HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 845Computer VisionarXiv:2608.06257

    MASS: Multiplayer World Models with Authoritative Shared State

    • It explicitly represents the world shared by multiple users as structured entity records, including coordinates, orientation, and whether an entity is alive, maintaining a single authoritative state.
    • It separates the Logic Engine, which updates state, from the Rendering Engine, which produces camera images, decoupling world-transition cost from the number of requested views.
    • In a matched Snake evaluation, it achieved state-reconstruction accuracy of 0.764, above the video-based baseline's 0.128, and progressed 1024 players for up to 10,000 steps while preserving structural validity.
    • It applied the same structure to multiple games through declarative schemas and also demonstrated clients predicting the shared state during network delays.
    • Current evidence is primarily limited to simple two-dimensional games and restricted entity interactions, and the authors left three-dimensional environments and more complex interactions as future work.
    AbstractPDF
  2. 844Computer VisionarXiv:2608.05811

    Energy-Guided Flow Matching

    • It replaced flow matching's fixed clean-image endpoint with an endpoint that moves from a low-frequency image to the original, making generation proceed from coarse structure to detail.
    • It calculated frequency energy for each image and automatically adjusted heat-diffusion time so that the missing energy is released at the same rate.
    • On ImageNet 256×256, it achieved FID 1.55 after 200 training rounds and 1.45 after 600, improving both quality and training efficiency.
    • It achieved FID 1.58 at 512×512, and GenEval 0.85 and DPG-Bench 83.9 for text-to-image generation, without architectural changes or additional losses.
    • It has not yet been evaluated on video and time-varying signals, joint text-image models, embodied decision making, or the largest recent foundation models.
    AbstractPDF
  3. 843Computer VisionarXiv:2608.05798

    KVAE: Family of Tokenizers for Multimodal Generative Models

    • It designed the KVAE tokenizer family with two compression ratios for video, eightfold compression for images, and 48kHz audio with a 50Hz latent representation.
    • In video and image reconstruction and generation evaluations, it reported results matching or exceeding public tokenizers such as Wan-2.2, HunyuanVideo-1.5, and FLUX.2.
    • The audio model has 166.9 million parameters, making it smaller than the larger comparison models, while showing preference advantages in multiple domains in human side-by-side listening evaluations.
    • It treated diffusability, whether the latent space is easy for a generative model to learn, as a design criterion alongside reconstruction quality, and released the code and models under the MIT License.
    • The audio FAD evaluator internally handles only 16–32kHz and therefore cannot measure full-band quality above 16kHz, while model-based aesthetic scores are also subject to evaluator bias.
    AbstractPDF
  4. 842Machine LearningarXiv:2608.06296

    On-Policy Self-Distillation without Any Supervision

    • It achieves self-distillation using only the model's own generations, without external answers, environment feedback, or a larger model.
    • It forms a pseudo-answer by majority vote and uses the shortest agreeing rollout as the teacher.
    • It applies distillation to the prefix of incorrect rollouts so that the model corrects the wrong part precisely.
    • On math benchmarks, it improves by 8.5 to 10.7 points over the base model without reasoning mode and by 3.2 to 2.3 points over OPSD.
    • Forward KL works best, reverse KL diverges during training, and JSD is ineffective.
    AbstractPDF
  5. 841Computer VisionarXiv:2608.06009

    Wan-Animate-2: Pushing the Application Boundaries of Character Animation

    • It removes the intermediate motion extractor, such as skeleton or SMPL, and feeds the driving video directly into the DiT.
    • A dual-branch design, temporal-alignment RoPE, and sparse reference attention balance fidelity and efficiency.
    • It introduces Viewpoint LoRA, which controls the output camera viewpoint from a text prompt.
    • Wan-Animate-2-Lite reaches 24 fps in real time through a three-stage training process of teacher forcing, error buffer, and self-forcing.
    • In user studies, it is rated equal to or better than commercial closed-source systems such as Dreamina and Kling.
    AbstractPDF
  6. 840AI / GeneralarXiv:2608.06301

    HarnessOpt-Bench: Evaluating LLMs at Harness Optimization

    • The authors design a trustworthy execution environment that turns harness optimization into a reproducible evaluation target.
    • They separate the model, harness, and task effects through 111 evaluations across five frontier models and four tasks.
    • Differences between optimization models are larger on average than differences between coding harnesses.
    • Broader modifications are associated with larger improvements, while detailed failure-trace analysis is not correlated with improvement.
    • A native harness is not always better than a shared harness, with 11 wins and 9 losses.
    AbstractPDF
  7. 839LLMs / NLParXiv:2608.05139

    Toward Skill-Native LLMs: Skill Entropy for Benchmarking and Training Long-Horizon Reasoning

    • It defines skill entropy as a directional metric that measures how hard it is to switch between two skills.
    • It builds Skill2-Bench, a benchmark with 558 skills across 9 domains.
    • Frontier models show monotonically decreasing accuracy on tasks with high skill entropy.
    • Skill entropy RL improves Qwen3-4B from 34.4% to 68.4% and Qwen3-1.7B from 14.6% to 40.1%.
    • It is a reusable training signal that can also be applied to existing datasets such as OpenR1-Math.
    AbstractPDF
  8. 838RoboticsarXiv:2608.06332

    GeniWorld: A Generalizable Interactive World Model for Robotic Manipulation via Visual Actions

    • We improved spatial accuracy by using a visual action representation based on URDF rendering instead of numeric actions.
    • By separating robot kinematics from environment dynamics, we reduced scene overfitting and improved generalization.
    • Autoregressive video prediction supports closed-loop interaction at 8 Hz on an H20 GPU.
    • With only limited real data, it achieves robust zero-shot generalization in OOD environments.
    • Policies trained on trajectories synthesized by the world model improve overall success from 40.8% to 69.0%.
    AbstractPDF
  9. 837Machine LearningarXiv:2608.05136

    The Loss Does Not See the Basis, but Adam Does

    • We proved that gauge equivalence, or rotational invariance, is necessary to preserve low-rank bias.
    • A classification of nine optimizers shows that GD, Muon, and Shampoo are equivalent, while Adam, RMSProp, and Lion are not.
    • In matrix sensing, equivalence classes achieve recovery error from 0.00 to 0.29, whereas coordinate-wise classes are at 0.42 to 0.57.
    • In Transformers, Adam separates gauge-equivalent initializations from the first step and creates a 56% difference.
    • Interpolating from coordinate-wise forms to equivalent scalars improves monotonically, proving that heterogeneity is the cause.
    AbstractPDF
  10. 836RoboticsarXiv:2608.05674

    JoyAI-RA 0.5: Scaling Robot Manipulation Learning via Dual Action Alignment

    • The authors built a world model, LAC-WM, that learns physical dynamics by extracting latent actions from human videos.
    • We propose an explicit alignment method that aligns human and robot trajectories in a unified physical action space.
    • With inner-loop and outer-loop reinforcement learning, it achieves both fast adaptation and improvement of the base policy.
    • As the scale of first-person human video increases, robot performance improves consistently and does not saturate.
    • It shows strong performance on both seen tasks and new variants on a real AgiBot robot.
    AbstractPDF
  11. 835AI / GeneralarXiv:2608.05466

    Recursive Synthesis for Long-Horizon Terminal Tasks

    • Recursive synthesis makes tasks progressively harder while preserving consistency across instructions, environments, answers, and verifiers.
    • After 15 rounds, it generates 37,484 tasks, and the synthesis yield does not fall, which means there is no apparent ceiling.
    • The intermediate answer length grows from 67 lines to 374 lines, which is a 5.6 times increase, and the number of instructions increases by more than 6 times.
    • When the synthetic data is used for SFT and PPO training, performance improves by up to 10 points on Terminal-Bench 2, Hard, and Long-Horizon tasks.
    • DeepSeek-V4-Pro pass@4 falls from 90 percent to 2.5 percent, which shows that the tasks become substantially harder.
    AbstractPDF
  12. 834RoboticsarXiv:2608.06375

    $ω$-0: A Latent Predictive World Action Model for Concurrent Humanoid Loco-Manipulation

    • We trained full-body cooperative motion in one model without separating locomotion and manipulation.
    • Instead of future video generation, we improved action quality using latent embedding prediction, a lightweight world model.
    • We built and released omega-HOME, a 40-hour real-home humanoid dataset.
    • It performs 11 household tasks with a single model and outperforms prior imitation learning, VLA, and WAM methods.
    • We propose a pipeline that converts human video data into robot-executable actions through simulation replay.
    AbstractPDF
  13. 833AI / GeneralarXiv:2608.05987

    AgentOPSD: Recursive Self-Distillation for Agentic Reinforcement Learning

    • It aligns token-level distillation gaps with environment transition points by aggregating them at the turn level.
    • It recursively updates a Bayesian belief state to compute credit that reflects the accumulated context of earlier turns.
    • On ALFWorld with Qwen2.5-7B, the method achieves an 89.1% success rate, outperforming GRPO and existing self-distillation methods.
    • The performance gap widens for longer horizons, while the effect is smaller on short tasks.
    • It is fully compatible with standard policy optimization without extra rollouts or a learned critic.
    AbstractPDF
  14. 832ResearcharXiv:2608.04850

    Two-dimensional Toda--Arnoldi correspondence: Holomorphic Krylov geometry and counterdiabatic transport

    • The paper proves that the diagonal and subdiagonal coefficients in the Arnoldi reduction of a non-Hermitian Hamiltonian exactly match the Flaschka variables of the two-dimensional Toda lattice.
    • This Toda structure closes without determining the remaining upper Hessenberg terms of the Arnoldi matrix.
    • The square of the subdiagonal coefficient determines the Fubini-Study metric and Berry curvature of the Krylov subspace.
    • Along the actual path, the Arnoldi frame generates exact isospectral transport and realizes counterdiabatic driving.
    • The paper confirms that the Hermitian limit reduces to the classical Toda-Lanczos correspondence.
    AbstractPDF
  15. 831AI / GeneralarXiv:2608.05102

    ABSeeker: Training Long-Horizon Search Agents via Answer-Backtracked Credit Assignment

    • The paper proposes the ABC framework, which backtracks from the answer to reconstruct intermediate clues and evaluates each search step against those clues.
    • The key idea is precise credit assignment: useful steps in failed trajectories are rewarded, while unnecessary steps in successful trajectories are penalized.
    • ABC-SFT uses step-wise loss weighting, and ABC-GRPO uses step-wise rewards to improve both conventional SFT and GRPO.
    • With a 4B model, it achieves state-of-the-art performance at the same scale, improving from 37.3 to 55.3 percent on BrowseComp and from 39.1 to 52.9 percent on BrowseComp-ZH.
    • Trained on only 8.5K examples, it also generalizes well to other benchmarks such as xbench and GAIA.
    AbstractPDF
  16. 830Computer VisionarXiv:2608.02980

    Qwen-3D: A Generalist 3D Vision-Language Model for Spatial Understanding

    • It compresses visual tokens spatially with 3D position information, including depth and camera pose, to process long videos efficiently.
    • With 3D Rotary Positional Embedding, attention operates directly in 3D space, which improves cross-view reasoning.
    • A structure where the language model and the segmentation decoder share the full visual token set greatly improves 3D visual grounding.
    • It outperforms existing 3D LMMs by 4% in 3D visual grounding at Acc@25 and by 13% in 3D instance segmentation mAP.
    • Its 2D vision-language performance stays at the level of Qwen2.5-VL, which shows that 3D training does not hurt 2D capability.
    AbstractPDF
  17. 829AI / GeneralarXiv:2608.03972

    ReflectRL: Learning from Golden Negative Trajectories via Reflective-to-Direct Reasoning

    • We prove that expert-model failure trajectories, GNT, are far more effective for iterative learning than failures from itself or from weaker models.
    • GNT includes both a valid prefix close to the correct answer and local error points, making it high-quality material for reflective learning.
    • Reflective-to-Direct Policy Transition gradually shifts capability from hint dependence to solving on its own.
    • It achieves consistent accuracy gains across nine benchmarks, four models from 1.5B to 8B, and four training methods.
    • It cuts response length by more than half while improving accuracy and preventing entropy collapse.
    AbstractPDF
  18. 828Computer VisionarXiv:2608.05070

    HelloWorld: Enabling Socially Interactive Characters in Video World Models

    • This is the first video world model to realize social interaction between a character and a user, such as waving, greeting, and speaking to the user.
    • The paper proposes a self-distillation pipeline that reuses interaction videos generated by the model itself as training data.
    • Without additional training, it controls the timing of interactions at the frame level using only cross-attention masks.
    • It builds the first social-interaction benchmark, HelloWorldBench, with 400 samples and 3 interaction metrics.
    • It substantially outperforms existing world models on interaction quality while maintaining SOTA-level video aesthetics and camera following.
    AbstractPDF
  19. 827Computer VisionarXiv:2608.04436

    ToolArtist: Tool-Using Unified Multimodal Models for Agentic Image Generation

    • It is the first fully agentic image generation model that unifies reasoning, external tools such as text and image search, and image generation under one UMM policy.
    • It develops an optical transformation method for SFT that hides the image-generation tool and leaves only the final image, forcing the UMM to draw directly.
    • It uses RAD-GRPO, which combines intent reward and quality reward, to reinforce the full reasoning-search-drawing trajectory.
    • With only 7,132 high-quality SFT trajectories, it outperforms partial agentic methods on WISE and WorldGenBench-Humanities.
    • It openly releases the training data and the full post-processing infrastructure.
    AbstractPDF
  20. 826AI / GeneralarXiv:2608.05144

    Argus: A General-Purpose Agentic Runtime for Long-Horizon Reasoning

    • It proposes a verification-gated runtime self-evolution structure that approves goal and method changes only through a verification gate.
    • The Manager, Planner, Engineer, and Reviewer roles each carry out bounded missions on durable project state.
    • On SWE-Bench Pro, it reaches 78% compared with 59% for Direct Copilot, while using 1.41x as many tokens.
    • At the mature stage, it uses 21% fewer solve tokens and shortens workflow time by 15% compared with the new Wave stage.
    • It is validated on seven GPT-5.5 benchmarks and produces practical results, including an upstream merge of the RWKV6 kernel.
    AbstractPDF
  21. 825Machine LearningarXiv:2608.04001

    Test-Time Scaling in Reasoning LLMs: Inference Regimes, Evaluation, and Reproducibility

    • It formalizes test-time scaling into three structural regimes: single-path, leaf-level, and prefix-level.
    • It emphasizes that the target of evaluation is the entire inference system, including prompts, decoders, verifiers, and stopping rules, not just the model weights.
    • It shows the risk of a poor selection criterion, with mean log probability accuracy dropping from 75.6% to 65.8% as the sample count increases.
    • It releases more than 2 billion inference traces, covering mathematics, science, and competitions, to provide a foundation for reproducible research.
    • It organizes the open-weight reasoning-model ecosystem by training method and interface control criteria.
    AbstractPDF
  22. 824Computer VisionarXiv:2608.05131

    OPD-V: Visual On-Policy Self-Distillation with Modality Balance

    • It quantifies modality imbalance in multimodal models and proposes a new paradigm that uses this text bias as privileged information.
    • It defines a modality-balanced trust region from the logit difference between a zoom-in image as the positive teacher and a masked image as the negative teacher, and it uses this region for selective distillation.
    • On Qwen3.5-4B, it improves the average score from 64.3% to 80.0%, which also beats the 397B model at 77.4%.
    • It shortens response length by 74.5% while improving accuracy, which reduces training time per sample by 25% to 32%.
    • It shows consistent gains across six benchmarks, four backbones, and five post-processing methods.
    AbstractPDF
  23. 823RoboticsarXiv:2608.05042

    BridgeVLA++: A Data-Efficient, Generalizable, and Memory-Augmented Vision-Language-Action Framework for 3D Manipulation

    • It is the first framework to integrate temporal memory, which remembers task order, and spatial memory, which remembers the location of occluded objects, into a 3D VLA model.
    • It maximizes data efficiency by converting point clouds into multi-view images for the VLM and predicting actions through heatmaps.
    • On memory-dependent benchmarks such as RMBench and MemoryBench, it achieves an average jump from 20 percent to 93.3 percent compared with the previous state of the art.
    • On real robots such as Franka and Dobot, it shows strong generalization with fewer than 10 training demonstrations and extends to dual-arm manipulation.
    • The heatmap intermediate representation, 2D pretraining, and input alignment are the main contributors to performance, while 3D position input actually hurts performance.
    AbstractPDF
  24. 822RoboticsarXiv:2608.04996

    DreamWAM: Beyond RGB Future Prediction for World Action Models

    • It extends future prediction beyond RGB to motion, depth, and semantic information, which substantially improves the robustness of robot action models.
    • The extra information is used only during training, while inference uses RGB alone, so deployment cost does not increase.
    • On standard LIBERO, performance improves slightly from 97.3 percent to 98.4 percent, while on unseen shifts in LIBERO-Plus it improves substantially from 51.4 percent to 63.4 percent.
    • Real-robot experiments show the same pattern, with standard settings improving from 90.8 percent to 96.7 percent and perturbed settings improving from 55.6 percent to 74.4 percent.
    • Among the three auxiliary signals, motion matters most, and depth and semantics work best when injected as residual information.
    AbstractPDF
  25. 821Computer VisionarXiv:2608.05000

    Towards Physics of Multimodal Pretraining: Knowledge Flow, Modality Synergy, Early Unification, and Recipes

    • It finds a strong asymmetry in which language and visual understanding act as priors that drive visual generation, while the reverse transfer is weak.
    • When data complexity is low, modalities are synergistic, but when complexity is high they compete, and a shared attention structure promotes synergy.
    • Early joint training is far more effective than late alignment, and late fusion causes a vision-laziness effect.
    • Using these principles, the authors create a training recipe that uses only 5 percent of the compute to train a 13.5B MoE model on 2T tokens and still achieves strong generation quality.
    • The synergy pattern remains consistent regardless of visual tokenizer design, which shows that architecture choice is the main driver.
    AbstractPDF
  26. 820Computer VisionarXiv:2608.02711

    Hunyuan3D-Buffalo 1.0: A Unified Multimodal Model for Scalable 3D Generation, Understanding, and Editing

    • It unifies 3D understanding, text-to-3D generation, instruction-based 3D editing, and text-based part generation in one architecture.
    • It builds one of the largest multimodal 3D training corpora, with 87 million examples in total, including 25 million for understanding, 50 million for generation, and 12 million for editing.
    • The Nano3D-v2 agent-based pipeline mass-produces geometrically consistent 3D editing data.
    • The combination of Hunyuan3D-VLM for understanding and DiT for generation achieves state of the art on both generation and editing benchmarks.
    • Experiments show cross-task synergy, where better generation and understanding improve editing quality.
    AbstractPDF
  27. 819Computer VisionarXiv:2608.02713

    Quo Vadis, World Modeling?

    • It expands the paradigm of world models from predicting physical state to transforming information into agent-usable information.
    • It classifies world proxies into six functional forms: dynamics, spatial, execution, memory, skill, and reward.
    • It systematizes capability enhancement in three stages: inference-time support, training-time optimization, and co-evolution between the agent and the proxy.
    • It analyzes how world proxies address the cost, risk, and limited parallelism of direct environment interaction.
    • It is a roadmap paper from a joint effort by Shanghai AI Lab, Zhejiang University, and NUS that sets the direction for world-model research.
    AbstractPDF
  28. 818AI / GeneralarXiv:2608.01735

    DAPD: Dual-Anchored Policy Distillation

    • It diagnoses the root cause of privilege illusion as information asymmetry between the teacher and the student.
    • It uses the self-conditioned distribution as a bridge to align reference and rollout behaviors along two information-matched paths.
    • It supervises both directions, from reference to rollout and from rollout to reference, to reduce dependence on privileged information while preserving accuracy.
    • It reduces wrong claims by 45% compared with OPSD while consistently improving reasoning performance.
    • The improvement persists as the model scales from 1.7B to 32B.
    AbstractPDF
  29. 817RoboticsarXiv:2608.03727

    Track4Action: Distilling World-Centric 3D Tracker into Vision-Language-Action Policies

    • It proposes a privileged-supervision approach in which a 3D tracker is used only during training and removed at deployment.
    • Track queries match the 3D tracker features extracted from Track4World and inject them into the VLA action head.
    • It reaches 82.3 percent zero-shot on LIBERO-Plus, improving by 7.6 percentage points over the unaligned version and by 3.0 points over LaMP.
    • It records 80.44 percent on RoboTwin 2.0 clean and 81.48 percent on randomized.
    • On four real dual-arm robot tasks, it reaches 67.5 percent, which is 25.0 percentage points higher than the unaligned version.
    AbstractPDF
  30. 816Computer VisionarXiv:2608.04010

    ParVL: Parallel Scaling and Expandable Compute Allocation for Multimodal LLMs

    • It proposes the first framework that creates separate parallel branches for ViT and LLM so that compute allocation can be adjusted independently.
    • A prefix-conditioned branch shares backbone parameters, so compute can be expanded without increasing parameters.
    • Systematic experiments on nine vision-language branch configurations show that the optimal allocation differs by task.
    • It confirms consistent gains over a single-branch model at 1B, 2B, and 8B scale.
    • It analyzes that parallel execution does not directly translate extra compute into wall-clock latency.
    AbstractPDF
  31. 815AI / GeneralarXiv:2608.02276

    Harness-R1: Learning to Edit Executable Runtime Harnesses from Agent Failure Trajectories

    • It proposes the first learning-based method that automatically generates executable patches for runtime harnesses from agent failure trajectories.
    • A 9B harness engineer is trained only with GRPO while the target agent weights remain frozen, which creates a stable pipeline.
    • It improves success rate by 9.3 percentage points on average across WebShop, ALFWorld, and DBBench.
    • Even after target-agent fine-tuning, it adds another 5.0 percentage points, which shows that the harness engineer and the agent can co-evolve.
    • Transfer experiments on 20 different target settings show an average improvement of 7.06 percentage points, which confirms generalization.
    AbstractPDF
  32. 814RoboticsarXiv:2608.03701

    LiLa-WAM: Lightweight Latent Reasoning World-Action Model for Robotic Manipulation

    • It trains a lightweight world-action model that unifies future-state prediction and action generation in a single stream on one 24 GB GPU.
    • It proposes the Visual Transition Token, or VTT, which specifies a task only through visual feature direction without language.
    • On 50 RoboTwin 2.0 tasks, it achieves a 90.48 percent success rate and shows strong efficiency relative to parameter count.
    • It also performs competitively with much larger models in LIBERO and real-robot experiments.
    • Using only a DINOV3 visual backbone, it shows that effective robot control is possible without dependence on a language model.
    AbstractPDF
  33. 813Computer VisionarXiv:2608.03979

    Video-DeepResearch: Towards the Next-Generation Multimodal Deepresearch Agent

    • We designed a deep-research agent pipeline that selects key scenes from video frames and performs visual search.
    • We found and fixed modality bias, where the model prefers text search only, and knowledge leakage, where it answers from memory without tools.
    • After SFT, we used GRPO, Group Relative Policy Optimization, to build autonomous exploration ability beyond imitation learning.
    • We newly built VideoDR-Bench, with 200 complex multi-hop VQA problems.
    • The 35B-A3B model reaches SOTA at 64.0%, beating Claude-4.5-Sonnet at 59.0% by five points.
    AbstractPDF
  34. 812Machine LearningarXiv:2608.02352

    Qwen-CUA: Native Computer Use for (almost) Everything

    • We trained a native computer-use agent that controls the keyboard and mouse from screenshots only, using a 397B-A17B MoE.
    • It keeps up to 20 screenshots and groups older ones into blocks to improve KV-cache efficiency.
    • We ran large-scale reinforcement learning with about 100,000 vCPUs, 40,000 verified tasks, and tens of thousands of concurrent environments.
    • It consistently outperformed Qwen3.7 across eight benchmarks, with 86.2% on OSWorld-Verified.
    • It also improved safety by reducing attack success rate from 36.6% to 16.4%.
    AbstractPDF
  35. 811Information RetrievalarXiv:2608.02738

    Knowledge-Geometry Decoupling: Refreshable Pretrained Transfer for Streaming Recommendation

    • [Withdrawn paper] The latest arXiv version, v2, states that Zixuan Wang has withdrawn this paper.
    • The official version history dates the withdrawal to 6 August 2026; the stated reason concerns industrial co-authors’ data sharing and privacy regulations.
    • Official reason: Withdrawn due to data sharing and privacy regulations of industrial co-authors.
    • Previously described methods and performance figures were pre-withdrawal claims and are not presented as currently valid research findings or grounds for recommending adoption.
    • Official withdrawal notice and version history: https://arxiv.org/abs/2608.02738 .
    AbstractPDF
  36. 810RoboticsarXiv:2608.03924

    ETA: A New Agentic Paradigm for Embodied Tasks

    • It proposes an agent paradigm for robots in a Planner-Tool-World loop, where the robot acts once, observes, and then decides again.
    • OpenETA provides interchangeable planners, composable tools, auditable memory, and replayable trajectories in open source.
    • Without a VLA, gpt-5.6-Sol alone solves 117 of 130 LIBERO tasks, or 90%, at PASS@5.
    • It includes a limited self-evolution mechanism that improves skills by learning from failed trajectories.
    • It demonstrates real hardware integration with a UR5e robot arm, showing applicability beyond simulation.
    AbstractPDF
  37. 809Computer VisionarXiv:2608.03974

    JoyAI-Video-Edit: Real-Time Open-Ended Video Editing with Autoregressive Diffusion

    • It achieves real-time 720p video editing at 30 FPS on a single B200 GPU with a 16B-parameter diffusion model.
    • SA-DMD, or source-anchoring distillation, reduces the diffusion process to two stages while preserving the fidelity of the original video.
    • Long-horizon autoregressive distillation suppresses accumulated color drift and appearance distortion in long videos.
    • It builds a new long-video editing benchmark called LongV2VBench for systematic evaluation.
    • It substantially outperforms existing streaming editors and produces results that are competitive with offline systems.
    AbstractPDF
  38. 808Machine LearningarXiv:2608.02585

    GradCuit: Credit-Assigned Gradient Flow Enables Robust and Interpretable Test-Time Latent Reasoning

    • It inserted the latent states to be optimized between prompt representations and generated-token representations at a selected transformer layer, while freezing the base model's parameters.
    • Through causal self-attention, reward-weighted gradients from all generated tokens flow directly to earlier latent states without a decoding detour.
    • Average accuracy across five base models, three reasoning benchmarks, and two answer formats was 64.5%, exceeding the strongest comparison method by 2.4 points and chain-of-thought prompting by 6.6 points.
    • It reduced accuracy variation across seven learning rates from 1.53 to 0.82, and gradient analysis showed that latent influence concentrates on reasoning connectives, also providing interpretability.
    • The effect was strongest in early-to-middle layers and disappeared in later layers. Evaluation was limited to three reasoning tasks, so layer selection and generalization to other tasks require verification.
    AbstractPDF
  39. 807AI / GeneralarXiv:2608.01418

    Reusing Rollouts under Policy Lag: Prefix-Normalized Policy Optimization for LLM Reinforcement Learning

    • PNPO multiplies likelihood ratios through each token position and then takes their geometric mean, retaining causal prefix information while narrowing the weight range.
    • Unlike GSPO, which assigns one weight to an entire response, it uses a different prefix average at each position, but it is a biased approximation that gives up exact state-action distribution correction.
    • Under fourfold reuse, its peak average across three mathematical benchmarks was 50.24, 3.00 points above GSPO, but the difference was inconsistent under single-use training.
    • Under the same budget of 2400 updates, using only 150 fresh rollout batches produced results similar to using 600 batches for single-use training, demonstrating the potential of rollout reuse.
    • The study used only one model, three mathematical tasks, and one run per setting, without separately ablating components, so the effect of prefix normalization alone and its effects on asynchronous and replay data have not yet been isolated.
    AbstractPDF
  40. 806Computer VisionarXiv:2608.02437

    InfiniSplat: Implicit Gaussian Decoding for Large-Baseline Monocular View Synthesis

    • It built an implicit decoder that places Gaussian support points on local surfaces estimated from monocular depth and predicts Gaussian attributes from image features queried at those positions.
    • Decoupling support-point placement and attribute prediction from a fixed pixel grid produced a structure that reduces surface cracks and scattered points even under large viewpoint shifts.
    • After training only on the synthetic indoor dataset Hypersim, it outperformed existing single-image feed-forward methods on ETH3D, ScanNet++, Tanks and Temples, and DL3DV.
    • A single forward pass produces a 3D representation that supports real-time rendering, showing potential for spatial-photo exploration and augmented- and virtual-reality displays.
    • Regions unseen in a single image remain unknown, and incorrect depth estimates propagate directly to support-point placement, leaving structural distortions on reflective, transparent, or thin objects and under extreme viewpoint shifts.
    AbstractPDF
  41. 805LLMs / NLParXiv:2608.02101

    Cross-Domain Hybrid OPD for Generalizable Search Agents

    • It used two-stage training: first learning autonomous planning and iterative search through agent reinforcement learning, then restoring general capabilities through on-policy distillation from experts in multiple domains.
    • Search capabilities were largely preserved after distillation, and several search tasks improved further over the reinforcement-learning stage, demonstrating a balance between specialization and generality.
    • Replacing distillation with reinforcement learning on mixed search and general data reduced HYEval3.1 logical reasoning from 46.90% to 36.31% and BBEH Mini from 47.86% to 35.22%.
    • Domain-specific teachers and a curriculum beginning with easy problems produced better results on difficult reasoning, coding, and mathematical tasks than a single generalist teacher or training without a curriculum.
    • The technical report's evidence focuses on Hunyuan3-based Yuanbao and the presented search and general benchmarks, so generalization to other base models and external operational environments requires additional validation.
    AbstractPDF
  42. 804ResearcharXiv:2608.02023

    SwanTale: Unified Multi-Speaker Speech and Audio Generation for Instruct and Zero-Shot Tasks

    • Speech preprocessing, multilayer captions, quality refinement, and synthesis of rare situations produced approximately 70 million training records describing environments, speakers, and content.
    • It combined SwanVAE and a flow-based transformer with quality conditioning, Engram, and a unified mixture of experts to handle reference speech and natural-language instructions in one model.
    • In single-speaker evaluation with reference speech, it achieved timbre similarity of 0.95 and content error of 0.086, improving over the earlier SwanVoice.
    • In instruction evaluations, it ranked 1st in Chinese with APS 86.1 and tied for 1st in English with APS 84.2, and could generate speech, ambient sound, and sound effects in one waveform.
    • Weaknesses remain in emotional transitions in complex background music, multi-speaker scenes longer than 2 minutes, continuous control of a particular speaker's emotion, stress, and pauses, and role acting.
    AbstractPDF
  43. 803AI / GeneralarXiv:2608.02163

    From Simple QA to Deep Research: A Verifiable Benchmark Constructed through Iterative Task Evolution

    • The benchmark consists of 500 deep-research tasks spanning 31 topics and 10 broad categories, and it separates different capabilities with three query types.
    • The construction loop works as follows: the Explorer solves tasks in a real environment, the Formalizer organizes the results into a DAG and scoring checkpoints, and the Challenger creates the next-round query using clues that have not yet been used.
    • As the rounds progress, the average DAG depth rises quickly from 2.5 in round 1 to 5.2 in round 5 and then levels off at about 7.0, while the number of nodes grows from 4.6 to 25.2, checkpoints from 13.6 to 123.2, and Explorer tool calls from 3.9 to 23.3. The average evolution rounds is 14.19.
    • The difficulty genuinely increases. For the two query types other than the one that gives ample information, the average round score falls from about 0.93 and 0.95 to about 0.29 and 0.40. Even so, the relative ranking of models is preserved in most rounds.
    • Removing tools lowers the score for all ten models. The overall score drops by an average of 0.14 and the evidence-based and analytical checkpoints drop by an average of 0.07 each, which means the tasks require long-tail information that is not well covered in pretraining.
    • The scoring is stable as well. On 100 manually checked tasks, repeated scoring by the same model or by different models both show high Pearson correlation with human evaluation, between 0.884 and 0.916.
    AbstractPDF
  44. 802RoboticsarXiv:2608.02326

    ChainVLA: Chaining Vision-Language-Action Queries through a Unified Execution State for Long-Horizon Manipulation

    • It divides execution state into two parts. The backward-looking part is progress, combining a recurrent task state and sparsely logged event memory. The forward-looking part is the unexecuted tail of the most recent prediction. This state is used only as revisable prior information, not as a fixed plan.
    • On RMBench, the 1.2B model scores 62.8 on average. Among methods that cover all tasks, the strongest Mem-0 uses 8B plus an extra 2B and gets 52.8.
    • The largest gain is on Put Back Block. This task requires putting back a block whose original location is no longer visible, and it scores 96, compared with 90 for Mem-0 and 50 for MemoryVLA.
    • The ablation results are asymmetric. Keeping only progress gives 11.2, keeping only the action tail gives 3.0, and removing both gives 1.6, while the full model gets 62.8. A half-state is not very different from no state.
    • A control that mimicked prior methods was insufficient. A fixed-length observation history combined with temporal ensembling, which smooths overlapping predictions from the back, reached 35.6, 27.2 points below the full model.
    • It shows that smoothness alone is not the cause. The linear extrapolation control has lower error than the full model on all four boundary metrics for Put Back Block and on three of four for Swap Blocks. Yet its success rates are 0% and 2%, while the full model reaches 96% and 74%.
    AbstractPDF
  45. 801Computer VisionarXiv:2608.02583

    UEmbed: Unified Sparse and Dense Multimodal Embeddings

    • It attaches 16 special tokens and partitions the vocabulary into 16 k-means clusters so that each token owns one chunk, and the number of special tokens N was swept across 2, 4, 8, 16, and 32, with stable performance up to 16 and a clear drop at 32.
    • It compresses the vocabulary first, merging tokens that become identical after accent removal, lowercasing, and whitespace normalization, which reduces the vocabulary size from 248,320 to 184,016, and scoring uses the maximum weight among merged tokens.
    • UEmbed-9B scores 71.8 dense and 71.0 sparse on MMEB-v2, making it the best dense retrieval model among those trained on public data and a new state-of-the-art for sparse retrieval.
    • For UEmbed-2B, using a combined score that sums dense and sparse results raises text from 53.6 to 53.9 and visual documents from 77.0 to 77.5, while natural images and videos see little benefit because they contain less surface vocabulary information.
    • On the agentic retrieval benchmark BrowseComp-Plus, sparse mode reduces the number of search rounds. UEmbed-2B drops from 39.19 dense rounds to 32.67 sparse rounds while recall rises from 53.47 to 57.04, and the 9B model keeps the same 49.76 accuracy while reducing rounds from 33.68 to 31.05.
    • The authors also state the limits honestly: modern LLMs with large vocabularies can activate odd tokens such as "_alt", and because the training corpus is biased toward English and Chinese, tokens from other languages are much less likely to appear.
    AbstractPDF
  46. 800Computer VisionarXiv:2608.01964

    LongHorizon-Harness: Advancing Long-Horizon Agents for Real-World Tasks

    • It identifies three main reasons for failure on long tasks: goals drift as errors accumulate, important information becomes hard to find as the record grows longer, and the system loses track of what it has done so far.
    • Its core structure is a Manage-Execute-Audit loop. The manager cannot access the environment directly and sees only state and audit reports; the executor does one subtask at a time in a fresh context; and the auditor checks the environment directly with read-only tools.
    • The only memory passed between rounds is the audit report, while the executor's raw trace is discarded. Termination happens when the audited state satisfies the original goal, when there are no more subtasks to do, when user input is needed, or when the round budget is exhausted.
    • A thin AgentAdapter swaps in a different backend without changing the original agent loop. It can mix harnesses such as Claude Code, Codex CLI, OpenClaw, and Hermes Agent, and it can mix models such as Claude Opus, GPT, and Qwen across the three roles.
    • On the combination of Qwen 3.7-Plus and Claude Code, the pass rate on WeaveBench rises from 51.8 percent to 80.7 percent, which is almost twice the 41.2 percent of the official best-reported Claude Opus 4.7 combination.
    • Across the full OSWorld 2.0 benchmark, the binary completion rate rises from 2.8 percent to 8.3 percent, about a threefold gain, and on 34 subsets it rises from 20.0 percent to 34.3 percent relative to Claude Opus 4.7.
    • In the case study, the baseline method gets 0.00 because it directly edits document XML and only makes the output look right, while the proposed method follows the specified GUI procedure and gets 0.89 after the auditor reparses the XML and checks 15 title styles.
    AbstractPDF
  47. 799RoboticsarXiv:2608.02547

    Why Does Action Chunking Improve Behavioral Cloning Performance in Robotic Control?

    • It tested the three existing hypotheses with controlled experiments and concluded that neither temporal consistency, reduced effective horizon, nor representation learning alone can fully explain the performance of action chunking.
    • A lag policy that predicts only one step ahead from past observations replaces much of chunking's ability to use the past and reduce accumulated error. The validation error was lowest for lags between 5 and 15.
    • The chunking policy learns relationships between actions at different time gaps from a single observation, so it behaves like an ensemble of several lagged policies. The authors call this an implicit ensemble.
    • In every simulated and real-robot environment they tested, deploying the chunking policy as a random lag ensemble matched the performance of action chunking without using action chunks.
    • An explicit ensemble built from multiple independently trained policies performed even better than chunking. They report about a 30% improvement on the Robomimic Transport task.
    • At control frequencies between 15 and 20 Hz, the lag policy worked, but it failed when raised to 50 to 60 Hz. Treating partial chunks of length 5 as a single action made it work again, and the authors explain this by noting that human vision-based behavior updates at 2 to 10 Hz.
    AbstractPDF
  48. 798LLMs / NLParXiv:2608.02046

    CompanionBench: A Theory-Anchored, Real-World-Grounded Benchmark for AI Emotional Companionship

    • The capability system measures 10 capabilities extracted from 25 theory papers on a 1-point to 10-point scale, with 42 sub-items in total, and it adds a reverse-layer list of 18 negative patterns with three severity levels.
    • It adds four capabilities that prior work did not score explicitly: tolerating ambiguity, self-object responsiveness, positive resonance, and appropriately calibrated challenge.
    • The second axis is self-disclosure gate replay, which measures how often disclosure is obtained and how deep it goes, and it also counts disclosures that were elicited without being originally given.
    • It is highly reproducible, with Spearman correlations of 0.996 in Chinese and 0.953 in English for the ranking of the 10 capability scores under independent reruns, and 8 of the 10 capabilities are at or above 0.93; the lowest are positive resonance at 0.829 and calibrated challenge at 0.863.
    • The gate behavior really follows the agent's behavior, because the rate of obtained disclosures and the rate of early self-disclosure are strongly negatively correlated, with Pearson correlations of -0.938 in English and -0.979 in Chinese.
    • Across the 28-model leaderboard, gpt-5.5 ranks first in English with an IRT score of 8.27, while the lowest, doubao-char-251128, scores 4.28; three models specialized for roleplay sit near the bottom, and the biggest deficits are in tolerating ambiguity, calibrated challenge, and acknowledgment.
    • The numbers show that warmth and substance are different dimensions, because claude-haiku-4-5 drops from 0.780 before penalty rules to 0.548 after penalties are applied, and the lowest model drops from 0.395 to 0.017.
    AbstractPDF
  49. 797Computer VisionarXiv:2608.02603

    WorldExam: Benchmarking World Models from Apparent Appearance to Inherent Reactivity

    • The benchmark spans three manipulation styles, four diagnosis stages, eight evaluation tasks, 1,474 test cases, and 20 evaluated models.
    • The eight tasks are Camera Control, Subject Control, Scene Revisit, Terrain Interaction, Object Interaction, Social Interaction, Physical Reaction, and Goal Completion. The first three test instruction following and spatial consistency. The next four test whether the world reacts on its own after seeing the scene, and the final Goal Completion task tests whether the model can plan its own execution steps from only a high-level goal.
    • It breaks the same behavior into atomic control units and converts them to the input format each model expects. Camera-based models receive SE(3) camera trajectories, action-based models receive discrete action sequences, and language-based models receive natural-language prompts.
    • In World Reactivity cases, the expected reaction is removed from the instruction. For example, the model is told only to move forward toward a staircase, and it must infer the vertical adaptation needed to climb the stairs from the scene itself.
    • To match model capabilities, the benchmark is split into two tracks. The static-scene track, where only the camera moves, evaluates all three modalities, while the dynamic-interaction track, where the subject and environment must interact, evaluates only the action and language modalities.
    • From the Table 1 comparison, existing benchmarks have many cases but narrow task coverage. WorldScore has 3,000 cases and one task, iWorld-Bench has 4,900 cases and two tasks, while WorldExam covers all eight tasks with only 1,474 cases.
    • The conclusion is that high image quality and good instruction following do not imply world reactivity. No model combined broad task coverage with consistently high performance.
    AbstractPDF
  50. 796AI / GeneralarXiv:2608.02287

    SKT: Skill-Use Training at Scale via Verified Synthetic Data Generation

    • From 2,000 public skills, it created 4,000 task packages and 27,164 verified execution trajectories, and it also built the SkillEval benchmark from a separate skill pool that does not overlap with the training data using the same pipeline.
    • As the skill budget used for training increases from 0 to 100, 500, 1000, and 2000, the Qwen3.5-9B SkillEval score rises monotonically to 55.24, 59.8, 67.4, 70.8, and 72.48.
    • The benefit transfers across execution environments: a model trained on OpenCode trajectories improves from 4.94 to 9.29 on SkillsBench in DeepAgents, and training on same-environment trajectories raises it to 13.62, meaning it retains 50.1 percent of that gain.
    • A single checkpoint trained on mixed trajectories from two environments beats the original models in all 8 paired comparisons between benchmarks and execution environments, with gains ranging from 2.64 to 18.81 points. SkillEval rises from 55.24 to 74.05 on OpenCode and from 51.62 to 69.96 on DeepAgents.
    • The mixed-trained model stays within an average of 2.71 points of the environment-specific model and even surpasses it in 3 of the 8 comparisons, which means separate models per environment are less necessary.
    • When the 77 SkillsBench tasks are grouped by the number of required skills, one-skill tasks improve from 9.54 to 16.94, two-skill tasks from 3.95 to 20.72, and tasks with three or more skills from 4.04 to 12.08, so the gain is not limited to single-skill tasks.
    AbstractPDF