adaptive local attention whose claimed memory and compute scale linearly with sequence length. The abstract emphasizes a systems prototype rather than training from scratch, but gives no concrete quality, throughput, or adaptation results.
The combination of CPU/offloaded full-parameter adaptation and linear-cost deformable video attention could materially change workstation-scale training, but the abstract lacks quantitative evidence that bandwidth, runtime, and model quality are viable.
-
Tight Sample Bounds for Renyi and Min-Entropy Estimation
This paper gives tight sample-complexity bounds for estimating Rényi entropy and min-entropy over a finite alphabet. It shows that min-entropy needs Θ(k log k) samples—much more than Shannon entropy—and establishes matching Θ(α k^{1−1/α}) bounds for a broad range of Rényi orders, correcting a prior characterization and extending earlier partial results.
The key result is a rigorous correction of the believed min-entropy sample complexity, plus tight bounds across Rényi orders; it is important foundational theory but has no immediate demonstrated impact on practical AI systems.
-
Taurus: Accelerating Out-of-Core Graph Neural Network Inference on Billion-Scale Graphs
Taurus is a single-machine system for exact or sampled GNN inference when graph features and embeddings exceed RAM. It reformulates inference as source-centric communication over sequential SSD scans, combined with GPU/CPU/SSD pipelining and storage-aware optimizations; on graphs up to 269M vertices, 4B edges, and 514 GiB of features, it reports 7–25× speedups over a strong layer-wise baseline and 40–140× over vertex-wise methods.
The sequential-scan reformulation and reported large single-machine speedups are a meaningful efficiency advance for billion-scale GNN inference, though it is a specialized systems result rather than a broadly new GNN capability.
-
Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos
The paper introduces an open audio-visual language model aimed at reasoning over long, complex videos rather than short clips. It combines a roughly 7-million-example training collection, a three-stage curriculum for increasingly long-horizon reasoning, and timestamp-grounded intermediate reasoning steps; the abstract claims strong results across more than 15 benchmarks and transfer to unseen tasks, but gives no quantitative details.
Long-horizon audio-visual reasoning with timestamp-grounded reasoning and an open training recipe is plausibly useful and somewhat novel, but the abstract provides insufficient numerical evidence to establish that the claimed gains are substantial or broadly reliable.
-
FVAttn: Adaptive Sparse Attention with Runtime Load Balancing for Video Generation
FVAttn is a training-free systems method for making adaptive sparse attention practical across multiple GPUs during video diffusion inference. It dynamically moves overloaded attention heads between ranks and uses otherwise idle compute capacity to process extra important blocks, reporting a 4.41× attention speedup and roughly 2.1× end-to-end DiT speedup on Wan2.2 I2V without apparent video-quality loss.
The runtime load balancing and slack-filling approach addresses a real distributed-inference bottleneck and produces a substantial reported speedup, but the evidence is limited to one video-generation setting and appears to be an optimization of an existing sparse-attention strategy rather than a broadly new modeling direction.
-
PalmClaw: A Native On-Device Agent Framework for Mobile Phones
PalmClaw is an open-source framework for running LLM agents directly on smartphones, with native device tools for capabilities such as apps, sensors, and data rather than relying mainly on long GUI action sequences. The authors report an 11.5% relative task-success improvement and a 94.9% reduction in completion time versus the strongest baseline, but the abstract gives little detail about the tasks, devices, or evaluation scale.
The native, explicitly bounded tool interface is a plausible step beyond GUI-only mobile agents, and the claimed speedup is striking, but the limited evaluation details make it hard to judge whether the result generalizes.
-
Confidence Scores in Open-Vocabulary Detection Are a Biased Mixture of Scale and Semantics
The paper shows that confidence scores in several open-vocabulary detectors reflect object scale and the specificity of the text query, not just the probability of a correct detection. Across COCO and LVIS, large objects receive inflated scores and generic queries are suppressed; a parameter-free temperature correction improves small-object Recall@10 by 19.6% without retraining, though it hurts pooled-ranking precision.
This is a useful cross-model diagnosis of a structural limitation in adapting image-level vision-language models to detection, backed by large-scale statistical tests and a practical correction, but the result is primarily calibration analysis rather than a major capability advance.
-
When the Reward Suite Is Leaky: A Preregistered Causal Contrast of Natural Verifier False Positives in RLVR
This preregistered study compares GRPO trained with ordinary MBPP tests against the same training with additional hardened tests, isolating the effect of persistent false positives in code-verification rewards. Leaky tests strongly concentrate rewarded wrong programs, including nearly half of audited rewarded false positives being genuinely incorrect, but hardening the suite produces little average capability change over the tested horizon; a static leakiness audit predicts exposure before training.
The causal, preregistered analysis and evidence that verifier errors reward pre-existing bugs rather than learned exploitation are genuinely interesting, but the capability result is narrow and limited to MBPP, GRPO, and a small set of replications.
-
The Equilibrium Is the Initialization: Lazy Identity Collapse in Physics-Structured Deep Equilibrium Reasoning
This paper audits a physics-structured deep equilibrium model and finds that its implicit solver often performs no meaningful computation: the equilibrium remains numerically identical to the learned initialization, while the decoder effectively ignores the solver output. Across 19 runs on entailment and graph-reachability tasks, bypassing the solver had no accuracy effect in 18 cases, and solver iterations were unrelated to problem difficulty; the authors identify gradient-starvation mechanisms and propose four tests for detecting such lazy equilibria.
The cross-seed substitution tests expose a non-obvious failure mode in DEQs and provide a practical diagnostic protocol, though the study is limited to one architecture and two relatively small tasks rather than demonstrating a broad field-level consequence.
-
Compile, Then Page: Executable SOP Programs and a Capability-Gated Runtime for Procedural LLM Agents
The paper compiles machine-readable SOPs into executable pseudo-code and uses a stack-machine runtime that exposes the active procedure frame to an LLM while hiding or paging the rest. Across six models and seven domains, strong models benefit substantially—on the Bank task, performance rises from 70.4% to 92.8% with perfect refusal correctness—while weaker models can be harmed; ablations suggest the benefit comes from better spontaneous state discipline rather than merely helping the model reconstruct context.
The separation of SOP compilation from capability-gated runtime guidance, plus evidence that paging helps disciplined models but hurts weaker ones, is a non-obvious and potentially useful design principle, though the claims are based on a relatively narrow benchmark and limited model set.
-
WALA Learning Executable Latent Actions from Action-Labeled Demonstrations and Action-Free Videos
WALA learns a latent action representation from action-free videos by predicting semantic and geometric changes between observations, then aligns those latent actions with executable robot actions using labeled demonstrations. It reports 75.2% average success on RoboCasa, along with improved generalization in real-world manipulation, suggesting that unlabeled videos can contribute useful dynamics supervision rather than only visual pretraining.
The combination of future feature/depth prediction with action-labeled policy learning is a meaningful approach to reducing robot-action annotation costs, but the abstract provides limited detail on baselines, scale, and the strength of the real-world evidence.
-
Lesioned Multimodal Language Models Reproduce Aphasic Picture-Naming Patterns
The authors perturb different layers and units of LLaVA 1.6 and test whether its picture-naming errors resemble those of 278 people with post-stroke aphasia. Six of seven error types appeared at clinically comparable rates, and searched perturbations matched at least six categories for 97.8% of individuals and all seven for 79.5%, with Monte Carlo tests suggesting this was not just matching category frequencies. The main limitation is that this demonstrates behavioral resemblance under flexible perturbation, not that the model lesions correspond to the brain mechanisms causing aphasia.
It offers an unusually quantitative bridge between lesioning multimodal language models and individual aphasic behavior, but the flexible parameter search and missing formal-paraphasia category make the claimed digital-twin interpretation premature.
-
Automated Synthesis of Facial Mechanisms for Conversational Animatronic Robots
The paper presents a parameterized linkage-based animatronic face whose mechanism, actuator layout, and motion constraints can be automatically adapted to a target face reconstructed from a single 2D portrait. It also adds a real-time audiovisual motion model that generates both speaking and listening facial behaviors for the synthesized hardware. The combination could reduce the manual redesign burden for customized social robots, but the abstract does not report concrete performance gains or deployment scale.
Automating both the mechanical synthesis and bidirectional conversational behavior of physically realizable animatronic faces is a substantive, potentially generalizable robotics direction, though the abstract provides too few quantitative results to justify a stronger verdict.
-
An Exact Instrument for State Usage in Selective State-Space Models, and the Input-Driven Migration It Reveals
The paper develops an exact per-mode attribution and pruning instrument for diagonal selective state-space models such as Mamba, allowing the output error from dropping any subset of modes to be computed offline. Applying it across Mamba-1, Falcon-Mamba, and Mamba-2, it finds that useful state modes migrate with the input, driven mainly by the input-dependent write map rather than the timestep, and that input-conditioned mode selection can substantially reduce pruning error. The method currently demonstrates headroom rather than actual inference savings because usage is measured in a first pass.
The exact attribution tool and evidence that mode usage migrates across inputs—primarily through the write map, contrary to the usual focus on timestep selectivity—are genuinely interesting, but the practical payoff remains unproven without deployed compute or memory savings.
-
Graph Feedback Controls Consensus and Clique Formation in Open-Weight Language-Model Populations
This paper studies how the communication graph among language-model agents affects whether they converge on a shared naming convention or split into persistent cliques. Across open-weight models from 1.1B to 32B parameters, it finds that similarity-based routing can preserve fragmentation, while routing across disagreeing groups can improve coordination—but only when agents retain interaction history; on ARC-Challenge and MMLU, routing changes error propagation without consistently improving accuracy.
The notable contribution is treating runtime routing as a causal determinant of collective behavior and separating exposure, memory, model response, and population composition with matched controls, though the demonstrated gains are mainly in synthetic coordination rather than task capability.
-
Partial Identification with Multiple Nonlinear Measurements of a Latent Regressor
The paper develops a partial-identification method for linear regression when a latent regressor is observed through several noisy, nonlinear measurements. By bounding cross-source curvature relative to slope, it derives a closed-form, loading-invariant interval for the structural coefficient and provides estimable bounds and uniformly valid confidence intervals. Applied to six AI-exposure measures and 8.88 million ACS person-years, it finds that some measures imply opposite post-2022 employment effects, while a five-source consensus estimate is negative but not presented as a causal displacement effect.
This is a genuinely nontrivial identification framework for reconciling conflicting AI-exposure measures, with sharp theory and a very large application, but its practical conclusions remain measurement reconciliation rather than a major new AI capability or causal finding.
-
Gradient-Free Topology Adaptation for Power Flow Surrogates via In-Context Whitening
The paper proposes adapting neural AC power-flow surrogates to unseen line-outage topologies by recalculating output mean and covariance, then applying ZCA whitening, rather than updating model weights with gradients. On IEEE 30-, 118-, and 300-bus systems, this reportedly reduces errors by 6–28× over frozen models and adapts 21–34× faster than gradient-based methods using a few hundred solved cases per topology.
The gradient-free, architecture-agnostic use of output-statistics adaptation is a genuinely interesting idea with broad system-level speed advantages, but the demonstrated scope is limited to simulated power-flow contingencies and the abstract does not establish robustness beyond these settings.
-
DiffRadar: Differentiable Physics-Aware Radar SLAM with Gaussian Fields
DiffRadar represents a scene with anisotropic Gaussian primitives and uses a differentiable radar forward model to jointly optimize pose and map directly from range-azimuth and Doppler-azimuth measurements. On Radarize and targeted stress tests, it reportedly reduces trajectory error, improves map consistency by more than 2×, and runs in real time at 70 FPS, with especially strong gains in feature-poor corridors and dynamic scenes.
The physics-aware, signal-domain Gaussian-field formulation is a meaningful departure from heatmap scan matching and targets important radar-SLAM failure modes, but the abstract gives no numerical trajectory results or detailed evidence to justify a stronger verdict.
-
Saturation Makes Quantization Error Additive: A Coverage Model with a Certificate
The paper studies whether the loss from quantizing multiple layers can be predicted from measurements of individual layers, rather than requiring interaction terms. Across 4-bit configurations, it finds that per-layer effects explain 85–93% of the variance and that a simple coverage-style model, fitted with one parameter per layer, largely preserves the ranking of configurations; on 30B–355B models it also reports better allocation quality than gradient-sensitivity methods, including at sub-4-bit budgets.
The potentially important result is that quantization interactions are mostly additive or saturating at deployment-relevant precisions, backed by broad experiments and an explicit error certificate, but the practical and generality claims need closer inspection.
-
Code-MUE: Measuring Code LLMs' Uncertainty through Execution-based Semantic Interaction Graphs
Code-MUE estimates uncertainty in code-generating LLMs by sampling multiple outputs, executing them, and representing their behavioral relationships in a Semantic Interaction Graph. It uses the graph’s Von Neumann entropy as a measure of semantic diversity, and reports correlations as strong as -0.98 between uncertainty and functional correctness across eight code LLMs, outperforming text-similarity baselines.
The execution-based replacement of lexical uncertainty measures is a meaningful and somewhat surprising idea for black-box code models, but the evidence is still limited to an empirical correlation study and the approach may be costly or dependent on test quality.
-
VistaVLA: Geometry- and Semantic-Aware 3D Gaussian-Grounded VLA for Robotic Manipulation
VistaVLA augments a vision-language-action policy with a persistent 3D representation built from semantic features attached to 3D Gaussian primitives. Its Merge-then-Query module compresses this representation by 99% into action-relevant context tokens, and the system reports 22.8% higher success across seven real-world tasks and 30.0% improvement over VLA-Adapter on out-of-distribution tasks.
The combination of semantically grounded 3D Gaussian representations with extreme token compression and real-robot OOD gains is a substantive direction, but the abstract lacks absolute success rates, task details, and comparisons to stronger modern VLA baselines.
-
Same Loss, Same Noise, Opposite Schedules: Noise Structure and Optimizer Normalization Jointly Determine Whether Learning-Rate Cooldown Helps
This paper argues that learning-rate cooldown helps normalized or sign-based optimizers under multiplicative gradient noise because their updates do not naturally shrink near a minimizer, while ordinary SGD self-anneals. It proves this on strongly convex quadratics, derives an exact signSGD stationary distribution, extends the result to higher dimensions, momentum, and heavy-tailed noise, and tests the predicted noise-regime distinction on a real classification task.
The joint role of gradient-noise structure and optimizer normalization offers a non-obvious, mechanistic explanation for when cooldown matters, with stronger-than-usual theoretical support, though its relevance to large-scale pretraining remains only indirectly demonstrated.
-
Local Maxima of the Entrywise $\ell_4$ Norm on the Orthogonal Group
This paper completely characterizes local maxima of the sum of fourth powers of the entries of a real orthogonal matrix. It proves that signed permutation matrices are the only local—and therefore global—maximizers, while every other stationary point has a rank-two direction of strictly positive curvature. The result also covers degenerate cases such as zeros, repeated entry magnitudes, reducible supports, and Hadamard-like stationary points.
The exhaustive landscape characterization and explicit escape direction are mathematically clean and somewhat surprising, with possible relevance to orthogonal matrix optimization, but the result is specialized and its direct practical impact on AI/ML is unclear.
-
Fisher Rank Inflation: A Spectral Signature of Memorization under Label Noise
The paper identifies a temporary rise and later collapse in the effective rank of per-example last-layer gradients when networks begin memorizing noisy labels. This “Fisher Rank Inflation” tracks corruption severity and enriches high-rank-contributing examples for mislabeled data across CNNs, ResNets, and ViTs, sometimes appearing before test performance degrades. The authors also derive a first-order spectral attribution method that approximates leave-one-out effects.
The spectral inflation–collapse pattern and its connection to early detection and attribution of label noise are a non-obvious, potentially useful diagnostic, supported by experiments across datasets and architectures, but the abstract does not establish a major practical improvement over existing memorization or noisy-label detection methods.
-
TSSM: Triaxial State Space Model for Global Station Weather Forecasting with Temporal-Variable-Historical Modeling
The paper introduces TSSM, a state-space forecasting model that scans across time, weather variables, and period-aligned historical observations, rather than relying only on a recent lookback window. It reports large improvements on global station-weather datasets, especially for long-horizon and iterative forecasts, extreme events, and missing-observation settings, though the abstract does not give enough detail to assess the comparisons or baselines.
The period-aligned historical-memory design and strong claimed gains under long horizons and severe missingness are potentially important, but the unusually large headline improvements are insufficiently substantiated in the abstract.
-
Gauge-Fixing the Forward-Forward Objective: A Whitened Goodness Derived from a Likelihood-Ratio Account
The paper argues that Forward-Forward training has a scale (gauge) pathology: layers can improve the pairwise goodness loss by amplifying activations rather than separating positive and negative examples. It derives a whitened, scale-invariant goodness objective and reports consistent 4–7 point linear-probe gains over the standard objective across corpora, depths, and widths, while showing that thresholding alone does not fix the underlying issue.
The combination of a likelihood-ratio account, an explicit scale pathology, and controlled evidence that invariance—not merely bounding activation growth—improves Forward-Forward training is a genuinely useful insight, though its practical significance is limited by the narrow Forward-Forward setting.
-
A Learning-Rate-Gated Failure of GRPO in a Small Language and Vision-Language Model Web Agent: A Controlled Null and Its Mechanism
This paper systematically tests GRPO on 4B–8B web agents and finds that it does not improve a strong supervised policy on tasks it already mostly solves; at moderate or high learning rates it can instead degrade or collapse performance. GRPO does improve performance by 22 points when the sampled policy has genuine headroom, suggesting that the method needs reachable successes rather than merely a strong checkpoint. The authors further localize degradation to attention/MLP blocks, distinguish it from collapse, and report that the relationship between late-layer effective rank and capability changes with model scale.
The controlled null, positive headroom condition, and mechanistic separation of degradation from collapse challenge the assumption that verifiable-reward RL reliably improves already-strong agents, though the evidence is limited to small web-agent models and one training setup.
-
Extractable Memorization From First Principles
The paper develops a principled way to test whether an LLM has memorized a training sequence, rather than merely generating predictable text. It compares generation probabilities for matched training and non-training sequences, using either a calibrated conformal test or a document-level census; experiments show substantial false positives in prior extraction setups and that some claimed-to-be-extractable book passages would require infeasible sampling budgets.
This is a meaningful methodological correction with concrete evidence that common extraction rates can substantially overstate memorization, though it mainly sharpens evaluation validity rather than introducing a new model capability.
-
What Makes a Representational Prior Work? Feature Families, Label-Free Invariances, and Critical Windows in Grokking
This paper studies why representational priors reduce the delay before grokking on synthetic algorithmic tasks. It finds that priors must use the task’s relevant feature family, but that label-free invariance signals from commuted input pairs are enough to accelerate generalization; applying them only early in training captures most of the benefit and greatly reduces sensitivity to weight norms. The effects replicate across modular multiplication, model depths, and normalization variants, though the claims are limited to grokking-style settings.
The combination of feature-family alignment, label-free invariances, and a sharply defined early critical window is a non-obvious mechanistic result with substantial supporting runs, but its importance beyond narrow synthetic grokking tasks is still unclear.