HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 895Computer VisionarXiv:2608.13556

    V-RAE: Rethinking Video Latent Spaces for Generation

    • Instead of learning the video latent space for pixel reconstruction, it builds it on top of the semantic feature space of a frozen visual backbone.
    • On K600 reconstruction, it achieves an rFVD of 2.13, which beats every large video VAE in the evaluation, and it preserves semantic information in the latent space much better, with 89.13 percent classification on UCF101 versus a VAE best of 30.83 percent.
    • Under the same training conditions, it also leads in generation quality, measured by gFVD, and converges up to 6 times faster.
    • It shows that reconstruction ranking, measured by rFVD, can differ greatly from generation quality ranking, and it proposes tFVD as a metric that better reflects generation suitability.
    • On Cityscapes future video prediction, it also outperforms the Wan 2.2 VAE latent space, with gFVD 111.36 versus 144.47.
    AbstractPDF
  2. 894Computer VisionarXiv:2608.13489

    DreamX-Phi 1.0: Action-Conditioned Video World Model for Robotic Manipulation

    • We inject the 3D motion path of a robot arm geometrically into attention, preventing the model from moving the wrong arm even when the video looks plausible.
    • A depth prediction branch and SAM3 masks, a technique that separates object regions, reduce mistakes where grasped objects disappear or get swapped during the video.
    • We enable fast deployment by distilling the multi-stage generator into a few stages.
    • On WorldArena 2.0, Track 1, it ranks first among 31 teams with EWMScore-P 60.65, and it ranks joint second on Track 2 with 67.19%.
    • On the WorldArena 1.0 offline evaluation, it scores 76.88, which is 3.24 points above the previous best of 73.64.
    AbstractPDF
  3. 893LLMs / NLParXiv:2608.12149

    Massive Activations in Hybrid Linear Attention Large Language Models: Pre-Attention Spikes and Inter-Spike Plateaus

    • The first pattern is a large spike in values immediately before a full attention layer. The second pattern is those values staying high across several linear attention layers between spikes.
    • The same patterns are observed in five linear-attention architectures, six mixture ratios, five input domains, and public models ranging from 1.2B to 397B parameters.
    • In a controlled 340M and 1.3B parameter 12-to-1 mixture, alignment between the layer right before full attention and the spike position was 99.4% to 100% overall.
    • In the 340M control model, both patterns appeared from the first measurement point, after 1B training tokens. The full-attention output gate reduced magnitude but did not eliminate the layer-wise position pattern.
    • The authors explain that fast cancellation makes spikes and slow cancellation makes plateaus, but they leave the cause of the cancellation timing and its actual computational role for future work.
    AbstractPDF
  4. 892RoboticsarXiv:2608.10232

    FACT: Failure-Aware Causal Training for World-Action Models

    • It predicts future video and progress conditioned on actions created first, and it uses failed actions only for outcome prediction, not imitation.
    • On 50 RoboTwin tasks, average success is 81.8% without video co-training, 85.6% with video co-training, and 87.5% when failure data are added.
    • On a real dual-arm robot, tasks seen during training improve from 82% to 89%, and unseen variants rise from 67% to 77%.
    • Video prediction PSNR on failure cases improves from 19.51 to 25.92, while success cases stay almost unchanged at 26.12 and 26.08, so normal prediction is preserved.
    • Evaluating four action candidates by progress gives 92% on seen tasks and 82% on new variants, but candidate evaluation requires extra computation.
    AbstractPDF
  5. 891Machine LearningarXiv:2608.11612

    Dion3: Full-Stack Orthogonal Updates

    • Gram Newton-Schulz rewrites existing methods into mathematically equivalent form and reduces compute by iterating on small symmetric matrices instead of large rectangular ones.
    • It adds a symmetric GPU kernel and a grouped-communication strategy for matrices of the same shape. In the 8-way setup for a 1-billion-parameter model, grouped communication alone cuts optimizer-step time from 80.7 milliseconds to 52.1 milliseconds.
    • At each step, it orthogonalizes only a subset of rows in the momentum matrix to reduce compute and communication, and it improves both speed and loss over the previous compression method, Dion.
    • In experiments training 3B to 14B models on 10 billion tokens, all models achieved lower validation loss, and the 14B model had 0.027 lower loss than NorMuon and 0.7 percentage points higher average accuracy.
    • The reported up to 6x speedup applies only to the optimizer step, not full training, and the authors analyze that this step accounts for roughly 1 percent to 17 percent of total training time depending on the setup.
    AbstractPDF
  6. 890Computer VisionarXiv:2608.12179

    Map-Det3D: Metric Feed-Forward 3D Reconstruction Prior for Multi-view 3D Object Detection from Streaming Inputs

    • It processes RGB frames from a short time window in order and applies the scale ratio predicted by the 3D reconstruction model to create object boxes in real-world units.
    • It removes the step that converts 2D detections into 3D and directly predicts boxes in 3D space by combining geometric information from multiple viewpoints.
    • On CA-1M, the AP15 of the single-frame baseline was 11.7, and using time information, camera information, and learnable multi-view processing raised it to 21.2.
    • Without any extra training, it achieved AP15 of 15.2 on ScanNet200, surpassing the previous monocular best of 11.7.
    • On CA-1M, it outperformed existing monocular and multi-view image methods, but it remained below methods that use real depth or 3D point data. It is also limited to detection that does not distinguish between indoor scenes and object categories.
    AbstractPDF
  7. 889ResearcharXiv:2608.11898

    On-Policy Self-Distillation for Multi-Dialect ASR: Mastering Dialects, Retaining Mandarin

    • The full training set is about 100,000 hours of standard Mandarin and dialect speech, and the final self-distillation stage uses about 5,000 hours of dialect speech with high error rates.
    • The student uses its own generated prefix as input, while the teacher also consults the ground-truth sentence and provides probabilities for each next-output candidate; only the student model is used in deployment.
    • Evaluation is performed on eight standard-Mandarin public datasets, five public dialect datasets, and 18 internal dialect datasets.
    • Compared with the base model, the average character error rate drops from 3.46 percent to 3.27 percent for standard Mandarin, from 15.37 percent to 12.79 percent for public dialects, and from 21.01 percent to 12.42 percent for internal dialects.
    • Continuing ordinary supervised learning on the same data raises the standard-Mandarin error rate to 4.43 percent, while self-distillation keeps it at 3.27 percent, and the overall average error rate is 10.12 percent instead of 10.74 percent.
    AbstractPDF
  8. 888Machine LearningarXiv:2608.11859

    Small-Scale Experiments: Are We There Yet?

    • The authors experiment from about 4 million to 268 million parameters, fitting the scaling law on models from 4 million to 34 million parameters and then validating it on larger models.
    • Testing only 4 or 16 size settings failed to reveal the law, 64 settings were inaccurate, and 256 settings recovered the correct law.
    • Small models were highly sensitive to training settings, but the sensitivity decreased as the model size grew, and near the optimum the number of influential hyperparameters fell to about one.
    • The law derived at small scale extrapolated well to models about 10 times larger, but the uncertainty grew sharply at farther scales.
    • With this method, they reproduce the large-scale result that pre-normalization becomes more advantageous as the model grows, although the mapping between loss and capability holds only when the composition of the pretraining data is kept the same.
    AbstractPDF
  9. 887RoboticsarXiv:2608.11739

    G0.5: One Autoregressive Stream for Robot Reasoning and Action

    • Reasoning and action use the same model weights and the same token stream, and there is no separate action-expert model.
    • It trains 14 robot embodiments in both real and simulated environments and maps their different action spaces into one 27-dimensional shared space and one token vocabulary.
    • The thought process includes subgoals, object locations, 2D movement paths, and action hints, while visual memory uses the most recent few seconds of video.
    • The average success rates are 76.7 percent on real robots, 82.5 percent on DROID's new environments and objects, 98.9 percent on LIBERO, 93.3 percent on RoboTwin 2.0, and 87.3 percent on SimplerEnv-Bridge.
    • The BEHAVIOR long-horizon score is 31.4 percent, which is higher than the comparison model's 26.3 percent, but visual memory is limited to a few seconds and lower-body actions are not evaluated separately.
    AbstractPDF
  10. 886Machine LearningarXiv:2608.12307

    AI4AI at Test-Time: Strong-to-Weak Capability Transfer via Harnesses

    • It combines prompts, problem-specific handling rules, calculation code, and answer-format checks without changing the target model's weights.
    • The hidden test set contained 3,900 questions across four theory-of-mind tasks, and 195 separate validation items were provided for improving the harness.
    • GPT-5.4-mini's base average score was 0.488, while the overall average with the automatically generated harness was 0.763 and the best score was 0.912.
    • Deterministic code, problem-specific path selection, and strict output formatting had a larger effect than long reasoning or extra sampling.
    • The improvement was large on weak target models but some performance degradation also appeared on strong target models, and the experiments were limited to theory-of-mind tasks.
    AbstractPDF
  11. 885Computer VisionarXiv:2608.12313

    AVA-Encoder: Towards Agent-Native Video Representation Learning

    • It organizes people, backgrounds, objects, style, camera, and audio states under story, event, and scene hierarchies, and links the related assets.
    • It turns reconstruction errors into natural-language correction instructions, then improves both the shared encoding policy and the video-specific knowledge graph in two stages.
    • Policy improvement used 6 videos, and evaluation was done on 18 non-overlapping videos, 129 scenes, and 246 key frames.
    • The overall reconstruction score is 49.0 percent, which is 20.7 points higher than the strongest external baseline at 28.3 percent.
    • When comparing policies only, it reaches 45.8 percent, beating the human-tuned policy at 44.4 percent, and it reduces the system prompt from 31,336 tokens to 8,052 tokens.
    AbstractPDF
  12. 884Computer VisionarXiv:2608.12314

    StateFlow: Building, Evolving, and Accessing 3D World States for Previsualization

    • It stores the world as an editable state composed of each object's 3D shape, position and pose, and semantic attributes.
    • Front-view images are used for object appearance and top-view images for spatial layout, and a vision-language model resolves conflicts when the two disagree.
    • Scene expansion, style changes, and object movement or pose changes are handled by editing parts of the state or swapping assets, which reduces the need to regenerate the whole scene.
    • It proposes camera paths first and then checks rendered results to correct occlusion, bad composition, collision risk, and unstable motion.
    • The average VBench score is 0.8484, the highest among the compared methods, and in a user study with 30 participants the combined scene and video scores were both 4.5 out of 5.
    AbstractPDF
  13. 883Machine LearningarXiv:2608.10703

    Your LLM, Your Style: Behavioral Mode Axes for LLM Behavioral Control

    • It builds 3,200 behavioral scenarios, 20 patterns times 4 situations, to overcome the limits of survey-based personality measurement.
    • It finds that an LLM's behavioral style is stable across models but still changes meaningfully with situation, such as first person, advice, or task mode.
    • It extracts behavioral mode axes from the activation space and demonstrates across seven models that adjusting only one direction can precisely control behavioral style.
    • BMA extracted from the thought process provides much more accurate control than BMA extracted from responses, revealing a trait drift phenomenon.
    • It shows that the behavioral control layer is concentrated in a specific region of the model rather than spread uniformly across the whole network.
    AbstractPDF
  14. 882RoboticsarXiv:2608.11204

    Surgical WAM: A World-Action Model for Data-Efficient Surgical Robot Learning

    • The authors propose a two-stage learning method that first learns visual dynamics in surgical scenes from unlabeled endoscopic video and then fine-tunes on a small amount of labeled action data.
    • Video pretraining raises the average success rate on four SurRoL benchmark tasks from 63.5 percent to 77.8 percent.
    • On PegTransfer, it improves by 20 percentage points, from 66 percent to 86 percent, and the effect is strongest on tasks that involve many contacts and two-handed cooperation.
    • The video-pretrained model reaches higher performance with only half the fine-tuning budget, which greatly improves training efficiency.
    • The same trend appears on real surgical robot data from JIGSAW, which confirms that the effect is real and not just a simulator artifact.
    AbstractPDF
  15. 881RoboticsarXiv:2608.10860

    Flex-$π$: A Multi-Stream World-Action Model with Compute Flexibility

    • It discovers a free lunch in which a video-generation model's VAE can encode not only RGB but also 3D point maps with almost no loss.
    • It builds a multi-stream world-action model that predicts RGB, 3D depth, and DINO object features together in a shared latent space.
    • During training it randomly drops modalities, or stream dropout, so that one checkpoint works with any combination at inference time.
    • On robot tasks in RoboTwin, with only 50 demos, it is 1.9x better than existing WAMs, and on a real dual-arm robot it achieves 2x to 7x higher success.
    • In action-only mode it is faster than π0.5 and still achieves higher success, at about 60 ms.
    AbstractPDF
  16. 880AI / GeneralarXiv:2608.11095

    Why Does CLAUDE.md Keep Growing? Catastrophic Remembering in Agentic Coding

    • It tracked 247,694 instructions across 1,867 GitHub repositories and showed that agent prompts grow by 226% over their lifetime while the deletion probability only keeps falling.
    • The cause is incomplete recall: once the reason for adding an instruction disappears, it cannot be safely deleted, so additions continue indefinitely, with O(2^|D|) complexity.
    • 76.8% of instruction deletions happen in a single commit that rewrites the entire file, so individual deletions are rare.
    • If you annotate the prompt with why each instruction is needed, you can remove 99.3% of unnecessary instructions, reducing growth from +211.3% to +1.4%.
    • On the real WildIFEval benchmark, prompt annotations improved instruction-following accuracy by up to 23.1%.
    AbstractPDF
  17. 879AI / GeneralarXiv:2608.10915

    ComBodied Agents: a New Paradigm of Human-Centric Agentic AI

    • It proposes the Combodied Agents paradigm, which goes beyond the digital-agent and robot-agent split by treating the human state trajectory itself as the primary modeling target.
    • It designs an integrated closed loop of multimodal perception, long-term memory, personal world models, adaptive intervention policies, and feedback.
    • It clearly distinguishes this from human digital twins: it uses uncertainty-aware representations tailored to the task and does not require full replication.
    • It organizes the full design space, including edge-native personal models, scenario-centered evaluation, autonomy-preserving metrics, and governance directions.
    • It argues that success should be defined not by replacement of human work, but by improvement of human capability and judgment.
    AbstractPDF
  18. 878AI / GeneralarXiv:2608.08888

    Full-bandwidth transformer

    • We introduce latent feedback decoding, which feeds the full top hidden state from the previous token directly into the next input.
    • It adds less than 1% extra inference cost while preserving the standard Transformer structure and KV cache.
    • At 1B parameters and 400B training tokens, it consistently improves validation loss, math and code generation, and instruction following.
    • It saves about 1.5x training tokens for the same performance and in some tasks approaches models trained with 5x more tokens.
    • When latent feedback is enabled, the reasoning trace becomes shorter while accuracy is maintained or improved.
    AbstractPDF
  19. 877Computer VisionarXiv:2608.10744

    Beyond Pixels: From Video Priors to 4D Worlds

    • It proposes Latent-to-4D, a framework that connects the VAE latent space of video generation models directly to a 4D decoder.
    • It skips the step of generating RGB images, which avoids information loss and error accumulation.
    • It is trained on about 1,000 existing reconstruction clips and can be reused across multiple video generators with a single checkpoint.
    • On the Text4D-200 benchmark, it improves DINO-F1 by 2.88 to 3.45 points, and on I4D-200 it improves by 5.81 points.
    • Human evaluation shows that it outperforms prior methods in geometric validity, completeness, and temporal stability.
    AbstractPDF
  20. 876Neural NetworksarXiv:2608.09888

    BDH-CQ: In-Context Learning with Recurrent Latent Reasoning

    • It proposes a 'latent reasoning' approach that performs recurrent reasoning in hidden state space without outputting tokens.
    • With a very small 150M-parameter model, it reaches 29.5% accuracy on ARC-AGI-1 at a cost of $0.0007 per task.
    • It sets a new cost-accuracy standard while cutting cost to about one-thousandth of the previous best system.
    • It systematically analyzes the strengths and limits of in-context learning from demonstrations alone, using experiments with color mapping, boundary propagation, rotation, and other compositions.
    • It designed an architecture that can scale naturally to 1 trillion parameters, and early experiments confirm Transformer-like scaling laws.
    AbstractPDF
  21. 875LLMs / NLParXiv:2608.09096

    Evo-Bench: Can Language Models Improve Agent Harness?

    • It creates the first benchmark for evaluating whether an LLM can autonomously improve the agent harness, meaning the execution code.
    • It introduces a strict design that avoids overfitting through harness-sensitivity-based task selection and separate validation and evaluation splits.
    • Evaluation on nine frontier models shows gains of up to 16.6 points and comes close to the human-designed baseline.
    • In the general domain, autonomous evolution beats the human design, but it shows limits in office workflows.
    • It confirms that the evolved harness acts as a transferable reasoning structure for other policy models.
    AbstractPDF
  22. 874AI / GeneralarXiv:2608.09703

    Matryoshka Language Model Suites

    • We propose the Matryoshka framework, which trains submodels of different sizes end to end in a single nested structure.
    • It reduces total training compute by 36% while matching the performance and validation and out-of-domain perplexity of independently trained models.
    • Free distillation from the largest model to the smaller ones happens automatically at every step.
    • The nested draft models improve speculative decoding throughput by 14% to 26%.
    • Extensive ablations validate the importance of width and depth settings, the junction mechanism, and the distillation loss.
    AbstractPDF
  23. 873Computer VisionarXiv:2608.09873

    Sci-VBench: Evaluating Knowledge- and Reasoning-Intensive Video Generation in Science Domains

    • We built 1,253 expert-annotated video-generation tasks across 60 courses in four fields: natural sciences, medicine, humanities and social sciences, and engineering.
    • We provide an expert-written evaluation rubric and step-by-step scoring criteria from 1 to 5 for every example.
    • Non-expert raters also achieve high agreement with experts when given the rubric, with Cohen's kappa = 0.842.
    • Evaluation of 16 frontier models reveals similar visual quality but large gaps in scientific accuracy.
    • Prompt improvements help somewhat, but the limits in spatial and temporal consistency come from the generator itself.
    AbstractPDF
  24. 872AI / GeneralarXiv:2608.09629

    Rethinking Self-Evolving Agents: Do We Still Need Prescribed Optimization Pipelines?

    • It is the first systematic attempt to question whether framework-defined improvement procedures, or optimization meta-policies, are still necessary.
    • Using GPT-5.5, OEO beats the SKILLOPT and GEPA pipelines in 12 of 14 comparisons, with 1 tie and 1 loss.
    • OEO uses only 34.3 percent of the median tokens used by SKILLOPT, which demonstrates its efficiency.
    • For mid-sized models, prescribed pipelines work better, which reveals a model-capacity boundary for adaptive division of labor.
    • The prescribed procedures make optimization paths more consistent, but the path differences are larger than the final action differences.
    AbstractPDF
  25. 871Computer VisionarXiv:2608.09438

    Unveiling the Secret of AdaLN-Zero in Diffusion Transformer

    • It breaks down the performance gains of adaLN-Zero into three factors: SE structure, zero initialization, and progressive updates.
    • Zero initialization is the most important factor, and the paper finds that the conditional modulation weights follow a Gaussian distribution after training.
    • It proposes adaLN-Gaussian, which initializes with a Gaussian distribution, to validate the analysis and obtain practical gains.
    • It proposes SE-adaLN-Zero, inspired by the SE structure, and achieves additional performance gains.
    • It confirms consistent improvements and generalization on four datasets, including ImageNet1K.
    AbstractPDF
  26. 870Machine LearningarXiv:2608.09228

    Privileged Solutions or Context-Induced Teacher Behavior? Dissecting On-Policy Self-Distillation

    • It designs a controlled experiment, OP2SD, to distinguish whether OPSD's gains come from answer transfer or context-induced behavior.
    • Even when the teacher is given solutions to other math problems instead of the correct answer, the improvement is similar to OPSD.
    • When the context comes from physics rather than math, the effect weakens, showing that domain alignment in the context is necessary.
    • It confirms the same pattern across three Qwen3 models, 1.7B, 4B, and 8B, and on the AIME and HMMT benchmarks.
    • It shows that OPSD should not be interpreted only as privileged-information transfer, because context-induced behavior is also an important factor.
    AbstractPDF
  27. 869Computer VisionarXiv:2608.09931

    Perception Before Supervision: Self-Contained Visual Distillation from Counterfactual Blind Spots

    • It proposes a three-stage verification criterion that finds visual blind spots using only the model's own counterfactual responses, without external models, annotations, or tools.
    • It performs contrastive distillation with an expanded view of the blind spot as the positive teacher and a removed view as the negative teacher.
    • On Qwen3-VL-8B, it improves all 12 benchmarks with no regression.
    • It records gains of +3.60 on OCRBench, +3.38 on fine-grained perception in MMStar, and +3.08 on logical reasoning.
    • It outperforms all six self-evolution baselines that depend on external GPT-4o supervision.
    AbstractPDF
  28. 868Computer VisionarXiv:2608.09730

    World Tokens: Enhancing Embodied Policies with Training-Time World Modeling

    • It proposes the World Tokens architecture, which uses a video world model only during training and removes it at deployment.
    • The World Adapter converts VLM features into 256 tokens and conditions both video prediction and action generation on them.
    • The action expert cannot see the full VLM sequence directly, which prevents it from bypassing the world-model supervision.
    • With a 2B model, it reaches 98.2 percent on LIBERO, 71.5 percent on SIMPLER-WidowX, and 82.1 percent on SIMPLER-GoogleRobot.
    • On the real R1 Pro robot, it improves success from 59.4 percent to 76.0 percent compared with the matching baseline.
    AbstractPDF
  29. 867AI / GeneralarXiv:2608.09836

    Mismatch Matters: On-Policy Distillation Beyond Token Agreement

    • We discovered a degenerate agreement phenomenon in which the student cheats by repeating tokens to achieve high token-level agreement with the teacher.
    • We distinguish between surplus tokens, which the teacher scores near zero, and missing tokens, which the teacher wants but the student does not produce.
    • We propose TIDE, which corrects surplus tokens with Hellinger-distance-based bounded suppression and missing tokens with analytical top-K injection.
    • It raises Avg@8 from 6.9% to 20.3% under strong teacher-student disagreement.
    • It also cuts average response length by 3.6x while greatly reducing format failures.
    AbstractPDF
  30. 866RoboticsarXiv:2608.09892

    XPolicyLab: A Unified Standard and Open Ecosystem for Robot Policy Evaluation and Deployment

    • It proposes a unified standard that reduces the cost of connecting robot policies to environments from O(NM) to O(N+M).
    • It integrates 42 robot policies, including VLA, WAM, and imitation learning, behind one interface.
    • It creates a structure that isolates policy servers and environment clients according to software dependencies.
    • It cuts adaptation time for a new policy from more than 5 hours to 30 minutes when packaged skills are used.
    • It supports RoboTwin and RoboDojo simulation, as well as real robot evaluation, through one adapter.
    AbstractPDF
  31. 865RoboticsarXiv:2608.09853

    RynnValue: Scaling Robotic Value Foundation Models with Temporal Distance

    • It changes the supervision target for reward models from progress to temporal distance, meaning cost-to-go to the goal, so scaling no longer requires preference labels.
    • It trains on heterogeneous robot data covering 7,000 hours and about 3 million clips.
    • Time-order shuffling and value-isolation attention keep the model from relying only on the learned values.
    • It surpasses the previous best preference-labeled model, 0.655, by reaching 0.675.
    • In real robot RL, it improves success rates from 52.5 percent to 72.5 percent online and from 63.8 percent to 82.5 percent offline.
    AbstractPDF
  32. 864Computer VisionarXiv:2608.09926

    Learning How the World Evolves: Extrapolative Video World Models via Latent Dynamics Reasoning

    • We propose LDR, the first video world model to explicitly infer the laws of motion in latent space.
    • The architecture enforces learning of physical laws by numerically integrating the lower-order dynamics and having the model regress only the third- and higher-order residual that drives the rollout.
    • It operates with 26x fewer parameters and 143x faster speed than video diffusion models.
    • Its out-of-distribution error gap is more than 20x smaller, showing strong extrapolation ability.
    • It accurately generalizes learned dynamics even when color, shape, and orientation change.
    AbstractPDF
  33. 863Computer VisionarXiv:2608.09449

    Sekai2: From World Exploration to Interactive World Modeling

    • It provides 2,826 hours of real-world video, 128,892 clips, and camera trajectories with hierarchical annotations across 113 countries.
    • It creates 649,597 time-level segments that separate subject motion, environmental change, static scenes, and camera behavior.
    • It offers 982 loop videos revisiting 360-degree panoramas, which provide supervision for long-range spatial consistency.
    • It cross-validates that the captions generated independently of camera annotations are aligned with actual camera motion.
    • It provides a scalable data foundation for video generation, camera control, and pretraining interactive world models.
    AbstractPDF
  34. 862AI / GeneralarXiv:2608.09119

    Motif 3: Technical Report

    • It uses an ultra-fine MoE architecture that selects only 8 experts per token out of 384, increasing capacity while reducing compute.
    • Grouped Differential Latent Attention, or GDLA, achieves both noise suppression and KV cache compression.
    • It is pretrained on 12.5 trillion tokens and then uses multi-teacher on-policy distillation, or MOPD, to unify diverse capabilities into one model.
    • It supports a 256K context length and stabilizes large-scale training with MXFP8 precision.
    • Among open-weight models, it shows competitive performance on agent tasks, math reasoning, and scientific knowledge.
    AbstractPDF
  35. 861Machine LearningarXiv:2608.09819

    Macaron-V1: Towards Open Continual Learning with Self-Improvement and Mixture-of-LoRA

    • The system uses a Mixture-of-LoRA structure that keeps the base model frozen and selects expert adapters for chat, agent, coding, or GenUI on each turn.
    • It proposes a framework for continual learning after deployment through joint model and harness design plus a recursive self-improvement loop.
    • It releases Venti, a 744B model based on GLM-5.2, and Tall, a 50B model based on Qwen3.6.
    • It also develops UI4A, a component-native GenUI harness, and MindForge, an agent RL framework.
    • It shows competitive performance with frontier models on Personal Intelligence, GenUI, and general capability benchmarks.
    AbstractPDF
  36. 860RoboticsarXiv:2608.09381

    JEPA-WAM: Learning Vision-Language-Action Policies with Joint-Embedding World Modeling

    • It proposed a latent world-action model that predicts relationships between current and future spatial structure in a pretrained V-JEPA representation space, without generating video directly.
    • It combined transition prediction and continuous action generation in a single predictor so that future supervision directly refines the core structure that forms action representations.
    • On LIBERO-Plus, the setting without prior robot-policy training achieved 79.2%, and the setting combined with π0.5 achieved 86.3%, each yielding the highest result.
    • It also generalized to changes in visual and spatial conditions on RoboTwin 2.0 and a real dual-arm robot, showing potential for robot control without the cost of video generation.
    • Because the transition objective learns shared visual structure independent of language, it may lack expressive power for tasks where different instructions produce substantially different futures from the same scene.
    AbstractPDF
  37. 859LLMs / NLParXiv:2608.09802

    SWE-Bench ProMax: Benchmarking Agents on Large-Scale Multilingual Code Refactoring

    • It quantifies the problem that 60% of existing benchmarks contain bad tests that are either too narrow or too broad.
    • It builds a multilingual refactoring benchmark across seven programming languages: Python, Java, TypeScript, Go, C, C++, and Rust.
    • The tasks are large, with an average of 11.4 files and 261.6 lines of code to change, which makes them more than three times more complex than prior benchmarks.
    • GLM-5 shows strong value for money by matching Claude Sonnet 4.6 at about one twentieth of the cost.
    • Failure analysis shows that the main problem is not modifying enough of the files that actually need to change.
    AbstractPDF
  38. 858AI / GeneralarXiv:2608.09867

    Stealing Reasoning Traces from Proprietary LLM APIs

    • It discovers a fundamental design flaw in encrypted reasoning blocks that makes them interchangeable across sessions, users, and models.
    • It demonstrates a cross-model attack that turns the reasoning of a strong model into something that can be decoded by a weaker model against OpenAI, Anthropic, and Google.
    • It recovers real sensitive information from public developer logs, including 62 API keys, 33 passwords, and 30 emails.
    • It also confirms an invisible prompt-injection attack vector that passes through encrypted blocks.
    • It highlights the fundamental limits of encryption-based approaches and proposes concrete mitigation measures for both vendors and users.
    AbstractPDF
  39. 857LLMs / NLParXiv:2608.06329

    Benchmarking the Benchmarks: Evaluating Benchmarks for Conversational Agents

    • It proposes an LLM-judge-based framework for evaluating the consistency, complexity, and policy coverage of conversational-agent benchmarks without reference labels.
    • It validates the framework with benchmarks generated by LLMs of different strengths and consistently distinguishes quality differences across benchmarks made by models with different capabilities.
    • The metrics also detect the expected drop in quality on deliberately degraded benchmarks.
    • Correlation tests against human annotation show a medium-to-strong statistically significant correlation.
    • It also applies the framework to manually curated benchmarks, such as tau3-bench, to demonstrate practical usefulness.
    AbstractPDF
  40. 856Computer VisionarXiv:2608.07409

    UniJEPA: A Unified Joint-Embedding Predictive Architecture for Task-Agnostic Visual World Modeling

    • It proposes a single objective that jointly learns image-level transformation prediction and video-level temporal prediction in one latent space.
    • End-to-end training is possible without EMA, stop-gradient, or a pretrained encoder, and Gaussian regularization prevents representation collapse in a mathematically grounded way.
    • Photometric prediction learns invariant structure, temporal prediction learns equivariant dynamics, and the balance between them controls the level of abstraction.
    • It achieves competitive or better performance than individual JEPA models on ImageNet, Something-Something-v2, Epic-Kitchens, and control benchmarks.
    • After action-conditioned post-training, it supports zero-shot planning and plans up to tens of times faster than generative world models.
    AbstractPDF
  41. 855Software EngineeringarXiv:2608.03392

    Self-Evolving Coding Agents

    • It clearly defines self-evolving coding agents as distinct from general coding agents and general self-evolving agents.
    • It classifies six evolution targets: framework, memory, skills and tools, model, workflow, and topology, which it calls an object-centered taxonomy.
    • It adds two auxiliary axes: when evolution happens, such as during a task, after a task, or step by step, and software-specific evidence, such as execution feedback, repository context, and coding traces.
    • It argues that executable feedback, repository-level context, and coding traces make software engineering a natural domain for agent self-evolution.
    • The paper's challenges are feedback reliability, benchmark overfitting, safety, maintainability, cost, and generalization, and the paper list is publicly available on GitHub.
    AbstractPDF
  42. 854Machine LearningarXiv:2608.07110

    Modular TTT: Rethinking Test-Time Training as Composable Modules

    • It expresses the internal learner of TTT as a DAG and builds a modular framework where loss functions, learning rates, weight decay, and normalization can be adjusted independently.
    • Small learning-rate initialization, weight decay, and single-layer nonlinearity, specifically SiLU, consistently improve performance.
    • In contrast, deep fast-weight networks and normalization hurt performance because activation values become too large. Residual connections and gating have little effect.
    • The best variant trains 410M and 1.45B models on 100B tokens and reaches performance on par with Gated DeltaNet.
    • Developed at ByteDance, it releases an optimized implementation that improves training throughput several times over prior TTT systems.
    AbstractPDF
  43. 853Computer VisionarXiv:2608.07468

    SimWAM: A Simple World Action Model for End-to-End Autonomous Driving

    • We propose SimWAM, a structure that uses video generation only as a training signal and removes it at inference time, so real-time video generation is unnecessary and inference is faster.
    • Because the video expert and action expert do not share parameters, it is easy to replace the video model with a better one or scale the action model.
    • On NAVSIM, it achieves 91.5 PDMS. It performs better than prior world-action-model-based planners while keeping inference latency much lower.
    • We additionally apply reinforcement learning (GRPO) to optimize complex driving rewards beyond trajectory imitation.
    • Zero-shot transfer to the nuScenes dataset also confirms generality.
    AbstractPDF
  44. 852AI / GeneralarXiv:2608.07449

    SkillProx: Self-Evolving Agent Skills via Proximal Textual Gradient Descent

    • It formulates skill modification as a forward-backward optimization problem. The forward pass verifies diagnosis, revision, and execution, while the backward pass estimates the usefulness of each knowledge unit for integration or removal.
    • It introduces a result-based verification loop that reruns the same task after revision and rolls back if performance drops.
    • In the backward pass, it measures the contribution of each knowledge unit with a leave-one-out scheme, and removing bad knowledge improves accuracy from 46 percent to 54 percent.
    • Across three backbone LLMs, it achieves an average accuracy gain of 3.0 percentage points over the previous best method.
    • Unlike prior methods that only keep adding skills, its key distinction is that it explicitly asks what should be removed.
    AbstractPDF
  45. 851Computer VisionarXiv:2608.07408

    Addressable Memory for Video World Models

    • We precisely diagnose the cause of long-term memory failure in video world models as distribution shift in positional encoding (RoPE).
    • WorldTrace assigns each memory summary a virtual position within the training range, restoring memory accessibility without extra training.
    • WorldTrace-Field improves temporal consistency by 15.5%, and WorldTrace-Landmark improves recall of previous scenes by 19.5%.
    • We propose LoopBench, a new benchmark that measures whether a model can recover after a long detour back to the original scene.
    • It was developed at NVIDIA and can be applied directly to existing models without retraining.
    AbstractPDF
  46. 850LLMs / NLParXiv:2608.06867

    LLMRouter: Unified Infrastructure for Developing, Evaluating, and Deploying LLM Routers

    • It proposes a unified framework that decomposes routers into five components, allowing more than 16 heterogeneous routers to be compared in the same language.
    • It automatically runs 18 models across a range of benchmark tasks to build routing evaluation data, xRouteBench, covering general NLP, vision, time series, and personalization.
    • The learned router improves by 14.6% relative to fixed model selection. However, when the cost constraint gets tight, the lighter router wins.
    • User-specific routing consistently produces personalization gains, but only when the user context is modeled well.
    • It releases the open-source LLMRouter library so that new routers can be added in a few lines.
    AbstractPDF
  47. 849LLMs / NLParXiv:2608.06370

    The Bitter Lesson of Tool Calling

    • On 11 of 14 models, the Python code method matches or beats the JSON method in accuracy, and the GPT-5.6 family improves by up to 10.6%.
    • The code method is overwhelmingly better for chained tool calls, reaching 18.8% higher accuracy than JSON when the chain length is at least 12.
    • JSON misses calls when more than 70 tools are invoked at once, while the code method keeps 100% accuracy up to 100 tools.
    • Under context rot, the code method remains stable, while JSON drops by 2.3% on average and the filesystem method drops by 32%.
    • PTC performance is determined more by model generation time than by model family.
    AbstractPDF
  48. 848ResearcharXiv:2608.05459

    Moduli of the Sourceless Framed Beltrami-Vekua Normal Form

    • It reduced the sourceless Beltrami-Vekua equation on a simply connected domain with boundary to a minimal form on the unit disk and precisely characterized the remaining symmetries.
    • When the coefficient field has no zeros, the pair ϑ and K completely classifies the equation, and it also determined the range of possible invariants.
    • In phase-integrable regions containing zeros, it expanded the classification data to the three quantities ϑ, dη, and d⋆η, separating vortex charge from the remaining phase curvature.
    • It explicitly constructed inequivalent equations with the same mass and charge, showing that those two numbers cannot capture the infinite-dimensional classification space.
    • Multiply connected domains require additional period data, and complete classification remains open for rough measurable coefficients or zeros that are not phase-integrable.
    AbstractPDF
  49. 847Machine LearningarXiv:2608.06216

    Continual Learning in Transition

    • It argues that continual learning used to mean only parameter updates, but now it should also include elements outside the model, such as memory, skills, and protocols.
    • It proposes a three-axis framework, When, Where, and How, so that dozens of existing methods can be compared on a single map.
    • It summarizes that on-policy learning, where the model learns from its own experience, is better than off-policy methods for reducing forgetting.
    • It analyzes verifiable reward as the key bridge that lets agents improve themselves without human help.
    • It separates areas where research is dense from areas that are nearly empty, pointing out that system-level integrated learning is the main challenge ahead.
    AbstractPDF
  50. 846Computer VisionarXiv:2608.05903

    Robust-WAM: Bridging Generative Pretraining and Semantic Foresight in World-Action Models

    • It proposed a post-training method that adds semantic information robust to appearance changes while preserving large-scale video pretraining in VAE-based world-action models.
    • It inserted training query tokens with temporal positions into the action stream and aligned them with DINOv3 semantic representations of future scenes, keeping additional inference cost low.
    • On LIBERO-Plus, it raised out-of-distribution success rates for FastWAM and GE-Act by 9.2 and 8.6 percentage points, respectively, while preserving performance under normal conditions.
    • With lighting-color changes on a real Franka robot, it raised GE-Act success from 57.3% to 80.0%, and the same method could be added to multiple world-action architectures.
    • The tested changes centered on lighting and color across two simulation benchmarks and one real robot, so robustness to more complex changes in objects, backgrounds, and actions needs further verification.
    AbstractPDF