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. 582Computer VisionarXiv:2607.07534

    Infinite Worlds with Versatile Interactions

    • It applies an anti-drift training paradigm to causal video generation, enabling more than one hour of uninterrupted generation that is robust to accumulated error.
    • A real-time distilled model renders infinite worlds at 60 fps and 720p resolution, including a 1.3B version that can be deployed on a single GPU.
    • It offers rich interactions, including combat, archery, spell casting, shooting, and weather changes, with far more action variety than existing world models.
    • It integrates an agent harness composed of a Pilot agent for action planning and execution and a Director agent for generating environment content.
    • It uses a three-source data engine with an egocentric-video plus game plus Unreal Engine synthetic multi-granularity annotation pipeline.
    • It is fully open source, including code and checkpoints, and it is the only system among the compared models, such as Matrix-Game-3.0 and Genie 3, to achieve hour-level infinite generation.
    概要PDF
  2. 581Computer VisionarXiv:2607.07675

    Scaling Mixture-of-Experts Video Pretraining for Embodied Intelligence

    • The MoE, or sparse expert, architecture improves the trade-off between inference efficiency and model capacity relative to dense models.
    • The data profiling engine jointly analyzes, filters, and rebalances internet video and robot manipulation, navigation, and egocentric data.
    • A multidimensional reward system aligns not only aesthetic quality but also physical rationality and task-completion signals during training.
    • A single-stream Diffusion Transformer with Multi-Modal 3D RoPE unifies T2I, T2V, and I2V within one framework.
    • QK-Norm and AdaLN-Single modulation stabilize deep transformers, and a cascaded refiner improves quality.
    • It is the first large-scale open-source MoE video model for embodied intelligence, and the code and checkpoints are released.
    概要PDF
  3. 580Machine LearningarXiv:2607.05339

    TREK: Distill to Explore, Reinforce to Refine

    • It diagnoses a limitation of GRPO: the bottleneck is hard prompts for which the student never samples a reward trajectory.
    • The output-only interface requires only validated trajectories, not the teacher's logits or internal states, so it is compatible with black-box teachers.
    • Prompt routing queries the teacher only for prompts whose student pass rate pS(x) is less than or equal to tau_low, and it selects the top-r student-near trajectories.
    • In math tasks, Qwen3-8B improves from 36.9 to 40.3 on AIME 2025 and from 47.9 to 51.1 on AIME 2024, both averaged at 16 samples, and the self-context variant also improves meaningfully.
    • In agent tasks, it improves from 75.8 to 82.8 on ALFWorld and from 12.5 to 26.7 on ScienceWorld, which shows especially large gains on difficult tasks.
    • It returns to standard GRPO after the forward-KL integration stage, where the validated mode is further refined and strengthened.
    概要PDF
  4. 579Computer VisionarXiv:2607.06291

    AlayaWorld: Long-Horizon and Playable Video World Generation

    • It performs real-time 720p generation at 24 fps using LTX-2.3 fine-tuning, with four denoising steps per chunk for about one second of video.
    • Prompt switching updates the text condition at chunk boundaries to support open-ended actions.
    • A 3D cache plus history compression preserves spatial consistency and demonstrates leave-and-return consistency.
    • The error bank stabilizes long-horizon generation by reusing residual artifacts accumulated during rollout as structured perturbations.
    • Few-step DMD distillation minimizes visual latency, while short chunks minimize semantic latency.
    • The full stack, from data preparation to training, inference acceleration, and deployment, is modularized and planned for open-source release.
    概要PDF
  5. 578Machine LearningarXiv:2607.04763

    Multi-Turn On-Policy Distillation with Prefix Replay

    • It reuses teacher RL rollouts without environment interaction, making the cost zero and requiring no separate collection as a free by-product.
    • It identifies the prefix trap as both a temporal problem, caused by error accumulation, and a distributional problem, caused by the student's occupancy shifting away from the teacher's confidence distribution.
    • Step-decay sampling gives higher weight to early steps with smaller shifts, prioritizing more reliable regions of the trajectory.
    • On math and Python tasks, when the teacher-student gap is large, ReOPD outperforms OPD, while on retrieval environments it matches OPD.
    • It is more than 4x faster per training stage than OPD, uses zero tool calls, and is easier to integrate across heterogeneous environments.
    • Prefixes generated by the teacher model itself work best, while prefixes from stronger models can actually hurt performance.
    概要PDF
  6. 577Machine LearningarXiv:2607.05394

    Weak-to-Strong Generalization via Direct On-Policy Distillation

    • The key idea is that the log ratio between a post-RL weak model and a pre-RL reference is equal to the implicit reward of KL-regularized RL.
    • The RL results of a weak 1.5B model transfer to stronger models such as Qwen3-1.7B/4B and R1-Distill-7B, improving all of them.
    • Qwen3-1.7B's AIME 2024 accuracy rises from 48.3% to 62.4% with 8x A100s over 4 hours, matching direct RL with Polaris, which uses 32x A100s for a week.
    • By contrast, simply imitating the final policy of the weak model, that is, ordinary OPD, actually hurts the stronger student's performance.
    • Response length and the KL coefficient are the key control variables for transfer reliability, and adaptive KL stabilizes average reward.
    • Consistent improvements are observed across several teacher pairs, including R1-Distill to JustRL and Nemotron to QuestA.
    概要PDF
  7. 576Computer VisionarXiv:2607.06560

    Vision as Unified Multimodal Generation

    • A single UMM unifies structured visual understanding, dense geometry prediction, segmentation, and multi-view 3D, without task-specific heads.
    • Text generation handles symbolic outputs, such as detection, OCR, keypoints, and camera pose, while image generation handles dense predictions such as depth, normals, masks, and point maps.
    • SenseNova-Vision Corpus is a large training corpus that converts heterogeneous vision annotations into instruction-response examples.
    • It outperforms leading systems on structured visual understanding and approaches the best specialized systems on dense geometry and segmentation.
    • Task variations can be defined freely in language, including new combinations of color, region, and category that were not seen during training.
    • It opens the model and corpus, providing an open-source foundation for reproducible research.
    概要PDF
  8. 575AI / GeneralarXiv:2607.05147

    DSpark: Confidence-Scheduled Speculative Decoding with Semi-Autoregressive Generation

    • A semi-autoregressive architecture injects token-to-token dependence with a parallel backbone and a lightweight RNN head, while keeping per-token probabilities as exact softmax values.
    • Confidence-scheduled verification uses position-wise prefix survival probability estimates and a hardware-aware scheduler to adjust verification length dynamically.
    • On Qwen3-4B, 8B, and 14B, it improves accepted length by 30.9%, 26.7%, and 30.0% over Eagle3, and by 16% to 18% over DFlash.
    • On real traffic for DeepSeek-V4-Flash, it delivers 60% to 85% speedups per user over MTP-1, and 57% to 78% in V4-Pro.
    • It automatically shortens verification length in highly concurrent environments, preventing wasted batch capacity and expanding the Pareto frontier.
    • It releases the DSpark checkpoint and the DeepSpec training repository as open source.
    概要PDF
  9. 574RoboticsarXiv:2607.06403

    From Foundation to Application: Improving VLA Models in Practice

    • It builds about 60,000 hours of pretraining data, made up of 50,000 hours of robot trajectories and 10,000 hours of egocentric video, across 20 robot configurations.
    • It supports control of the full body, extending beyond the arms to the head, waist, mobile base, and active hands.
    • It introduces dual-query distillation for predictive dynamics modeling, using a video representation model as the semantic prior and a depth estimation model as the geometric cue.
    • On the GM-100 dual-arm benchmark, it achieves a 55.0% average success rate across four tasks using MeanStd normalization and L2 loss, with the best task reaching 76.8%.
    • The relative action target, relQpos, compresses the action scale to 0.34x compared with the absolute target and improves training stability.
    • It applies token-level adaptive expert routing with a MoE-based action expert structure.
    概要PDF
  10. 573AI / GeneralarXiv:2607.05391

    LLM-as-a-Verifier: A General-Purpose Verification Framework

    • It computes continuous scores as the expected value of the probability distribution over scoring-token logits, greatly reducing ties.
    • It shows three verification scaling axes, score granularity g, repeat evaluation k, and criterion decomposition c, each improve accuracy.
    • It reaches state of the art on Terminal-Bench V2 at 86.5%, SWE-Bench Verified at 78.2%, RoboRewardBench at 87.4%, and MedAgentBench at 73.3%.
    • The fine-grained verification score can also be used for task progress estimation and as a continuous RL reward signal.
    • In LIBERO, using SAC reward shaping improves sample efficiency by about 1.8x, and in MATH with GRPO by about 1.1x.
    概要PDF
  11. 572RoboticsarXiv:2607.04988

    InternVLA-A1.5: Unifying Understanding, Latent Foresight, and Action for Compositional Generalization

    • Foresight tokens are supervised only during training by the video generation model WAN2.2-5B and removed at inference time to preserve real-time speed.
    • We keep strengthening semantic ability by continuously training the VLM backbone with VQA, subtask prediction, and discrete action-token objectives.
    • It ranks first across six simulation benchmarks, including LIBERO at 98.9%, RoboTwin at 93.2%, and SimplerEnv at 80.8%.
    • It achieves 84.8% zero-shot on LIBERO-Plus and 27.7% zero-shot on DOMINO, demonstrating robustness to distribution shift and dynamic object interaction.
    • Removing foresight tokens drops DOMINO to 23.8%, confirming the value of transferring dynamic priors.
    概要PDF
  12. 571Computer VisionarXiv:2607.05373

    PixWorld: Unifying 3D Scene Generation and Reconstruction in Pixel Space

    • It optimizes 3DGS directly in pixel space through differentiable rendering without VAE or RAE, eliminating information loss.
    • It unifies generation and reconstruction in a single forward pass by splitting views into clean and noisy subsets.
    • The geometry perception loss provides geometric supervision in the feature space of a 3D foundation model.
    • It outperforms existing latent-space generation methods on RealEstate10K, DL3DV-10K, and WorldScore, while matching state-of-the-art reconstruction quality.
    • Removing the geometry perception loss causes a 1.13 dB drop in PSNR and a 12.5 percent relative drop in AUC@5, confirming its effectiveness.
    概要PDF
  13. 570LLMs / NLParXiv:2607.02770

    Gemma 4 Technical Report

    • It consists of dense models from 2.3B to 31B and a 26B-A4B MoE, and every model includes a thinking mode.
    • The 12B model removes the vision and audio encoders and uses an encoder-free architecture that projects raw patches directly into the LLM embedding space.
    • KV-cache sharing, keys-as-values, and p-RoPE reduce the global KV cache by 37.5 percent.
    • With QAT quantization, the 31B model shrinks from 64 GB to 19.2 GB, and the audio encoder shrinks from 390 MB to 87 MB, which is a 78 percent reduction.
    • It is the leading dense open model on Arena Elo at 1451 for the 31B model and reaches 96.4 percent accuracy on RULER 128k.
    概要PDF
  14. 569Machine LearningarXiv:2607.05378

    CompactionRL: Reinforcement Learning with Context Compaction for Long-Horizon Agents

    • It integrates context compression, or summarization, into the RL training loop and makes summarization part of a learnable policy.
    • Token-level loss normalization corrects segment length bias, and cross-trajectory GAE assigns credit across compression boundaries.
    • GLM-4.5-Air, a 106B model, reaches 66.8% on SWE-bench Verified, which is a 7.0 point gain, and 24.5% on Terminal-Bench 2.0, which is a 3.1 point gain.
    • GLM-4.7-Flash, a 30B model, reaches 56.0% on SWE-bench Verified, a 5.5 point gain, and 20.2% on Terminal-Bench 2.0, a 6.8 point gain.
    • Including summary training increases summary length, produces richer compressed states, and also increases reasoning tokens, which effectively expands the usable context.
    概要PDF
  15. 568LLMs / NLParXiv:2607.04425

    UI-MOPD: Multi-Platform On-Policy Distillation for Continual GUI Agent Learning

    • MOPD, multi-teacher on-policy distillation, is introduced for the first time to continual cross-platform learning for GUI agents.
    • Uni-GUI is a dataset of about 10K high-quality cross-platform GUI interaction trajectories.
    • A platform router dynamically selects a teacher based on the environment, mixing behavior rules and preventing catastrophic forgetting.
    • It achieves relative gains of 38.2% on OSWorld and 12.0% on MobileWorld, or 12.7% and 55.8% depending on the baseline, showing that a single 8B model improves on both platforms at once.
    • GUI grounding ability is also preserved, with 43.14% on ScreenSpot-Pro and 90.88% on ScreenSpotV2, showing minimal loss relative to the base model.
    概要PDF
  16. 567LLMs / NLParXiv:2607.05155

    EdgeBench: Unveiling Scaling Laws of Learning from Real-World Environments

    • Environment learning average performance follows a log-sigmoid scaling law, with R² = 0.998, and the pattern is stable across all 6 task families.
    • It studies 134 real-world tasks, each requiring 12 or more continuous hours of execution, with large tasks that take human experts an average of 57.2 hours.
    • Frontier agent learning speed roughly doubles every three months, based on models released after September 2025.
    • Continuous experience is stronger than independent restarts, and accumulated experience strongly determines long-horizon performance.
    • The theory derives the log-sigmoid shape as a frontier-expansion process on a latent task graph.
    概要PDF
  17. 566AI / GeneralarXiv:2607.04439

    ResearchStudio-Idea: An Evidence-Grounded Research-Ideation Skill Suite from ML Conference Outcomes

    • It extracts 15 ideation patterns, 31 subpatterns, and 28 research domains from 1,947 machine learning conference papers across oral, highlighted, and rejected papers.
    • The pattern cards are designed as operational tools, not simple classifications, with a structure that contrasts success conditions and failure modes.
    • IdeaSpark is a multi-step reasoning pipeline of gap finding, pattern matching, candidate generation, conflict search, and audit decision making.
    • Rejected papers were also mapped into the same 15-pattern space, showing that they can be used as contrastive evidence.
    • Automatic evaluation shows improved idea quality, although human blinded evaluation remains a future task.
    概要PDF
  18. 565Computer VisionarXiv:2607.05247

    Vision Pretraining for Dense Spatial Perception

    • It reframes boundaries as dense training signals rather than outputs and introduces boundary modeling into self-supervised pretraining.
    • Masked boundary modeling uses the teacher's boundary prediction as the masked target, and geometric routing resolves conflicts between semantics and boundaries.
    • A categorical reparameterization of the boundary field enables stable dense self-distillation.
    • A 1B LingBot-Vision model outperforms 7B DINOv3 on NYU-Depth v2 accuracy and reaches state-of-the-art on 14 depth-completion benchmarks.
    • A 0.3B distilled model matches the NYU-Depth v2 accuracy of 7B DINOv3 while using about 23 times fewer parameters.
    概要PDF
  19. 564RoboticsarXiv:2607.04434

    RoboDojo: A Unified Sim-and-Real Benchmark for Comprehensive Evaluation of Generalist Robot Manipulation Policies

    • It evaluates five capability dimensions, namely generalization, memory, precision, long-horizon execution, and open-ended instruction, across 42 simulation tasks and 18 real-world tasks.
    • It greatly improves evaluation throughput through heterogeneous parallel simulation based on Isaac Sim.
    • RoboDojo-RealEval provides a reproducible real-world evaluation system with standardized hardware, scene resets, and evaluation protocols.
    • It builds a public leaderboard by integrating 30 policies into XPolicyLab and performing cross-evaluation in simulation and in the real world.
    • Stability is confirmed by maximum GPU-to-GPU standard deviation of 1.1 percentage points in simulation and a maximum repeat-evaluation standard deviation of 1.3 percentage points in the real world.
    概要PDF
  20. 563AI / GeneralarXiv:2607.02032

    PACE: A Proxy for Agentic Capability Evaluation

    • The core idea is to predict expensive agentic evaluation with cheap single-turn benchmark scores, and 100 proxy instances are enough.
    • It combines SVD leverage, which measures geometric information content, with rank correlation, which measures target relevance, using target-specific weights.
    • Leave-one-out cross-validation gives a MAE of 3.80 percent, Spearman of 0.81, and pairwise accuracy of 84.4 percent across 14 models and 4 targets.
    • It is about 100 times cheaper than random sampling with the same quality, which makes agentic evaluation practical in everyday model development.
    • The selected instances are interpretable because their capability distribution clearly reflects the abilities required by each benchmark, such as IF plus verification for GAIA and planning plus testing for SWT.
    概要PDF
  21. 562AI / GeneralarXiv:2607.02255

    AgenticSTS: A Bounded-Memory Testbed for Long-Horizon LLM Agents

    • The key idea is that memory is not storage but a decision visibility contract, so per-decision state is composed through typed retrieval instead of raw transcript accumulation.
    • Five slots, L1 through L5, can each be ablated independently, and the L5 trigger skill contributes the most, moving performance from 3 out of 10 to 6 out of 10 with directional p around 0.37.
    • The prompt stays around 5K tokens, while STS2MCP expands to about 527K tokens per call in a single run, which is an order-of-magnitude to two-order-of-magnitude gain in token efficiency.
    • It uses Slay the Spire 2 A0 as the testbed, where human win rate is 16% and frontier LLMs score zero, so the task is hard but not saturated.
    • It releases 298 trajectories, SHA-anchored L4 and L5 snapshots, prompt logs, and Wilson and bootstrap scripts.
    • A cross-backbone probe across Gemini, Qwen, and DeepSeek verifies the same ablation surface.
    概要PDF
  22. 561RoboticsarXiv:2607.01804

    VLA-Corrector: Lightweight Detect-and-Correct Inference for Adaptive Action Horizon

    • The core problem is that the longer H is, the lower the policy call cost, but the more open-loop blind spots accumulate and cause errors.
    • A 40M LVM detects visual dynamics deviations in latent space in real time, so no backbone retraining is needed and the method is plug-and-play.
    • The adaptive action horizon keeps long chunks in confident regions and cuts them immediately, then reruns OGG when deviation is detected.
    • With a horizon of 50 on pi0.5, success rate rises from 48.7 percent to 58.7 percent and average calls fall from 5.15 to 4.98, which improves success per call by 24.6 percent.
    • In real-world disturbance recovery, performance improves from 40.0 percent to 68.3 percent, and precision alignment tasks improve by 16.6 percent.
    • Cuts happen mostly in critical phases, 83.7 percent of them during grasping and alignment, which shows that the monitor correctly detects the meaningful moments.
    概要PDF
  23. 560Computer VisionarXiv:2607.02515

    PointDiT: Pixel-Space Diffusion for Monocular Geometry Estimation

    • It is the first pixel-space approach to apply diffusion directly to raw 3D point maps without going through a latent VAE space.
    • Direct x-prediction of the clean point map is far better than v-prediction, with Rel 35.44 versus 9.29.
    • Injecting DINOv3 4-layer features raises boundary sharpness to BF1 13.47, which is better than MoGe-2 at 11.75 and DA3 at 12.58.
    • The generative approach clearly beats deterministic regression on boundary quality, with BF1 13.92 versus 10.90, and on handling ambiguous regions.
    • It was accepted to ICML 2026 in Seoul, showing that pixel-space diffusion can extend beyond natural images to 3D and 4D geometry.
    概要PDF
  24. 559RoboticsarXiv:2607.02466

    Learning to Move Before Learning to Do: Task-Agnostic pretraining for VLAs

    • The Decomposition Hypothesis separates how to move, physical control, from what to do, semantics, and only the latter requires language.
    • An inverse-dynamics objective enables learning of a physical prior from unrelated trajectories and autonomous-driving data.
    • SIMPLER matches models trained on more than 1 million expert examples and beats behavior cloning by an absolute 10%.
    • On a real WidowX robot, it retains 25% success under camera perturbation, compared with 0% for an internet-scale baseline.
    • Partial success rises from 31.8% to 45.8%, confirming the transferability of the physical pretraining.
    • Stage 1 scale determines the performance ceiling, while extending Stage 2 has diminishing returns.
    概要PDF
  25. 558LLMs / NLParXiv:2607.01233

    Measuring the Gap Between Human and LLM Research Ideas

    • We introduce a two-axis taxonomy of research taste: seven opportunity patterns times six method paradigms, based on NSF, NIH, and DARPA.
    • LLMs score 47% to 64% on bridge opportunity versus humans at 12.1%, and 22% to 39% on synthesis and unification versus humans at 5.1%.
    • Human ideas have consistently higher normalized entropy on both axes, meaning they are more broadly distributed.
    • Cosine similarity between LLMs, 0.83, is higher than human LLM similarity, 0.72 to 0.78, showing model-to-model convergence.
    • LLMs emphasize the integrate and unify archetype at 34.2% versus humans at 2.35%, while humans focus on replace and decouple.
    • The key takeaway is that LLM ideas are reasonable but distributionally narrow, and the task is to align research taste.
    概要PDF
  26. 557Machine LearningarXiv:2607.01179

    QuasiMoTTo: Quasi-Monte Carlo Test-Time Scaling

    • It reparameterizes autoregressive sampling as inverse-CDF sampling, converting QMC uniform numbers into LM samples.
    • Each sample is marginally exact, meaning it comes from the same LM distribution, while negative correlation across the batch dimension increases coverage.
    • On pass@k, it achieves the same performance with 25% to 47% fewer samples across four reasoning benchmarks.
    • On GRPO RL, it reaches the same pass@1 with 50% fewer training steps and reduces the zero-variance group ratio.
    • Coverage increases in the order lattice, stratified, Sobol, and i.i.d., quantifying the freedom-versus-coverage tradeoff.
    • The limitation is that it was validated only on 1B to 2B models and short outputs, so extension to long CoT remains future work.
    概要PDF
  27. 556AI / GeneralarXiv:2607.02234

    Purified OPSD: On-Policy Self-Distillation Without Losing How to Think

    • OPSD fails consistently on four long-CoT models: Qwen3-8B, Qwen3-4B, R1-Distill-7B, and OLMo-7B.
    • Decomposing the teacher signal shows that the reference-induced component dominates both the update direction and size, while the question-conditioned component is orthogonal or opposite.
    • Epistemic markers such as 'Wait' and 'Maybe' explode or collapse during OPSD training, hurting reflective reasoning ability.
    • The PMI target distribution, log πT - log πref, removes non-transferable components and is robust to β and c.
    • On AIME 2024 and 2025, it consistently improves over OPSD-Standard and even outperforms the base model.
    • The method remains stably improved across soft-clipping thresholds c in {5, 10, 20} and correction strengths β in {0.5, 1, 2}.
    概要PDF
  28. 555Machine LearningarXiv:2607.01763

    Denser $\neq$ Better: Limits of On-Policy Self-Distillation for Continual Post-Training

    • SDPO is effective for single-domain specialization, but it is weak in OOD generalization and continual learning.
    • Compared with GRPO, SDPO causes larger changes in both parameter drift, measured by NSS and Principal Angle, and response embedding shift.
    • At 5% SDPO, it exhibits an infinite repetition collapse of the boxed token, showing an artifact amplification mechanism in token-level supervision.
    • Theorem 1 proves that SDPO's KL drift is always greater than or equal to that of the GRPO KL-minimal Razor policy.
    • The key lesson is that on-policy data does not equal forgetting mitigation; the density of the objective function determines stability.
    • Reordering data or masking tokens only partially alleviates the collapse, indicating a fundamental limitation.
    概要PDF
  29. 554Computer VisionarXiv:2607.02375

    Representation Distribution Matching for One-Step Visual Generation

    • It systematizes RDM along two axes, distribution comparison and representation space, and derives the best design for each axis.
    • It shows that MMD becomes a strong objective when estimated with Nyström approximation and large batches of 2,048 or more.
    • A single encoder can be gamed, so the paper proposes SWr14, an evaluation metric using 14 encoders that is resistant to gaming and independent of training loss.
    • iRDM reaches a one-stage state-of-the-art SWr14 of 1.30 on ImageNet and a PickScore win rate of 71.2%.
    • It compresses FLUX.2[klein] from four stages to one stage, reaching a GenEval score of 0.826 versus 0.794 for the original model with 90 H200 GPU hours.
    • Proportional Lagrangian optimization keeps the encoders balanced and improves the weakest encoder.
    概要PDF
  30. 553RoboticsarXiv:2606.32027

    Freeform Preference Learning for Robotic Manipulation

    • It resolves the ambiguity of binary preferences by defining natural-language axes such as speed, safety, and precision, and collecting pairwise preferences along each axis.
    • It extends Bradley-Terry to multiple dimensions so that a single reward model conditioned on language labels outputs axis-specific scalar rewards.
    • A VLM-based reward model using Qwen VL 3.5 4B combines visual and language understanding for reward learning.
    • It achieves a 38-point gain over sparse reward and also a significant improvement over binary preference.
    • It exhibits emergent compositionality, combining fast speed with new target combinations that were not present in the training data.
    • By changing the reward condition at test time, it can steer policy behavior without retraining.
    概要PDF
  31. 552RoboticsarXiv:2607.01051

    AutoSpeed: Annotation-Free Stage-Adaptive Motion Speed Learning for Robot Manipulation

    • The first framework to infer stage-adaptive motion speeds implicitly through end-to-end learning, without speed or stage annotations.
    • DCT-based frequency-domain scaling enables acceleration and deceleration by non-integer factors while preserving high-frequency action details.
    • Multi-target selective optimization uses the composite cost J = E/w + log(h) to manage the trade-off between prediction error and horizon length.
    • The model-agnostic design applies to both non-generative policies based on MLPs and generative policies based on diffusion or flow.
    • Across four real-robot tasks, the method cut execution time by a factor of 1.78 on average. On the Place the Toy task, it improved the success rate by 10 percentage points.
    • The inferred speed ratios showed a strong correspondence with task stages such as reaching, assembly, and precision insertion.
    概要PDF
  32. 551Machine LearningarXiv:2607.02512

    Program-as-Weights: A Programming Paradigm for Fuzzy Functions

    • The paper proposes a new paradigm that compiles a natural-language specification into a hybrid neural program composed of pseudo-program text and LoRA weights.
    • It releases FuzzyBench-10M, a fuzzy-function dataset with 29 topic variants, more than 800 categories, and 10 million examples.
    • A 0.6B interpreter plus the PAW program outperforms direct prompting of Qwen3-32B in accuracy, 73.78 percent versus 68.70 percent.
    • When quantized on a MacBook M3, it runs offline at 30 tokens per second with a 430 MB shared base and 23 MB per-function LoRA.
    • By replacing only the compiler in a vision-language model, the approach can extend to image-conditioned fuzzy tasks.
    • The compiled programs are reusable artifacts that support versioning, package-manager distribution, and two-line Python or JavaScript API calls.
    概要PDF
  33. 550Machine LearningarXiv:2607.01693

    A Mathematical Introduction to Diffusion Models

    • The convergence guarantees of Langevin dynamics and ULA and MALA are proved through KL contraction under log-Sobolev inequalities.
    • The Tweedie identity and the reverse-time SDE form the theoretical backbone of diffusion models, while the same Gaussian channel is reinterpreted through Polchinski flow and probabilistic localization.
    • The discretization error of DDPM is decomposed into three terms, early stopping, KL telescoping, and score error, for analysis.
    • Discrete diffusion models replace the noising SDE with a continuous-time Markov chain and reframe the error analysis in finite state spaces.
    • It systematizes guidance, reward tilting, path-space control, and inference-time reinforcement learning as ways to steer trained models at inference time.
    • More than 50 references and appendices on Itô calculus and Gaussian tools make the work self-contained.
    概要PDF
  34. 549Distributed SystemsarXiv:2607.01120

    Next-Generation Agentic Reinforcement Learning Systems Enable Self-Evolving Agents

    • The core claim is that the bottleneck in self-evolving agents lies not in the RL algorithm, but in the system layer, including ATDP, data proxies, and the evolution control surface.
    • ATDP proposes a vendor-neutral protocol that includes step-level decision context, actions, outcomes, delayed rewards, and governance metadata.
    • The evolution control surface automatically chooses among memory, skills, harnesses, weights, and rollback based on trajectory statistics, evaluation scores, user correction rates, cost, and safety constraints.
    • AREAL2.0 is a prototype that reorganizes existing RL frameworks into an online agent-service-oriented RL loop, and it folds LLM calls from agent services into the RL training loop by swapping only the gateway.
    • It cites OpenClaw as an example of a personal self-evolving agent and analyzes the gap to enterprise-level governance.
    • It defines self-evolution not as a simple weight update, but as a multidimensional improvement of a composite policy made up of the LLM, harness, memory, and tools.
    概要PDF
  35. 548Computer VisionarXiv:2607.02517

    WorldDirector: Building Controllable World Simulators with Persistent Dynamic Memory

    • It explicitly separates motion planning for dynamic objects from video synthesis, which lets it preserve both physical logic and appearance stability.
    • An LLM orchestrator generates 3D trajectories and then projects them into 2D to provide location conditioning for the video generation model.
    • Appearance Binding preserves visual consistency when an object reappears, regardless of how long it was off screen.
    • It achieves SOTA on every reconstruction metric, with PSNR 18.127, SSIM 0.502, and LPIPS 0.359.
    • It is far ahead on dynamic object consistency as well, with DSC_DINO at 0.769 and DSC_CLIP at 0.917.
    • Promptable World Events lets users freely design the timing, motion, and trajectories of new objects that do not appear in the initial frame.
    概要PDF
  36. 547Computer VisionarXiv:2607.01202

    World from Motion: Generative Dynamic Gaussian Reconstruction from Monocular Video

    • It is the first method to condition a video generation model on a dynamic 3DGS representation, rendering geometry, appearance, and 3D motion according to reference and target views.
    • It achieves state-of-the-art results on the DyCheck benchmark with mPSNR 19.98, mSSIM 0.716, and mLPIPS 0.178.
    • In the optimization step that distills generated samples back into dynamic 3DGS, it also improves 3D tracking accuracy, with PCK@0.05 rising from 0.824 for MoSca to 0.862.
    • Reconstruction quality improves monotonically as the number of virtual cameras increases and converges after 8 cameras, with mPSNR rising from 18.69 to 19.78.
    • It outperforms both generation-based methods such as CAT4D and Vista4D and reconstruction-based methods such as WorldTree and ViDAR.
    • It can perform visual outpainting on in-the-wild video, dynamic correction, and out-of-view dynamics inference.
    概要PDF
  37. 546AI / GeneralarXiv:2607.00248

    Seed2.0 Model Card: Towards Intelligence Frontier for Real-World Complexity

    • Three sizes, Pro, Lite, and Mini, let users choose a performance-speed tradeoff, and Mini costs $0.03 per million for prefill and $0.31 per million for decode.
    • Compared with GPT-5.2 at $1.75/$14 and Claude Opus 4.5 at $5/$25, Seed2.0 Pro is far more cost-efficient at $0.47/$2.37.
    • In China MaaS traffic, the internet industry dominates, while manufacturing, automotive, and telecom each account for less than 1%, and unstructured information processing is the main scenario.
    • For agentic coding requests, frontend development and bug fixing dominate, with Vue.js used more than three times as often as React.
    • The goal is to achieve competitiveness on research-level reasoning tasks such as Erdős problems and scientific coding.
    • Seed2.0 acknowledges that there is still a gap to Claude and Gemini on coding tasks such as SWE-Evo and NL2Repo, as well as on long-tail knowledge such as SuperGPQA.
    概要PDF
  38. 545AI / GeneralarXiv:2607.01224

    AutoMem: Automated Learning of Memory as a Cognitive Skill

    • The file system, including read, write, search, append, and create, is promoted to a first-class memory action equivalent to task actions, so every memory decision is traceable.
    • In Outer-loop 1, a meta-LLM reviews trajectories of up to 100,000 steps and automatically improves the agent scaffold, including the code, prompts, and memory schema.
    • In Outer-loop 2, good memory decisions from the agent are extracted and used to LoRA fine-tune a dedicated memory specialist model, while the task model remains frozen.
    • Memory optimization alone yields a 2x to 4x performance gain without changing task-model weights, showing that memory has more leverage than model scale.
    • The optimized 32B agent surpasses 72B and reaches the level of Claude Opus 4.5 and Gemini 3.1 Pro Thinking.
    • This is the first case that turns the cognitive-science concept of metamemory into a concrete optimization target for LLM agents.
    概要PDF
  39. 544RoboticsarXiv:2607.00272

    ASPIRE: Agentic /Skills Discovery for Robotics

    • The closed-loop robot execution engine exposes multimodal traces for each primitive, including perception overlays, grasp candidates, trajectories, and collision feedback, enabling autonomous failure diagnosis and repair by the agent.
    • It accumulates validated repair patterns in a skill library and reuses them as in-context guidance on later tasks, improving adaptation speed as the task set grows.
    • Evolutionary search goes beyond single-trajectory repair by exploring multiple candidate programs in parallel, which further improves success rates.
    • It achieves a 31 percent zero-shot success rate on LIBERO-Pro Long, compared with 4 percent for the previous method, demonstrating the generalization effect of the growing skill library.
    • When sim skills are transferred to a real dual-arm robot, inference tokens are reduced by about 10x and success rates also improve.
    • Performance improves step by step from a macro average of 14 percent without the execution engine to 62 percent with the engine and 72 percent with evolutionary search.
    概要PDF
  40. 543Computer VisionarXiv:2607.00784

    LeVLJEPA: End-to-End Vision-Language Pretraining Without Negatives

    • This is the first proof that stable large-scale vision-language pretraining is possible with only cross-modal prediction and SIGReg regularization, without negative pairs.
    • When used as a VLM backbone, it outperforms both CLIP and SigLIP, with GQA at 44.6%, VQAv2 at 54.2%, and POPE at 66.9% on the Llama-1N setup.
    • For semantic segmentation, it provides better dense semantic features than contrastive learning, preserving structured information at the patch-token level.
    • In zero-shot classification, contrastive learning still leads, but as a backbone the zero-shot score is actually inversely correlated with performance.
    • The entire training recipe removes the need for negatives, temperature, momentum, and teacher-student mechanisms, simplifying training and reducing cost.
    • It is well suited to object-centric representation learning and is more robust to backgrounds than contrastive methods.
    概要PDF
  41. 542Machine LearningarXiv:2607.01232

    Is One Layer Enough? Training A Single Transformer Layer Can Match Full-Parameter RL Training

    • Training just a single transformer layer can recover up to 114% of the full RL gain, while the weakest layers contribute less than 30%.
    • High-contribution layers consistently concentrate in the middle 40% to 60% of the network depth, regardless of model scale, family, RL algorithm, or dataset.
    • The layer contribution ranking is stable, with Spearman rho equal to 0.76 across datasets and 0.59 across tasks.
    • Raising the learning rate or selectively training the high-contribution layers adds further gains over full RL, including +43% for Qwen3-1.7B, +27% for 4B, and +32% for 8B.
    • Models trained on different layers show complementary behaviors, and majority voting can improve performance further.
    • Although the weight change magnitude is relatively uniform across layers during full training, the contribution is extremely uneven, which shows that parameter subspace efficiency differs sharply.
    概要PDF
  42. 541AI / GeneralarXiv:2606.31651

    FARS: A Fully Automated Research System Deployed at Scale

    • The system uses a fully automated four-stage pipeline: Ideation, Planning, Experiment, and Writing, with a specialized agent at each stage and a shared workspace.
    • It generated 166 papers across 67 fine-grained AI and ML topics and preserved the full output, including successful, failed, and negative results, rather than only selected samples.
    • It collected 282 structured volunteer reviews covering 140 papers, including ratings, soundness, presentation, contribution, integrity checks, and disclosure of LLM use.
    • In a cross-comparison with the Stanford Agentic Reviewer, FARS scores 5.00 out of 10 across 165 papers, which surpasses existing systems such as DeepScientist at 4.13 and ScientistOne at 4.00.
    • Human reviews average 3.23 out of 10, so they are stricter than automated reviewers but still aligned in direction, and 11.4 percent of papers reach the accept threshold under automated review criteria.
    • Repeated failure modes include narrow experimental scope, methodological limitations, and integrity issues such as claim-artifact mismatch, which makes contribution quality, experimental sufficiency, and faithful reporting the true bottlenecks rather than polished presentation.
    概要PDF
  43. 540Computer VisionarXiv:2606.31986

    CoLT: Teaching Multi-Modal Models to Think with Chain of Latent Thoughts

    • The problem is that text CoT must generate thousands of tokens autoregressively, which causes large inference latency, and the natural-language form can erase alternative reasoning paths.
    • It uses three forms of step-level supervision: forward decoding from latent to text, backward decoding from text to latent, and internal supervision for consistency between latent steps.
    • At inference time it removes the decoder and internal supervision, so reasoning is completed with only 3 latent vectors and no extra parameter overhead.
    • Across 8 benchmarks on Qwen3-VL-8B, including MMStar, MathVista, and SeedBench, it consistently outperforms both text CoT and previous latent reasoning methods.
    • For speed, text CoT takes 7.24 seconds and 142 tokens, whereas CoLT takes 0.32 seconds with 3 latent vectors, which is a 22.6x decoding speedup and a 10.1x end-to-end speedup.
    • It is also robust to noise: under visual and textual noise, Direct Answer drops by 7.2 to 19.0 percent, Text CoT drops by 5.0 to 13.8 percent, but CoLT drops only 2.8 to 9.0 percent.
    概要PDF
  44. 539LLMs / NLParXiv:2606.30015

    Parametric Skills

    • The core problem is that in-context skills are limited by the model's ability to understand and follow instructions, and performance drops sharply, especially for open-source models or long contexts.
    • The solution is to feed a text skill into a hypernetwork that generates a LoRA adapter in a single forward pass, applying the skill without consuming context window.
    • It builds a 45.8K-scale high-quality skill library from web crawling and summaries of real agent trajectories, covering 13 domains.
    • On six SWE tasks, ParametricSkills scores 64.09 versus 57.65 for in-context learning and 48.48 for SHINE, averaged over LLM judges. SHINE fails to faithfully compress both the skill content and how to use it.
    • Rank-concatenation merging gives multi-skill composition performance equal to or better than a single LoRA, and it outperforms factor-wise linear merging.
    • In HumanEval-based self-evolving experiments, five rounds of iteration improve accuracy from 75% to 84.76%, and online continual merging further improves it from 29% to 51.6%.
    概要PDF
  45. 538Computer VisionarXiv:2607.00678

    ABot-M0.5: Unified Mobility-and-Manipulation World Action Model

    • It identifies three structural bottlenecks: temporal resolution mismatch between coarse video and fine-grained control, action-space mismatch between navigation and manipulation, and context mismatch between training and inference.
    • It introduces an intermediate latent action that bridges video latent representations and embodiment-specific control while capturing local visual state transitions.
    • A dual-level Mixture-of-Transformers separates modality representation from heterogeneous action subspaces, such as base movement and arm manipulation, so that each can be optimized separately.
    • Dream forcing trains dynamics on model-predicted dream video, which aligns autoregressive inference with training conditions and reduces long-horizon rollout error accumulation.
    • On RoboCasa365 mobile manipulation, it reaches 90.8 percent, 75.4 percent, and 58.4 percent across difficulty levels, and on real-robot peg insertion it achieves a 70 percent success rate and a 96 percent process score, compared with 50 percent and 90 percent for pi0.5 and 30 percent and 77 percent for Fast-WAM.
    • On real robot long-horizon tasks such as plate sorting, fruit sorting, and flower arrangement, it achieves 60 to 80 percent success, which is a large improvement over Fast-WAM at 20 to 40 percent.
    概要PDF
  46. 537Machine LearningarXiv:2606.32017

    TRIAGE: Role-Typed Credit Assignment for Agentic Reinforcement Learning

    • A core problem in agent RL is that GRPO applies trajectory success or failure to every action equally, which punishes useful exploration in failed rollouts and rewards regression in successful rollouts.
    • It introduces a four-role taxonomy, Decisive, Exploration, No-progress, and Regression, and applies differentiated credit rules by role.
    • It proves that role labels alone can support MSE-optimal segment correction and always reduce advantage estimation error within the judge's trusted range.
    • On Qwen3-1.7B, it improves ALFWorld by 18.4 points and WebShop by 7.9 points, while reducing environment turns per completed rollout by 10.4% to 14.8%.
    • Regression suppression is the main driver of the gains, with exploration bonuses adding only 0.6 to 1.7 points as a stable auxiliary benefit.
    • It outperforms scalar process rewards and outcome-supervised value baselines, so the key benefit comes from role typing itself rather than simply adding dense rewards.
    概要PDF
  47. 536Information RetrievalarXiv:2606.31984

    GR2 Technical Report

    • Semantic ID mid-training with at least 99% uniqueness lets the LLM directly reason about catalog items.
    • Trace distillation from a strong 32B teacher model installs reasoning priors in the 8B student model.
    • DAPO-based RL with multiple rewards, format, AUC/NDCG, and LLM-as-judge, optimizes re-ranking.
    • It improves industrial traffic by +18.7% in R@1, +7.1% in R@3, and +9.6% in N@3.
    • On-Policy Distillation (OPD) recovers about 82% of the 32B teacher gains with a 1.7B model that is only 5% of the size.
    • Conditional reward design to prevent reward hacking, such as preserving input order and exploiting position bias, is essential.
    概要PDF
  48. 535Software EngineeringarXiv:2606.29957

    SWE-Together: Evaluating Coding Agents in Interactive User Sessions

    • It extracts 109 verifiable tasks from 11,260 real sessions, for a conversion rate of 0.97 percent.
    • The state-conditioned LLM user simulator preserves the original intent while adapting feedback to the trajectory.
    • It evaluates along multiple dimensions using final accuracy, user correction, meaning the number of correction turns, and intent coverage.
    • Stronger agents achieve higher success rates with fewer interventions, suggesting an improved user experience.
    • It highlights the limits of static SWE-Bench style evaluation, which misses real interaction by scoring only single-turn commands and final code.
    • It uses four upstream data sources: DataClaw, Pi-staging, Hyperswitch, and SWE-chat.
    概要PDF
  49. 534Computer VisionarXiv:2606.29148

    GPC: Large-Scale Generative Pretraining for Transferable Motor Control

    • It learns a discrete motion representation based on Finite Scalar Quantization (FSQ), which avoids the codebook problems of VQ-VAE.
    • It models motion tokens with a GPT-style autoregressive transformer using next-token prediction.
    • It is trained on a dataset with more than 600 hours of motion data and achieves a 99.98% tracking success rate.
    • It shows emergent behaviors such as disturbance recovery and fall recovery strategies.
    • It supports adaptation to new tasks through Parameter-Efficient Fine-Tuning (PEFT).
    • It is a SIGGRAPH 2026 paper from NVIDIA and Simon Fraser University.
    概要PDF
  50. 533AI / GeneralarXiv:2606.30639

    Self-Evolving World Models for LLM Agent Planning

    • We present a self-evolving world model that changes only deployment-time context without parameter updates.
    • Episodic memory uses retrieval-based simulation of real action transitions.
    • Semantic memory extracts persistent heuristic rules from prediction-observation mismatches.
    • Selective foresight filters low-confidence predictions before integrating them into the agent's reasoning context.
    • It achieves the best predictive accuracy across all three backbones on ALFWorld and ScienceWorld.
    • Oracle experiments show that noisy foresight harms action accuracy, proving the importance of confidence filtering.
    概要PDF