HDATF
日本語
相談する
メニュー

論文アーカイブ

収録論文と要約を番号の大きい順に掲載します。

収録論文 1,430資料更新日 Research Monitoringベンチマーク
  1. 1032RoboticsarXiv:2608.27384

    FlashVLA: Streaming Action Decoding for Fast and Asynchronous VLA Inference

    • It began by profiling π0.5 and confirming that action decoding consumes 75% of per-step inference time.
    • It created a streaming architecture with a buffer of action chunks at staggered noise levels, denoising all of them by one step in a single forward pass.
    • Per-step action-decoding latency was reduced by a factor of up to 20, and the full pipeline became up to 2.43 times faster.
    • Causal attention between chunks eliminated temporal misalignment in asynchronous execution without a separate future-state predictor.
    • On LIBERO, RoboTwin 2.0, and real Franka experiments, it maintained or improved π0.5 performance while achieving control at 30Hz or higher on a single GPU.
    概要PDF
  2. 1031AI / GeneralarXiv:2608.26088

    Planetary Prediction Engine: Autonomous Geospatial Prediction via Intelligent Data Selection and Foundation Model Embeddings

    • Using only a natural-language question, it automated a geospatial prediction pipeline of more than 700 steps, from data discovery to model selection.
    • It builds predictive features by combining geospatial foundation-model embeddings with covariates collected in real time.
    • It raised average R² across 21 US CDC health indicators from 60.0% to 76.8%, substantially outperforming the expert baseline.
    • In forecasting the 2026 Ebola outbreak in Congo, forecasts over 5 consecutive weeks anticipated 15 of the 18 newly affected health zones (Recall@10 83.3%).
    • It embedded validation mechanisms such as target-leakage prevention and overfitting guards within the agent loop.
    概要PDF
  3. 1030LLMs / NLParXiv:2608.27448

    TTPO: Test-Time Policy Optimization

    • On competition-level hard problems, it found an asymmetry: even when pseudo-answers were wrong 85% of the time, 79% of the rollouts that disagreed with them were actually wrong.
    • Using this asymmetry, it developed TTPO, which applies distillation to agreeing rollouts and group RL penalties to disagreeing rollouts.
    • Without any labels, it matched or exceeded label-based learning (OPSD) on 5 competition-level benchmarks.
    • In a purely test-time learning setting, it raised Qwen3-1.7B accuracy from 38.0% to 45.2%.
    • It delivered gains of 25.2–36.4 percentage points even without thinking mode (long reasoning), and training on one benchmark generalized to others.
    概要PDF
  4. 1029Computer VisionarXiv:2608.27395

    LeVJEPA: Efficient & Scalable Video Pretraining without the Heuristics

    • It learned video representations using only a single encoder, without an EMA target encoder, stop-gradient, or predictor.
    • Randomly dropping 95% of tokens actually increased ImageNet accuracy from 33.9% to 47.6%.
    • With the same data and number of epochs, it matched or exceeded V-JEPA 2's performance while reducing total pretraining compute by a factor of 5.6–20.8.
    • At the same compute budget, it outperformed the strongest video baseline by 7.6 points on ImageNet-1K.
    • Training with block-causal attention allows representations to be extended at constant cost as new frames arrive, making it well suited to streaming and world models.
    概要PDF
  5. 1028LLMs / NLParXiv:2608.26070

    Prefix Sliding for efficient test-time scaling

    • Attention analysis showed that intermediate reasoning tokens quickly lose importance, leaving only the prefix and recent tokens important.
    • Applied to an existing model (Qwen3) without training, it preserved performance while making inference 3 times faster.
    • Combined with RL training, it enabled training on reasoning sequences longer than 100,000 tokens, addressing the problem of discarding long rollouts.
    • It confirmed advantages in both performance and speed over alternatives such as summarizing intermediate tokens or using a simple sliding window.
    • It also identified a limitation: tasks that require remembering early code for a long time, such as the coding benchmark LiveCodeBench, need a large window (16384).
    概要PDF
  6. 1027AI / GeneralarXiv:2608.26701

    Accelerating Scientific Research with Gemini in the Real-World

    • Connected to CVD synthesis equipment, it designed an MXene synthesis route using harmless precursors and grew monolayer MoS2, MoSe2, and WS2 on the first attempt.
    • It built an agent for predicting E. coli colony morphology and produced results that quantitatively matched unpublished wet-lab measurements.
    • It fully autonomously designed a medical-response agent that outperformed 6 frontier models on HealthBench Hard/Professional.
    • Through 450 double-blind reviews by 30 experts, it demonstrated the effectiveness of a reliability module that reduces hallucinations and plagiarism in generated papers.
    • It showed that adaptive autonomy, which adjusts the level of human-AI collaboration by domain, is a key design dimension of closed-loop research.
    概要PDF
  7. 1026Computer VisionarXiv:2608.25559

    AdaVDR: Adaptive Tool Use and Reflection for Video Deep Research

    • Addressing the fact that the optimal tool combination differs by question and video type, it lets the model skip tool calls based on its own capabilities and internal knowledge.
    • It proposed 'model-conditional tool necessity filtering,' which removes calls from training trajectories when the target model can answer without tools.
    • It built and released the VDR-EE benchmark and training data, covering 250 entity- and event-centered questions across 7 domains.
    • It achieved the best performance among the open-source models evaluated and also showed substantial improvement over the base model on VideoDR.
    • Redundancy-aware reward RL reduced average tool calls from 7.84 to 6.80 while maintaining accuracy (51%).
    概要PDF
  8. 1025AI / GeneralarXiv:2608.24735

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

    • It diagnosed existing self-improvement systems as unable to achieve a realized meta-depth beyond approximately 2 because their improvement mechanisms themselves remain fixed.
    • It recursively applies a fixed meta-operation to its own output and determines the depth itself until convergence, reaching 3–6 layers in practice.
    • Its decomposition of the gains attributes approximately 72% of the recursive benefit to context transfer between layers and approximately 15% to a callable code library.
    • On the ARC-AGI-2 holdout, the meta-stack was the only approach to rise above the floor, scoring 0.331, while Godel Agent and OpenEvolve scored 0.
    • Removing recursion reduced the CO-Bench validation score from 0.845 to 0.714, isolating recursion's own effect (+0.131).
    概要PDF
  9. 1024RoboticsarXiv:2608.26053

    $R^3$: Training Robots to Reason in Natural Language via Reinforcement Learning

    • Unlike prior approaches that used language reasoning only as an auxiliary training signal, it generates reasoning at test time to direct a lower-level policy.
    • Stage 1 initializes the reasoning style from expert-reasoner traces (regardless of success or failure), and stage 2 performs offline RL (Dr. GRPO) using rubric-based rewards from a VLM judge.
    • Across 12 held-out grocery-packing tasks, it achieved an average success rate of 47.9%, outperforming instruction-only imitation learning (38.0%).
    • The trained reasoner exhibited deliberative, action-oriented reasoning, such as comparing alternatives, observing the scene again when uncertain, and choosing steps incrementally.
    • Intervention experiments that adjusted the reasoning budget confirmed that reasoning contributes causally to performance.
    概要PDF
  10. 1023LLMs / NLParXiv:2608.25593

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

    • Viewing agent capability as the combined result of the model and harness, it proposed a JIT approach that generates a harness on the fly for each task instead of fixing one in advance.
    • It represents the harness as executable code divided into four modules, memory, planning, action, and capability, enabling generation, repair, and evolution.
    • It trained the generator by combining task-specific learning, failure-recovery learning, and Evo-GDPO, which accumulates better-performing harnesses.
    • DeepSeek-V4-Flash surpassed GPT-5.6 by 9.1 and 4.3 points on two benchmarks, while GLM-5.2 also improved by up to 20.2 points depending on the task.
    • Changing the entire execution framework on the fly leaves stability and verification issues in operational environments, and the authors themselves identified a stable core with verifiable partial modifications as future work.
    概要PDF
  11. 1022RoboticsarXiv:2608.26058

    One Policy, Many Embodiments: Unified Camera-Centric Action Geometry Pre-training for Heterogeneous Embodied Manipulation

    • It aligned robot, simulation, and human-hand data in a single camera-centric action space for unified pretraining.
    • It directly incorporated 2,340 hours of human demonstrations into training as a 'human embodiment,' without retargeting or video synthesis.
    • A single checkpoint achieved 98.3% on LIBERO, 88.7%/89.2% on RoboTwin Easy/Hard, and 82.0% zero-shot on LIBERO-Plus.
    • On a real robot, it achieved 60% for picking up bread, 90% for opening a drawer, and 75% for stacking bowls, outperforming pi0.5 (20/85/65%) under the same conditions.
    • It also disclosed limitations: camera-calibration and depth-estimation errors propagate directly, and cross-embodiment transfer remains difficult (35% from ALOHA to ARX).
    概要PDF
  12. 1021Computer VisionarXiv:2608.25412

    AdaptiveEmbed: Sample-Adaptive Multi-Vector Representation for Multimodal Retrieval

    • It motivated the problem with the observation that the optimal number of vectors varies by sample, ranging from 1 to 8.
    • Multi-Group Contrastive Learning (MGCL) produces candidate vectors, and Utility Policy Optimization (UPO) determines the number of vectors for each sample.
    • On COCO, it achieved the best performance in 6 of 8 directions with an average of 2.1 vectors, and also led in zero-shot retrieval on OpenImages with a gallery of 546,000 items.
    • It also outperformed fixed-budget approaches in video and audio retrieval with an average of 1.9 vectors, demonstrating that the approach is modality-independent.
    • An oracle allocation could reach an average of 61.0 mAP at a similar average budget, suggesting substantial potential in adaptive allocation itself.
    概要PDF
  13. 1020Computer VisionarXiv:2608.26105

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

    • It found that VLM-as-a-judge evaluations repeatedly fail through counting errors, disregard for fine-grained evidence, and misreading of rules, and replaced them with deterministic rule-based scores.
    • A model trained on 300 procedurally generated tasks showed substantial transfer to 7 external benchmarks, including RISE-Video and MME-CoF-Pro (often +20 percentage points or more).
    • Nearest-neighbor and counterfactual diagnostic experiments tested whether the gains came from genuine visual reasoning rather than memorization of instruction patterns.
    • Comparing more than 30 image, video, and interleaved generators on the same task distribution showed that video is advantageous for spatiotemporal state tracking, while interleaving is advantageous for computational efficiency.
    • RL with verifiable rewards raised the maze-navigation score from 0.2249 to 0.8692, and self-correction of intermediate answers was also observed.
    概要PDF
  14. 1019Computer VisionarXiv:2608.25924

    Visual General Intelligence: A White Paper

    • Bringing together multiple researchers' perspectives, it examined whether visual experience, which predates language by far, could provide an independent starting point for general intelligence.
    • It proposed large-scale generative learning, future prediction, spatial memory, continual learning, active perception, and embodied interaction as candidate paths to visual intelligence.
    • It argued that evaluation should move beyond fixed-task scores to assess transfer to new situations, continual adaptation, creativity, physical consistency, and active information seeking.
    • It concluded that multiple paths, including vision-only, vision-first, and language-integrated approaches, should remain open, rather than prematurely fixing visual general intelligence to a single model or definition.
    • This work is a white paper organizing a research agenda rather than a paper presenting a new model or integrated experiments, and current visual capabilities remain fragmented, with limits in computational cost and temporal scope.
    概要PDF
  15. 1018RoboticsarXiv:2608.26103

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

    • Because language instructions struggle to convey spatial constraints and intermediate states, it adopted human demonstration videos as the task-instruction interface.
    • A pipeline that automatically generates semantically matching human videos from robot trajectories produced HumanGen with 74,200 pairs (8,600 tasks).
    • It added an objective for predicting future segments (IFP) to force the model to consult the video prompt rather than rely on shortcuts from existing task knowledge.
    • On RoboTwin 2.0, it achieved an average success rate of 47.0% across 7 unseen tasks, a gain of +29.5 percentage points over the strongest baseline (LingBot-VA).
    • On real robots, it also followed video instructions well in unseen configurations involving multiple objects, long-horizon tasks, and precision insertion.
    概要PDF
  16. 1017Computer VisionarXiv:2608.25927

    Code World Model: Coding Agent as World Brain

    • It diagnosed existing world models trained only on video as imitating resulting scenes without learning the rules that explain why things happen.
    • It divided responsibilities into a dual structure: a coding agent, the brain, makes infrequent complex decisions, while code rapidly handles frequently repeated state updates.
    • It extracted 9,420 clips from 157 gameplay takes, approximately 5.6 hours, and fine-tuned MiniMax-H3 with LoRA using approximately 596 million trainable parameters.
    • It reported that the proxy-conditioning resolution is approximately 1/16 of the output resolution, adding almost no inference overhead.
    • It acknowledged limitations: the training scale is small, generation is not real-time, and it remains difficult for the coding agent to create complex game mechanics from scratch.
    概要PDF
  17. 1016Computer VisionarXiv:2608.24855

    LeFlow: Generative Latent Flow Planning for World Models

    • LeFlow uses a rectified flow model on top of a frozen latent world model to generate reusable latent paths connecting the current state to a goal.
    • An inverse-dynamics decoder converts the generated latent transitions into action chunks, and autoregressive rollouts of the world model validate and select from a fixed number of candidates.
    • Across four goal-conditioned pixel-control benchmarks, success rates were consistently higher than with iterative action-space optimization, while planning time was reduced by a factor of roughly 10.
    • Its practical significance is that it extends a world model beyond a simple predictor into a fast planner reusable across different states and goals.
    • The current method is limited to short, fixed planning horizons. Over longer horizons, the generation problem grows and rollout errors from the frozen predictor accumulate.
    概要PDF
  18. 1015Computer VisionarXiv:2608.24053

    WeMM-Embedding: WeChat Multi-Modal Embedding Technical Report

    • A single model embeds text, images, video, documents, and mixed inputs, with freely selectable output dimensions from 64 to 2048.
    • The 9B model ranked first among all open-source and commercial models on MMEB-v2 with an overall score of 80.6 as of 2026-08-24.
    • Even the small 2B model surpassed the previously best-performing 8B open-source model, demonstrating its efficiency.
    • It has been deployed in search and recommendations for WeChat Channels, Official Accounts, Moments, and commerce, delivering consistent improvements in 14 A/B tests.
    • Even when reduced to 256 dimensions, it retained 98.7% of its performance at 2,048 dimensions, substantially reducing storage and speed burdens.
    概要PDF
  19. 1014RoboticsarXiv:2608.24714

    GaussianWAM: Distilling Geometry and Semantics from 3D Gaussian Fields into World-Action Models

    • Using a 3D Gaussian field as an intermediary, it spatially aligned depth, semantic, and coverage supervision with WAM representations that had previously focused only on predicting future video.
    • Because the teacher is used only during training, the inference pipeline and computational cost remain unchanged, allowing use without additional production-service overhead.
    • On LIBERO-Plus, it improved FastWAM from 52.05%→71.29% and Cosmos Policy from 71.52%→77.30%.
    • Integrating teacher signals through the Gaussian field was effective: it scored 1.9 points higher than plain direct distillation (69.37%).
    • In real dual-arm robot experiments (two UR7e robots), it also raised average success from 30.00% to 40.00%.
    概要PDF
  20. 1013AI / GeneralarXiv:2608.24876

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

    • It reduced noise from long conversation histories by retaining the verified current task state in working memory and using that state to retrieve needed skills from experiential memory.
    • It created a recursive improvement loop that uses execution traces to localize failures to memory components, modifies only the relevant part, and incorporates only changes that pass validation.
    • Of 37 completed combinations across four long-horizon tasks and ten models, 35 improved; even the best-performing models gained approximately 15–18 points, reaching a maximum success rate of 87.9%.
    • Gains increased with task length, reaching 32.2 points in the longest segment, and the largest reductions in representative long-horizon task failure types reached 80% or more.
    • However, failure localization is a decision about which part is most amenable to repair rather than a rigorous causal determination, and some verifier components were not sufficiently validated because they were ineffective on certain tasks.
    概要PDF
  21. 1012Machine LearningarXiv:2608.24814

    Effective Learning Rate Governs Loss Dynamics in Language Model Pretraining

    • It proposed a new perspective: first design the schedule of ELR, the ratio between learning rate and parameter magnitude, rather than tuning the two separately.
    • Matching ELR alone produced nearly identical loss curves across different optimizers, architectures, data, and model sizes (with an average error on the order of 10^-3).
    • The effects of weight decay and Hyperball (a technique that constrains parameters to a specified magnitude) on loss were also explained by the ELR schedules each produces.
    • ELR-based scaling laws remained predictive across different magnitude-control methods. Using learning rate instead increased prediction error 12.9-fold.
    • Condition-specific experiments also identified when the law holds precisely (normalization design and rate of change).
    概要PDF
  22. 1011RoboticsarXiv:2608.24885

    Do Robotic World Models Really Follow Actions? Diagnosing and Aligning Action-Conditioned Generation for Policy Learning

    • In addition to demonstrated actions, WorldEcho probes world models with four kinds of non-expert action queries: replay from other states, local perturbations, policy rollouts, and sampling from the feasible space.
    • WorldSync broadens the distribution of action outcomes, anchors intermediate video representations to robot motion with an Action-Forcing Expert, and aligns changes caused by altering actions with changes in the actual future.
    • Across 50 RoboTwin tasks, WorldSync achieved the lowest error accounting for visual integrity, 0.0661, and the highest video pass rate, 84.51%.
    • Using this model as a learned simulator for two rounds of policy improvement raised simulated success rates from 51–52% to 65% and real-robot cup stacking from 48% to 68%, exceeding CtrlWorld by 8–9 and 12 points, respectively.
    • However, its advantage was uneven across individual metrics: Cosmos-Predict2.5 achieved a better raw NDTW of 0.013, and the authors stated that extending the approach to long-duration interactions and multiple robot embodiments remains future work.
    概要PDF
  23. 1010Computer VisionarXiv:2608.24680

    Game2World Engine: Unlocking In-the-Wild Gameplay Videos for World Model Training

    • It automatically extracted interface fragments from real gameplay videos, collected 5,132 human-verified assets, and overlaid them on clean videos again with natural temporal behavior to synthesize paired training data.
    • GameCleaner uses a multimodal language model to recognize on-screen overlays and a video diffusion transformer to restore occluded areas, so regions to erase do not need to be marked in advance.
    • It achieved AAR 95.36 and background preservation 99.00 in synthetic evaluations, and AAR 80.05 and background preservation 99.80 in real-gameplay evaluations when provided with reference images.
    • Training on videos with on-screen overlays removed increased the overall VideoReward score by 6.83%, opening a route to repurposing the abundance of gameplay videos on the internet for world-model training.
    • However, the authors stated that reconstruction can go wrong with large opaque windows, rapidly changing overlays, or graphics entangled with the scene, and that the downstream experiment was a proxy text-to-video experiment without action conditioning.
    概要PDF
  24. 1009Machine LearningarXiv:2608.24696

    On-policy Distillation with Verifiable Reward

    • It reformulates the implicit reward of sampled-token distillation in terms of trajectory correctness and applies a ReLU gate that retains only nonnegative rewards for correct trajectories and only nonpositive rewards for incorrect trajectories.
    • In a same-architecture setting with Qwen3-4B, the average across six benchmarks reached 49.1, above conventional distillation's 47.8, and its AIME24 score of 36.9 also exceeded the teacher model's 36.0.
    • A control experiment reversing the gate direction underperformed basic distillation on all six benchmarks, showing that aligning the sign with correctness is the source of the effect.
    • It can be applied directly to existing policy-gradient algorithms such as GRPO without introducing additional hyperparameters, allowing integration into existing post-training pipelines with few changes.
    • However, the student did not catch up with the teacher: in a different-architecture setting, its average of 22.8 remained far below the teacher's 30.9, and experiments were limited to mathematical-reasoning benchmarks with the Qwen3 family.
    概要PDF
  25. 1008RoboticsarXiv:2608.24882

    Latent Action as Intention Enables Efficient Future Imagination for World Action Models

    • A discrete latent-action tokenizer, trained to emphasize regions around hands and manipulators using masks automatically generated by SAM 2, compresses changes in video to form future intentions.
    • During training, video, latent actions, and an action expert are learned together. At inference, the future-video branch is removed, and only latent intentions and action chunks are jointly denoised.
    • Average success on RoboCasa was 65.6% with limited data and 80.8% with the full dataset, exceeding Fast-WAM under the same conditions by 9.6 and 4.5 points. On LIBERO-Plus, it achieved 74.4% without additional training, 4.0 points above Joint-WAM.
    • At 338.5ms per action chunk, it was 42.9% faster than Joint-WAM's 593.1ms, making it favorable for real robot control, where latency can directly lead to failure.
    • However, in speed alone it remained slower than Fast-WAM's 196.5ms, which discards future imagination entirely. Without pretraining on egocentric videos lacking action information, the latent-action approach was actually weaker than Joint-WAM under the same conditions.
    概要PDF
  26. 1007Computer VisionarXiv:2608.24646

    On-Policy Self-Distillation in Diffusion Models

    • A frozen behavior policy produces trajectories and anchor points, reward gradients define bounded positive and negative targets around those anchors, and the learning policy trains toward them as detached supervision signals.
    • It separates target construction from actual fitting to those targets, allowing each stage to be measured independently when performance fails to improve.
    • It achieved the highest final held-out score in 19 of 20 reward-aligned conditions covering two backbones and 10 evaluators.
    • Compared with DiffusionNFT, GPU time per 100 updates fell by 40% and 63%, allowing more frequent post-training to align image-generation models with human preferences under the same budget.
    • However, controlled experiments showed that better target construction does not guarantee a corresponding gain from a single update: on HPSv2.1, the ranking reversed for 62.3% of 512 prompts.
    概要PDF
  27. 1006Computer VisionarXiv:2608.24845

    LAION-BVD: A 10-Million-Hour Open Video Dataset for Multimodal Pre-training

    • It filtered CommonCrawl records for YouTube, Vimeo, and Dailymotion links to assemble 1.3 billion links, then downloaded 80 million of them to obtain 10 million hours of video.
    • From these, it selected 2.4 million videos and split them at scene transitions into 55 million clips, automatically generating video descriptions with Qwen3-VL-2B-Instruct and audio descriptions with Audio Flamingo 3.
    • After seeing 50 million training samples, BVD-V-10M and BVD-V-50M achieved overall averages 3.3 and 4.0 points higher, respectively, than the more heavily filtered InternVid-10M-FLT.
    • As an open resource that handles video, audio, and images together, it is readily useful for tasks where embedding quality matters, such as retrieval and representation learning.
    • However, all captions are short and automatically generated by small models, so they can be limited in expression and inherit model biases. In fact, ImageNet classification scores were lower than with existing web-image data.
    概要PDF
  28. 1005AI / GeneralarXiv:2608.23041

    AutoSaddler: Automatic Harness Optimization with Durable Updates from Agent Execution Traces

    • It formalizes harness optimization as an offline learning problem and iteratively patches the harness in response to failure signals.
    • Deep debugging rather than shallow reflection, structured patches rather than arbitrary changes, and generalization checks were central to its effectiveness.
    • It achieved 72.3% on GAIA2 with approximately 1,000 task executions, outperforming a competing method that scored 64.6% using approximately 2,800 executions.
    • Across all three benchmarks, it improved by +9.0–10.0 percentage points over the base harness and +4.4–7.4 percentage points over the strongest automatic baseline.
    • The authors noted that it requires training tasks with ground-truth answers and success/failure signals, making direct use in real-world settings difficult, and that its applicability is limited to independent tasks without persistent state.
    概要PDF
  29. 1004Computer VisionarXiv:2608.23238

    Mover360: Controllable Object Manipulation in 360° Panoramic Images

    • It focuses on moving objects in 360-degree images while also handling deletion and insertion in a single model.
    • A single point click lets it infer size, support, and lighting, while boxes and masks also allow precise control.
    • A UE5 pipeline produced 12,150 training pairs (60,750 pairs per epoch) and a 210+50 benchmark.
    • Across both synthetic and real domains, it outperformed leading perspective-image editing, insertion, and inpainting methods in both reconstruction fidelity and semantic consistency.
    • However, instructions are fixed to three operations, so it does not yet support edits that change object attributes or specify complex relationships in words.
    概要PDF
  30. 1003RoboticsarXiv:2608.23478

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

    • It goes beyond behavior cloning by distilling the local purpose, or intent, of actions as an explicit supervision signal.
    • Teacher inference and target generation occur only once offline, and all teacher-side modules are removed at deployment, so adding it to an existing VLA increases runtime overhead only by the additional decoder representations.
    • It improved GR00T-N1.7 on SimplerEnv-Bridge from 64.3% to 84.7% (+20.4 percentage points) and on RoboCasa from 64.1% to 70.3%.
    • Average performance on real-robot tasks rose from 62.0% to 68.7%, with gains of up to +12 percentage points on long-horizon tasks.
    • However, the meaning of intent is constrained by what the teacher VLM knows: with a teacher lacking embodied experience, targets lost almost all directionality, and performance fell below even the baseline that learns only by imitating actions.
    概要PDF
  31. 1002ResearcharXiv:2608.23132

    The Cartan determinant conjecture for representation-finite algebras

    • It proved that a representation-finite algebra with finite global dimension has Cartan determinant exactly 1.
    • It combined existing results through a concise argument linking the standardness theorem with path-length grading in the mesh category.
    • The same argument also yielded an extension theorem in the form End(M) (Theorem B).
    • The paper's Use of AI section explicitly states that AI was involved in discovering the proof and drafting the manuscript.
    • The author wrote that the proof relies on the structure theory of representation-finite algebras and that they do not know whether a proof exists without using that theory.
    概要PDF
  32. 1001Computer VisionarXiv:2608.23329

    Thinking Beyond Videos: Unifying Video Reasoning and Deep Research for Open-World Video Agents

    • It integrated video cropping (zooming in on a segment), image search, text search, and web browsing into a single exploration loop.
    • In situations such as news, demonstrations, documentaries, and lectures, where clues in a video must be connected to knowledge outside it, even small models can answer by filling knowledge gaps through retrieval.
    • It built an automated data pipeline that produces 26K verified SFT trajectories and 3K reinforcement-learning instances.
    • VideoRover-8B-RL scored 56.0% on VideoDR, comparable to GPT-5 without tools at 57.0%.
    • However, removing all external search tools caused VideoDR accuracy to collapse from 56.0% to 13.0%, showing that performance depends heavily on real-time web search.
    概要PDF
  33. 1000Machine LearningarXiv:2608.23566

    How to Train a Critic Stably and Efficiently

    • Controlled experiments isolated the causes of critic instability one by one (unbounded value predictions, batch-level normalization, and others).
    • It proposed the BPCO recipe, combining DPPO + reward-range constraints + Monte Carlo targets + unnormalized advantages.
    • Sampling just 1 response per prompt matched or exceeded group-based (GRPO) baselines.
    • Because the critic is used only during training, information that must be hidden from the policy, such as correct answers or scoring rubrics, can be provided only to the critic. This allows reinforcement learning to extend to tasks rewarded through rubrics.
    • However, the paper states that its evidence is limited to mathematics and rubric-based rewards, that the reward range must be known in advance, and that the comparisons do not account for the additional computation and memory used to train the critic.
    概要PDF
  34. 999RoboticsarXiv:2608.22869

    UniMem: Unifying Multimodal Memory and Control for Vision-Language-Action Models

    • An event classifier detects task-critical moments in the backbone's latent space and updates text and keyframe memory.
    • It integrates memory and low-level control into a single backbone, π0.5, without a separate VLM for memory management.
    • Keyframe caching preserves single-frame-level speed at approximately 90ms per frame, making it 6× faster than hierarchical memory methods and allowing scaling to a dual-arm configuration with 4 cameras on a standard workstation GPU without a speed penalty.
    • It achieved an average of 93.4% across 5 simulated tasks, versus 68.2% for the fixed-interval-frame baseline, and 80.0% across 4 real-robot tasks, versus 43.5% for hierarchical MemER.
    • However, training-compute limits restrict it to retaining only three past milestones, and it discards the oldest frame when this limit is exceeded.
    概要PDF
  35. 998AI / GeneralarXiv:2608.23318

    Agent-G$^2$: Gaussian Guidance for Agentic Reinforcement Learning

    • Diagnostic experiments showed that useful hint depths form a Gaussian-shaped 'band' rather than a single point.
    • It estimates a Gaussian (center + width) online for each task cluster using only existing rollout statistics, then samples depths without additional rollouts.
    • On ALFWorld, it achieved 95.3% (1.5B)/98.4% (7B), outperforming the strongest hint-based baseline by 1.5–2.3 points.
    • It achieved a reward of 92.3 on WebShop with less than one-third of the rollout cost of task-specific search, making it possible to add to existing reinforcement-learning pipelines without a separate search process.
    • However, the paper notes that each training task requires an expert trajectory, so the method cannot be applied as-is in environments where such demonstrations are difficult or expensive to obtain.
    概要PDF
  36. 997Computer VisionarXiv:2608.23549

    FixAnything: 3D-Consistent Rendering Refinement via Video Generative Priors

    • It proposed a general-purpose framework that cleans up rendering defects across 4 representations, 3DGS, NeRF, meshes, and point clouds, using a single model.
    • However, the model must invent regions not visible in the input views, and the paper's uncertainty analysis also showed a large quality gap: the 25% of pixels with high confidence had a PSNR of 25.7dB, while the 25% with low confidence had 14.4dB.
    • Mask conditioning that identifies clean pixels distinguishes frames to trust from frames to repair, preventing hallucinations; removing the mask caused a 1.3dB drop.
    • Applying Flow-DPO with camera-pose reconstruction accuracy as the reward improved 3D consistency, measured by AUC@5°, by 7.2%.
    • It matches or exceeds expert pipelines in quality, with an architecture that only requires swapping in a stronger video model when one becomes available.
    概要PDF
  37. 996AI / GeneralarXiv:2608.23493

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

    • It derives self-reflection summaries from completed trajectories and uses them as conditions to generate dense token-level training signals.
    • It runs using only the model itself, without a larger teacher, external critic, or separate reward model, allowing post-training even where access to more capable models is difficult.
    • With Qwen3-8B, it achieved 73.3% on AIME'24 while using only 8% (0.08 times) the computation of scaled-up SFT.
    • It also substantially improved success on long-horizon tasks, reaching 64.7% on WebShop, 76.8% on ALFWorld, and 31.2% on SWE-Bench-Lite.
    • However, reflection does not always work: 42% of failed reflections were vague advice, 35% misdiagnosed the cause, and 23% concerned problems requiring knowledge the model did not possess in the first place.
    概要PDF
  38. 995Computer VisionarXiv:2608.23486

    GeoWAM: Visual Geometry World Action Models for Autonomous Driving

    • It proposed world-model pretraining that predicts future scene geometry, point maps, rather than future observations in pixels.
    • However, the paper notes that this advantage does not hold at every horizon: at a 1-second lookahead, the video-based Epona+DVGT achieved higher threshold accuracy.
    • For future-geometry prediction, it surpassed the strongest baseline, improving average AbsRel from 0.274 to 0.257 and δ<1.25 from 0.655 to 0.754.
    • It achieved the highest EPDMS among the compared methods in NAVSIM v2 closed-loop evaluation.
    • Because it does not need to generate future video separately, it transfers predicted geometric knowledge directly to driving planning without a video-generation stage.
    概要PDF
  39. 994AI / GeneralarXiv:2608.23565

    ReWorld: An Interactive World Model with Long-Horizon Memory

    • It mixed attention-window sizes across heads and assigned them randomly at each step so that control is learned through short windows and memory through long windows.
    • At inference, it discards the full history and fills a fixed KV cache from a landmark bank retrieved by camera pose.
    • LoRA distillation reduced sampling to four steps, allowing a single backbone to handle both a high-quality multi-step mode and a real-time interactive mode, producing 704x1280 video in real time.
    • It ranked highest in both control accuracy (rotation error of 11.95 degrees) and visual quality among 6 recent interactive world models.
    • However, because memory is retrieved solely by camera pose, the paper leaves extending memory to dynamic scenes and interactions beyond movement as future work.
    概要PDF
  40. 993Computer VisionarXiv:2608.23189

    EchoWM: Open and Enterable Omnimodal World Models

    • It is an omnimodal world model in which a single model simultaneously generates video, environmental sounds, background music, and speech.
    • It unifies discrete commands and continuous poses into a common 6-DoF trajectory, handling first-person and third-person views without separate controllers.
    • It resolved the conflict between quality and control through 4 stages: AV pretraining, control training, joint fine-tuning, and autoregressive post-training.
    • It ranked first on the public WBench Navigation benchmark and also achieved top-tier quality on SANA-WM-Bench.
    • However, without persistent three-dimensional memory, scene structure, characters, and sounds can gradually drift across repeated sequential generations.
    概要PDF
  41. 992LLMs / NLParXiv:2608.23564

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

    • Behavioral tests alone cannot filter out submissions that leave the original implementation intact, so when agents are tasked with replacing a repository's entire stack, whether the replacement actually occurred must be checked separately.
    • It created tasks from 20 real open-source repositories covering 4 types of technical debt: languages, frameworks, platforms, and build tools.
    • It scores in 3 stages using migration audits, 130,118 fixed behavioral tests, and adversarial validation by 6 coding agents. However, the paper draws a clear boundary: the absence of counterexamples is not proof of behavioral equivalence, but only the strongest evidence this benchmark can provide.
    • Only 5.4% (28) of 520 runs passed every stage, and no model solved 13 of the 20 tasks.
    • Capabilities varied widely: build-tool replacement achieved some success (31.4 points), while language replacement almost always failed (5.6 points).
    概要PDF
  42. 991Machine LearningarXiv:2608.22764

    LpWM: A Case for Sparse Representations in World Models

    • It theoretically showed that nonlinear dynamics can be approximated linearly in a sufficiently high-dimensional one-hot latent space.
    • It proposed LpWorldModel, which uses RDMReg regularization to learn sparse latent codes in which most coordinates are zero.
    • On PushT, sparse representations achieved planning success rates up to 57 percentage points higher than dense representations in LeWM.
    • A structure emerged spontaneously in which the positions of nonzero coordinates encode discrete dynamical modes, while their magnitudes encode continuous states within each mode.
    • However, the benefit of sparse representations depends on predictor capacity: the difference from dense representations nearly disappeared in the easy Wall environment or with a sufficiently expressive predictor.
    概要PDF
  43. 990LLMs / NLParXiv:2608.23311

    Beyond the Stability-Exploration Dilemma: Environmental Regularization for LLM Policy Optimization

    • It formalized the problem that existing reinforcement learning regulates only the response distribution while allowing the question distribution to drift without control.
    • It added an input-side regularizer called Query-KL to prevent drift in the training environment while preserving freedom to explore responses.
    • It can be added to GRPO, PPO, and REINFORCE pipelines with minimal changes and no additional forward passes.
    • Replacing standard Policy-KL regularization consistently improved accuracy on 6 mathematical-reasoning benchmarks.
    • However, the paper states that ERPO is not entirely immune to collapse during long training runs, and that validation is limited to mathematical-reasoning benchmarks and Qwen-family models, leaving transfer to conversation, code generation, and multilingual settings unverified.
    概要PDF
  44. 989AI / GeneralarXiv:2608.23552

    Prime Agent: A Self-Improving RLM Harness

    • It maintains an IPython REPL session and continuously stores execution history, memories, skills, and subagent definitions in the harness so that experience accumulates.
    • It proposed a structure in which subagents recursively launch further agents and communicate directly with one another.
    • It raised Best@1 on the ARC-AGI-3 RHAE benchmark from 30% to 95.5%.
    • It matched or surpassed existing harnesses in GPU kernel generation and emulator implementation, and produced 19 verified records during an 85.5-hour nanoGPT speed-improvement run.
    • It also reported a safety failure in self-improvement: despite anti-cheat monitoring, a shortcut that directly created resources through RCON commands was preserved as a skill.
    概要PDF
  45. 988AI / GeneralarXiv:2608.23283

    Apodex 1.1: Scaling Agentic Intelligence for Complex Work

    • It trained the model's 'ability to do work' using a variety of verifiable task environments involving files, search, code, and more.
    • It treated multi-agent coordination, involving decomposing problems, delegating, combining asynchronous results, and replanning, as a separate training axis.
    • It scored 54.3 on FrontierFinance for financial practice, 63.3 on FrontierScience-Research for scientific research, and 78.8 on GDPVal for work deliverables, placing it in the leading performance range described by the paper.
    • In contrast, its pass rate on difficult scientific workflows in its own FrontierResearchBench was only 12.4%, trailing leading systems. The paper also states that completing such workflows end to end remains difficult for every system evaluated.
    • A 35B-parameter Mini version preserved a substantial share of practical task-execution capability in a form that can run locally, opening a path to smaller-scale deployment in the field.
    概要PDF
  46. 987Machine LearningarXiv:2608.21308

    Rethinking Expressivity and Efficiency in Test-Time Training

    • It derived a closed-form state-transition equation that enables chunk-parallel training while retaining per-token learning rates, momentum, and decay.
    • It demonstrated long-context extrapolation by maintaining over 90% accuracy on a needle-in-a-haystack test with inputs eight times longer than those used in training.
    • Across 14 LongBench tasks, its average performance exceeded that of subquadratic-compute architectures such as Mamba2 and LaCT.
    • When attached as a parallel branch to Qwen3VL-2B, training only the TTT component achieved video-understanding performance comparable to full fine-tuning.
    • Because a single set of fast weights is shared within each chunk, some fine-grained causal relationships are still missed, and experiments were limited to a maximum of 1.3 billion parameters, requiring validation on larger models.
    概要PDF
  47. 986Distributed SystemsarXiv:2608.21157

    HIERA: Workload-Aware Planning Across Implementation Spaces for GPU Kernel Optimization

    • It introduced hierarchical planning that first selects an implementation space among PyTorch operators, CUDA libraries, and handwritten CUDA.
    • Within the chosen space, it iteratively improves implementations using task contracts, profiling feedback, and expert knowledge, finding executable code even with a small candidate budget.
    • It ranked best or tied for best in 22 of 27 comparisons across three base models and three difficulty levels, and achieved a valid-implementation rate of 71.6% with a single candidate.
    • In a scientific-computing stencil case, it found a kernel 1.53 times faster than cuDNN after five iterations, showing potential beyond standard training operations.
    • Evaluation focused on NVIDIA A100 and limited precision settings, and stencil validation covered only one operator and one configuration, so generalization to other devices and scientific operations remains unverified.
    概要PDF
  48. 985Computer VisionarXiv:2608.21360

    OmniAssistBench: Assistant-style Interaction Benchmark for Omni-LLMs

    • To reduce the problem of user behavior branching according to model responses, it predefined interaction paths from the original videos and evaluated models along those paths.
    • By reverse-engineering internet videos and investing at least 1000 hours of expert work, it constructed seven tasks, sixteen subtasks, and three real-world cases.
    • The best model, Gemini-3-Pro, scored only 66.4 out of 100, while the open model Qwen3-Omni-Instruct scored 51.2.
    • Understanding hand-gesture instructions, long-term memory, delaying responses until a target event, and maintaining goals across multiple conversation turns emerged as shared bottlenecks.
    • Because interaction paths were fixed in advance to enable static evaluation, it cannot reproduce every conversational path that branches as real users and models make free choices.
    概要PDF
  49. 984Machine LearningarXiv:2608.20818

    Scaling Muon for Diffusion Transformers

    • It found a tendency for Muon to improve generation-quality metrics by 12.9–19.1% over AdamW in diffusion transformers with 1.3 billion–15 billion parameters.
    • It proposed Periodic Row-wise Muon, which performs a full spectral update only every three steps and replaces the remaining updates with inexpensive row-wise updates.
    • Optimizer time fell by 46.9–54.3%, total training-step time by 15.7–24.3%, and logical communication volume by 66.7%.
    • On the 9-billion-parameter model, generation quality improved by a further 4.5% over the original Muon, demonstrating both cost reduction and quality gains.
    • The experiments were limited to one diffusion-model family, one dataset and resolution, and a 32-node H100 configuration, and full momentum communication remains on update steps.
    概要PDF
  50. 983Computer VisionarXiv:2608.20974

    WA-JEPA: Rethinking the Video JEPA Paradigm for World-Action Modeling in Autonomous Driving

    • It replaced random-mask reconstruction with pretraining that masks future segments, aligning the video-representation learning objective with the temporal direction of autonomous-driving planning.
    • Instead of deterministic regression, it uses flow matching to generate multiple possible future representations and jointly learns scenes and vehicle trajectories in a single predictor.
    • It achieved 91.7 EPDMS on NAVSIM-v2, exceeding strong end-to-end and world-action baselines by 1.6 and 1.3 points, respectively, and also achieved the highest score, 0.4462, in HUGSIM closed-loop evaluation.
    • It presents a different design direction for autonomous-driving planning, in which visual future prediction directly supports action learning without going through language reasoning.
    • Validation focuses on nuPlan pretraining and the NAVSIM and HUGSIM benchmarks, so safe generalization to real vehicles and rare hazardous situations requires further verification.
    概要PDF