HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 494Computer VisionarXiv:2606.24253

    TuringViT: Making SOTA Vision Transformers Accessible to All

    • TuringViT introduces Turing linear attention (TLA) to efficiently process high-resolution and dynamically resized visual sequences.
    • Its VISTA-Curation pipeline curates image and video data to improve data efficiency and support native dynamic-resolution training.
    AbstractPDF
  2. 493Machine LearningarXiv:2606.24020

    You Don't Need to Run Every Eval

    • The authors build a public score matrix over 84 frontier models and 133 benchmarks, with 2,604 cells and 23.3% coverage, and show that it is effectively rank-2.
    • They design BENCHPRESS, a rank-2 ALS matrix completion method in logit space, to recover hidden scores within 4.6 points.
    AbstractPDF
  3. 492Computer VisionarXiv:2606.25041

    Wan-Streamer v0.1: End-to-end Real-time Interactive Foundation Models

    • Wan-Streamer models language, audio, and video as both inputs and outputs with a single transformer, and uses block-causal attention to support incremental streaming.
    • It jointly learns cognition, reasoning, generation, response timing, turn management, and cross-modal synchronization in one model instead of relying on separate VAD, ASR, language, TTS, avatar, and video generation modules.
    AbstractPDF
  4. 491Computer VisionarXiv:2606.24805

    DDStereo: Efficient Dual Decoder Transformers for Stereo 3D Road Anomaly Detection

    • This is the first stereo 3D detector to achieve real-time inference at 23.5 ms, at the level of monocular methods.
    • The dual decoder separates foreground binary detection using disparity features from 3D attribute regression using left-view features.
    • Shared object queries align the two decoders and produce anomaly scores without NMS or anchor matching.
    • On the KITTI-AR-OoD open-set benchmark, it outperforms S3AD with 78.09% OoD AP3D on the R40 metric.
    • It achieves SOTA compared with all existing stereo methods on the pedestrian and cyclist classes.
    • Using only an additional 10% of synthetic data, it still reaches 54.65% OoD AP3D, far exceeding S3AD's 24.45% and showing strong generalization.
    AbstractPDF
  5. 490LLMs / NLParXiv:2606.24530

    NatureBench: Can Coding Agents Match the Published SOTA of Nature-Family Papers?

    • It covers 90 carefully selected tasks from 5,500 Nature-family papers across six scientific domains: cell omics, proteins, biological modeling, physical modeling, molecular design, and relational reasoning.
    • NatureGym is an automatic pipeline that builds a containerized task environment directly from a paper, which ensures reproducibility.
    • Even the strongest agent, Claude Opus 4.7, exceeds the published SOTA only 17.8 percent of the time at g greater than 0.1 and matches it 47.8 percent of the time, which shows that the benchmark relies more on translation than discovery.
    • In 45.5 percent of successful paths, the key step was methodological translation rather than scientific invention, meaning conversion into familiar supervised-learning form.
    • The main failure causes are wrong method selection at 45.1 percent and insufficient compute budget at 24.4 percent, not lack of task understanding.
    • An information firewall hides the original authors' methodology, which forces independent discovery instead of reproduction.
    AbstractPDF
  6. 489LLMs / NLParXiv:2606.23525

    Self-Compacting Language Model Agents

    • Fixed token-threshold compression can trigger in the middle of reasoning or retrieval and accidentally delete verified facts.
    • The method combines a compression tool that the model can call directly with a lightweight rubric that says to compress when a subtask is solved and suppress compression during reasoning.
    • When the tool is provided without the rubric, usage varies widely across models and is often invoked at the wrong time, which shows that the rubric is essential.
    • On math benchmarks such as IMO-Answerbench and HMMT, it improves by up to 18.1 points over the no-compression baseline.
    • On agentic search tasks such as BrowseComp, it improves by 5 to 9 points while reducing token cost by 30 to 70 percent per question.
    • Adaptive compression is possible at inference time alone, without any fine-tuning or external supervision.
    AbstractPDF
  7. 488RoboticsarXiv:2606.24742

    World Value Models for Robotic Manipulation

    • This is the first approach to reuse the spatiotemporal prior knowledge of world models as the basis for robotic value estimation.
    • It combines a video-generation stream and a value DiT through a Mixture-of-Transformers, or MoT, to minimize representation interference.
    • It formulates the value function as a flow-matching-trained distributional chunk, which provides dense learning signals.
    • It releases Suboptimal-Value-Bench, which consists of 800 suboptimal trajectories plus human labels.
    • It achieves SOTA on both the expert VOC and Suboptimal-Value-Bench, and it also improves downstream policy learning consistently.
    AbstractPDF
  8. 487Computer VisionarXiv:2606.24888

    DiffusionBench: On Holistic Evaluation of Diffusion Transformers

    • It builds NanoGen, a unified training and evaluation framework that can switch from ImageNet to text-to-image training with only 12 lines of configuration changes.
    • After training 21 latent diffusion models, it confirms almost no correlation between ImageNet rank and text-to-image rank, with Pearson coefficients from -0.377 to -0.580.
    • It provides empirical evidence that better ImageNet FID does not translate into better text-to-image quality.
    • It releases a unified codebase that supports RAE, VAE, pixel-space, and MeanFlow.
    • DiffusionBench is a holistic benchmark that combines ImageNet and text-to-image results and is recommended as the default reporting standard for future DiT research.
    AbstractPDF
  9. 486AI / GeneralarXiv:2606.24855

    OpenThoughts-Agent: Data Recipes for Agentic Models

    • The authors run more than 100 controlled ablation studies on a six-stage SFT data pipeline to quantify the effect of each stage.
    • Instruction selection is the most important factor in the pipeline, which shows that the strongest model is not necessarily the best teacher.
    • Fine-tuning Qwen3-32B on 100K examples yields 54.0 percent on SWE-Bench Verified and 26.2 percent on Terminal-Bench 2.0.
    • The average score is 44.8 percent, which is 3.9 points better than the previous best open-data agent model, Nemotron-Terminal-32B at 40.9 percent.
    • It shows scaling behavior that beats competing open datasets across all training-set sizes.
    • It confirms that an 8B model trained with SFT and RL in two stages outperforms both pure SFT and pure RL.
    AbstractPDF
  10. 485AI / GeneralarXiv:2606.23595

    SPIRAL: Learning to Search and Aggregate

    • It is the first framework to jointly optimize sequential, parallel, and aggregation compute within a single reinforcement learning pipeline.
    • Set reinforcement learning trains parallel traces so that they are collectively useful for aggregation.
    • It achieves up to 11x scaling efficiency in pass@k versus GRPO and 15 percent better performance when parallel and aggregation scale together.
    • It points out that existing training paradigms optimize only sequential reasoning, leaving a gap with test-time scaffolds.
    • Experiments with Qwen3-4B-Instruct validate the method, and the authors identify 8B-plus parameter scaling as future work.
    AbstractPDF
  11. 484Computer VisionarXiv:2606.23688

    Lift4D: Harmonizing Single-View 3D Estimation for 4D Reconstruction In-the-Wild

    • We apply causal latent conditioning to a monocular 3D reconstruction model to ensure temporal consistency across frames.
    • We convert temporally consistent initialization into a deformable 3D Gaussian Splatting representation.
    • We use occlusion-aware rendering supervision, which localizes occluded regions with depth cues and handles visible and invisible regions separately.
    • We apply a view-conditioned diffusion prior to the unseen regions to complete plausible appearances.
    • On the Consistent4D synthetic benchmark, we achieve SOTA with LPIPS 0.116 and FVD 592.44.
    • On DAVIS in-the-wild videos, we demonstrate superior motion accuracy over prior methods, with CLIP 0.715 and EPE 0.161.
    AbstractPDF
  12. 483Computer VisionarXiv:2606.23489

    MeshFlow: Mesh Generation with Equivariant Flow Matching

    • It directly represents meshes as triangle soup, avoiding the limitations of autoregressive serialization such as slow speed and error accumulation.
    • It designs an Equivariant DiT architecture that preserves both face permutation invariance and vertex cyclic invariance.
    • It introduces a coupling loss based on nested optimal transport that removes learning signals that violate symmetry.
    • It achieves the best 1-NNA on three of the four ShapeNet categories and matches the quality of state-of-the-art autoregressive models.
    • It speeds up inference by 18.55x, enabling sub-second mesh generation compared with autoregressive methods.
    • A post-processing denoiser reduces self-intersection rate, Ri, by about 56% with negligible additional latency.
    AbstractPDF
  13. 482RoboticsarXiv:2606.24884

    InSight: Self-Guided Skill Acquisition via Steerable VLAs

    • It automatically splits human demonstrations into VLM planning segments and end-effector trajectories, making the VLA controllable at the primitive-action level.
    • The VLM identifies missing primitive actions in a new task and builds a flywheel that tries them through low-level control, collects success data, and retrains.
    • It achieves 92% and 96% end-to-end success rates on twisting and following without human demonstrations, compared with 32% and 16% for CaP-X.
    • It reaches 80% success on long-horizon tasks chained from 14 primitive actions, such as twisting followed by following.
    • Adding new primitive actions does not reduce existing pick-and-place skill, confirming that forgetting is not a problem.
    • In a Mars-robot scenario, it learns a sweeping behavior autonomously from only mopping demonstrations and succeeds on all 5 evaluations.
    AbstractPDF
  14. 481AI / GeneralarXiv:2606.22902

    Agent-as-a-Router: Agentic Model Routing for Coding Tasks

    • Experiments show that the performance bottleneck in LLM routers is information deficit, not a lack of reasoning ability.
    • When performance statistics are provided, it outperforms heuristic routers by 15.3% relative improvement.
    • The Context-Action-Feedback loop continuously accumulates execution-based experience in streaming tasks.
    • ACRouter reaches 49.98% AvgPerf in distribution and 62.50% AvgPerf on out-of-distribution agentic programming.
    • Static routers such as classifiers and bandits collapse out of distribution, but ACRouter preserves generalization.
    • CodeRouterBench is a regret-based router evaluation environment with about 10K tasks, 8 LLMs, and execution-verified scores.
    AbstractPDF
  15. 480World Models / AgentsarXiv:2606.24597

    Qwen-AgentWorld: Language World Models for General Agents

    • It presents language world models, Qwen-AgentWorld-35B-A3B and Qwen-AgentWorld-397B-A17B, across seven simulated domains.
    • It uses a three-stage training process, CPT for world-model capability injection, SFT for activating next-state prediction, and RL for improving simulation accuracy.
    • It introduces AgentWorldBench and shows superior performance across five frontier models and nine benchmarks.
    • In Paradigm A, the separated approach, thousands of environment simulations for agent RL outperform training in the real environment.
    • In Paradigm B, the integrated approach, world-model training is used as warmup and improves performance on seven agent benchmarks.
    • It trains on more than 10M environment interaction trajectories and supports reasoning based on long chains of thought.
    AbstractPDF
  16. 479ResearcharXiv:2606.22832

    Linear optical Bell state measurement for rotation-symmetric cat codes

    • It implements RS-cat code Bell state measurement with only a half beam splitter (HBS) and photon-number-resolving detectors (PNRDs).
    • It distinguishes the four Bell states using discrete rotational symmetry, photon-number remainder information, and phase information.
    • In the lossless and large-amplitude regime, it enables deterministic Bell state measurement for any symmetry order N.
    • Numerical evaluation of photon loss shows that higher-order RS-cat codes are advantageous in certain loss regimes.
    • Postselection can further increase success probability, and the required photon-number resolution is within practical reach.
    AbstractPDF
  17. 478LLMs / NLParXiv:2606.23321

    Tmax: A simple recipe for terminal agents

    • TMAX-15K includes 14,600 RL environment instances, more than 2.5 times the size of the largest existing terminal dataset.
    • A 9B model reaches 27% on Terminal-Bench 2.0, which is the best result among open models under 30B.
    • It uses difficulty control, personas, and validator diversity to synthesize high-quality environments.
    • DPPO, an FP32 LM head, and a large group size stabilize RL for long-horizon agents.
    • It confirms cross-task and cross-harness generalization, including a +5-point gain on SWE-Bench Verified, showing that RL is not just harness fitting.
    AbstractPDF
  18. 477Computer VisionarXiv:2606.22568

    SeFi-Image: A Text-to-Image Foundation Model with Semantic-First Diffusion

    • Semantic-First Diffusion injects semantic encoder features asynchronously and reaches fast convergence even in high-fidelity VAE spaces.
    • Training the 5B model requires only 125K A800 hours of compute, which is 10 to 20 percent of the compute used by Z-Image.
    • It matches or exceeds Qwen-Image and Z-Image on many benchmarks, including GenEval, DPG, LongTextBench, OneIG, and CVTG-2K.
    • Three scales, 1B, 2B, and 5B, plus a DMD2-distilled turbo variant cover different hardware and latency requirements.
    • It is trained on 450M image-text pairs plus 28M synthetic text-rendering samples, and the code and weights are released.
    AbstractPDF
  19. 476LLMs / NLParXiv:2606.21906

    Deeper is Not Always Better: Mitigating the Alignment Tax via Confident Layer Decoding

    • The Guess-Refine-Perturb pattern shows that early layers guess, middle layers refine, and final layers are disturbed by alignment bias.
    • It uses entropy-based conservative backward search to dynamically select the best nearby final layer on a token-by-token basis.
    • It improves reasoning benchmarks such as GPQA-Diamond, Omni-MATH, and HLE consistently across both dense and MoE architectures.
    • It can be dropped into existing inference pipelines with no extra memory and less than 2 percent latency overhead.
    • The stronger the alignment, the larger the disturbance in the final layer, which means the effect grows with model size.
    AbstractPDF
  20. 475Machine LearningarXiv:2606.23664

    MAS-PromptBench: When Does Prompt Optimization Improve Multi-Agent LLM Systems?

    • A benchmark dedicated to prompt optimization for MAS, covering five topologies, three communication protocols, and team sizes from 2 to 10 agents.
    • The gains range dramatically from as much as +24 points to as much as -16 points, proving that blind application is risky.
    • Optimization works better when communication protocols are more structured: Freeform < Semi-structured < Structured.
    • As team size grows, coordination complexity increases and optimization gains shrink, from +2.4 points for 2 agents to -2.1 points on average for 10 agents.
    • The natural extension of single-agent optimization tools to MAS is not sufficient, and MAS-specific algorithms are needed.
    AbstractPDF
  21. 474RoboticsarXiv:2606.20781

    World Action Models: A Survey

    • The survey defines WAM as a model that predicts actionable futures from observation and action history together with task context, and then uses those predictions to generate action.
    • It organizes WAMs along two classification axes: what should be generated, such as rendered, latent, or free, and how the system is decomposed into substrate, backbone, coupling, and deployment.
    • It identifies five core properties: interactability, causality, persistence, physical plausibility, and generalization.
    • It systematically categorizes more than 200 studies into five backbone families: diffusion, autoregressive, joint, optional-future, and related variants.
    • Its core insight is that WAMs are evolving toward a 'generate less, preserve only what is needed for control' design philosophy.
    AbstractPDF
  22. 473Machine LearningarXiv:2606.23670

    Tapered Language Models

    • It applies a cosine decay to layerwise MLP width, such as 1.5x at the front and 0.5x at the back, to concentrate capacity early and improve perplexity by up to 1.84 points over uniform allocation.
    • It shows consistent gains across four architectures, Transformer, Gated Attention, Hope-attention, and Titans, and three model sizes, 440M, 760M, and 1.3B.
    • It improves downstream benchmark performance without increasing the parameter count or FLOPs.
    • The mechanism analysis shows that deeper MLP outputs align more strongly with the residual stream, which means they reinforce existing information more than they update it.
    • It is a general design lever that is not tied to one architecture and can also be applied to other foundation models such as vision transformers and diffusion models.
    AbstractPDF
  23. 472Image ProcessingarXiv:2606.21588

    Unsupervised Susceptibility Distortion Correction of EPI without Calibration Scans via Image Translation-Based Registration

    • An invertible neural network, or INN, based image translation narrows the contrast gap between BOLD and T1w images.
    • The method applies structural preservation constraints through the MIND-SSC, a modality-independent neighborhood descriptor.
    • Test-time adaptation, or TTA, provides out-of-distribution robustness to scanner and population shifts.
    • The network is trained with a single contrast-similarity objective, so no distortion-correction ground truth is needed.
    • It consistently outperforms prior calibration-free methods on one in-distribution and two out-of-distribution datasets.
    AbstractPDF
  24. 471AI / GeneralarXiv:2606.22883

    CLI-Universe: Towards Verifiable Task Synthesis Engine for Terminal Agents

    • The inside-out approach defines tasks from a structured taxonomy first, then grounds them in evidence-based research rather than reusing existing artifacts.
    • It uses a three-stage verification process: rubric-gated testing, hint-conditional filtering to remove trivial tasks, and fail-to-pass checks.
    • It discards about two-thirds of all candidates, keeping only high-quality tasks to maximize training-signal density.
    • Training Qwen3-32B on CLI-Universe-6K, which contains 6,000 trajectories, raises TB 2.0 to 33.4 percent and makes it the best open-source model at or below 32B.
    • It improves generalization across benchmarks, with gains of +11.3 on BFCL v4 and +11.6 on VitaBench.
    AbstractPDF
  25. 470RoboticsarXiv:2606.23079

    AdaReP:Adaptive Re-Planning under Model Mismatch for Neural World-Model Predictive Control

    • It defines an adaptive replanning threshold based on dynamic regret theory, computed from prediction deviation and local dynamics sensitivity.
    • It requires no training and can be applied simply by wrapping existing world models and planners without modifying them.
    • On VP2 and RoboDesk, it reduces NFE by 59 percent, and on the DMC Suite with TD-MPC2, it reduces NFE by 54.5 percent.
    • On a real Franka robot, it cuts queries by more than 80 percent over 50 trials while keeping the success count at 36 out of 50, compared with 34 out of 50 for the baseline.
    • In sensitive hinge-like regions, replanning frequency increases automatically, while in smooth regions the system reuses cached plans for longer.
    AbstractPDF
  26. 469RoboticsarXiv:2606.23565

    HoloAgent-0: A Unified Embodied Agent Framework with 3D Spatial Memory

    • A three-layer structure: Embodied AgentOS for planning, scheduling, and monitoring; a memory layer with a 3D semantic map; and a skill layer for navigation, manipulation, and motion.
    • A ROS2-based command and state interface supports heterogeneous skill composition and real-time failure recovery.
    • It integrates HoloNavi for navigation, HoloBrain for VLA-based manipulation, and HoloMotion for whole-body motion.
    • Its 3D scene-graph-based spatial memory provides long-horizon task context through a room-object hierarchy.
    • It demonstrates object search, multi-robot collaboration, and closed-loop mobile manipulation on real robot hardware.
    AbstractPDF
  27. 468Computer VisionarXiv:2606.23050

    Unlimited OCR Works

    • Reference Sliding Window Attention (R-SWA) keeps the KV cache constant by giving reference tokens full attention and output tokens a sliding window of 128.
    • On OmniDocBench v1.5, it improves by 6% over DeepSeek OCR, reaching 93%, with an MoE structure that uses only 0.5B active parameters.
    • It can process more than 40 pages at once and is 35% faster when generating 6,000 tokens.
    • Replacing full standard attention with R-SWA does not reduce single-page OCR accuracy.
    • It was developed by Baidu, and the code and weights are released on GitHub.
    AbstractPDF
  28. 467RoboticsarXiv:2606.19836

    World Engine: Towards the Era of Post-Training for Autonomous Driving

    • A four-stage pipeline: uncover long-tail safety events from real logs, turn them into high-quality 3DGS simulations, generate variations with an action world model, and use them for RL post-training.
    • World Engine post-training delivers larger safety gains than simply doubling the pretraining data, while remaining competitive even against about 10x more data.
    • When applied to Huawei's production autonomous-driving system with 80,000 hours of driving data, it reduced simulation collision rates by up to 45.5%.
    • It achieved zero disengagement in on-road tests over 65 km of Shanghai urban driving plus 70 km of nighttime driving.
    • On the 58.3M-parameter small model, repeated post-training caused policy instability, so whether large models can support regular iterative retraining remains unresolved.
    • Beyond autonomous driving, it proposes a safety-learning template for general physical AI such as robot manipulation, walking, and surgical robots.
    AbstractPDF
  29. 466AI / GeneralarXiv:2606.19980

    ENPIRE: Agentic Robot Policy Self-Improvement in the Real World

    • It is the first integrated framework that lets coding agents such as Codex, GPT-5.5, Claude Code, and Kimi K2.6 autonomously improve policies in real robot environments.
    • It has a four-stage structure: Environment with automatic reset and verification, Policy Improvement with heuristics, BC, and RL, Rollout on the real robot, and Evolution where the agent analyzes logs and improves code.
    • In Pin Insertion, it reaches 100% success faster than human-in-the-loop state-of-the-art methods, and 99% on tasks such as PushT and zip tie cutting.
    • When scaled to a multi-robot fleet from 1 to 8 agents, it shows shorter wall-clock time, but token cost grows superlinearly and spikes at MTU 8 agents.
    • It introduces Mean Robot Utilization (MRU) and Mean Token Utilization (MTU) as two physical resource efficiency metrics.
    • It automatically discovers the synergy between VLA, GR00T, and procedural tool calls, improving VLA accuracy in RoboCasa and successfully transferring to the real world.
    AbstractPDF
  30. 465LLMs / NLParXiv:2606.19857

    Large Language Models Do Not Always Need Readable Language

    • It first proposes the concept of BabelTele, a compressed representation that preserves meaning without human readability, retaining 99.5% of meaning at only 27.9% of the original length.
    • It is zero-shot compatible across heterogeneous models such as Gemini, GPT, Claude, Qwen, and DeepSeek, so one model can decode text compressed by another without fine-tuning.
    • In multi-agent communication, it saves 39% to 44% of tokens while preserving task scores between 96.6% and 99.7%.
    • On the LoCoMo agent memory benchmark, it achieves higher accuracy than Summary, 96.48% versus 94.20%, while using fewer tokens.
    • On LongBench v2 code repository QA, BabelTele compression outperforms simple truncation when the context window is exceeded, improving Qwen3.6-Max from 55.17% to 62.07%.
    • BabelTele interpretability is not monotonic with model size, so robustness to the compression format is the key variable.
    AbstractPDF
  31. 464Software EngineeringarXiv:2606.19605

    FAPO: Fully Autonomous Prompt Optimization of Multi-Step LLM Pipelines

    • The Claude Code-based optimization loop follows a pipeline of evaluation, intermediate-step analysis for failure attribution, scoped change proposal, safety review, and verification.
    • The prompt-first strategy prioritizes prompt edits and escalates to chain-structure or parameter changes only when attribution identifies a structural bottleneck.
    • It represents the pipeline as a stateful graph built with LangGraph, and it ensures reproducibility through tenant-isolated workspaces.
    • Across 6 benchmarks and 3 models, meaning 18 comparisons, GEPA wins 15 times with an average gain of 14.1 percentage points, and it wins all 6 structure-change comparisons on HoVer and IFBench with an average gain of 33.8 percentage points.
    • On the security task CTIBench-RCM, it improves GPT-5 by 4.0 points, Foundation-Sec-8B-Instruct by 7.1 points, and Foundation-Sec-8B-Reasoning by 2.0 points.
    • A step-attribution subagent separates prompt-editable issues from structural bottlenecks, which prevents unnecessary structural changes.
    AbstractPDF
  32. 463Machine LearningarXiv:2606.20008

    VIMPO: Value-Implicit Policy Optimization for LLMs

    • It derives a closed-form expression under the optimality conditions of KL-regularized RL by expressing the value function as a policy-reference log-ratio, which removes the need for a critic.
    • It constructs a critic-free value-optimization objective with a terminal boundary condition V*(s_T) = 0 and a Monte Carlo group estimator.
    • From the same derivation, it obtains a PPO-style actor advantage, separating reward injection through the value loss from policy improvement through the actor update.
    • It shows consistent gains over GRPO on MATH-500, AIME 2024, AIME 2025, and OlympiadBench, with especially large improvements on contest-style problems.
    • It remains consistently better than GRPO even under 25 percent reward-flip noise, showing stronger robustness to reward noise.
    • It validates the method for math RLVR with a 4B base model, and shows that tuning the actor coefficient c_A and value coefficient β determines the tradeoff between learning speed and stability.
    AbstractPDF
  33. 462Computer VisionarXiv:2606.20545

    Current World Models Lack a Persistent State Core

    • WRBench defines camera motion as an intervention on observability and evaluates it through a 6-dimensional hierarchical diagnostic chain, from requested-camera to visual integrity to re-observed consistency.
    • Natural-25 uses a design with 25 scene families and four event stages to separate spatial movement from state change orthogonally.
    • Across 23 models and four control paradigms, the common failure is to restore only the state from the time the object was left when the camera returns, without reflecting events that happened while it was unobserved.
    • Scaling Wan from 1.3B to 14B actually worsened re-observed state from 0.66 to 0.62, showing that scaling alone does not solve the problem.
    • LingBot-World has the strongest visible-state consistency, at 0.874 and 0.719, but the lowest requested-camera precision at 0.468, revealing a tradeoff between controllability and state consistency.
    • The key diagnosis is that all public models currently lack what-memory, meaning a record of hidden changes, and an endpoint-persistence training objective.
    AbstractPDF
  34. 461Computer VisionarXiv:2606.20521

    HumanScale: Egocentric Human Video Can Outperform Real-Robot Data for Embodied Pretraining

    • Egocentric pretraining shows a clear scaling law, with validation loss decreasing monotonically as more data is added.
    • At the same 5,000-hour scale, egocentric data reduces validation loss by 24% versus robot data, improves in-distribution success from 40% to 92.5%, and improves OOD success from 0% to 90%.
    • The comparison is based on a World-Action Model with a MoT architecture, using a framework that combines video generation and action prediction.
    • Egocentric video has lower normalized jerk, meaning smoother motion, and a lower idle fraction, which makes the data more efficient.
    • In real-robot experiments on an AgiBot bi-manipulator, the egocentric pretraining model maintained a 90% success rate even on OOD objects, while the baseline collapsed to 0%.
    • Robot data shows almost no improvement in unseen-task generalization even when scaled up, demonstrating the fundamental limitations of real robot data.
    AbstractPDF
  35. 460LLMs / NLParXiv:2606.19847

    AtomMem: Building Simple and Effective Memory System for LLM Agents via Atomic Facts

    • Fact Executor: A lightweight SFT-tuned LLM extracts only the core information from raw conversations as atomic facts, using Qwen3-14B with LoRA rank 128.
    • Hierarchical memory structure: It organizes atomic facts into event memory for context and temporal profiles for tracking user state.
    • Association memory graph: Retrieval uses entity duplication, shared events, and conversation continuity to connect fragmented memories.
    • On the LoCoMo benchmark's Multi-Hop J-score, it improves by 31.1% over LightMem and reduces tokens by 61.4% versus Mem0.
    • Even the simple flat variant, AtomMem-Flat, improves Multi-Hop F1 from 20.97 to 37.03, a 76.6% gain over the baseline, showing that memory representation quality is the key.
    • The retrieval parameter k_f = 10 is the best balance of accuracy, cost, and latency, while k > 20 hurts performance because of noise.
    AbstractPDF
  36. 459Agents / World ModelsarXiv:2510.04542

    Code World Models for General Game Playing

    • The core method uses the LLM not as a policy that directly chooses moves, but as an inductive engine that translates rules and trajectory data into Python code. The generated Code World Model (CWM) is an executable simulation engine that includes state-transition, legal-move enumeration, and terminal-check functions.
    • On top of the CWM, it places MCTS for perfect-information games or ISMCTS for imperfect-information games to perform deep search-based planning. In addition, the LLM is asked to generate a heuristic value function and a hidden-state inference function as code.
    • For imperfect-information games, it proposes a code-based autoencoder paradigm. The inference function serves as the encoder from observation to hidden state, the CWM serves as the decoder from hidden state to observation, and the game rules act as a structural regularizer.
    • It was evaluated on 10 games, five with perfect information and five with imperfect information, including four OOD games newly created for the paper. CWM-MCTS and CWM-ISMCTS performed nearly as well as players based on ground-truth models, and they beat Gemini 2.5 Pro in 9 of the 10 games or tied with it.
    • In perfect-information games, CWM transition accuracy reached 1.0 in most cases, while in imperfect-information games hidden-state inference remained difficult and accuracy dropped to around 52% on the hard game Gin Rummy.
    • The limitations are that learning remains difficult in imperfect-information games with complex rules such as Gin Rummy, and extending the approach to open-world or vision-based games remains future work.
    AbstractPDF
  37. 458Machine LearningarXiv:2606.20560

    How Transparent is DiffusionGemma?

    • The paper decomposes transparency into variable transparency, which is how well intermediate states can be understood, and algorithmic transparency, which is how well the process can be reconstructed.
    • Opaque serial depth drops from 28.6 times the initial Gemma 4 level to 1.1 times when mapped to a token bottleneck.
    • Compressing self-conditioning between denoising steps into O(c) tokens causes only a small downstream performance drop.
    • On monitoring evaluation, DiffusionGemma performs about the same as Gemma 4 in terms of CoT monitoring quality.
    • The paper finds diffusion-specific phenomena such as non-sequential reasoning, token and sequence smearing, and intermediate-context reasoning.
    • It provides a framework and benchmark for auditing the transparency of future latent reasoning models.
    AbstractPDF
  38. 457RoboticsarXiv:2606.19419

    Playful Agentic Robot Learning

    • During play time, the system autonomously proposes tasks, executes them, verifies the results, diagnoses failures, and accumulates a skill library.
    • It uses a novelty-learnability rule to propose curiosity-driven tasks, with curiosity play improving from 24.7 percent under random play to 32.3 percent.
    • On LIBERO-PRO, it improves by 20.6 points over CaP-Agent0, from 23.2 percent to 43.8 percent, and exceeds every VLA baseline, including pi0.5 at 12.8 percent.
    • When the learned skills are plugged into CaP-Agent0, they improve RoboSuite by 8.9 points and real-robot performance by 8.8 points.
    • Skills learned in a single-arm environment also transfer to two-arm tasks, such as two-arm lifting, with a 24.0-point gain.
    • Curiosity play and execution-system improvements are complementary, and both together reach the best score of 44.3 percent.
    AbstractPDF
  39. 456RoboticsarXiv:2606.20562

    MemoryWAM: Efficient World Action Modeling with Persistent Memory

    • The memory design combines a short-term sliding window, long-term gist tokens, and anchor frames at event boundaries.
    • It reduces time and space complexity from O(N) to O(N/d), where d is the compression ratio, making real-time control feasible.
    • On RMBench, it reaches a 83.0 percent average across nine tasks, outperforming LingBot-VA by 4.8 points and FastWAM by a wide margin.
    • On the real robot, it achieves 18 out of 20 on Shell Game and 15 out of 20 on Look and Press.
    • Removing gist tokens causes the largest performance drop, from 92.5 percent to 40 percent, which shows their critical role in long-term memory.
    • Even at 1,600 frames, it remains more efficient than full attention approaches such as RNN and TTT while maintaining 87 percent performance.
    AbstractPDF
  40. 455Computer VisionarXiv:2606.19531

    ImageWAM: Do World Action Models Really Need Video Generation, or Just Image Editing?

    • It points out three limitations of video-generation WAMs, namely cost, irrelevant details, and error accumulation, and proposes image editing as an alternative.
    • It passes the KV cache from the editing backbone to a flow-matching action expert, enabling action prediction without image decoding.
    • It achieves a 98.4% success rate on LIBERO and an 84.5% success rate on real robots, outperforming π0 at 55.8% and FastWAM at 79.0%.
    • It achieves FLOPs reduction from 63.65 to 9.72, a 6x reduction, and latency reduction from 1081 ms to 263 ms, about 4x faster.
    • It can use a variety of editing backbones such as FLUX.2, OmniGen2, and Ovis-U1, so it is not tied to a specific model.
    • Attention analysis confirms that the editing cache focuses on task-relevant changed regions.
    AbstractPDF
  41. 454Computer VisionarXiv:2606.20515

    S-Agent: Spatial Tool-Use Elicits Reasoning for Spatial Intelligence

    • It reframes spatial reasoning as the accumulation of spatiotemporal evidence rather than frame-by-frame prediction.
    • It uses a three-layer tool stack: 2D grounding, 3D geometric lifting, and spatial knowledge aggregation.
    • Scene Memory, which stores object state, and Agent Memory, which stores reasoning history, support stateful reasoning.
    • In a training-free setting, it improves MMSI-Bench by 4.5 percent over GPT-5.4.
    • S-AGENT-8B trained on 300K trajectories improves by 10.5 percentage points over Qwen3-VL-8B, from 31.1 percent to 41.6 percent.
    • The key idea is that a level-3 expert transforms raw 3D data into task-oriented measurements.
    AbstractPDF
  42. 453Machine LearningarXiv:2606.17551

    Reversal Q-Learning

    • It applies an extended MDP framework to offline RL and generates virtual on-policy trajectories through a reverse flow.
    • Multi-step returns reduce the effective horizon from T×F to T, solving the curse of horizon.
    • It uses first-order gradients of the value function without BPTT, outperforming regression-based methods such as FAWAC and IFQL.
    • It achieves the best average performance among 18 baselines on 50 robot tasks in OGBench.
    • It is especially strong on long-horizon tasks such as antmaze-giant and humanoidmaze-large.
    AbstractPDF
  43. 452RoboticsarXiv:2606.18363

    Guava: An Effective and Universal Harness for Embodied Manipulation

    • ReAct-style closed-loop execution is decisively better than one-shot planning on long-horizon tasks.
    • A semantic action space, such as grasp, align, and move, reduces the reasoning burden on the VLM compared with low-level joint control.
    • Training on fewer than 2,000 simulation trajectories and then transferring zero-shot to the real world succeeds on three backbones, with 86 percent on in-distribution tasks and 92 percent on out-of-distribution tasks.
    • RL post-training raises shell game success from 6.7 percent to 60 percent and place-all-red-objects success from 0 percent to 93.3 percent.
    • The system exhibits error awareness and recovery behavior even in joint-limit or unreachable situations that were absent from training.
    • It achieves a 75.6 percent overall success rate, outperforming GPT-5.4 at 70.2 percent and CaP-Agent0 at 62.7 percent.
    AbstractPDF
  44. 451LLMs / NLParXiv:2606.18216

    Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradients

    • It addresses both the mode-seeking bias of teacher-logit imitation and the drift problem caused by injecting teacher response gradients.
    • BCQ places teacher-correct and student-wrong answers as anonymous candidates so the student has to discriminate them while staying on-policy.
    • NCQ aggregates the student's failed rollouts and presents common failure patterns in the prompt.
    • A prompt replay buffer recycles hard questions to amplify learning within the student's zone of proximal development.
    • At 0.8B, it improves average performance by +4.9 points for VLMs, +4.4 points for LLMs, and +2.3 points for video models compared with GRPO, with a large +16.9-point gain on GPQA-D.
    • It confirms that the gains from ZPPO grow with teacher size and shrink with 4B and 9B teachers.
    AbstractPDF
  45. 450RoboticsarXiv:2606.18610

    SC3-Eval: Evaluating Robot Foundation Models via Self-Consistent Video Generation

    • Joint training of forward and backward dynamics corrects autoregressive rollout drift using backward signals.
    • Cross-view inpainting training preserves multi-camera consistency without explicit memory.
    • The backward dynamics mode is reused as a per-chunk uncertainty signal for early stopping when drift appears.
    • Across seven real VLA policies, it achieves a closed-loop Pearson correlation of 0.929 and MMRV of 0.119, outperforming Ctrl-World, IRASim, and Cosmos-Predict 2.5.
    • It classifies the real failure modes of policies, including language understanding, grasping, and releasing, into four categories and achieves more than 70 percent recall.
    • It is trained on 381 hours of real table-bussing data and shows generalization to out-of-distribution tasks.
    AbstractPDF
  46. 449RoboticsarXiv:2606.19333

    Do as I Do: Dexterous Manipulation Data from Everyday Human Videos

    • It proposes a guided diffusion method that repurposes the SAM 3D generation model as a video object tracker, achieving tracking performance 67% better than FoundationPose.
    • It achieves state-of-the-art reconstruction accuracy on the DexYCB and HOI4D benchmarks, with F-5 of 0.71 and 0.72 and CD of 0.66 and 0.49.
    • Three novel components, warmup, perturbation, and transition reward, improve retargeting success from 25% to 71%.
    • It also shows effectiveness on clean data, improving OakInk2 MoCap performance from 72% to 81%.
    • Analysis of the 100DOH dataset shows that only about 4% of online human videos can be directly used for dexterous manipulation learning.
    • It validates real execution on a two-arm system with a 22-DoF Sharpa Wave hand and UR3e arm across 10 tasks, including stirring, pouring, and wiping.
    AbstractPDF
  47. 448LLMs / NLParXiv:2606.18195

    Learning from the Self-future: On-policy Self-distillation for dLLMs

    • It is the first on-policy self-distillation framework for dLLMs that overcomes the limitation of existing OPSD, which was designed only for autoregressive models.
    • By exploiting the non-autoregressive nature of dLLMs, it uses self-generated answers as suffix-conditioned inputs to transfer more novel knowledge.
    • It aligns dLLM's iterative denoising structure with token-level to step-level divergence supervision.
    • It achieves better reasoning performance with only about 10% of the optimization steps used by RLVR, dramatically improving sample efficiency.
    • It consistently outperforms RLVR and SFT baselines on four reasoning benchmarks.
    • It proposes a self-future-experience learning paradigm inspired by the human idea of going back to the past.
    AbstractPDF
  48. 447RoboticsarXiv:2606.18112

    Qwen-RobotNav Technical Report: A Scalable Navigation Model Designed for an Agentic Navigation System

    • It uses a two-axis parameter interface, with task mode and observation parameters such as token budget and camera weight, to support diverse navigation tasks in a single model.
    • Without changing the architecture, randomizing only the training-time parameters makes the Qwen3-VL backbone robust to arbitrary inference-time settings.
    • Training on 15.6M samples and co-training with vision-language data prevents collapse into a reactive action-sequence mapper.
    • It achieves new state-of-the-art results on VLN-CE RxR at 76.5%, EVT-Bench at 90.0%, and NAVSIM at 91.4 PDMS.
    • In agentic system composition, it improves HM-EQA by 10.8% and EXPRESS-Bench by 15.4%, while reducing navigation steps by 77%.
    • It scales favorably from 2B to 8B and shows strong zero-shot generalization in real robot environments.
    AbstractPDF
  49. 446AI / GeneralarXiv:2606.18206

    Fixed-Point Reasoners: Stable and Adaptive Deep Looped Transformers

    • It solves signal propagation problems that grow with loop depth by using pre-norm and residual scaling.
    • It introduces fixed-point convergence through an end-to-end halting mechanism, so no external halting module is needed.
    • It guarantees adaptive computation, with the number of iterations automatically adjusted to task difficulty.
    • On Sudoku Extreme, it matches TRM performance with 10 percentage points higher accuracy and about 27 percent less compute.
    • It shows consistent gains on combinatorial reasoning benchmarks such as Maze, state tracking, and ARC-AGI.
    • It provides theoretical justification for the halting decision based on mathematical fixed-point theorems.
    AbstractPDF
  50. 445LLMs / NLParXiv:2606.18246

    Variable-Width Transformers

    • It systematically explores non-uniform width allocation by breaking the assumption that every transformer layer must have the same width.
    • An X-shaped structure, wide at the ends and narrow in the middle, consistently outperforms equal-width baselines with matched parameter counts.
    • Parameter-free residual resizing implements dimensional reduction and recovery without projection layers.
    • With the same number of parameters, it improves perplexity by about 3 percent, reduces KV cache by about 10 percent, and reduces FLOPs by about 3 percent.
    • Under a loss-matched scaling curve, it reduces FLOPs by 22 percent and extends to MoE transformers.
    • It mitigates representation collapse in the middle layers and encourages qualitatively different representations at different depths.
    AbstractPDF