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. 632Computer VisionarXiv:2607.13125

    Boogu-Image-0.1: Boosting Open-Source Unified Multimodal Understanding and Generation

    • It developed four models, Base, Turbo, Edit, and Edit-Turbo, combining a stronger multimodal encoder with agentic prompt rewriting.
    • It jointly refined data quality, training procedures, and inference-time scaling so that improvements in understanding translated into image generation and editing.
    • Using 208.62 million unique images and a theoretical base-model training cost of approximately $400,000, it matched or exceeded several public models on benchmarks.
    • It released weights, code, and training methods under Apache 2.0, allowing reuse in unified image-generation and editing research with smaller budgets.
    • Although it approached leading closed systems, it did not report outperforming them overall, and the comparisons are limited to the scope of published standard evaluations.
    原文摘要PDF
  2. 631LLMs / NLParXiv:2607.12395

    Ring-Zero: Scaling Zero RL to a Trillion Parameters for Emergent Reasoning

    • It stabilized zero reinforcement learning for a 1-trillion-parameter model by combining importance-ratio clipping, training-inference ratio correction, and mixed-precision control.
    • Scaling the model to 1 trillion parameters increased sample efficiency and the performance ceiling, with training progressing from a discovery phase to a refinement phase.
    • Behaviors such as structured formatting, self-verification, parallel reasoning, and context anxiety emerged without explicit rules.
    • It proposed a framework for evaluating reasoning processes through understandability, reproducibility, and efficiency as well as accuracy, providing a tool for analyzing large-scale reasoning training.
    • Empirical evidence is concentrated on seven math benchmarks and a 1-trillion-parameter configuration, leaving open whether the results generalize directly to other tasks and smaller computation budgets.
    原文摘要PDF
  3. 630RoboticsarXiv:2607.13960

    GigaWorld-Policy-0.5: A Faster and Stronger WAM Empowered by AutoResearch

    • It uses an action-centric WAM that jointly learns future visual changes and actions during training, but decodes only actions at inference without generating video.
    • It combines AC-WM and WAM pretraining and assigns visual changes and action generation to separate Transformer experts.
    • In a local RTX 4090 environment, it reduced action-inference latency to 85 milliseconds while retaining the policy-learning benefits of future visual supervision.
    • Agent-based AutoResearch automates the search over training configurations, potentially reducing the development burden of real-time closed-loop robot control.
    • The 85-millisecond figure was measured in a specific RTX 4090 deployment environment, so it does not guarantee latency or performance on other hardware or a wider variety of robot tasks.
    原文摘要PDF
  4. 629Computer VisionarXiv:2607.12800

    UniVR: Thinking in Visual Space for Unified Visual Reasoning

    • VR-GRPO combines overall outcome rewards with rewards for error-prone steps to learn logical and physical consistency in visual reasoning.
    • It built the VR-X benchmark from 16 sources, bringing long-horizon manipulation, spatial puzzles, and physical reasoning together in a purely visual format.
    • UniVR improved by up to 25% on VR-X, and stronger visual reasoning also improved performance on multiple multimodal-understanding benchmarks.
    • It released code, data, and models for research on planning and reasoning in visual space without textual explanations.
    • The 34B model and long-video training require substantial computational resources, and step rewards also depend on a general-purpose VLM that lacks detailed physical knowledge.
    原文摘要PDF
  5. 628RoboticsarXiv:2607.12992

    ChunkFlow: Towards Continuity-Consistent Chunked Policy Learning

    • It divides each action chunk into fixed, revisable, and future regions and deterministically blends the overlapping portions during execution.
    • It trains with seam loss and first- and second-order continuity losses, using history corruption, scheduled sampling, and AWAC fine-tuning to improve robustness to execution errors.
    • In CALVIN, LIBERO, and real-robot experiments, it improved the balance between success rate and motion stability while maintaining low inference latency.
    • By reducing jitter at chunk boundaries during training rather than through post-processing filters, it can be applied to real-time VLA robot control.
    • The theoretical result that seam discrepancies decrease with overlap length holds under mild smoothness assumptions, and the experiments are limited to three environments.
    原文摘要PDF
  6. 627RoboticsarXiv:2607.13033

    DenseReward: Dense Reward Learning via Failure Synthesis for Robotic Manipulation

    • It automatically synthesized trajectories of physical failures such as collisions, failed grasps, dropped objects, and recovery in simulation without human labels.
    • It trained a dense reward model that takes visual observations and language instructions and scores task progress at every frame.
    • In both simulated and real manipulation, it predicted dense rewards more accurately than general-purpose VLMs and existing robot reward models.
    • These rewards provide fine-grained progress signals for model-predictive control and reinforcement learning and can be used to improve robot policies after imitation learning.
    • The current scope focuses on relatively short manipulation tasks, leaving tool use, long-horizon work, and rewards reflecting human preferences for future work.
    原文摘要PDF
  7. 626LLMs / NLParXiv:2607.13027

    PalmClaw: A Native On-Device Agent Framework for Mobile Phones

    • It implements an open-source framework that directly manages sessions, memory, skills, tools, and the agent loop on a smartphone.
    • It reduces lengthy screen interactions by exposing device capabilities as tools with explicit arguments, structured results, and clear execution boundaries.
    • On MobileTask, its task success rate was 11.5% higher in relative terms than the strongest baseline, and completion time was 94.9% shorter.
    • It can run mobile automation close to user data and sensors on the device while controlling permissions and task scope at the tool level.
    • The results were verified on a specific Android device and MobileTask configuration, and using new device capabilities requires separately providing safe, dedicated tools and permission boundaries.
    原文摘要PDF
  8. 625AI / GeneralarXiv:2607.12463

    Function-Aware Fill-in-the-Middle as Mid-Training for Coding Agent Foundation Models

    • It designed function-aware FIM mid-training by exploiting the similarity between a coding agent's action-observation-continuation structure and the function call-return structure.
    • It selected functions to mask using program-dependence graphs and criteria for complexity and inferability, then trained on 2.6 billion tokens from 968 GitHub repositories.
    • SWE-Bench-Verified improved by 2.8, 3.0, and 3.2 points on Qwen-family 7B, 14B, and 8B models, respectively.
    • The effect persisted across multiple agent post-training procedures and also mitigated capability degradation in general coding and non-coding tool use.
    • The training corpus and evaluations focused on Python, and generating reasoning explanations depended on a teacher model, so other languages and teacher-free settings remain unverified.
    原文摘要PDF
  9. 624Computer VisionarXiv:2607.12477

    Self in Space: Benchmarking Self-Awareness and Spatial Cognition in UAV Embodied Intelligence

    • SIS-Bench evaluates UAV self-spatial understanding by crossing the two axes of spatial cognition and self-awareness with three stages: perception, memory, and reasoning.
    • Through expert validation, it created 13 tasks and 4,856 question-answer pairs from 1,646 real UAV videos.
    • Current MLLMs were weaker at self-awareness than spatial cognition, and their performance declined progressively from perception to memory and reasoning.
    • SIS-Motion, which combines optical flow and visual features, improved perception and memory along both axes, as well as downstream UAV decision-making performance.
    • The evaluation mainly remains at the level of video-based question answering and does not cover agents that integrate multiple sensors to carry out actual long-duration autonomous flight.
    原文摘要PDF
  10. 623Computer VisionarXiv:2607.12592

    WanToFight: Real-Time Generative Game Engine for Multi-Player Combat Interaction

    • It implemented streaming autoregressive generation by adding block-causal attention and a rolling KV cache to the Wan-1.3B video diffusion model.
    • A player-binding module associates each keyboard input with a character, and a local causal key-input module is trained through a staged curriculum.
    • A 4-step DMD-distilled model and a pruned VAE generated 512×384 video at 30 frames per second throughout a complete match on a single RTX 5090.
    • It shows the potential to handle adversarial interaction between two players and complex physics together in one real-time generative engine.
    • It is currently limited to a single game, KOF '97, with remaining issues in fine visual detail for complex actions and accumulated error in sessions longer than 2 minutes.
    原文摘要PDF
  11. 622ResearcharXiv:2607.12264

    Heisenberg Scaling in Many-Body Kinetic Uncertainty Relation via Quantum Feedback

    • It analyzes the counting precision of a superradiant spin ensemble using a many-body kinetic uncertainty relation and mean-field equations incorporating feedback.
    • Feedback after quantum jumps sustains transient superradiant activity, maintaining collective effects that depend on the particle number N.
    • Both analytical results and numerical simulations showed Heisenberg-like scaling, with the variance of the counting observable decreasing as 1/N².
    • It proposes a design principle for turning collective dissipation from noise into a resource for precision in quantum clocks and counting devices.
    • The results are based on theory and numerical simulations of a specific superradiant spin model, and no experimental validation on actual devices was presented.
    原文摘要PDF
  12. 621ResearcharXiv:2607.12772

    Quantum States Prepared by Wormholes: Long-Wavelength Deviations from Bunch-Davies

    • It calculated the quantum states of tensor perturbations and test-scalar-field perturbations prepared when a wineglass-shaped wormhole creates a new universe.
    • It showed that the Bunch-Davies vacuum is reproduced at short wavelengths, while wormhole-charge-dependent deviations arise at long wavelengths.
    • These deviations leave small shifts and oscillations in the angular power spectrum produced by the subsequent inflation.
    • As the charge approaches 0, scalar and tensor deviations are suppressed, providing evidence that the topological transition to a no-boundary instanton is smooth.
    • These are cosmological predictions derived from idealized wormhole geometry and test-field calculations, not detections of signals in actual observational data.
    原文摘要PDF
  13. 620Computer VisionarXiv:2607.12894

    Hy-Embodied-VLM-1.0: Efficient Physical-World Agents

    • It organizes pretraining and post-training data around a three-stage capability taxonomy: understanding action-relevant states, reasoning about action transitions, and sequential and adaptive reasoning.
    • It limits active parameters to 3B using the Hy3-A3B language backbone, the Hy-ViT2 visual encoder, and a mixture-of-experts architecture.
    • It achieved the highest performance among similarly sized models on 19 of 38 benchmarks and improved by an average of 8.4% over the previous Hy-Embodied-0.5.
    • By supporting multi-turn interaction and long-horizon reasoning with few active parameters, it can serve as a foundation model for latency-sensitive physical-world agents.
    • The model with 3B active parameters only approaches the previous generation's model with 32B active parameters rather than surpassing it on every evaluation, and its top performance was limited to half of the 38 benchmarks.
    原文摘要PDF
  14. 619RoboticsarXiv:2607.13017

    FlowWAM: Optical Flow as a Unified Action Representation for World Action Models

    • It represented actions as optical flow in the same format as RGB video and jointly diffusion-modeled RGB and flow within a shared video generator.
    • In policy mode, it predicts actions from flow; in world-model mode, target flow guides future video. Videos without action labels are also used for pretraining.
    • Success rates on RoboTwin were 92.94% for Clean and 92.14% for Random, and it achieved an EWMScore of 63.71 on WorldArena.
    • By connecting robot policies and future-scene prediction through a single video-friendly representation, it can use motion information from large-scale general video for control learning.
    • The reported figures come from RoboTwin, WorldArena, and a limited set of real-robot tasks, so generality across other robot embodiments and contact-centered tasks requires further validation.
    原文摘要PDF
  15. 618AI / GeneralarXiv:2607.11185

    SCALECUA: Scaling Computer Use Agents with Verifiable Task Synthesis and Efficient Online RL

    • VeriGen automatically synthesizes verifiable computer-use tasks through Docker interactions and feedback from multiple agents.
    • It uses Frontier Sampling, which allocates executions to each task's current capability frontier, and Visual Context Segmentation, which processes only recent screens in separate segments.
    • It created more than 24,000 verifiable tasks and approximately 3,000 high-quality reinforcement-learning tasks, increasing training speed by 2.83 times.
    • ScaleCUA scored 68.7% on OSWorld and 54.0% on ScienceBoard and can be used for large-scale online training of open computer-use agents.
    • Each execution is limited to 50 turns, and evaluation centers on the Ubuntu desktop, so generalization to very long-horizon tasks and Windows or macOS has not been verified.
    原文摘要PDF
  16. 617ResearcharXiv:2607.11762

    Operational Concealment of Measurement Incompatibility by Quantum Channels

    • It defined operational concealment as the phenomenon in which measurement incompatibility can no longer be certified when all input states pass through a quantum channel.
    • In finite dimensions, it proved that some binary POVM pair is exactly concealed only if the channel's Heisenberg adjoint has a nontrivial kernel.
    • It distinguished injective dynamics that only contract from cases where information rank decreases, and derived an exact expression for concealment robustness in qubit channels.
    • Using quotient-space duality, it interpreted this robustness as an experimentally accessible reward advantage in channel-output games.
    • The analysis is limited to finite-dimensional, finite-outcome measurements, exact knowledge of the channel, complete input tomography, and single use; it does not cover adaptive multiple uses.
    原文摘要PDF
  17. 616ResearcharXiv:2607.11637

    From Circuits to Hardware: Benchmarking Standard and Qubit-Efficient Quantum Optimization on Real Hardware

    • It compares VQE, CVaR-VQE, several QAOA variants, PCE, and QRAO using the same real-hardware procedure for four NP-hard problems.
    • On IBM Heron r1 and r2, it records transpiled circuit sizes, execution results, and gate-count-based fidelity estimates for 247 method-problem combinations.
    • At intermediate Heron-r2 error rates, noise began to dominate execution around F_est≈0.1, corresponding to approximately 770 two-qubit gates.
    • No method produced a feasible solution for QAP, and qubit-efficient methods also ran larger problems only within the fidelity budget.
    • The conclusions concern the tested implementations and hardware; they do not prove asymptotic inferiority of the entire QAOA family or quantum advantage over classical methods.
    原文摘要PDF
  18. 615Computer VisionarXiv:2607.09450

    Robustifying Vision-Language Models via Test-Time Prompt Adaptation

    • RITA reduces adversarial outliers by using optimal transport to align the feature distribution of multiple augmentations of a test image with the text-prototype distribution.
    • A dynamic cache accumulates reliable cues from the test stream to continually refine prompts online.
    • Across multiple experiments, it improved VLM robustness to adversarial attacks without reducing clean-image accuracy.
    • Because it requires neither model retraining nor access to labeled data, it can be added to existing CLIP-family classification systems as a test-time defense.
    • Current validation centers on image classification. Extension to generative tasks such as image captioning remains open, and the method does not eliminate every risk from new attacks.
    原文摘要PDF
  19. 614Computer VisionarXiv:2607.11588

    FoundationGeo: Learning Spatial Pixel-Wise Fields for Monocular Metric Geometry

    • Stage 1 uses DINOv3 initialization and a 10.2M multi-domain corpus to learn affine-invariant, high-quality relative geometry.
    • Stage 2 applies a pixel-wise scale field and ray-direction correction field for spatially varying metric correction.
    • We identify focal-length distribution mismatch as the key bottleneck in zero-shot metric generalization.
    • By augmenting with Blender-based synthetic data over a wider range of focal lengths, we improve robustness to intrinsics shift.
    • Across seven zero-shot benchmarks, it leads by an average of more than 5.2%, preventing large drops across domains.
    • It outperforms existing metric methods such as DepthPro, Metric3D V2, UniDepthV2, and MoGe-2.
    原文摘要PDF
  20. 613Computer VisionarXiv:2607.11862

    Evidence-Backed Video Question Answering

    • E-VQA defines a new task that requires three outputs: an answer, temporal evidence, and spatial evidence in the form of masklets.
    • ST-Evidence is the first human-verified spatiotemporal grounding benchmark, with both generative and objective variants.
    • It shows that state-of-the-art models such as Qwen3-VL, OpenAI-o3, and Gemini-2.5-Pro have decoupled QA accuracy and grounding ability.
    • An automatically generated 160K-scale ST-Evidence-Instruct pipeline bridges the gap between reasoning and grounding.
    • Fine-tuning UniPixel 7B brings large gains, with t-mean up by 27.2 and J&F up by 13.8.
    • Scaling alone is not enough; the task requires fundamental advances in architecture, data, and training objectives.
    原文摘要PDF
  21. 612ResearcharXiv:2607.11738

    Qwen-Audio-VAE Technical Report

    • We reduce downstream DiT training cost by using a 12.5 Hz low-bit-rate latent space.
    • An asymmetric encoder-decoder with latency-aware encoder pruning encodes 32 minutes in 541 ms, a 3.62x speedup.
    • Trained on 5 million hours of multi-domain audio, it generalizes across speech, music, and sound-effect domains.
    • Four discriminators, multi-period, multi-resolution STFT, multi-scale STFT, and sub-band CQT, provide adversarial supervision.
    • It achieves robust reconstruction quality across domains on the public LibriSpeech, AudioCaps, and SongDescriber benchmarks.
    • It outperforms the previous best VAE on STFT distance, SI-SDR, and related metrics.
    原文摘要PDF
  22. 611LLMs / NLParXiv:2607.11487

    LightMem-Ego: Your AI Memory for Everyday Life

    • It aligns visual and audio streams on a timeline and segments them into events.
    • It incrementally organizes memory into three levels: current, short-term, and long-term.
    • It designs a memory router that dynamically routes queries according to their temporal scope and intent.
    • It demonstrates four core scenarios: finding objects, recalling conversations, daily summaries, and routine discovery.
    • It uses a lightweight architecture that can be deployed on smartphones and AI glasses.
    • It outperforms baselines without explicit multimodal memory in retrieval, QA, and latency.
    原文摘要PDF
  23. 610Computer VisionarXiv:2607.11562

    MonkeyOCRv2: A Visual-Text Foundation Model for Document AI

    • It builds MonkeyDoc v2, the largest document image pretraining corpus, with 113 million images across 17 languages.
    • It jointly learns image-to-text generation and pixel-level document reconstruction to achieve character-level visual recognition.
    • It improves CRNN recognition accuracy from 58.7 percent to 67.3 percent, and the 110M UniMERNet-T outperforms the 325M UniMERNet-B.
    • A 0.7B document parsing model achieves open-source SOTA on MDPBench, beating the previous best, the 3B dots.mocr, by 2.8 percent.
    • With the vision encoder frozen, it achieves stronger document understanding than large VLMs such as Qwen3-VL-235B and GPT-5.2.
    • It shows consistent gains over CLIP, DINO, and SAM across 8 benchmarks.
    原文摘要PDF
  24. 609RoboticsarXiv:2607.11689

    From World Action Models to Embodied Brains: A Roadmap for Open-World Physical Intelligence

    • It defines WAMs as action-conditioned predictive models and classifies them as models that explicitly predict the results of interventions.
    • It identifies three structural limitations: the model-representation gap, the objective-standardization gap, and the ecosystem-system gap.
    • An embodied brain is the central model that integrates physical observations and compares interventions to output intended state transitions.
    • A physical harness converts the brain's outputs into execution tools and controllers, validates them, and records traces.
    • Shared contracts make heterogeneous data, tasks, and components mutually comparable.
    • Closed-loop post-training turns verified interactions into reusable experience.
    原文摘要PDF
  25. 608Computer VisionarXiv:2607.11886

    Read It Back: Pretrained MLLMs Are Zero-Shot Reward Models for Text-to-Image Generation

    • We use image-conditioned prompt log-likelihood from an MLLM as the reward, which removes the need for preference labels and fine-tuning.
    • Self-SpectraReward is a closed-loop self-improvement scheme in which the understanding branch of a unified model (UMM) evaluates its own generation branch.
    • We validate it at scale with 9 MLLM backbones from 4B to 235B, 4 model families, 3 RL algorithms, and 2 diffusion models.
    • A larger reward MLLM is not always better; reward-policy alignment matters more than reward-model size.
    • Self-SpectraReward outperforms a 235B external model, demonstrating the importance of policy-reward alignment.
    • It consistently outperforms prior MLLM-derived reward methods on five OOD text-to-image benchmarks.
    原文摘要PDF
  26. 607LLMs / NLParXiv:2607.11881

    Metacognition in LLMs: Foundations, Progress, and Opportunities

    • It defines metacognition as a loop of monitoring, meaning state evaluation, and control, meaning strategy adjustment, and uses that lens to classify LLM research.
    • It organizes a range of metacognitive measurement methods, including psychology-based, neurofeedback-based, confidence-based, and interpretability-based approaches.
    • The literature shows that metacognition helps reduce hallucination, improve reliability, and improve human-AI collaboration.
    • It analyzes metacognitive applications in reasoning models, agents, multi-agent systems, memory, retrieval, and tool use.
    • It systematically studies the effects of model size, post-training, and sampling temperature on metacognitive ability.
    • It identifies gaps in metacognitive evaluation and proposes future research directions, including domain generality and architecture design.
    原文摘要PDF
  27. 606ResearcharXiv:2607.11699

    Qwen-Music Technical Report

    • It compresses an entire song into a 25 Hz single-codebook Music Semantic Token so the LLM has room to predict the sequence.
    • The Melody-CoT mechanism plans melody as an explicit intermediate representation before token generation, which improves musicality and structural consistency.
    • Qwen-Music-Render combines a semantic-conditioned DiT, Spec-VAE, and Band-Mode Refiner to synthesize 48 kHz stereo audio.
    • It is trained on more than 5 million hours of multilingual music data with a curriculum by quality level and multi-stage post-training from SFT to DPO to GSPO.
    • In blind tests by expert evaluators, it achieves a 66.7% win rate against MiniMax 2.6 and a 58.3% win rate against Mureka V8.
    • It reaches state of the art on 13 of 16 objective metrics, including SongBench, SongEval, and AudioBox-Aesthetic.
    原文摘要PDF
  28. 605Machine LearningarXiv:2607.11505

    Proxy Exploration and Reusable Guidance: A Modular LLM Post-Training Paradigm via Proxy-Guided Update Signals

    • By separating exploration, meaning proxy, and alignment, meaning primary, it restructures post-training into a modular, asynchronous, and reusable pipeline.
    • It supports weak-to-strong improvement by transferring relative improvement signals rather than the absolute distribution of the proxy.
    • Exploration signals from a small proxy such as Qwen3-1.7B improve the larger primary model in a controllable way.
    • Compared with GRPO, OPD-based distribution alignment converges within 70 steps, which demonstrates exploration efficiency.
    • The transfer strength can be controlled with a scaling factor, which gives robust performance across different settings.
    • Caching, reuse, and cross-model transfer of exploration signals greatly improve cost efficiency.
    原文摘要PDF
  29. 604Computer VisionarXiv:2607.11673

    ABot-3DWorld 0: A Universal World Model to Explore Any 3D Space

    • SGP, a single representation combining panorama and point cloud, processes all input modalities in a unified way.
    • We improve 3D consistency in a 14B VACE-based panorama video generator through 3DRL post-training.
    • Rich inputs, including multi-view and video, are handled through a strict geometric pipeline to minimize hallucination.
    • An iterative FLUX-based repair module resolves panorama resolution limits and view mismatches.
    • The ABot-3DGS data engine generates fully aligned training data of RGB, point maps, and poses.
    • It supports indoor, outdoor, and aerial domains in a single pipeline and can be linked with mapped POIs.
    原文摘要PDF
  30. 603RoboticsarXiv:2607.11643

    Xiaomi-Robotics-U0: Unified Embodied Synthesis with World Foundation Model

    • It trains five generative tasks, image generation, editing, embodied scene generation, transfer, and video generation, in a single 38B-parameter autoregressive framework.
    • A structured control format separates workspace, background, foreground, target, and lighting as independent control dimensions to preserve multi-view consistency.
    • GPT-Image-2.0 ranks first in human evaluation for embodied scene generation and transfer, and first in World Arena embodied video generation.
    • Synthetic data generated by the model improves the OOD success rate of the π0.5 robot policy from 36.9% to 63.2%, a 26.3-point gain.
    • Sequential world modeling turns static scene generation into a scalable trajectory generation engine.
    • It releases code and checkpoints so the model can serve as an embodied-intelligence data engine.
    原文摘要PDF
  31. 602Computer VisionarXiv:2607.09544

    The Count Is There, but Misaligned: Understanding and Correcting Counting Failures in VLMs

    • It trains three probes over VLM internal activations to predict ground-truth count, output count, and whether an error is present, and the probes reliably predict mistakes.
    • SVCCA shows that the correct-count probe and the output-count probe share a partial subspace, but the readout is misaligned, which is a structural cause of error.
    • Causal steering confirms the cause, because strengthening the count direction improves counting performance while random directions hurt it.
    • Detector-guided self-correction applies selective reprompting only when an error is predicted, which improves accuracy by up to 15.6 absolute points.
    • It improves performance through inference-time intervention alone, so the method requires no parameter update and is practical to deploy.
    • There is a strong positive correlation, with rho equal to 0.803, between probe F1 and correction gains, which means that internal error-signal quality determines correction effectiveness.
    原文摘要PDF
  32. 601RoboticsarXiv:2607.09648

    B-spline Policy: Accelerating Manipulation Policies via B-spline Action Representations

    • It replaces discrete action chunking with continuous B-spline curves to accelerate execution through temporal rescaling.
    • Inference-time segment alignment ensures smooth transitions between adjacent B-spline segments, which is essential for high-speed execution.
    • On Table Cleaning, 4x speed reduces completion time by 50%, from 23.57 seconds to 11.80 seconds, while maintaining success rate at 13 out of 20 versus 14 out of 20.
    • It integrates as a drop-in replacement for both Diffusion Policy and ACT backbones, requiring minimal pipeline changes.
    • On PushT simulation, 4x speed improves the score from 0.59 to 0.73 and reduces time from 7.19 seconds to 2.87 seconds.
    • It automatically allocates control granularity to match task stages, from fast reaching to precise grasping.
    原文摘要PDF
  33. 600AI / GeneralarXiv:2607.08964

    Long-Horizon-Terminal-Bench: Testing the Limits of Agents on Long-Horizon Terminal Tasks with Dense Reward-Based Grading

    • It covers 46 long-horizon terminal tasks across 9 categories, including experimental reproduction, software engineering, multimodal analysis, games, and scientific computation.
    • It uses subtask-based dense rewards that provide partial scores instead of a binary pass or fail signal, making intermediate progress visible.
    • The benchmark averages 9.9 million tokens, 231 episodes, and 85.3 minutes per task, making it far more demanding than Terminal-Bench or SWE-Bench.
    • Even the strongest model, GPT-5.5, reaches only 15.2 percent pass@1 at R ≥ 0.95, while the average pass@1 is just 4.3 percent, leaving substantial room for improvement.
    • The main failure mode is partial progress followed by timeouts and early false finishes, with weak self-verification emerging as the key bottleneck.
    • Dense rewards are essential for exposing false-finish patterns, whereas binary evaluation misses them.
    原文摘要PDF
  34. 599Computer VisionarXiv:2607.09225

    Glob3R: Global Structure-from-Motion with 3D Foundation Models

    • It uses a frozen 3D foundation model, Pi3X, plus a lightweight dense matching head to turn feed-forward prediction into an optimizable geometric constraint.
    • A keyframe-based sliding-window strategy propagates tracks and relative poses across windows, which overcomes chunk-level limits.
    • On KITTI, it achieves a mean RMSE of 13.21 m, which is a 10 to 50 percent improvement over recent methods such as Scal3R at 14.55 and LoGeR at 18.65.
    • On ETH3D, it achieves R@1 of 91.58 and T@1 of 73.27, which substantially improves rotation and translation accuracy over the latest learning-based SfM method, AMB3R.
    • It improves PSNR by 2 to 3 dB in NeRF and neural rendering, which demonstrates the practical value of refined poses.
    • Global motion averaging and bundle adjustment resolve scale mismatch and recover dense geometry.
    原文摘要PDF
  35. 598Machine LearningarXiv:2607.08940

    TSRouter: Dynamic Modality-Model Selection for Time Series Reasoning

    • It is the first modality plus model co-routing framework that uses the complementarity of LLMs for numerical accuracy and VLMs for global patterns.
    • It models complex interactions with a heterogeneous graph over task, query, modality, and model and uses a GNN.
    • On four time-series reasoning tasks, it achieves 16% to 46% relative improvements and builds a Pareto frontier with cost-aware optimization.
    • It greatly outperforms the best baseline even when using only 10% of the training data, showing data efficiency.
    • GPT-5 is used for node profiling and initialization from pretrained embeddings, enabling zero-shot generalization to new models and tasks.
    • The user-defined alpha parameter allows real-time control of the performance-cost tradeoff.
    原文摘要PDF
  36. 597ResearcharXiv:2607.09627

    A complete ultrametric on von Neumann's incomplete tensor products

    • We define a convergence-index-based pseudo-ultrametric d on the equivalence classes Gamma of infinite tensor products, and prove the strong triangle inequality and completeness.
    • We define a gauge-invariant variant d~, which is insensitive to topological changes and consistent with von Neumann weak equivalence.
    • We prove that the product of any unitary U with 1 not in sigma(U) moves every equivalence class to the maximum distance 1.
    • We interpret d~ as a decoherence exponent: when monitoring N environmental degrees of freedom, the branching rate decays as exp(-N d~).
    • In the Everett many-worlds model, the universe is an infinite qubit tensor product, worlds are sectors, and time is a discrete unitary step.
    • The hierarchical structure of the ultrametric space corresponds exactly to the leaf space of the branching tree.
    原文摘要PDF
  37. 596Computer VisionarXiv:2607.09661

    PanoWorld: Real-World Panoramic Generation

    • It uses the rotational equivariance of panoramic ERP to treat rotation as a geometric transform and model only translation explicitly.
    • DPRC uses dense conditioning on panorama rays for accurate motion control, and GMA uses geometry-aware memory augmentation to maintain scene persistence.
    • The World360 dataset contains 70K real UAV clips and 50K AirSim360 simulation clips, including multi-altitude flight trajectories.
    • Compared with Matrix-3D, it improves FID from 27.64 to 16.93 at 720p, and it improves PSNR trajectory accuracy by 2 to 4 points.
    • Real-time generation with causal forcing produces 161 frames in 8 seconds on a single H20 GPU, versus about 4 minutes and 48 seconds previously.
    • A three-stage training pipeline progressively optimizes modeling, memory, and control.
    原文摘要PDF
  38. 595Computer VisionarXiv:2607.09657

    Scalable Visual Pretraining for Language Intelligence

    • The paper proposes a VP framework that pretrains directly from raw visual patches, without parsing documents into text, through next-latent prediction.
    • It shows consistent gains over TP across four backbones, with up to +3.22 on GPQA Diamond and +2.1 on MMLU-Pro.
    • It reaches equal or better performance with only 25 percent of the token budget, which shows that visual representations are more compression-efficient than text.
    • The gains are largest on visually dense documents such as figures, equations, and tables, which indicates recovery of information that text conversion would lose.
    • VP is effective not only for multimodal models but also for language-only models, which shows that it is architecture-agnostic.
    • A frozen vision encoder plus a trainable projection layer are enough to integrate the visual stream into an autoregressive backbone.
    原文摘要PDF
  39. 594Machine LearningarXiv:2607.09375

    Mach-Mind-4-Flash Technical Report

    • A 35B MoE model with 3B active parameters reaches 100B-plus-level performance through post-training alone, without scaling pretraining compute.
    • MOPD, or Multi-Teacher On-Policy Distillation, trains domain-specific RL experts independently and then merges them with routed reverse-KL to remove the seesaw effect of mixed-reward RL.
    • HMPO is a single-stage token-efficient optimization method that compresses reasoning chains by 19 to 46 percent while keeping accuracy loss below 0.7 percentage points.
    • It reports AIME'26 at 92.70, BFCL-v4 at 75.80, and ClawBench at 84.20, showing that 3B active parameters can compete with 1T-scale models.
    • SonicMoE GEMM kernels and segmented shared-expert fusion speed training up by 17 percent end to end.
    • It also reshapes the Pareto frontier for token efficiency by sharply reducing token usage at similar accuracy.
    原文摘要PDF
  40. 593Computer VisionarXiv:2607.09024

    Video Generation Models are General-Purpose Vision Learners

    • It proposes video diffusion models as a pretraining paradigm for general-purpose vision recognition, serving a catalytic role for vision analogous to next-token prediction in NLP.
    • It converts iterative denoising into a single forward pass that can control depth, normal, segmentation, pose, keypoints, and more through text instructions.
    • It matches or exceeds specialized models such as D4RT and VGGT-Ω while requiring 7x to 500x less training data.
    • It experimentally shows that diffusion-based pretraining outperforms prior pretraining methods such as V-JEPA and VideoMAE V2.
    • Training on synthetic human videos alone still yields zero-shot generalization to real videos and OOD objects, evidence that world models are internalized.
    • Without dedicated heads or losses, the unified backbone, head, and loss structure lets tasks expand simply by changing the data format.
    原文摘要PDF
  41. 592AI AgentsarXiv:2607.01942

    Atomic Task Graph: A Unified Framework for Agentic Planning and Execution

    • The core idea is to represent agent task solving as a DAG composed of atomic tool-use nodes and input-output dependency edges.
    • Interface-preserving recursive graph compilation recursively decomposes coarse tasks down to atomic units while preserving the parent node's input-output interface, making graph evolution traceable.
    • Dependency-aware execution runs independent branches in parallel, shortening the average number of steps versus ReAct from 31.42 to 18.36 on ALFWorld and from 47.35 to 29.72 on ScienceWorld.
    • Minimal necessary subgraph repair preserves verified regions and repairs only the smallest affected subgraph when errors occur, and removing the repair step lowers Mistral-7B's ALFWorld score by 7.72 points.
    • A pre-execution thought experiment uses lightweight verification before execution to detect risky plans in advance, finding 24.6% to 27.4% risky-plan detection with over 74% precision.
    • The hallucinated action rate is 12.14%, a 71.7% relative reduction versus ReAct's 42.86%, and the average success rate across three benchmarks rises from 25.51 to 56.10 with Mistral-7B.
    原文摘要PDF
  42. 591Computer VisionarXiv:2607.08127

    Understanding and Mitigating the Video-Action Generalization Gap via Temporal Ratio

    • The VAG gap is the phenomenon in which the compositional prior of video foundation models is lost after robot behavior fine-tuning, and we define and analyze it systematically.
    • Temporal Ratio (TR) measures how much the action head's attention focuses on future latent rollouts compared with the current frame, and it predicts compositional generalization ability.
    • TR changes dynamically across task phases: it rises during planning and falls during precise manipulation, which is the natural attention pattern.
    • TR-Adaptive Guidance is an inference-time method that adaptively amplifies compositional conditioning signals during planning phases.
    • On the LIBERO benchmark, it improves average OOD success by more than 5x over the previous VAM baseline, from about 10% to 59.4% with guidance.
    • It also achieves an average of 83.3% on real bimanual robot tasks in YAM, far above pi0 at 36.7% and pi0.5 at 55.0%.
    原文摘要PDF
  43. 590AI / GeneralarXiv:2607.08393

    Towards Mechanistically Understanding Why Memorized Knowledge Fails to Generalize in Large Language Model Finetuning

    • It formalizes the Knowing-Using Gap, the phenomenon in which memorization is nearly perfect but generalization, especially multi-step reasoning, lags far behind.
    • It introduces self-patching, a causal diagnostic tool that recovers generalization failures immediately by moving a representation from one layer to another.
    • It proposes the knowledge-circuit misalignment hypothesis, where memorized representations exist in early and late layers but are not routed into the middle reasoning layers, causing failure.
    • The paper validates this across six models, from Qwen-2.5 1.5B to 7B and LLaMA-3.2 1B to 8B, and across two domains, biomedical and academic.
    • A fixed heuristic using only two layer-pair settings, 0.8L to 0.5L and 0.1L to 0.5L, recovers 58 to 75 percent of the oracle improvement.
    • Self-patching is far better than CoT prompting or irrelevant patching, which supports the claim that the root cause is structural.
    原文摘要PDF
  44. 589AI / GeneralarXiv:2607.08716

    Remember When It Matters: Proactive Memory Agent for Long-Horizon Agents

    • It defines the concept of behavioral state decay, a failure mode in which information is present in context but still fails to affect behavior.
    • It uses a two-stage plug-and-play architecture in which a separate memory agent maintains a structured memory bank and decides whether to intervene selectively.
    • On Terminal-Bench 2.0, it improves Sonnet 4.5 from 37.6% to 45.9%, a gain of 8.3 points, and on τ²-Bench from 55.0% to 61.8%, a gain of 6.8 points.
    • The +2.4 to +2.5 point improvement also holds for the stronger Opus 4.6 model, so it remains effective even on strong models.
    • Catenation experiments show that active intervention outperforms passive memory exposure, always-on injection, advisor models, and Mem0 retrieval.
    • An open-weight memory agent trained from Qwen3.5-27B with SFT and GRPO also achieves a +3.5-point transfer gain on Terminal-Bench.
    原文摘要PDF
  45. 588LLMs / NLParXiv:2607.08046

    What LLM Forecasters Know but Don't Say: Probing Internal Representations for Calibration and Faithfulness

    • Internal probes cut ECE in half, from 0.093 to 0.044, compared with verbalized confidence, greatly improving calibration.
    • When CoT evidence is removed, predictions change while the reasoning trace stays the same in 23% of cases, revealing insincerity (rho = 0.22).
    • The probe's accuracy at tracking behavioral changes (rho = 0.57) far exceeds CoT (rho = 0.22), and it can detect 84% of the hidden effects with the correct direction.
    • Forced-answering experiments confirm that many predicted answers are already decided before reasoning even begins.
    • Using the entropy of the pre-reasoning answer distribution, we classify questions into three levels and save 30% to 47% of tokens without accuracy loss.
    • Probe-only calibration also works on GLM-4.7-Flash and GLM-4.5-Air, showing that calibration can improve without weight updates.
    原文摘要PDF
  46. 587Machine LearningarXiv:2607.07386

    Sparse Delta Memory: Scaling the State of Linear RNNs through Sparsity

    • Sparse addressing with product-key memory expands the state memory about 4,000 times beyond GDN at the 8B scale.
    • With the same FLOPs and the same number of parameters, it achieves better training loss than both GDN and full attention.
    • On long-context recall in RULER, it improves from 20.0 percent for GDN to 31.2 percent for SDM at 128k on the 1.4B model, which is an absolute gain of 11.2 percentage points.
    • A learned initial state, or learned M0, stores pretrained knowledge and further improves common-knowledge and reasoning tasks.
    • At the 8B scale, decoding is 1.49 times slower than GDN but 6 times faster than full attention.
    • The read and write distributions change adaptively with capacity: writes become concentrated, while reads become more distributed.
    原文摘要PDF
  47. 586Computer VisionarXiv:2607.07033

    AnchorPrune: Relevance-Anchored Contextual Expansion for Visual Token Pruning

    • Its sequential design first anchors relevance and then expands context, which prioritizes non-redundant query evidence.
    • It automatically chooses the anchor size from each token's novelty profile rather than using a fixed ratio.
    • The architecture-aware design applies to both CLIP-aligned and unaligned models, as well as image and video VLMs.
    • On LLaVA-NeXT-7B, it retains 97.6 percent of full performance with only 160 of 2,880 tokens, or 5.6 percent.
    • At extreme compression with 32 tokens, it outperforms prior relevance-diversity integrated methods by 3.7 to 5.7 percentage points.
    • It is a lightweight framework that requires no training and can be applied directly to existing VLMs without architectural changes.
    原文摘要PDF
  48. 585RoboticsarXiv:2607.08639

    Native Video-Action Pretraining for Generalizable Robot Control

    • Semantic visual-action tokenizer SemVAE aligns visual representations and actions in a shared latent space.
    • Causal pretraining avoids catastrophic forgetting that appears when applying a bidirectional architecture, and it trains from scratch.
    • A Sparse MoE backbone expands model capacity while keeping the active compute per step unchanged.
    • Foresight Reasoning asynchronous inference achieves 225 Hz real-time closed-loop control, a 6.5x speedup over the 35 Hz baseline.
    • It achieves few-shot generalization with 10 to 15 demos and outperforms π0.5 and LingBot-VA on both simulation and real evaluation.
    • Multi-chunk prediction, or MCP, encourages trajectory-level dynamics learning instead of short-term visual continuity.
    原文摘要PDF
  49. 584RoboticsarXiv:2607.07459

    EmbodiedGen V2: An Agentic, Simulation-Ready 3D World Engine for Embodied AI

    • It proposes a unified sim-ready representation that combines metric geometry, collision geometry, physical parameters, and affordance annotations in a single asset.
    • Its model-agnostic asset pipeline connects TRELLIS, SAM3D, and Hunyuan3D as plugins and automates mesh repair, convex decomposition, and texture baking.
    • The task-driven scene graph decomposes a natural-language task into spatial constraints over the robot, background, objects to manipulate, and distractors, then generates physically stable 6-DoF placements.
    • Vibe Coding implements a stateful agent-skill harness that turns natural-language commands into physically validated bounded edits and then exports them to the simulator.
    • It supports cross-simulator export to MuJoCo, Isaac Sim, Genesis, SAPIEN, and Bullet in USD, MJCF, and URDF formats.
    • Online RL validation shows that training VLA policies on the generated environments raises simulation success from 9.7 percent to 79.8 percent and real-robot success from 21.7 percent to 75.0 percent.
    原文摘要PDF
  50. 583Machine LearningarXiv:2607.07508

    Single-Rollout Asynchronous Optimization for Agentic Reinforcement Learning

    • We replace GRPO's group-wise sampling with single-rollout to remove waiting latency and reduce off-policy drift in asynchronous RL.
    • Direct double-sided IS uses rollout log-probabilities directly to remove the cost of tracking the old policy.
    • skip-observation GAE skips observation tokens and computes advantage between actions in multi-turn agent trajectories.
    • We update the value model more frequently than the actor and fine-tune frozen attention to stabilize single-rollout training.
    • Deployed in the GLM-5.2 (750B-A40B) RL pipeline, it achieves stable training for about 1,000 steps and consistent gains over GRPO.
    • In a simulated online-learning environment, it adapts quickly when reward criteria change, demonstrating robustness to nonstationary environments.
    原文摘要PDF