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. 782LLMs / NLParXiv:2607.27652

    Harness-G: A Graph-Structured Harness for Search Agents

    • It finds a retrieval-equivalence collapse phenomenon, where different queries increasingly retrieve the same information as training progresses.
    • It replaces free-form query generation with graph-based finite action selection over Select, Lookup, and Answer.
    • Structured Non-myopic Credit, or SNC, assigns credit by comparing alternative actions from the same state.
    • It improves average F1 by +10.74 points for Graph-R1 at 1.5B and +3.98 points at 3B, reaching the best score on six QA benchmarks.
    • Graph construction costs $0 when done programmatically, far cheaper than LLM-based graph construction, which costs $2.81 to $4.14.
    概要PDF
  2. 781AI / GeneralarXiv:2607.28336

    Correcting What You Cannot See: Credit Assignment for Perception Distillation in Multimodal Reasoners

    • It formalizes the credit-assignment problem in multimodal distillation by separating perception failures from reasoning failures.
    • A soft AND gate, formed by multiplying the sub-failure rate, or 1-PSR, and the teacher-student mismatch, or KL, increases correction strength only when both signals are present.
    • It separates perception and reasoning rollouts, distilling only the perception segment while leaving the reasoning segment to RL.
    • When moving from 32B to 8B, the average score across 8 benchmarks rises from 56.94 to 61.22, and when moving from 8B to 2B it rises from 44.50 to 47.28.
    • Removing the PCD weight lowers performance by 2.22 points, and removing the separated rollout lowers it by 0.88 points, confirming the ablation.
    概要PDF
  3. 780Machine LearningarXiv:2607.27372

    Explorative Modeling: Unlocking a Third Pretraining Axis and End-to-End Generation

    • It discovers a third scaling axis in generative model training called exploration. In addition to parameters and data, more exploration improves performance.
    • It improves FLOP efficiency by 4.1x, sample efficiency by 6.2x, and parameter efficiency by 47%. The effect grows with data size, from 7% to 36%.
    • On ImageNet 256x256, it achieves an FID of 1.43 without guidance, near state of the art.
    • It enables end-to-end generative models, where inference matches training, and gets similar performance with 16x to 256x fewer steps than diffusion.
    • It observes that as exploration increases, generalization also improves.
    概要PDF
  4. 779AI / GeneralarXiv:2607.28227

    Qwen-UI-Agent Technical Report: Toward Next-Generation Real-World Centric Foundation GUI Agents

    • It trains and evaluates in a real-device mobile environment with more than 100 actual smartphones and more than 150 apps.
    • It uses a mixed GUI and CLI action space, which allows multiple actions to be batched in one turn, and more than 40% of the actions are batched.
    • It uses online RL for tasks that run for more than 100 turns, and it accelerates rollouts with 10,000 concurrent environments.
    • It reaches 82.1% on MobileWorld, 79.5% on OSWorld-Verified, and 73.6% on WebArena.
    • It supports notification-driven spontaneous service starts and cross-platform workflows between mobile and PC.
    概要PDF
  5. 778Computer VisionarXiv:2607.28627

    ReToken: One Token to Improve Vision-Language Models for Visual Retrieval

    • It finds that the attention, or query-key, scores of a VLM are poorly suited for visual information retrieval, with an average recall@1 of only 5.1 percent.
    • It adds a single learnable token, ReToken, to explicitly retrieve relevant frames in value space.
    • On Visual Haystacks, it improves Qwen3VL-8B by 13.4 points and InternVL3.5 by 12.4 points, which is more than a 20 percent relative gain.
    • Even when trained only on image QA, it transfers zero-shot to long videos and improves LVBench by 8.0 points.
    • It is lightweight enough to train and run on a single H100, and the added retrieval latency is about 0.4 seconds.
    概要PDF
  6. 777Computer VisionarXiv:2607.27380

    VideoCoCo: Code-as-CoT for Physically-Consistent Video Generation via an Agentic Dual-Engine System

    • It uses executable code, Blender, as a process-level thought chain to preserve physical consistency.
    • The code agent programs the scene and a simulator renders a draft video, and then the video editor composes the final result realistically.
    • It builds the VideoCoCo-3K dataset, which contains draft, instruction, and target pairs, for editor adaptation training.
    • It improves PhyGenBench from 0.475 to 0.558 and VBench-2.0 from 52.18% to 77.88%, which are both state of the art.
    • Without tuning, the draft alone improves from 0.475 to 0.506, and LoRA tuning raises it further to 0.558, which separates the contribution of each stage.
    概要PDF
  7. 776Computer VisionarXiv:2607.28625

    ACE-Data-0: Human-Centric Ambient Capture as Embodied Data Engine

    • The ACE system turns real home environments into recording studios, with both table-scale and room-scale settings.
    • It is a large dataset with 150 hours, 17 million frames, 75,000 episodes, 50 participants, and 200 task categories.
    • It synchronizes egocentric and multi-view exocentric video, full-body and hand motion, 6-DoF object trajectories, audio, and touch at the millisecond level.
    • It supports zero-shot motion transfer from simulation to the real world and exposes the limitations of more than 30 existing methods through benchmarking.
    • It can serve as training data for imitation learning, world models, and VLA systems.
    概要PDF
  8. 775Computer VisionarXiv:2607.28611

    Chimera: Designing and Chinchilla-Scaling Hybrid Visual Diffusion Transformers

    • It designs a hybrid diffusion backbone that combines KDA, linear attention, MLA, global attention, and unit consistency for local patterns.
    • Without positional embeddings, it processes text, images, and video in a single stream using only raster order.
    • HeteroP automatically transfers hyperparameters across scales based on functional fan-in for each tensor.
    • With 11B total parameters and 2B active parameters, it achieves 7.3x better compute efficiency than Wan-2.1 at 2B, and training takes about 600 H100 days.
    • It is trained on 5-second clips and can extrapolate to 30-second video generation in zero shot, with only a 6 percent FID drop at 6x longer length.
    概要PDF
  9. 774LLMs / NLParXiv:2607.26760

    Metis: Memory Foundation Model

    • It is the first prototype to implement native memory state inside model weights instead of relying on external RAG.
    • Memory storage, forgetting, and updating are all performed automatically by the model's forward pass, which enables gradient-free updates.
    • It introduces a hyper-memory block and a local memory block inspired by Fast Weight Programming.
    • At 128K context, it is 2.4 to 2.67 times faster than full context and uses 67 times less storage per session, 16.79 MB versus 1118 MB.
    • Its limitation is compression loss on long tasks, and in some cases it also causes semantic confusion.
    概要PDF
  10. 773ResearcharXiv:2607.27197

    The Maxwell Conjecture is False

    • It refutes Maxwell's 1873 conjecture that the number of critical points of n point charges is at most (n-1)^2 by giving a counterexample with 5 charges.
    • A configuration of an equilateral triangle plus two axial charges produces 21 critical points around the center, for a total of 24 nondegenerate critical points.
    • By iterating the construction, 3 plus 2m charges can produce at least 4 plus 20m critical points, which is a ratio of 10 critical points per charge.
    • The idea was suggested by GPT-5.6 Sol and then verified in Mathematica and Maple before the authors wrote the mathematical proof.
    • The construction improves the previous best ratio of 25/7, or about 3.57, to 10.
    概要PDF
  11. 772Computer VisionarXiv:2607.28624

    PhiZero: A World Model Built Around Physical Language

    • We discovered a discrete compressed representation called a physics language in natural videos through self-supervised learning. It symbolizes patterns of state change rather than pixels.
    • A reasoning-then-rendering architecture explicitly predicts physical evolution and then synthesizes video with a diffusion model.
    • It substantially improves physical consistency over prior world models on video generation and physics-understanding benchmarks.
    • It demonstrates zero-shot action transfer by transferring human motions to humanoid robots or robotic hands.
    • It learns physical common sense from video data alone, without physics laws or simulators.
    概要PDF
  12. 771Computer VisionarXiv:2607.28595

    Beacon: Knowing When and How to Perform Agentic Visual Reasoning

    • It defines two problems in tool use: Mode Adaptiveness, knowing when a tool is needed, and Tool Effect, knowing whether the tool actually helps.
    • Existing models gained little from tools because the gains on hard problems were offset by errors on easy ones.
    • The Necessity-Aware Adaptive Reward (NAAR) reduces unnecessary tool calls, and Hint-Conditioned Exploration (HCE) strengthens tool use on hard problems.
    • It achieves a best-in-class average of 58.98% across 13 benchmarks, a +6.07-point gain over the base model, and the largest tool gain-minus-gap improvement at +3.14 points.
    • It recovers about 40% of the all-wrong group through hint rollouts, effectively reusing the learning signal.
    概要PDF
  13. 770AI / GeneralarXiv:2607.28272

    MemHarness: Memory Is Reconstructed, Not Replayed

    • It applies the human memory principle of reconstruction to AI agents, meaning that past experiences are critically revised and edited to fit the current situation.
    • It uses GRPO reinforcement learning to let the agent learn memory reconstruction naturally, without extra human labeling.
    • It achieves 85.2 percent on ALFWorld and 75.6 percent on WebShop, outperforming both pure RL and static-memory methods.
    • When the memory reconstruction module is removed, performance drops to 74.6 percent, confirming that reconstruction itself is the key contribution.
    • Learning to reconstruct memory during training also improves base reasoning ability at inference time even without memory, which the paper calls a latent guidance effect.
    概要PDF
  14. 769AI / GeneralarXiv:2607.27191

    Can AI agents conduct open-ended AI research? Early evidence from two case studies

    • It proposes a new method called shadow evaluation, in which AI is given the research questions from unpublished papers and the original authors do the grading.
    • The same failure patterns were reproduced for both Opus 4.8 and GPT-5.6 Sol.
    • The agents were good at engineering work, such as coding and running experiments, but they lacked strong judgment and creativity in research design.
    • The five core failures were weak research-level standards, failure to back away from dead ends, poor resource awareness, non-creative responses to feedback, and drift away from the instructions.
    • They used less than half of the API budget, still had time left over, and never received a pass in AI self-review.
    概要PDF
  15. 768LLMs / NLParXiv:2607.28568

    Frontis-MA1: Training an AI4AI Model towards Recursive Self-Improvement in Machine Learning Engineering

    • It builds the full-stack OpenMLE system for AI-for-AI improvement and creates 5,758 executable task environments.
    • It trains the meta-evolution agent Frontis-MA1-35B by using reinforcement learning to learn four evolutionary operators: drafting, improving, debugging, and crossover.
    • On MLE-Bench Lite, the medal average rises from 39.4 percent to 71.2 percent, which beats GPT-5.5 plus Codex at 68.2 percent.
    • A separate benchmark, NatureBench Lite, also confirms transfer learning, which shows that the result is not simple benchmark memorization.
    • It releases the model weights, task environments, and training code, providing a reproducible basis for recursive self-improvement research.
    概要PDF
  16. 767LLMs / NLParXiv:2607.25852

    AngelSpec: Towards Real-World High Performance Inference with Speculative Decoding

    • The authors empirically show that the best drafter depends on the task type in real service settings, such as chat versus code or math.
    • DFly improves block diffusion quality with a hybrid target-conditioned backbone and an autoregressive head conditioned on predecessors.
    • D-cut treats verification as a shared batch-level resource and adjusts verification depth adaptively based on load.
    • On Hy3-A21B, it increases average accepted length by about 30 percent and achieves the highest throughput across concurrency levels from 4 to 64.
    • It records 1.98 to 2.40 times the throughput of autoregressive decoding and 10.5 to 11.8 percent higher throughput than DFlash.
    概要PDF
  17. 766RoboticsarXiv:2607.26991

    RL$^2$-VLA: Adaptive RL Latent Compositional Steering with Test-Time Scaling for Vision-Language-Action Models

    • It applies compositional steering with a lightweight RL policy learned on top of VLA latent representations.
    • It finds different scaling laws in success and failure states, and the optimal intervention is to act only on failures.
    • It integrates a SAFE failure detector so that RL steering is activated only when failure is predicted.
    • It achieves up to a 17.3% success-rate improvement on the out-of-distribution SIMPLER and PolaRiS benchmarks.
    • It also transfers to the real PiperX robot, where it improves success by 17.5%.
    概要PDF
  18. 765Computer VisionarXiv:2607.27110

    FreqForcing: Autoregressive Long Video Generation via Spectral Self-Anchoring

    • It analyzes autoregressive video-generation error accumulation in the frequency domain and shows that low-frequency energy drift is the root cause.
    • Spectral Self-Anchoring, or SSA, fuses low-frequency content from the first frame with high-frequency content from the most recent frame to preserve both stability and dynamics.
    • Without extra training, it can extend a 5-second training model to stable generation for up to 2 minutes, which is a 24x horizon expansion.
    • On the VBench-Long metric, it outperforms previous training-free methods and remains competitive with training-based methods.
    • It adds only about 16.5% inference-latency overhead over Self-Forcing.
    概要PDF
  19. 764Computer VisionarXiv:2607.27194

    VidMap: Exploiting Temporal Structure for Video-Based Structure-from-Motion

    • We combine SLAM's sequential tracking with SfM's global optimization to overcome the limitations of both.
    • We use temporal order as the primary signal to prevent visual aliasing errors.
    • We add a monocular depth prior to global optimization to address scale drift and degenerate configurations.
    • We outperform existing SLAM, SfM, and learning-based methods by a wide margin on a challenging dataset with extreme motion.
    概要PDF
  20. 763AgentsarXiv:2607.26598

    Living-Harness Is an Interactive-Agent Evolver

    • It turns agent failures into persistent procedural repair rather than one-off lessons and stores them permanently.
    • It accumulates knowledge along two axes: episode memory, including trigger conditions, failure patterns, and recovery actions, and a state graph, including state nodes, repair edges, and transition rules.
    • It structures what to change and how to change it through a domain-level guide called Evolution-SOP.
    • It improves by an average of 10.07 points on tau^2-Bench and 9.91 points on MultiWOZ-2.4.
    • Even when the harness state evolved by GPT-5.2 is transferred to other models such as Gemini 3 Pro, GLM-5, and Qwen3-max through retrieval alone, performance still improves.
    概要PDF
  21. 762RoboticsarXiv:2607.26055

    $π\mathbf{R}^2$: Reactive Real-time Flow Policies

    • It separates joint-state information in a fast channel and vision-language information in a slow channel asynchronously, so it can respond in real time without waiting for vision latency.
    • A latency-adaptive noise schedule ensures smooth behavior regardless of hardware delay.
    • On a real robot based on GR00T-N1.7, with xArm6 plus XHand, it achieves 25 Hz recomputation, about 4x faster than before.
    • On contact-rich real-world tasks such as organizing books, inserting boxes, and catching falling objects, it improves success by up to 30%.
    • It can be applied with only minimal changes to existing VLA architectures, which makes it practical.
    概要PDF
  22. 761Machine LearningarXiv:2607.26246

    Weak-to-Strong On-Policy Distillation

    • It proposes a new distillation method that separates the capability direction from the logit difference of a weak model pair, positive and negative, and adds it to a strong student.
    • It proposes three settings: RL before and after model pairs, model pairs of different sizes, and answer versus hint pairs.
    • Across four math and three coding benchmarks, it achieves relative improvements over existing OPD of 11.4% in math and 12.0% in coding.
    • Even though it learns only from weak models, it outperforms the actual weak teacher models themselves.
    • It found that the RL-based pair and the size-based pair provide complementary signals, strengthening the reasoning skeleton and the solution procedure, respectively.
    概要PDF
  23. 760ResearcharXiv:2607.27011

    Qwen-Audio-3.0-Gen-Preview Technical Report

    • It unifies speech, music, sound effects, and mixtures into a single generation path with a non-autoregressive structure.
    • It designs a shared VAE that compresses 48 kHz stereo into a 25 Hz latent representation.
    • It uses only about 10 percent of the music data used by specialized prior models, yet still leads in 3 of the 7 SongBench tasks.
    • On Seed-TTS-Eval, it achieves the highest speaker similarity, and it also outperforms Seed-Audio-1.0 in multi-speaker consistency.
    概要PDF
  24. 759Computer VisionarXiv:2607.27180

    HumanCLAW: Can Vision-Language Models Act Through a Body?

    • We built a framework that separates behavioral decision making in VLMs from motor control, so it evaluates reasoning ability alone.
    • We built HumanCLAW-Bench, consisting of 1,218 egocentric exploration and interaction episodes across 41 indoor scenes.
    • Even the best of nine state-of-the-art VLMs reaches only a 16.8% success rate.
    • The core failure is not perception but a lack of self-awareness, meaning the model does not know where its body is, whether it has arrived, or whether it has collided.
    • The conclusion is that current VLMs only describe the world; they do not feel the body they control.
    概要PDF
  25. 758ResearcharXiv:2607.27199

    Settling the Optimal Exponent Relating Sumsets and Difference Sets

    • It is the first proof that the optimal universal exponent relating sumsets and difference sets is exactly 2.
    • An AI agent, Hyra, discovered the key structure through 24 hours of autonomous search, and a human then verified it rigorously in mathematics.
    • It also provides a Lean 4 formal proof, which makes the result machine-checkable.
    • The core construction combines base-12 digit techniques, a three-state carry automaton, and the Chinese remainder theorem.
    概要PDF
  26. 757Computer VisionarXiv:2607.27205

    TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM

    • It proposes a direct vision-plus-language-to-action approach by removing the LLM-centered architecture.
    • With 0.2B parameters, the model is only 6% the size of π0.5, which has about 3.4B parameters, yet matches or exceeds its performance with a 97.7% success rate on LIBERO.
    • It achieves 32 Hz inference, 31.2 ms, on a consumer RTX 4090 and uses less than 1 GB of VRAM, making edge deployment possible.
    • It outperforms π0.5 consistently on four real-robot tasks with an AgileX Piper.
    • Removing language drops the success rate to 70.8%, proving that language conditioning is still important even at the execution level.
    概要PDF
  27. 756RoboticsarXiv:2607.25918

    DC-WAM: Dynamic-Centric Visual Supervision and Reasoning for World-Action Models

    • We redistribute visual supervision from appearance reconstruction to interaction-centered dynamics in RGB video prediction, so PSNR drops but policy performance improves.
    • A temporal difference flow-matching objective and tracker-guided weighting focus on robot-object contact motion rather than static background and texture.
    • The DynaRoute attention bias predicts dynamic-relevant tokens and steers visual attention toward them.
    • On LIBERO-Plus OOD, it reaches 60.9%, a 9.4-point gain over FastWAM, and reduces the ID-OOD gap from 46.1 to 37.2.
    • On real robot bimanual tasks, it achieves up to a 30-point improvement under lighting changes.
    概要PDF
  28. 755Computer VisionarXiv:2607.25537

    Visual prompt engineering for video models

    • It systematically proposes and validates visual prompt engineering, or VIPE, for video models.
    • On Veo 3.1, converting sketches into photorealistic scenes raises physical reasoning accuracy from 41.3 percent to 59.3 percent, a gain of 18 points.
    • It uncovers a realism bias, meaning video models reason much better on realistic scenes, so abstract benchmarks underestimate their ability.
    • VIPE is often more effective than test-time scaling methods such as text-prompt optimization or self-consistency.
    • Stepwise experiments confirm that scene consistency improves whenever realism is increased by one level.
    概要PDF
  29. 754Computer VisionarXiv:2607.25389

    HOME: Robust Hough-space Matching Method for Structured and Textureless Videos

    • It uses the maxima in Hough space as keypoints to achieve stable feature-point detection on straight-line structures.
    • A one-dimensional radial descriptor mathematically guarantees rotation and translation invariance, which removes the need for separate orientation estimation.
    • It is training-free and outperforms line-segment matching methods based on LSD and LBD on every metric.
    • It runs in real time at 16.5 ms total, and although the Hough transform takes most of the time, 7.7 ms, it can be improved with GPU acceleration.
    • It remains robust even on nearly textureless images with alpha equal to 0.95, finding more than 400 inlier matches.
    概要PDF
  30. 753Computer VisionarXiv:2607.26004

    Parallel Decoding Distillation for Fast Image and Video Generation

    • It trained a 'parallel decoder' that predicts multiple denoising steps in parallel from a single forward pass, speeding up generation.
    • It solved the mode collapse problem by using only a simple regression objective, without VSD or GAN losses.
    • On Wan2.1 14B text-to-video, it ranked first on VBench and outperformed the baseline on diversity metrics.
    • On Qwen-Image 20B, it achieved SOTA scores with only 4 to 8 NFE.
    • On the LTX-2.3 22B model, it can generate a 720p 10-second video with audio in only 8 NFE.
    概要PDF
  31. 752Computer VisionarXiv:2607.24904

    Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model

    • The codec-native tokenizer, Mage-ViT, reduces visual tokens by more than 75% by encoding only dynamic patches.
    • Trained from scratch on just 560M unlabeled images, it matches SigLIP2, which uses billions of image-text pairs, demonstrating pretraining data efficiency.
    • A dual-system design, with a System 1 event gate for fast reaction and a System 2 causal decoder for deep reasoning, supports real-time streaming understanding.
    • With 4B parameters, it matches Qwen3-VL-4B on static tasks and surpasses even the 15B Phi-4 model on video understanding and 2D/3D spatial reasoning.
    • It achieves up to 3.5x inference speedup and reports seven key empirical findings.
    概要PDF
  32. 751RoboticsarXiv:2607.26056

    INTACT: Isomorphic Intent-to-Action Learning for Search-Free World Models

    • It integrates a JEPA prediction architecture with an intent-to-action interface so it can generate actions directly without retrieval.
    • It processes physical next states and target intents with one shared predictor, while designing the endpoint gradient asymmetrically.
    • In direct mode without retrieval, it reaches 85.78% to 100% success on four LeWM tasks, with 2.9 to 5.5 ms inference time.
    • With selective CEM search, it achieves 96.86% macro success while using 23.44x fewer samples than pure CEM.
    • Joint training with a single encoder over four tasks yields 89.39% direct macro success.
    概要PDF
  33. 750LLMs / NLParXiv:2607.26057

    Pass the Baton: Trajectory-Relayed On-Policy Distillation

    • It discovers a teacher-student asymmetry in direction switching: the teacher reflects and changes direction, but the student just keeps going straight.
    • It experimentally confirms that even 0.35% teacher tokens can raise accuracy by +7.23 points.
    • It limits the number and length of teacher interventions through a relay budget so the student policy does not drift too far away.
    • It cuts the average training trajectory length by more than 50%, from 4,658 to 2,296 tokens, while still improving accuracy.
    • A single regression objective makes stable training possible without VSD or GAN losses.
    概要PDF
  34. 749Computer VisionarXiv:2607.25948

    MODUS: Decoder-Only Any-to-Any Modeling of Diverse Modalities

    • It is the first model to process 15 modalities with a single decoder, without modality-specific heads, losses, or task pipelines.
    • Unlike prior any-to-any models, it uses a large pretrained decoder, BAGEL, to greatly reduce training cost.
    • Uniform timestep sampling avoids modality confusion and enables stable training.
    • It naturally supports chained generation, converting through intermediate modalities, and cross-modal self-verification, checking its own outputs through other modalities.
    • It releases the 29-million-sample MODUS-DATASET and two checkpoints as open source.
    概要PDF
  35. 748RoboticsarXiv:2607.25895

    HiFi-UMI: Learning Deployable Manipulation Policies from High-Fidelity UMI Data Alone

    • It overcomes the fidelity limits of prior UMI data with 3 mm trajectory accuracy, microsecond synchronization, and a field of view of about 200 degrees.
    • Zero-robot post-training means it is trained only on HiFi-UMI data and can then be deployed directly on real robots, which is validated on all three backbones.
    • The strongest policy reaches an 85 percent success rate on precision insertion tasks, even though HiFi-UMI contains no trajectories for those tasks.
    • With 4,000 hours of pretraining, action error decreases by 41 percent on 10 unsupported tasks, and one backbone gains an additional 18.1 percentage points in success rate.
    • HiFi-UMI-2K, a high-quality dataset containing 2,000 hours of microsecond-synchronized data, was released as open source.
    概要PDF
  36. 747ResearcharXiv:2607.24136

    How to read Mathematics? A study guide

    • Mathematics should be read in terms of time per page, not pages per time, so read slowly, very slowly.
    • When reading a definition, you must create examples and counterexamples. A definition without examples, and an example without a definition, is not enough.
    • When reading a proof, first skim it, then follow it line by line, and finally summarize the key idea.
    • Practice problems should be looked at as early as possible, and the problem itself should be read like a mathematical text.
    • It warns against overreliance on AI, citing EEG research that suggests heavy dependence on LLMs weakens brain connectivity.
    概要PDF
  37. 746Software EngineeringarXiv:2607.25886

    RSIBench-Data: Benchmarking Data-Centric Research for Recursive Self-Improvement

    • It is the first benchmark that fixes the training, evaluation, and serving infrastructure so that only pure data research ability is measured.
    • Among four frontier agents, including Codex and Claude Code, none is consistently dominant. Codex wins on non-SWE tasks, while three SWE tasks are won by different agents.
    • After feedback, 78% of runs still drop in score on the final attempt even after reaching a peak, which reveals a discovery-reliability gap.
    • Increasing max effort helps find stronger candidates earlier, but it reduces the number of exploration attempts, which reveals a depth-width trade-off.
    • A self-improvement experiment on the same model family, kimi-k2.6, improves from 8% to 21% but still falls short of the original 33%.
    概要PDF
  38. 745Computer VisionarXiv:2607.26037

    Wonder: Video World Model Done Better

    • It converts camera motion into rendered visual cues rather than abstract data, which prevents control accuracy from collapsing after distillation.
    • A sparse full-fidelity memory preserves every past frame while keeping attention cost constant.
    • A mixture-of-students over time and camera-aware adversarial regularization preserve both visual quality and control accuracy during distillation.
    • It achieves meaningful improvements over the strongest baseline on rotation position error, going from 0.0174 to 0.0132 for translation and from 0.1155 to 0.0784 for rotation.
    • It also supports video input, so it can re-shoot existing videos by changing the camera path in real time.
    概要PDF
  39. 744Machine LearningarXiv:2607.24645

    Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

    • It measures the geometry of how output logits change when a feature is removed.
    • Most features do not provide one stable control direction.
    • Value-like features more often show low-dimensional structure, but their effects are spread across multiple directions.
    • Pointer-like features produce mostly scattered effects.
    • The paper compares ReLU, TopK, and Matryoshka Batch TopK on Gemma-2-2B.
    • Even when a feature is interpretable and causal, that does not guarantee stable control.
    概要PDF
  40. 743AI / GeneralarXiv:2607.24280

    From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search

    • It uses commercial models only as offline teachers, so inference cost is zero.
    • The structured JSON convention created by the multi-agent system links teacher and student.
    • It strips away surface style to reduce stylistic contamination and hallucination.
    • It achieves average success rates of 39.4% with Qwen3-1.7B and 44.4% with Qwen3-4B.
    • Claude, GPT, and Gemini teachers are all within 2 points of each other, so API swapping is easy.
    • Convention synthesis is a one-time cost of about $1,454.
    概要PDF
  41. 742LLMs / NLParXiv:2607.24223

    A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

    • It redefines relevance as the order in which grep explores the corpus, rather than as a top-k document selection problem.
    • Document ranking provides global relevance, while match reranking provides local relevance.
    • It seeds the retrieval start points with question-matched passages so that decisive evidence is found earlier.
    • On GPT-5.4-mini, it reaches 84 percent accuracy with only 23.9 tool calls, which is far fewer calls than competing methods.
    • It retains 79 percent accuracy even when scaled to one million documents, showing only a mild performance drop.
    概要PDF
  42. 741Information RetrievalarXiv:2607.24617

    LaRec: Unleashing LLM-based Latent Reasoning for Generative Recommendation

    • It greatly reduces response latency by reasoning in latent space instead of through long text reasoning.
    • During latent pretraining, step-level alignment and process direction alignment provide rich supervisory signals to intermediate reasoning steps.
    • A user-specific Gaussian mixture distribution is used to sample the starting point of reasoning within the relevant latent space.
    • GRPO is combined with a sparse-hit reward and a dense semantic-similarity reward to align reasoning with the recommendation objective.
    • The latency of latent reasoning is close to no-reasoning and much lower than chain-of-thought.
    • An online A/B test on a real advertising platform records improvements such as a 2.93% conversion rate.
    概要PDF
  43. 740RoboticsarXiv:2607.23969

    LeapBot-WA: World-Anchor Action Models via Predictive Latent Alignments

    • It shifts from pixel-level video reconstruction to predicting abstract physical changes in latent space.
    • It uses a co-embedding prediction structure as a world anchor, focusing on state changes rather than visual appearance.
    • An isotropic semantic autoencoder bridges the mismatch between the predicted features and the diffusion prior.
    • The heavy prediction branch is used only during training and is removed at inference, so it runs without extra cost.
    • It achieves state-of-the-art results on RoboTwin 2.0 at 91.46 percent and on LIBERO at 97.3 percent without large-scale trajectory pretraining.
    • Isotropic regularization raises latent rank from 38.1 to 92.3 and improves success rate to 71.3 percent.
    概要PDF
  44. 739RoboticsarXiv:2607.24159

    DeVA: Decoupled Video-Action Model with physical guidance for robot policy learning

    • The method separates video prediction and action prediction into different experts so that policy learning becomes more manageable.
    • It passes representations from multiple layers of the video backbone to the action expert through cross-attention and bridge tokens.
    • It injects auxiliary supervision for object-manipulation regions and relative depth, which strengthens physical information.
    • With 50 demonstrations per RoboCasa task, it reaches a 72.0 percent success rate and improves by up to 22 points over the same budget.
    • On three real-world dual-arm robot tasks, it averages 74 percent and outperforms GR00T-N1.6 at 48 percent and Cosmos Policy at 34 percent.
    • It converges faster with up to 20 times fewer training examples than Cosmos-Policy.
    概要PDF
  45. 738ResearcharXiv:2607.24528

    On Feige's conjecture

    • It constructed a short proof by connecting Vlassis and Thomas's distribution-free p-value result to the random-variable-sum problem in Feige's conjecture.
    • It showed that the probability that the sum of n independent nonnegative random variables, each with expectation 1, is less than n+1 is at least (n/(n+1))^n.
    • Because this lower bound is at least 1/e, it establishes Feige's conjecture, posed in 2004.
    • It shows how resolving Gaffke's conjecture in statistics can lead to a proof of a probability inequality, and GPT-5.6 Sol assisted the proof search.
    • However, the central theorem focuses on Feige's conjecture with threshold n+1, rather than resolving the full range of more general small-deviation parameters.
    概要PDF
  46. 737ResearcharXiv:2607.23980

    Sharp small-deviation inequalities for sums of independent nonnegative random variables

    • It combined the exact Dirichlet calibration theorem of Vlassis and Thomas with results from convex geometry, including Grünbaum's centroid theorem.
    • Treating 0<δ<1 and δ≥1 separately, it proved explicit lower bounds on small-deviation probabilities for sums of independent nonnegative random variables.
    • For δ≥1, the lower bound is optimal for every n and is at least 1/e, settling Feige's conjecture.
    • The result provides a sharp reference for analyzing the lower tails of sums of independent nonnegative random variables.
    • However, optimality for every n is established only for δ≥1, and the lower bound in the range 0<δ<1 has not been proved equally sharp.
    概要PDF
  47. 736RoboticsarXiv:2607.24485

    τ: Learning Touch-Augmented Vision-Language-Action Models from Future Visual Supervision

    • It learns spatiotemporal tactile representations using a self-supervised objective that predicts future visual features from current touch and subsequent actions.
    • It combines the learned tactile representations with pretrained visual and language features to generate robot actions, while using the prediction branch only during training.
    • It outperformed vision-only VLAs and existing tactile policies on four contact-intensive manipulation tasks, including plug and USB insertion.
    • Removing the additional branch at inference allows it to improve contact-critical robot manipulation without increasing deployment cost.
    • Evaluation focuses on TacAura's four representative tasks and limited task-specific data, leaving validation across a wider variety of robots and contact situations outstanding.
    概要PDF
  48. 735AI / GeneralarXiv:2607.24459

    From Execution to Capability: Scientific Experience Consolidation via Procedural Knowledge Synthesis

    • It defines the problem of turning validated execution experience into reusable procedural knowledge and persistent model capability.
    • It extracts procedures by contrasting successes and failures, then keeps only the ones that pass development validation gates.
    • It increases training data by synthesizing new questions from failures even when no ground truth is available.
    • It empirically shows the gap between abstraction and execution, where weaker models cannot use the abstract procedures directly.
    • A stronger model concretizes the procedures into executable code and turns them into standard supervised training data.
    • The 9B student improves by 5.62 points on substeps and 11.25 points on full problems even without the original procedures.
    概要PDF
  49. 734Computer VisionarXiv:2607.23855

    OmniVAE: An Audio-Video VAE with Cross-Modal Alignment for Joint Generation

    • Training audio VAE and video VAE separately causes their latent spaces to diverge, making synchronization difficult.
    • OmniVAE is an audio-video VAE that finely aligns the two spaces through joint training.
    • It matches temporal semantics with segment-level contrastive learning.
    • It distills semantic features from a pretrained encoder to improve the trainability of each latent space.
    • Both objectives are used only during training, so inference cost does not increase.
    • On Verse-Bench, both synchronization, DeSync dropping from 0.884 to 0.570, and audio quality improve.
    概要PDF
  50. 733LLMs / NLParXiv:2607.24720

    The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation

    • Making the world model explicit as a chain of thought greatly improves long-horizon generalization, from 93.1 versus 0.3.
    • Atomic skills alone do not generalize compositionally, but adding only 5% long-horizon data causes a sharp jump.
    • Mixing in bad trajectories causes error accumulation and collapses long-horizon performance.
    • On-policy distillation has a wider effective range than GRPO under long-horizon and low-quality conditions.
    • Multi-teacher distillation converges on shared planning patterns, while conflicting patterns cause severe forgetting.
    • The experiments use full-parameter supervised learning based on Qwen2.5-100M.
    概要PDF