HDATF
简体中文
开始交流
菜单

论文档案库

按编号从大到小列出全部收录论文及摘要。

收录论文 1,430资料更新日期 Research Monitoring基准测试
  1. 682Computer VisionarXiv:2607.19064

    Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing

    • At 4B parameters, it matches the performance of competitors in the 6B to 80B range while keeping memory usage down to about 18 GB.
    • The Mage-VAE tokenizer reduces compute per pixel by 12 to 22 times compared with conventional VAEs.
    • CUDA kernel fusion improves training throughput by about 2.5 times.
    • The turbo variant generates a 1024 by 1024 image in 0.59 seconds and performs editing in 1.02 seconds on a single A100.
    • Its cost efficiency makes it a practical open baseline for researchers who want to experiment and iterate.
    原文摘要PDF
  2. 681Computer VisionarXiv:2607.19343

    Masked Visual Actions for Unified World Modeling

    • It proposes an interface called Masked Visual Actions, which represents robot actions as pixel-level masks.
    • A single checkpoint supports both forward simulation and inverse action generation.
    • It works well across different environments and robot embodiments even when it is fine-tuned on only 15 hours of real robot data.
    • The imagined rollouts from the video model are strongly correlated with real execution, with r = 0.982.
    • Given a desired object motion, the inverse model synthesizes the corresponding robot action and reaches a 90 percent success rate.
    原文摘要PDF
  3. 680Software EngineeringarXiv:2607.18161

    TRIM: Reducing AI-Generated CodeSlop via Agent Trajectory Minimization

    • It defines CodeSlop by functional behavior rather than surface style, such as verbosity or duplication. Any edit that can be removed while tests still pass counts as CodeSlop.
    • The cause is the agent's search process. Once it finds a correct solution, there is no reason to roll back the attempted edits, so they remain in the final patch.
    • Telling the agent to reduce its own patch was unstable. In 3.8% to 44.9% of cases it broke behavior or made the patch larger.
    • TRIM uses the task history, not the final patch, as the search space because the temporal order of edits approximates dependency structure.
    • It uses a hierarchical approach that tries to remove large chunks first, confirms success, and then splits more finely if needed, so the verification cost is about half that of Delta Debugging.
    • On SWE-Bench-Verified, it preserved correctness on 327 of 330 successful patches, or 99.1%, and 18 patches became exactly identical to the human-written reference patch.
    原文摘要PDF
  4. 679Distributed SystemsarXiv:2607.18002

    ExpertPlex: A High-Goodput Disaggregated Serving System for MoE LLMs with Adaptive Persistent Kernels

    • More than 95 percent of the weights in an MoE language model belong to the expert part, so duplicating the full model for both prefill and decode wastes a lot of memory.
    • The key idea is a hybrid structure in which the heavy experts are shared across the two stages, while less than 5 percent attention is assigned separate GPUs by stage.
    • An adaptive persistent kernel, or APK, schedules at matrix-multiplication tile boundaries so that urgent decode work can take over and free capacity can return to prefill without relaunching kernels or involving the CPU.
    • A one-way MoE communication path initiated from the attention side removes MoE polling, avoids deadlock and cross-stage network interference, and overlaps communication in one stage with computation in the other.
    • In serving experiments with MiniMax-M2.7 and GLM-5.1-FP8, effective throughput rises by 5.65 times over chunked prefill, by 2.01 times over instance-level separation, and by 1.66 times over Green Context colocation.
    • Because the preemption interval depends on tile execution time rather than token count, even a prefill with thousands of input tokens can be interrupted by short decode jobs.
    原文摘要PDF
  5. 678RoboticsarXiv:2607.18016

    Closing the Loop in Humanoid VLA: Persistent 3D Object Tokens for Verifiable Loco-Manipulation

    • It points out that robots fail on long tasks not because they cannot understand instructions, but because the object state used for action differs from the object state used for verification.
    • POT binds objects to role slots such as target, destination, support surface, and handoff partner, and refreshes the 3D position grounding after each action chunk.
    • Object records are turned into fixed-format tokens containing 33 features per slot across 8 slots and inserted into the input sequence of the whole-body action model.
    • The same record is used to verify grasp, placement, and release through geometry-based checks, and if it fails, the system triggers re-observation, retry, and replanning.
    • On eight task groups with a real Unitree G1, it improves the same-condition GR00T-N1.7 baseline from 39/80 to 71/80, and cup stacking jumps from 1/10 to 8/10.
    • In ablation, tokens alone improve from 15/40 to 31/40, verification alone gives 22/40, and using both gives 34/40, showing that token conditioning is the main driver of performance.
    原文摘要PDF
  6. 677RoboticsarXiv:2607.18060

    RoboHarness: Memory-Driven Orchestration of Heterogeneous Robot Policies for Long-Horizon Planning

    • Instead of solving long tasks with a single policy, we reformulate the problem as coordinating heterogeneous policies with different capability boundaries.
    • Existing planning methods assume homogeneous skills with fixed, non-overlapping capabilities, but we point out that the capability boundaries of real policies change with context.
    • We use three groups of skills, understanding, memory, and self-evolution, to handle context interpretation, capability-boundary estimation, and experience-based correction during execution.
    • The memory bridge retrieves trajectories associated with the next policy, estimates that policy's normal input region, and guides the robot into that region to make handoff stable without retraining.
    • In ablation tests on 500 long-horizon tasks, the full system reaches 86.0% success, removing the memory bridge drops it to 60.4%, removing understanding skills drops it to 37.6%, and using only a single policy yields 0%.
    • In 135 real-robot trials, success drops from 86.7% to 66.7% under block re-obscuration perturbations, but remains 80.0% even after partial dismantling of the structure.
    原文摘要PDF
  7. 676Machine LearningarXiv:2607.17247

    Distilled Reinforcement Learning for LLM Post-training

    • Reinforcement learning gives one reward per response, so it is hard to know which intermediate reasoning step to praise, and it mostly reinforces what the student can already do rather than adding new knowledge.
    • On-policy distillation forces the model to follow the teacher distribution through KL divergence, creating a dilemma in which there is little to learn if the teacher is too similar and too much noise if it is from a different family.
    • Distilled RL does not add a separate imitation objective; instead, it redistributes the learning signal of RL token by token using the reverse importance ratio between the teacher policy and the old student policy.
    • It consists of three parts: reverse importance sampling with ratio clipping, negative-sample reset that restores the weight of negative-advantage responses to 1, and response-level geometric normalization that sets the geometric mean of ratios within a response to 1.
    • Without negative-sample reset, mean pass@1 drops by 8.81 points on Qwen3-4B and 6.39 points on DSQW-1.5B, making it the most important of the three parts.
    • The gains are especially large when the teacher and student belong to different families. In that setting, on-policy distillation remains below RL, while Distilled RL beats all baselines.
    原文摘要PDF
  8. 675Computer VisionarXiv:2607.18227

    FlowMimic: Mask-free Visual Editing and Generation with Pixel-pair Warped Flow Field for Online Video Editing Data Generation and Modality Mimicry

    • The core claim is that video editing training does not need truly natural video pairs, only pixel correspondences from the first frame that remain temporally consistent.
    • It converts image-editing samples into video-editing samples in real time by turning both the input image and the target image into 3D grids and building a 4D temporally warped flow field over pixel pairs through grid inverse sampling.
    • It treats an image as a one-frame video and uses two modality-mimic losses so that T2I mimics the realism of T2V and V2V editing mimics I2I editing, which converges faster.
    • It introduces sense-related tasks such as referring expression segmentation, together with edit-region-aware latent-level and attention-level losses, so the model can locate the edit region without a separate MLLM or mask input at inference time.
    • It continues training from Wan2.1-T2V-1.3B, and when the learning rate is warmed up from 5e-6 to 1e-5, video-editing behavior appears within a few thousand steps.
    • Without using any specially made video-editing data, it still handles a wide range of editing tasks and even shows SAM3-like ability to segment and track objects specified by language in video.
    原文摘要PDF
  9. 674Computer VisionarXiv:2607.17967

    Fine-Detail Monocular Geometry Estimation with Self-Guided Sparse Volumetric Refinement

    • It diagnoses the limitations of existing monocular geometry estimation models as a structural mismatch rather than a capacity problem. When 3D information is decoded in a 2D coordinate system, features blend across depth boundaries.
    • The solution is to lift geometric modeling from the 2D image plane into 3D space, defining feature neighborhoods by 3D distance instead of screen distance.
    • It converts the point map into (X/Z, Y/Z, log Z) so that only log depth needs to be corrected, and quantizes this value to resolution D to build a sparse voxel shell.
    • A sparse 3D U-Net predicts log-depth residuals at each iteration, and the result redefines the voxel structure for the next iteration. Because of this self-guided scheme, computation is proportional to the number of occupied voxels.
    • In replacement experiments with a parameter-matched 2D U-Net, local metrics were similar to or worse than the base model, showing that the improvement comes from 3D inductive bias rather than capacity.
    • It was trained with K=3, but even when K was increased to 7 at inference time, the metrics held steady or improved slightly, indicating stable convergence of the learned correction.
    原文摘要PDF
  10. 673Computer VisionarXiv:2607.17585

    Pixel-Space Diffusion Transformers

    • It points out that latent diffusion has an information bottleneck because the VAE's lossy compression irreversibly removes high-frequency texture and sharp boundaries.
    • It identifies the core problem as a mismatch between the objectives of a compressor trained for reconstruction and a diffusion model trained for generation, which makes end-to-end optimization difficult.
    • Moving to pixel space enables single-stage training, but it introduces new challenges in noise schedules, loss weighting, token efficiency, and scalable architecture design.
    • Recent methods are introduced as x0 prediction approaches that directly predict the clean original image and speed convergence by keeping the training path on the data surface.
    • It summarizes the trend toward splitting global structure and fine texture across different components through multi-branch designs, frequency separation, and finer tokenization.
    • It suggests that if images, text, and task conditions are placed in a shared token space and trained with a single transformer, the field can move toward unified vision foundation models that do both understanding and generation.
    原文摘要PDF
  11. 672ResearcharXiv:2607.18238

    Color superconductors and holon metals from doping a Fractional Chern insulator

    • Using the parton construction c(r)=f1(r)f2(r)f3(r), the paper places the metallic and superconducting phases of a doped fractional Chern insulator within a unified framework.
    • Each color has three hole pockets across three valleys, so the charge minus e over 3 holons form nine pockets in total.
    • The SU(3) gauge field mediates pairing in the color-antisymmetric channel, allowing a gas of charge-e/3 anyons to enter a charge-2e superconducting state directly without an intermediate binding step.
    • If all nine pockets are weakly paired, c- becomes minus 3 over 2, which differs from the BCS expectation for an f-if superconductor, so an anyonic superconductor can break the usual relation between central charge and pair angular momentum.
    • Adding a particle-hole Higgs field yields two Z3 orthogonal metals with one or three pockets, and reducing valley symmetry to a triangular-lattice space group allows a three-pocket U(1)^2 holon metal that preserves C3.
    • In a chemical-potential-tuned FCI, a continuous transition to superconductivity is impossible in Z3 metals but possible in SU(3) or U(1)^2 holon metals, so the authors argue that the nine-fermion theory is a natural starting point.
    原文摘要PDF
  12. 671Computer VisionarXiv:2607.18217

    HOMIE: Human-object Centric Video Personalization via Multimodal Intelligent Enchancement

    • The problem setting unifies inter-subject cases, where the reference images are of different targets, and intra-subject cases, where they are multiple views of the same target.
    • It points out the limits of two existing approaches: aligning multimodal model features to the text-embedding space weakens the control power of the text encoder, while replacing the text encoder entirely with a multimodal model introduces a large realignment cost.
    • GMG, the first proposal, adds global representations extracted from the multimodal model to the video tokens during query and key computation, which injects knowledge into self-attention.
    • MRE, the second proposal, uses learnable embeddings that distinguish token modality and reference index, which prevents errors such as duplicating the same object or text map as separate entities.
    • In performance, text recognition accuracy improves by 21.8 percent relative to SkyReels-V3, DINOrec reaches 0.696 and is best among the compared methods, and 40 human raters prefer the overall quality 64.7 percent of the time.
    • Ablation studies show that removing GMG drops face similarity from 0.786 to 0.697, and removing MRE drops text accuracy from 0.452 to 0.376.
    原文摘要PDF
  13. 670LLMs / NLParXiv:2607.18213

    SWE-Pruner Pro: The Coder LLM Already Knows What to Prune

    • The core claim is that there is no need to recreate pruning signals outside the model, because the backbone already encodes that judgment in its representations when it reads tool outputs.
    • As evidence, the authors froze Qwen3-Coder-Next and ran a logistic-regression probe over line-averaged hidden states, reaching an AUC of 0.83 and a best F1 of 0.63 on held-out data.
    • The method uses the final-layer hidden states from the prefill stage, and a small head reads them to score tokens and aggregate them into line-level keep or delete decisions.
    • Two design choices improve performance: a length-aware embedding that includes the number of output lines, and a balanced focal loss that keeps the keep and delete ratios aligned per sample.
    • The result is 39% token savings on SWE-QA-Pro and 30% on Oolong across two open-weight backbones and four multi-turn benchmarks, and with MiMo-V2-Flash it also improves Oolong accuracy by 2.2 points and SWE-Bench Verified solve rate by 3.8 percentage points.
    • The extra cost is limited. In a replay setup with the head attached inside the inference engine, it adds only 15.0% wall-clock time over total generation time and requires no separate model call.
    原文摘要PDF
  14. 669Computer VisionarXiv:2607.17423

    TimeLens2: Generalist Video Temporal Grounding with Multimodal LLMs

    • It defines video evidence finding as a problem over a variable-size set of temporal segments rather than a single interval, so it can handle long videos, repeated scenes, first-person video, and question-form requests in one model.
    • It solves data quality by stepwise verification instead of labeling everything at once, using candidate proposal, independent re-search, agent agreement, semantic verification, and boundary refinement.
    • Its key technique is the temporal Wasserstein reward, which treats the predicted and ground-truth intervals as uniform distributions and computes the exact 1D W1 distance to give dense feedback without explicit matching.
    • Overlap-based rewards such as tIoU give zero signal when the predicted interval does not overlap the ground truth at all, while Wasserstein reward distinguishes near misses from far misses in that case.
    • Empirically, the fraction of groups with all equal rewards and no learning signal drops from 13.8% to 3.6%, and 75.8% of the groups that were all zero regain a meaningful ranking.
    • The practical takeaway is that a much smaller model beats much larger ones, which is strong evidence that making answers temporally grounded and human-checkable works.
    原文摘要PDF
  15. 668Machine LearningarXiv:2607.18110

    LLM-as-a-Coach: Experiential Learning for Non-Verifiable Tasks

    • A single score is an information bottleneck because it cannot distinguish answers that receive the same top score, especially at the upper end.
    • The solution is to use the same feedback model as a coach rather than a judge. The coach turns answer evaluation into generalized, reusable advice, or experiential knowledge.
    • That advice is given to the teacher model as context and internalized by the policy through on-policy context distillation, which minimizes token-level reverse KL divergence between the teacher and the policy output. At inference time, neither the coach nor the advice context is needed.
    • The motivation is illustrated by feedback bandwidth. A 1 to 10 integer score carries 3.3 bits per sample, while a 1024-token piece of advice over a 150,000-word vocabulary theoretically carries 17,600 bits.
    • Across two policy families, EL outperformed rubric-based reinforcement learning on held-out and novel benchmarks, whether the feedback model was the policy itself or a commercial API model.
    • While RL is prone to reward hacking because it learns to chase reward on the training set, EL is a distribution-matching method that follows the teacher distribution and therefore reduces that failure mode.
    原文摘要PDF
  16. 667RoboticsarXiv:2607.18231

    FM-VLA: Force-based Memory for Vision-Language-Action Models in Contact-Rich Manipulation

    • It identifies a problem with contact-rich repetitive tasks: because the screen changes little, visual memory alone cannot track progress.
    • It first trains a variational autoencoder on force and torque time series with reconstruction as the objective, then freezes the encoder and uses the latent representation as memory tokens.
    • Using force memory alone makes pre-contact motion unstable, so it supplements it with a short record of joint states.
    • It achieves an average success rate of 83.3% over three tasks, outperforming pi-zero without memory at 27.8%, TA-VLA with a short force window at 22.2%, and visual-memory pi-MEM at 53.7%.
    • Inference latency increases by only 3.3 ms, from 60.7 ms to 64.0 ms, whereas visual memory grows from 100 ms to 190 ms depending on the number of frames.
    • Eight memory tokens were optimal, while increasing the count to 32 exceeded the token budget of the pretrained action module and hurt performance.
    原文摘要PDF
  17. 666RoboticsarXiv:2607.18236

    Patch Policy: Efficient Embodied Control via Dense Visual Representations

    • The core claim is simple: dense visual features, not a giant generative model, are what visual motor policies really need.
    • The method is a minimal extension. It concatenates all patch tokens from a frozen ViT and uses a block causal mask that lets tokens within a frame see each other while blocking information from the future across frames.
    • The policy head can be swapped. It works with both VQ-BeT and Diffusion Policy, and the encoder is not fine-tuned.
    • The results show a 40% relative improvement over global features on four simulation tasks, Push-T, LIBERO Goal, BlockPush, and Cube, and on three real Franka robot tasks.
    • Efficiency is a major strength. The VQ-BeT-based system runs in 10.99 ms at inference even with DINOv2 patches, compared with 61.71 ms for OpenVLA-OFT, and training takes 6.5 GPU hours versus 16 GPU hours.
    • Reducing the number of patches hurts performance. On Push-T, 256 patches score 0.69, while 64 patches score 0.52 and 1 patch scores 0.48, which shows that spatial resolution is the key to precise manipulation.
    原文摘要PDF
  18. 665RoboticsarXiv:2607.17977

    RynnBrain 1.1: Towards More Capable and Generalizable Embodied Foundation Model

    • The core claim is that it is not enough to understand vision well; the representation and output must be immediately usable for robot manipulation.
    • As a method, it adds contact-point prediction, meaning a grasp center and gripper-plane rotation angle, and native 3D localization as pretraining tasks, and 3D localization is included in the 2B and 9B models.
    • The architecture is a decoder-only vision-language model based on Qwen3.5, and it uses DeepStack and Interleaved MRoPE to handle images, multi-view images, and video over long contexts.
    • To unify multiple robots under one policy, it uses a shared action space aligned by body part and masks only the axes that each robot supports.
    • As a result, the average success rate over three real-robot tasks is 86.67 percent for RynnBrain-VLA, 73.33 percent for GR00T N1.7, 65.00 percent for π0.5, and 60.00 percent for the Qwen-based model.
    • The generalist version trained on multiple tasks and multiple robots performs better than task-specific training, with success rising from 86.67 percent to 91.67 percent, which supports the idea that heterogeneous robot data can complement one another.
    原文摘要PDF
  19. 664Computer VisionarXiv:2607.17965

    Exploration Matters for Escaping the Blur Trap in 3D Gaussian Splatting

    • The core argument is that blur in 3DGS comes not from insufficient data or model capacity, but from bias in the optimization dynamics themselves.
    • Theoretically, it proves that 3D position updates derived from 2D gradients are orthogonal to the camera-to-Gaussian viewing ray, and reports that the 2D position component dominates the other two components by two to three orders of magnitude throughout training.
    • It splits blur into two kinds: far-side blur caused by missing depth signals and near-side blur caused by alpha blending suppressing the signal to increase density.
    • The remedy is simple: periodically seed candidate Gaussians at random positions, prune bad candidates by opacity pruning, and randomly split some Gaussians regardless of accumulated gradients.
    • Across five datasets, Mip-NeRF 360, Tanks and Temples, Deep Blending, OMMO, and DL3DV, and in the 4DGS extension, image quality improves, with PSNR on Tanks and Temples rising from 23.73 to 24.37 and DL3DV from 27.16 to 28.43.
    • Ablation studies show that forcing more splits by lowering the density threshold only inflates the number of points without materially improving quality, confirming that the gain comes from exploration itself rather than from simply increasing the count.
    原文摘要PDF
  20. 663AI / GeneralarXiv:2607.15901

    DSWorld: A Data Science World Model for Efficient Autonomous Agents

    • It is the first to propose the concept of a data science world model that predicts state transitions in data science tasks.
    • It introduces cost-aware routing, where lightweight tasks are executed directly and heavier tasks are predicted by an LLM simulator.
    • It speeds up RL agent training by about 14 times and retrieval-based reasoning by 3 to 6 times.
    • On the transition-prediction task, it achieves accuracy that is 35.6 percent higher than the strongest LLM baseline.
    原文摘要PDF
  22. 661Information RetrievalarXiv:2607.15591

    RecGPT-V3 Technical Report

    • The Memory Hub manages user behavior histories as memory units compressed by 94.5%, so the system does not have to reprocess the full history on every request.
    • It solves the tag-item information bottleneck with a hybrid multimodal model in which the LLM directly infers Semantic IDs.
    • It compresses 3,000-token reasoning into 15 tokens using latent CoT tokens, a 200x reduction, while keeping the content recoverable when needed.
    • In live A/B tests, it improved IPV by 1.28%, CTR by 1.00%, and GMV by 3.97%, while cutting GPU cost by 52.4%.
    原文摘要PDF
  23. 660RoboticsarXiv:2607.15330

    Xiaomi-Robotics-1: Scaling Vision-Language-Action Models with over 100K Hours of Real-World Trajectories

    • An automatic VLM-based labeling pipeline is applied to more than 100,000 hours of real UMI manipulation trajectories.
    • The pretraining stage shows a clear scaling law, with performance improving consistently as both data and model size increase.
    • On RoboCasa365, it reaches 57.6 percent, which beats the previous best of 46.6 percent.
    • On a real robot, it learns new tasks from less than 10 hours of data and reaches an average success rate of 75 percent.
    原文摘要PDF
  24. 659Computer VisionarXiv:2607.16107

    Audio-Visual Flamingo: Open Audio-Visual Intelligence for Long and Complex Videos

    • It releases the AV-Skills dataset, consisting of about 7 million caption and QA instances.
    • It designs a three-stage curriculum that progresses from short-video recognition to reasoning over long videos of up to 15 minutes.
    • TAVIT precisely maps intermediate reasoning steps to video and audio timestamps.
    • It achieves 60.2% on MMOU, the best performance among comparable open-source models.
    原文摘要PDF
  25. 658LLMs / NLParXiv:2607.16051

    Loop the Loopies!

    • It introduces a layer-loop method and shows that it is more efficient for MoE than the existing model-loop approach.
    • It outperforms a vanilla 30B-A3B model under the same compute budget.
    • It records state-of-the-art math reasoning with 92.10% on AIME and 94.21% on AMC.
    • It achieves gold-medal-level performance at IMO 2025 and IPhO 2025 without tools.
    原文摘要PDF
  26. 657Machine LearningarXiv:2607.16097

    Understanding Reasoning from Pretraining to Post-Training

    • It runs 36 pretraining-RL combinations with models from 5M to 1B parameters and discovers a scaling law in which pretraining loss predicts post-RL performance.
    • RL increases the probability of existing correct answers on easy problems and surfaces answers that were almost absent during SFT on hard problems.
    • The result shows that allocating a larger share of the total compute budget to RL becomes optimal as the overall budget grows.
    • The same pattern appears not only in chess but also in math, across a 1B language model and 10B to 200B tokens, where longer pretraining improves RL efficiency.
    原文摘要PDF
  27. 656AI / GeneralarXiv:2607.13095

    Full-Pipeline Inference Optimization for MiMo-V2.5 Series: Pushing Hybrid SWA Efficiency to the Limit

    • It separates the KV cache for full attention and sliding window attention, compressing SWA storage to the theoretical limit of O(W), which is about 7x more efficient than before.
    • It builds GCache, an RDMA-based distributed cache infrastructure that shares KV caches efficiently across multiple nodes.
    • It redesigns the prefix cache tree to be SWA-aware and fixes accuracy bugs that the old prefix matching approach caused under SWA.
    • It also optimizes the multimodal pipeline end to end, including GPU image preprocessing, parallel video decoding from 156 seconds to 23 seconds for a 1-hour video, and cross-request batching for the encoder.
    • It improves MTP, or multi-token prediction, so that it also works in prefill and speeds up the first 128 decoded tokens by 2.3x.
    • It shows that resolving only the NUMA contention improves end-to-end performance by about 10%, which means system-level details matter a lot.
    原文摘要PDF
  28. 655Machine LearningarXiv:2607.13188

    Concurrent Image Understanding and Generation: Self-Correcting Coupled Markov Jump Processes

    • It generates text and images simultaneously in a single decoding process, with each side verifying the other's decisions.
    • It proposes CO2Jump, a training-free sampler that outperforms prior methods on image editing and visual reasoning tasks.
    • It adds a remasking mechanism so that tokens once fixed can be hidden again when evidence changes, allowing the model to correct contradictions on its own.
    • It builds and plans to release three large multimodal datasets: JEdit-1M, JMaze-200K, and JNono-200K.
    • Performance improves steadily as the number of denoising steps increases, showing that the benefits of cross-modal coupling accumulate.
    原文摘要PDF
  29. 654Machine LearningarXiv:2607.15207

    BadWAM: When World-Action Models Dream Right but Act Wrong

    • We are the first to uncover a new vulnerability in WAMs, showing that imagination and action can be separated.
    • An attack that targets action only reduces robotic task success from 96.5% to 43.1%, cutting it by more than half.
    • We design a stealth attack that preserves imagination while quietly corrupting only action, demonstrating that existing safety monitors can be bypassed.
    • The attack transfers from one WAM variant to another, confirming that the vulnerability is not specific to a single implementation.
    • Simple preprocessing defenses are not enough, which suggests a new safety approach that directly checks synchronization between action and imagination.
    原文摘要PDF
  30. 653Machine LearningarXiv:2607.13491

    DeepLoop: Depth Scaling for Looped Transformers

    • The authors mathematically prove that the gradient accumulation in looped transformers with shared layers makes the usual DeepNorm scaling insufficient.
    • Simply increasing the exponent from one quarter to one half, with alpha equal to (2N) to the one-half power and beta equal to (8N) to the negative one-half power, stabilizes training as the number of loops grows.
    • On GPT-2 small and medium scales, increasing the number of loops consistently improves performance over prior methods.
    • When applied to the hierarchical reasoning model, or HRM, on the ARC-AGI reasoning benchmark, accuracy rises from 36.5 percent to 39.75 percent, which is a gain of 3.25 percentage points.
    • Because it changes only the scaling formula and adds no extra parameters, the method is essentially free to apply.
    原文摘要PDF
  31. 652Computer VisionarXiv:2607.15278

    Hierarchical Denoising For Multi-Step Visual Reasoning

    • We solve the limitations of both fast but dumb streaming and smart but slow bidirectional video methods with a tree-structured approach.
    • Across six multi-step reasoning tasks, including maze solving, Sokoban, and Hanoi towers, success rises from 34.22 to 60.29, a 76.2% improvement.
    • The streaming version runs at 0.70 seconds per frame, 54x faster than bidirectional diffusion, while still achieving higher reasoning accuracy.
    • Using only 2% of the training data preserves 82.9% of the performance of the full data, showing strong data efficiency.
    • Real robotic arm maze experiments also show robust success, confirming applicability in the physical world.
    原文摘要PDF
  32. 651Machine LearningarXiv:2607.15161

    On-Policy Delta Distillation

    • It argues that using the difference between a teacher's pretraining and post-training states, rather than the teacher's final output, transfers reasoning ability more accurately.
    • It consistently outperforms existing on-policy distillation methods across 14 benchmarks in math, science, and code.
    • It is validated broadly on recent models such as Qwen3, from 1.7B to 8B, and Gemma4, which makes the result credible.
    • The additional compute cost is only 8% to 28%, so the method can be added to existing distillation pipelines easily.
    • Its analysis shows that the delta signal concentrates on reasoning tokens such as hence and however, while suppressing exploratory tokens such as try and verify.
    原文摘要PDF
  33. 650LLMs / NLParXiv:2607.15178

    T^2MLR: Transformer with Temporal Middle-Layer Recurrence

    • It shows that if only the middle layers, the transformer’s thinking layers, are connected recurrently, reasoning improves more than when the entire stack is repeated.
    • Recursing through only 20% of the full stack is enough, which maximizes reasoning capability while minimizing inference latency.
    • Even just adding the recurrent path and fine-tuning an already trained 1.7B model greatly improves mathematical reasoning accuracy, with GSM8K up 4.1 points and MATH500 up 5.2 points.
    • The inference overhead is only about 8%, so it can be used without much burden in production.
    • Training cost increases by roughly 2x to 4x, but it is a meaningful tradeoff given the gains in inference efficiency and performance.
    原文摘要PDF
  34. 649AI / GeneralarXiv:2607.15257

    SearchOS-V1: Towards Robust Open-Domain Information-Seeking Agent Collaboration

    • It formalizes information retrieval as relational schema completion plus citation evidence, so progress can be measured at the entity and attribute level.
    • SOCM, with Frontier Task, Evidence Graph, Coverage Map, and Failure Memory, externalizes implicit progress into an explicit shared state.
    • Pipeline parallel scheduling assigns the next task immediately when an agent slot becomes free, which maximizes utilization.
    • The Search Tool Middleware Harness handles evidence extraction, budget enforcement, and loop detection outside the agent prompt.
    • It reaches 76.5 F1 on GISA, which is 13.4 points better than the strongest baseline, and 80.3 F1 on WideSearch, where it takes first place on every metric.
    原文摘要PDF
  35. 648Machine LearningarXiv:2607.14530

    xHC: Expanded Hyper-Connections

    • It is the first method to extend Hyper-Connection beyond N=4 to N=16, because prior mHC methods suffered strong diminishing returns above N=4.
    • Temporal feature augmentation enriches write-back with nearby token information, and sparse residual updates at k=4 or N=16 balance cost and efficiency.
    • On an 18B MoE model, it delivers a +4.0 point gain over mHC while adding only 4.1% more FLOPs than vanilla.
    • The scaling law shows that vanilla requires 1.50x compute and mHC requires 1.19x compute to reach the same loss, using xHC as the reference.
    • xHC-Flash keeps sublayer memory traffic at 40C even with N=16, which is similar to mHC at N=4 with 34C.
    原文摘要PDF
  36. 647Computer VisionarXiv:2607.14749

    WanSong v1.0 Technical Report

    • It abandons the autoregressive pipeline and uses pure diffusion to generate a five-minute song with vocals and background music as dual stems in a single pass.
    • Dual-stem tokens model vocals and accompaniment separately, so both can be generated at high quality without the usual quality drop from classifier-free guidance.
    • A 1024-compression VAE, which is half the compression of the previous 2048 setting, preserves fine acoustic detail and improves SI-SDR by 2.86 dB.
    • RLHF is used to optimize musicality, lyric accuracy, and prompt alignment, yielding a musicality score of 5.49 compared with 4.18 for Suno and 3.83 for Mureka.
    • Stable long-song generation is achieved through 6 million hours of multilingual data and three-stage curriculum training from 90 seconds to 300 seconds to SFT.
    原文摘要PDF
  37. 646Computer VisionarXiv:2607.15273

    MeanFlowNFT: Bringing Forward-Process RL to Average-Velocity Generators

    • It is the first combination of forward-process RL with MeanFlow, which generates 1 to 4 steps by predicting interval-averaged velocity.
    • It applies the DiffusionNFT objective through an instantaneous velocity predictor derived from MeanFlow identity, while sampling still uses average velocity.
    • It theoretically inherits the policy improvement guarantee of DiffusionNFT.
    • On SD3.5-M, it beats SOTA on 6 of 8 metrics, and on Wan2.1 video it uses 4 steps to beat LongCat-Video RL's 50-step score of 82.57 with 84.33 VBench.
    • It shows consistent gains under test-time scaling as the number of steps increases.
    原文摘要PDF
  38. 645RoboticsarXiv:2607.15065

    DriftWorld: Fast World Modeling through Drifting

    • It is the first application of a drifting generative model, which generates in a single forward pass, to a robotic world model and achieves 30-plus fps, 17 times faster than diffusion.
    • It proposes three adaptations: an action-conditioned drifting field, a loss in the DINOv2 feature space, and frame-wise action conditioning.
    • On Push-T, it raises IoU from 0.635 to 0.781, which greatly improves action-exploration quality, and it shows consistent gains across 5 environments.
    • It is also valid as an offline policy-evaluation metric and achieves a Pearson correlation of 0.92 to 0.99 with actual performance.
    • It proves that fast and accurate imagination directly helps both robot planning and policy evaluation.
    原文摘要PDF
  39. 644AI / GeneralarXiv:2607.13285

    Harness Handbook: Making Evolving Agent Harnesses Readable,Navigable, and Editable

    • It defines the behavior localization problem as finding where in the code a behavior described by a change request is implemented.
    • Harness Handbook reorganizes code by behavior rather than by file or function, and links each behavior directly to source code.
    • Behavior-Guided Progressive Disclosure (BGPD) guides exploration step by step from high-level behavior descriptions to relevant implementation details.
    • On two open-source harnesses, Codex and Terminus-2, it outperforms the baselines with win rates of 38.3% and 45.6%, compared with 28.3% and 26.7%.
    • It delivers especially large gains on multi-file edits, rarely executed code paths, and interactions across modules.
    原文摘要PDF
  40. 643RoboticsarXiv:2607.15163

    Scaling Behavior Foundation Model for Humanoid Robots

    • It adopts motion tracking as a unified training paradigm, combining diverse whole-body control problems under reference-motion imitation.
    • Experiments show that the synergy between the quantity of on-policy rollouts and the diversity of reference motions is the key to scaling.
    • It introduces the Humanoid Transformer, a scalable architecture in which structured action representations emerge naturally.
    • Compared with existing humanoid controllers, it reduces MPKPE by 10% locally and 82% globally, and reaches an overall accuracy of 0.9836.
    • It is validated in both simulation and on real robots, confirming its practicality.
    原文摘要PDF
  41. 642Computer VisionarXiv:2607.15038

    Video = World + Event Stream

    • It proposes a decomposition in which video equals world plus event stream, and uses it to define a general-purpose pretraining task.
    • It adds free-form actions, expressed as natural-language instructions in parentheses, to the event stream so that the agent can perform open-vocabulary actions rather than a fixed action set.
    • It maintains real-time duplex audio-visual dialogue with 200 ms model response latency and 550 ms end-to-end interaction latency.
    • It demonstrates flexible transfer from pretraining to downstream tasks, suggesting that the same streaming logic can adapt to a range of real-time problems.
    原文摘要PDF
  42. 641RoboticsarXiv:2607.15275

    RoboTTT: Context Scaling for Robot Policies

    • It integrates Test-Time Training, fast weights, into a VLA model to achieve an 8K time-step context without increasing inference latency.
    • With just one human video demonstration, it gains one-shot imitation of a new behavior, succeeding in 6 out of 10 cases while all baselines fail.
    • DAgger Distillation enables real-time performance improvement, reaching 83% success under disturbances, compared with a previous best of 53%.
    • The overall task success rate improves by 87% over the single-step baseline, and it completes a 5-minute, 10-step assembly all the way through, which no previous model could do.
    • Expanding context length from 1K to 8K yields 62% better performance, showing that context is a scaling axis for robots as well.
    原文摘要PDF
  43. 640Computer VisionarXiv:2607.14935

    VideoChat3: Fully Open Video MLLM for Efficient and Generalist Video Understanding

    • The 3D vision Transformer, I3D-ViT, processes frames as spatiotemporal bundles rather than separate images, which greatly reduces token count.
    • An adaptive mechanism that adjusts frame resolution by importance improves efficiency for real-time streaming.
    • It releases 3 million high-quality multimodal instruction examples, including Academic2M, LV116K, and OL617K.
    • With a 4B model, it outperforms previous open-source models of the same class on 12 benchmarks, including MotionBench, VideoMME, and LVBench.
    • It releases the model, code, training strategy, dataset, and pipeline, which provides a reproducible foundation.
    原文摘要PDF
  44. 639LLMs / NLParXiv:2607.14777

    SEED: Self-Evolving On-Policy Distillation for Agentic Reinforcement Learning

    • It creates a self-evolving loop that automatically extracts natural-language skills from completed trajectories and feeds them back into the policy.
    • It converts action-probability changes with and without a skill into a dense token-level supervision signal, OPD, and combines that with RL.
    • On ALFWorld, it reaches 91.8 percent, which is 7.4 points better than the previous best, and it beats GRPO trained on the full dataset even when using only 40 percent of the data.
    • It generalizes strongly to unseen environments, confirming that it learns real strategy rather than memorization.
    • It improves performance using only the trained policy, without extra prompts or external memory at inference time.
    原文摘要PDF
  45. 638Computer VisionarXiv:2607.13639

    OvisOCR2 Technical Report

    • Given a document-page image, it generates text, equations, tables, and visual regions as Markdown in a natural reading order in one pass.
    • It combines real document annotations with synthetic pages whose images and ground truth are generated from the same HTML, applying supervised learning, reinforcement learning, on-policy distillation, and model merging.
    • The 0.8B model achieved the highest scores on both benchmarks: 96.58 on OmniDocBench v1.6 and an Avg3 of 75.06 on PureDocBench.
    • By outperforming pipeline approaches with a single small model, it suggests the possibility of simplifying deployment configurations for document-retrieval and knowledge-building systems.
    • However, validation is limited to two public benchmarks and an in-house benchmark for converting page images to Markdown, so it does not guarantee performance in other document-processing workflows.
    原文摘要PDF
  46. 637Computer VisionarXiv:2607.12450

    Let RGB Be the Language of Vision

    • It converts structured visual information such as masks, depth maps, and poses into RGB images, representing multiple tasks as a common RGB-to-RGB editing problem.
    • A single general-purpose image-editing backbone handles understanding tasks and conditional-generation tasks without adding task-specific encoders or decoders.
    • It showed competitive zero-shot performance without separate fine-tuning on tasks including segmentation, depth estimation, and pose-conditioned generation.
    • By connecting visual understanding and generation through the same input-output format, it can serve as a common interface for general-purpose vision-language systems.
    • A limitation is that performance depends heavily on a web-scale-trained image-editing backbone, and gaps remain relative to dedicated models on some specialized tasks.
    原文摘要PDF
  47. 636AI / GeneralarXiv:2607.13034

    Do AI Agents Know When a Task Is Simple? Toward Complexity-Aware Reasoning and Execution

    • It defined ACRR to measure unnecessary agent work and proposed E3, which estimates scope, executes the minimal path, and expands only when verification fails.
    • Across 121 edits in MSE-Bench's controlled simulator, it maintained the same 100% success rate as the strongest baseline.
    • E3 reduced cost by 85%, tokens by 91%, and files inspected by 92%, and was also 16% better than the adaptive-search baseline.
    • It can be applied as an execution policy that reduces the waste of reading an entire repository for a small change while retaining verification as a safeguard.
    • The main figures come from a controlled simulator rather than a deployed agent, and real-model validation was limited to a gpt-4o case study, so broader validation is needed.
    原文摘要PDF
  48. 635Computer VisionarXiv:2607.14076

    From Pixels to States: Rethinking Interactive World Models as Game Engines

    • It views interactive world models through the action-state-observation loop of traditional game engines and classifies related research along four dimensions.
    • The four dimensions are player-action control, game-state transitions, long-term persistence of states and observations, and real-time generation.
    • It presented a data engine that collects more than 90 hours of gameplay in Black Myth: Wukong together with frame-level actions, actual game states, and visual observations.
    • The framework and data can serve as reference material for research on next-generation generative game engines that track explicit states beyond predicting screen images.
    • The paper focuses on classifying methods and collecting data; it does not present a complete world model that simultaneously satisfies rule adherence, persistence of long-term consequences, and real-time generation.
    原文摘要PDF
  49. 634Computer VisionarXiv:2607.14088

    VideoRAE: Taming Video Foundation Models for Generative Modeling via Representation Autoencoders

    • It compresses features from multiple layers of a frozen video foundation encoder through a lightweight one-dimensional self-attention projector.
    • The continuous latent representations can be used in a diffusion Transformer, while discrete tokens produced through multi-codebook quantization can be used in an autoregressive model.
    • On UCF-101, the autoregressive and diffusion generators achieved gFVD scores of 40 and 93, respectively, and converged approximately 5× faster than competing autoencoders.
    • It can accelerate video-generation training by converting semantically rich frozen video representations into a reconstructable generative latent space.
    • Generation results were verified primarily on UCF-101 and in controlled 2B-scale text-to-video experiments, leaving validation on larger models and more varied video domains outstanding.
    原文摘要PDF
  50. 633Machine LearningarXiv:2607.13988

    TRACE: Turn-level Reward Assignment via Credit Estimation for Long-Horizon Agents

    • At each state on a tool-call boundary, it computes the log-probability of the correct answer using a frozen reference model, then uses temporal differences in log-ratio state values as per-action rewards.
    • Without a separate critic or process labels, it provides finer-grained turn-level credit than outcome rewards, and values cancel across unnecessary tool calls.
    • On BrowseComp-Plus, Qwen3-4B improved from 7.2 to 35.6, and Qwen3-30B-A3B from 8.4 to 42.6.
    • By teaching useful intermediate actions in long search processes, it accelerated convergence in pure reinforcement learning and transferred behavior to public-web evaluations.
    • Current value estimation is tailored to search with short, known correct answers, so whether the same log-probability criterion is suitable for long code patches or open-ended answers remains unverified.
    原文摘要PDF