HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 444AI / GeneralarXiv:2606.16222

    Latent Thought Flow: Efficient Latent Reasoning in Large Language Models

    • This is a new approach that solves the language-space bottleneck of Chain-of-Thought, meaning the token-decoding overhead, by reasoning in latent space.
    • A continuous GFlowNet learns a distribution over latent reasoning trajectories proportional to reward, preserving diverse accurate and efficient paths.
    • Entropy-Weighted Subtrajectory Balance propagates terminal rewards to intermediate steps, which addresses the sparse-supervision problem.
    • Reference-prior regularization provides stable early exploration, and annealing then shifts the model toward reward-based trajectories.
    • Compared with CoLaR and ReGuLaR, it improves fine-tuning accuracy by 12.9 percent and shortens reasoning length by 34.5 percent.
    • In transfer-learning settings, it also improves accuracy by 6.0 percent and shortens reasoning length by 19.9 percent, which demonstrates generality.
    AbstractPDF
  2. 443Computer VisionarXiv:2606.18180

    EgoCS-400K: An Egocentric Gameplay Dataset for World Models

    • The dataset contains more than 400K first-person videos and more than 10,000 hours of gameplay data built from professional CS and CS2 demos.
    • It uses tick-level precise alignment so that keyboard and mouse input, camera motion, weapon and movement state, and game events are all synchronized.
    • The hierarchy goes from player-view sequence to segment to action chain to atomic action to per-tick state trace.
    • It covers more than 1,000 matches, more than 40,000 rounds, 13 maps, and up to 10 player viewpoints per round.
    • Its replay-grounded structure makes parsing, replaying, rendering, and temporal alignment all possible, which distinguishes it from simple video capture.
    • It supports action-conditioned future prediction, state-aware scene rollout, and egocentric action understanding.
    AbstractPDF
  3. 442ResearcharXiv:2606.18050

    A Semi-Analytical Loss Cone Theory for Tidal Disruption Event Rates Around Kerr Black Holes

    • It is the first semi-analytical incorporation of the spin-dependent loss-cone boundary of Kerr black holes into TDE rate theory.
    • Using tidal tensor formalism, it derives tidal disruption and direct capture thresholds as a function of orbital inclination.
    • It derives a closed-form capture fraction for the one-dimensional loss-cone problem with nested disruption and capture boundaries.
    • It formulates a two-dimensional Fokker-Planck equation that describes simultaneous diffusion in angular momentum magnitude and direction.
    • Perturbation analysis proves a first-order bias favoring retrograde star disruption, but shows that the global TDE rate is nearly insensitive to spin.
    • It provides an analytical foundation for population-level TDE studies that include spin and orbital inclination.
    AbstractPDF
  4. 441Computer VisionarXiv:2606.16295

    VisualClaw: A Real-Time, Personalized Agent for the Physical World

    • A cascade gate made of perceptual hash, a 128-dimensional CPU encoder, and an adaptive change gate compresses one hour of streaming video into only 5 to 20 API calls.
    • Hot and cold top-k skill injection prevents prompt cost from exploding as the skill bank grows.
    • An offline evolver based on failures automatically updates memory and the skill bank, achieving self-evolution without weight updates.
    • Across four video-QA benchmarks, it cuts API cost by an average of 98 percent and raises accuracy by an average of 3.85 percent, with EgoSchema peaking at +15.80 percent.
    • It builds VisualClawArena with 200 scenarios, supporting workspace-style agent evaluation through a five-stage curation pipeline.
    • With Codex, macro accuracy rises by 2.9 percent for GPT-5.5 and by 3.2 percent for Claude Code Sonnet 4.6, while cost falls by 9.5 percent.
    AbstractPDF
  5. 440Computer VisionarXiv:2606.18249

    Unified Multimodal Autoregressive Modeling with Shared Context-Visual Tokenizer is Key to Unification

    • A single bitwise visual tokenizer truly unifies the representation space for understanding and generation for the first time.
    • Multi-level feature fusion uses shallow layers for low-level detail and deep layers for high-level semantics at the same time.
    • Lookup-free BSQ quantization has the effect of a 2^64 codebook, exponentially expanding the vocabulary without an explicit codebook.
    • Parallel bitwise prediction achieves a 32x visual compression ratio and predicts only 256 tokens for 1024x1024 images.
    • It reaches SOTA on image generation benchmarks such as GenEval and ImgEdit while staying competitive on multimodal understanding benchmarks.
    • It trains efficiently with a total of about 33k GPU hours, using a lightweight design with a 400M visual tokenizer and a 2.5B decoder.
    AbstractPDF
  6. 439RoboticsarXiv:2606.17846

    Qwen-RobotManip Technical Report: Alignment Unlocks Scale for Robotic Manipulation Foundation Models

    • It proposes a three-dimensional alignment framework for representation, motion, and action, using canonical state, camera-frame delta pose, and in-context adaptation.
    • It builds a human-to-robot synthesis pipeline that converts human egocentric video into trajectories for 15 robot platforms.
    • It builds a pretraining corpus of about 38,100 hours of manipulation data using only open-source data, with no proprietary data collection.
    • In out-of-distribution evaluation on LIBERO-Plus, RoboCasa365, and EBench, it outperforms π0.5 and GR00T-N1.7 across all axes.
    • It ranks first on the RoboChallenge Table30-v1 generalist track and achieves a 20 percent relative improvement.
    • It has been validated on four real robot platforms: AgileX ALOHA, Franka, UR, and ARX.
    AbstractPDF
  7. 438LLMs / NLParXiv:2606.17861

    GameCraft-Bench: Can Agents Build Playable Games End-to-End in a Real Game Engine?

    • It is the first work to define three requirements for engine-based game generation evaluation: Engine Grounding, Artifact Completeness, and Interactive Verification.
    • It introduces a real engine benchmark with 140 Godot tasks across 15 game genres.
    • It uses a three-stage interaction-grounded evaluation pipeline that goes from headless build to play video to rubric-based multimodal judging.
    • The strongest agent, Opus 4.7 high, reaches 41.46%, while most agents remain below 40%, which shows that end-to-end game generation is still unsolved.
    • Agents can implement mechanics, but they fall far short on content depth, visual feedback, and presentation polish.
    • Mechanics, content, and visuals are moderately correlated at r=0.5 to 0.6, but art and presentation are weakly coupled at r=0.11, which shows that the capability splits only partially.
    AbstractPDF
  8. 437Machine LearningarXiv:2606.18208

    Looped World Models

    • It is the first application of a looped transformer to a world model and proposes iterative latent depth as a new scaling axis.
    • By repeating a shared transformer block, it achieves up to 100 times better parameter efficiency than prior designs.
    • A spectrally constrained state-retention parameterization guarantees stability over arbitrary rollout lengths.
    • Adaptive computation automatically assigns more loops to complex transitions such as collisions and contacts and fewer loops to simple dynamics.
    • It achieves 66 to 88 percent EM and 72 to 98 percent F1 on scientific simulation tasks such as ScienceWorld.
    • With only 1B parameters, it demonstrates better predictive performance on scientific simulation than Gemini and Qwen.
    AbstractPDF
  9. 436Machine LearningarXiv:2606.17024

    ExpRL: Exploratory RL for LLM Mid-Training

    • It proposes two ExpRL variants, ExpRL-Outcome and ExpRL-Process, that use reference answers as scoring rubrics for an LLM judge rather than as imitation targets.
    • On Qwen3-4B, ExpRL-Outcome reaches 34.2 percent pass@1 on AIME-25, far exceeding GRPO at 27.1 percent and SFT at 5.7 percent.
    • The improvement in pass@k confirms that coverage, meaning the diversity of solution strategies, also improves.
    • It observes more reasoning behaviors, such as verification, self-correction, and backtracking, than in the base model.
    • For judges of 4B or larger, the matched reference for a problem shows the lowest misplacement rate, confirming that the judge validates based on the reference.
    • Mixed-domain experiments across math, coding, and science suggest that ExpRL can scale beyond a single domain.
    AbstractPDF
  10. 435AI / GeneralarXiv:2606.16733

    A First-Principles Derivation of LLM Policy Optimization: From Expected Reward to GRPO and Its Structural Extensions

    • It unifies all policy gradient methods into two axes, trajectory side and reward side, for J(theta) = E[R(tau)].
    • It explains the evolution from REINFORCE to PPO to GRPO to post-GRPO variants through the failure modes and interventions at each stage.
    • It extends the same two-axis view to agentic RL and hybrid GRPO-OPD.
    • It identifies coupled failures that cannot be solved by changing only one axis and points to directions for next-generation algorithm design.
    AbstractPDF
  11. 434Computer VisionarXiv:2606.16255

    UniDDT: Unifying Multimodal Understanding and Generation with Decoupled Diffusion Transformer

    • A noisy ViT encoder serves as a unified semantic encoder for both understanding, with clean inputs, and generation, with noisy inputs.
    • A separate diffusion decoder separates text decoding from visual generation, which minimizes modality interference.
    • The model constructs a dual data structure from the same image-text pair so that it can exploit the duality between understanding and generation.
    • It achieves 0.87 on GenEval, 86.9 on DPG, 1699.5 on MME, and 76.5 on SEEDbench, which shows competitive performance on both understanding and generation.
    AbstractPDF
  12. 433RoboticsarXiv:2606.17055

    T-Rex: Tactile-Reactive Dexterous Manipulation

    • The architecture uses variable-rate MoT control, separating a low-frequency action expert at about 5 Hz from a high-frequency tactile expert at about 20 Hz.
    • A spatio-temporal VQ-VAE-based tactile encoder compresses raw tactile feedback into a compact representation.
    • The authors release a 100-hour tactile-synchronized bimanual manipulation dataset, the T-Rex Dataset, covering more than 200 everyday objects and 22 motor primitives.
    • On 12 tactile-reactive tasks, it achieves a 65 percent average success rate, which is 30 percent better than the strongest baseline.
    AbstractPDF
  13. 432Computer VisionarXiv:2606.15956

    You Don't Need Strong Assumptions: Visual Representation Learning via Temporal Differences

    • It is the first approach to self-supervised learning that does so with only a causality assumption and without strong inductive biases such as augmentation, masking, or cropping.
    • It trains the frame encoder and motion encoder jointly so that current frame plus motion delta equals the next frame representation.
    • It experimentally verifies the hypothesis that weaker inductive bias is better as data scale increases.
    • It achieves performance comparable to DINO and iBOT on ADE-20K semantic segmentation and MPI-Sintel optical flow.
    AbstractPDF
  14. 431RoboticsarXiv:2606.17011

    ROVE: Unlocking Human Interventions for Humanoid Manipulation via Reinforcement Learning

    • It decomposes the hesitation, adaptation, and recovery stages of human remote teleoperation intervention for humanoid robots, which supports mixed-quality data modeling.
    • Optimistic Value Estimation combines TD bootstrapping and expectile regression to extract high-value actions first.
    • It adds cross-embodiment human experience videos to critic training, which strengthens long-tail failure modes.
    • Three real task intervention rounds improve success rates from 45 to 80 percent on the whiteboard task and from 56.7 to 86.7 percent on the toaster task.
    AbstractPDF
  15. 430AI / GeneralarXiv:2606.14502

    From Chatbot to Digital Colleague: The Paradigm Shift Toward Persistent Autonomous AI

    • It proposes a four-stage LLM evolution framework: Chatbot, Thinking LLM, Agent, and OpenClaw with Workspace and Skill.
    • It summarizes the data paradigm shift from instruction-response pairs to State-Action-Observation trajectories.
    • It describes the evaluation paradigm shift from static benchmarks to task closure rate, sandbox auditing, and self-evolving ecosystems.
    • It defines workspace persistence, skill reuse, verification loops, and governance as the core elements of a digital colleague.
    AbstractPDF
  16. 429AI / GeneralarXiv:2606.16140

    VibeThinker-3B: Exploring the Frontier of Verifiable Reasoning in Small Language Models

    • With only 3 billion parameters, it reaches 94.3 on AIME26 and 97.1 with CLR, at the level of DeepSeek V3.2 671B and GLM-5 744B.
    • On the latest LeetCode contests from 2026-04 to 2026-05, it achieves a 96.1% acceptance rate, outperforming GPT-5.2 at 95.3% and Kimi K2.5 at 90.6%.
    • We propose a staged post-training pipeline consisting of curriculum SFT, MGPO RL, offline self-distillation, and Instruct RL.
    • We introduce the Parametric Compression-Coverage Hypothesis, which says that verifiable reasoning can be compressed, while open-ended knowledge requires coverage.
    AbstractPDF
  17. 428LLMs / NLParXiv:2606.15079

    Ling and Ring 2.6 Technical Report: Efficient and Instant Agentic Intelligence at Trillion-Parameter Scale

    • It mixes Lightning Attention and MLA in a 7:1 ratio, which substantially reduces long-context FLOPs and KV-cache burden.
    • Evo-CoT, LPO, and shortest-answer distillation improve token efficiency by about 4x compared with the 2.0 generation.
    • KPop asynchronous agentic RL enables stable environment-based learning even at the 1T scale.
    • Ring-2.6-1T records strong results on PinchBench at 87.60, ClawEval at 63.82, GAIA-2 Search, and tau2-Bench Telecom.
    AbstractPDF
  18. 427Computer VisionarXiv:2606.14703

    Gaze Heads: How VLMs Look at What They Describe

    • About 100 of the 1,152 attention heads in a VLM, or 8.7 percent, act as gaze heads that track the image region currently being described.
    • A single attention-mask intervention achieves 83.1 percent accuracy on comic panel redirection, whereas a random head fails.
    • The same mechanism appears across models with 2B to 32B parameters and across VLMs such as Qwen2-VL, Ovis1.5, and InternVL3.5.
    • Frozen-encoder families such as LLaVA and Bunny-3B do not show gaze heads, which suggests that joint training of the vision encoder is necessary.
    AbstractPDF
  19. 426RoboticsarXiv:2606.15768

    LaWAM: Latent World Action Models for Efficient Dynamics-Aware Robot Policies

    • LaWM is a 230M-parameter latent world model that reuses the decoder from the latent action model, using 95 percent fewer world-modeling parameters than pixel-space WAMs.
    • It predicts latent visual subgoals in a single forward pass and reaches 187 ms through non-iterative inference, up to 24 times faster than pixel-space WAMs.
    • The two-stage training pipeline first trains LaWM on about 3,000 hours of robot data and 1,500 hours of egocentric human video, then trains the policy through latent-action distillation.
    • In real-world evaluation, it achieves 93.3 percent on Pick-and-Place, 86.7 percent on Drawer Opening, and 90.0 percent on Towel Folding, for an overall average of 90.0 percent and first place among all baselines.
    AbstractPDF
  20. 425RoboticsarXiv:2606.17046

    Geometric Action Model for Robot Policy Learning

    • The GFM middle layers are split so that shallow layers contain the observation encoder and causal future predictor, while deep layers are reused as the action and geometry decoder.
    • A single autoregressive token sequence and a single forward pass generate both action tokens and future-scene tokens at the same time.
    • Compared with the existing foundation-model-scale baseline, it is 55 times faster at inference while using fewer parameters and matching or exceeding performance.
    • It is 9.7 percentage points better in the LIBERO-Plus camera perturbation setting, which reflects the direct effect of the 3D geometric prior.
    AbstractPDF
  21. 424Computer VisionarXiv:2606.14702

    OmniVideo-100K: A Dataset for Audio-Visual Reasoning through Structured Scripts and Evidence Chains

    • Entity-Anchored Scripting uses the main entity list as a global prior to keep references consistent across segments and recover audio-visual associations.
    • Clue-Guided QA Generation first mines cross-segment multimodal clues and then generates QA pairs from those clues.
    • OmniVideo-100K contains 100K automatically generated audio-visual QA pairs and OmniVideo-Test with 505 human-verified samples, covering 10 AV tasks.
    • Fine-tuning gains are +20.59% for VITA-1.5, +17.82% for Qwen2.5-Omni-7B, and +13.86% for Qwen3-Omni-30B on OmniVideo-Test.
    AbstractPDF
  22. 423Computer VisionarXiv:2606.16993

    DreamX-World 1.0: A General-Purpose Interactive World Model

    • E-PRoPE applies PRoPE's projection camera geometry to spatially reduced tokens, reducing inference latency by about 30%.
    • Memory-Conditioned Scene Persistence maintains scene consistency when revisiting previously observed areas through camera-geometry-based retrieval.
    • Causal forcing, DMD distillation, and long-rollout training reduce autoregressive chunk accumulation drift, including style and color drift.
    • On 8x RTX 5090, it streams at up to 16 FPS using mixed-precision DiT, residual reuse, 75% pruned VAE decoding, and asynchronous pipeline parallelism.
    AbstractPDF
  23. 422Computer VisionarXiv:2606.17030

    Qwen-RobotWorld Technical Report: Unifying Embodied World Modeling through Language-Conditioned Video Generation

    • It standardizes more than 20 embodiments and more than 500 action categories into a single language interface by using natural language as a unified action interface.
    • A double-stream MMDiT with 60 layers fuses semantic information from Qwen2.5-VL and video-VAE latents through layer-wise joint attention.
    • The EWK dataset contains 8.6 million video-text pairs and more than 200 million frames, including 5.9 million manipulation episodes, 200K autonomous-driving episodes, and more than 6K navigation episodes.
    • It ranks first on EWMBench with 4.60, ahead by 33 percent over HSD, first on DreamGen Bench with 4.952, first among open-source models on WorldModelBench with 8.99, and scores 0.804 on PBench.
    AbstractPDF
  24. 421Computer VisionarXiv:2606.14048

    WAM4D: Fast 4D World Action Model via Spatial Register Tokens

    • The spatial register token handles depth readout only during training and is removed at inference time, so it adds no latency cost.
    • Causal Mixture Attention defines modality-specific visibility among video, action, and geometry tokens in the MoT backbone.
    • On RoboTwin 2.0, it improves by 7.5 points over pi0 and by 6.1 points over LingBot-VA on the selected 10 tasks.
    • Its inference latency is 525 ms, which is slower than the 64 ms of VLA but still fast among WAM systems, and removing the depth branch leaves it at 9.71 GiB of VRAM.
    • On real-world long-horizon four-step manipulation tasks, it improves by 13.4 points over pi0.
    • The trainable pretrained geometric head is more efficient when it is unidirectional rather than bidirectional, and both beat random initialization.
    AbstractPDF
  25. 420AI / GeneralarXiv:2606.13669

    Agents-K1: Towards Agent-native Knowledge Orchestration

    • The five-module schema structures full-text papers into metadata, explicit mention, implicit abstraction, citation intent, and inter-entity relation.
    • A 4B model trained with GRPO and a rule-based reward outperforms an 8B open-source model on 10 benchmarks and matches a 32B model on NER.
    • After processing 2.46 million papers, it builds Scholar-KG and releases 1 million papers across six fields: CS, chemistry, biology, earth science, physics, and materials.
    • GraphAnything CLI provides a unified interface that combines web search, multimodal graph retrieval, and cross-document traversal.
    • On FrontierScience-Research, it improves Gemini-3 from 7.9 percent to 24.6 percent and GPT-5.2 from 25.2 percent to 39.4 percent.
    • It reaches SOTA over nine graph-RAG baselines on HotpotQA, 2WikiMultiHopQA, and MuSiQue.
    AbstractPDF
  26. 419AI / GeneralarXiv:2606.14249

    HarnessX: A Composable, Adaptive, and Evolvable Agent Harness Foundry

    • We compose harnesses in a type-safe way using a 9-dimensional processor taxonomy, covering context, tools, skills, control, memory, and more, plus a substitution algebra.
    • AEGIS is a four-stage evolution pipeline, Digester, Planner, Evolver, Critic, that defends by mapping RL pathologies into symbolic space.
    • Variant isolation prevents interference through independent evolution per task cluster on heterogeneous benchmarks, yielding +13.6 on GAIA.
    • Cross-harness GRPO jointly loops harness evolution and model learning, adding another +4.7 over harness-only training.
    • On ALFWorld, Qwen3.5-9B improves by +44.0%, showing that harness improvements matter more when the model is weaker.
    • The average gain is +14.5% across 15 settings, with positive improvements in 14 of them, and the lower the baseline, the larger the gain.
    AbstractPDF
  27. 418RoboticsarXiv:2606.13769

    $μ_0$: A Scalable 3D Interaction-Trace World Model

    • TraceExtract automates 3D trace extraction by selecting semantic keypoints through DINOv2 entity clustering, expanding the scale by about 8x over prior work.
    • It represents 3D traces with B-spline control points and trains a VLM backbone plus a permutation-equivariant trace expert with flow matching.
    • It pretrains from video only, without action labels, then freezes the model and trains a separate action expert for cross-embodiment transfer.
    • It reaches a 30.25% average on RoboCasa 8 tasks, beating π0's 25.25% by 5.0 points, and achieves 91.7% on three real-world UR3 tasks.
    • A +18.4-point gap over a VLM plus action expert without trace features shows that trace representations provide real motion information.
    • Event-centric captioning aligns language with trace segments, and depth input is not needed at inference time.
    AbstractPDF
  28. 417Computer VisionarXiv:2606.14700

    RepFusion: Leveraging Multimodal Priors for Denoising in Representation Space

    • It extends the MLLM's MLP projector in the RAE latent space to noisy inputs and conditions the DiT with a frozen LLM backbone.
    • At similar inference FLOPs, RepFusion's 1.3B DiT outperforms TextEmbed at 8B DiT and Transfusion at 8B joint.
    • When switching from VAE to RAE, RepFusion improves by 30 absolute points, from 0.54 to 0.70, which is a larger gain than TextEmbed's 21 percent and Transfusion's 11 percent.
    • It performs better when the MLLM is frozen than when it is fine-tuned, which shows that preserving the pretraining prior is important.
    • Test-time compute can be used efficiently through stepwise MLLM reconditioning.
    • RepFusion-SFT reaches state-of-the-art-level performance with GenEval at 0.87, GenEval2 at 34.9, and DPG at 85.11.
    AbstractPDF
  29. 416Computer VisionarXiv:2606.13674

    RepWAM: World Action Modeling with Representation Visual-Action Tokenizers

    • We identify a problem with the reconstruction-oriented tokenizer in existing WAM: it misses manipulation semantics, so we propose a semantic visual-action tokenizer.
    • RepViTok aligns video autoencoder latents with DINOv2 and induces latent actions in the same space.
    • A two-stage training process, pretraining on visual plus latent actions and then adapting to robot behaviors, outperforms joint prediction.
    • On RoboTwin 2.0, it reaches 86.6 on Easy and 83.1 on Hard, a large improvement over the WAN2.2 VAE-based baseline across an average of 50 tasks.
    • Its best performance occurs at CFG scale 1.0, showing that semantic alignment greatly reduces dependence on video CFG.
    • In latent-action visualizations, it shows patterns more focused on manipulation-related changes than LAPA does.
    AbstractPDF
  30. 415RoboticsarXiv:2606.14409

    Hy-Embodied-0.5-VLA: From Vision-Language-Action Models to a Real-World Robot Learning Stack

    • It collects 10,000 hours of sub-millimeter-precision egocentric UMI data in-house and reuses it for both pretraining and post-training.
    • It combines the Hy-Embodied-0.5 MoT backbone with a flow-matching action expert and a compact memory encoder, using a delta-chunk representation to achieve kinematics independence.
    • FlowPRO is an offline RL post-training method that learns directly from success-failure pairs without a critic or reward model and reaches near-ceiling success rates.
    • It uses two SFT tracks, Track A for adaptation to a target robot and Track B for UMI-only cross-embodiment transfer, to deploy across diverse robot platforms.
    • It implements high-frequency closed-loop control with asynchronous inference and cubic Bézier trajectory smoothing, and it plans to release a 2,000-hour UMI data subset.
    AbstractPDF
  31. 414Computer VisionarXiv:2606.13644

    Surflo: Consistent 3D Surface Flow Model with Global State

    • It compresses variable-view input into K latent tokens using a frozen VGGT backbone and Perceiver compression, producing a fixed-size representation independent of the number of views.
    • A flow-matching decoder transports each point from noise to surface independently, which supports arbitrary resolution from thousands to one million points.
    • Inference-time photometric and monodepth guidance resolves local inconsistencies that arise from independent decoding.
    • It is the only feed-forward method that overcomes the limitations of both per-view pointmap duplication and fixed-resolution global-latent approaches.
    • On eight 3D reconstruction benchmarks, it reaches the top feed-forward level and is more than 10x faster than optimization-based methods.
    • It plans to build and release a dataset that adds watertight meshes to about 10.5K DL3DV scenes.
    AbstractPDF
  32. 413Computer VisionarXiv:2606.13673

    SpatialClaw: Rethinking Action Interface for Agentic Spatial Reasoning

    • The agent loop is designed around a stateful Python kernel that lets the system observe intermediate results and modify code at each step.
    • It overcomes the limitations of both single-pass code with pre-commitment and structured tool calls with restricted composition.
    • Across 20 spatial reasoning benchmarks, including static and dynamic 3D and 4D settings, it reaches 59.9 percent on average, which is 11.2 points higher than prior spatial agents.
    • It improves performance consistently across six VLM backbones, from Qwen 27B to 397B and Gemma4, without benchmark-specific or model-specific tuning.
    • The largest gains appear in dynamic 4D video reasoning and multi-view inference.
    • It is training-free and is achieved only through the action-interface design, without changing the backbone model or system prompt.
    AbstractPDF
  33. 412Machine LearningarXiv:2606.13657

    Dense Supervision, Sparse Updates: On the Sparsity and Geometry of On-Policy Distillation

    • OPD updates are small and coordinate-sparse, and they are distributed mainly across FFN layers, with consistency across ten model pairs.
    • Training only the discovered sparse subnetwork is enough to recover the full OPD reasoning performance.
    • SGD underperforms AdamW, which indicates that even sparse updates need adaptive scaling because the gradient scale is heterogeneous.
    • The OPD delta is numerically full-rank, but its spectrum is concentrated, with the top 16 singular values accounting for about 20 to 30 percent of the energy.
    • The updates avoid the principal directions of the source weight and concentrate on low-magnitude coordinates, which resembles the pattern seen in RLVR.
    • Even with dense supervision, OPD looks more like sparse on-policy editing than dense rewriting.
    AbstractPDF
  34. 411AI / GeneralarXiv:2606.13361

    Can I Buy Your KV Cache?

    • KV cache reuse is token-exactly identical to prefill from scratch, with a 24 out of 24 greedy match and logits-level verification.
    • On Qwen3-4B, reused compute is 9 to 50 times cheaper than prefill, and the gap widens as documents get longer because of the L2 attention term.
    • The break-even point is effectively immediate, at N approximately 1, so reuse pays off from the second read onward.
    • KV is almost impossible to compress losslessly, so provider-side hosting is optimal and follows the same principle as production prompt caching.
    • Serving a 3,774-token document to 80M agents costs 0.5M with prefill versus 0.03M with reuse, a 49.7x reduction.
    • int8 quantization halves the size but breaks token-exactness, so it is unsuitable when losslessness is required.
    AbstractPDF
  35. 410LLMs / NLParXiv:2606.13681

    EvoArena: Tracking Memory Evolution for Robust LLM Agents in Dynamic Environments

    • EvoArena is a dynamic agent benchmark in which terminal, software, and social environments evolve across versions.
    • Even the current strongest agents achieve only 39.6 percent on average, which shows that adapting to dynamic environments is still a major challenge.
    • EvoMem is a git-like memory that stores memory changes as patch history, including pre-state, post-state, the reason for the change, and the evidence.
    • EvoArena delivers consistent performance gains of +1.5 percent, chain-level +3.7 percent, GAIA +6.1 percent, and LoCoMo +4.8 percent.
    • In software regression analysis, the PASS_TO_PASS failure rate drops from 9.09 percent to 6.32 percent on average.
    • The biggest improvement, +5.2 percent, appears in time-trajectory and multi-pattern synthesis questions, which confirms the value of evidence preservation.
    AbstractPDF
  36. 409AI / GeneralarXiv:2606.13662

    EurekAgent: Agent Environment Engineering is All You Need For Autonomous Scientific Discovery

    • The key claim is that as models get stronger, the bottleneck shifts from the workflow to environment engineering.
    • It defines four forms of environment engineering: permissions for evaluation separation, artifacts for Git and file-sharing memory, budget for cost-aware exploration, and human-in-the-loop control.
    • It achieves a new SOTA of 2.635999 on 26-circle packing at a cost of $11 in API usage.
    • It improves TriMul kernel optimization by 4.3% over the former leaderboard leader and by 10.8% over TTT-Discover.
    • On seven MLE-Bench Lite tasks, it earns medals in 85.71% of cases and outperforms systems built on commercial models.
    • ResearchClawBench provides evidence that general-purpose CLI agents outperform domain-specific systems.
    AbstractPDF
  37. 408Computer VisionarXiv:2606.13515

    MaskWAM: Unifying Mask Prompting and Prediction for World-Action Models

    • It unifies mask prompting and prediction by using masks both as input, in the form of a first-frame visual prompt, and as output, in the form of future mask prediction.
    • Future mask prediction provides object-centric semantic supervision, improving performance over RGB-only methods.
    • A first-frame mask prompt resolves language ambiguity and is overwhelmingly better than text coordinate prompts.
    • It uses a Mixture of Transformers (MoT) to jointly predict RGB, masks, and actions with unified flow matching.
    • It achieves SOTA results of 98.4% on LIBERO and 92.2% on RoboTwin, and 84.3% on clear-language real-robot tasks and 84.9% on ambiguous tasks.
    • Without mask prediction, performance drops to 21.6%, showing that the prediction objective is essential for using visual prompts.
    AbstractPDF
  38. 407Computer VisionarXiv:2606.13652

    World Tracing: Generative Pixel-Aligned Geometry Beyond the Visible

    • A pixel-aligned multilayer pointmap predicts L camera-space 3D points for each pixel, completing even occluded surfaces.
    • WT-DiT combines a frozen MoGe encoder with three-way decomposed attention over layer, ray, and global dimensions.
    • The depth-filling strategy forward-fills empty values in occluded layers and learns dense XYZ without mask prediction.
    • It handles objects, scenes, and dynamic clips with a single architecture and does not require camera intrinsics as input.
    • It outperforms existing depth predictors on visible surface accuracy and image-to-3D generators on Chamfer distance.
    • It also supports text-based 3D scene editing, geometry-conditioned novel-view video synthesis, and training-free texture mesh generation.
    AbstractPDF
  39. 406Machine LearningarXiv:2606.12507

    Rubric-Guided Self-Distillation: Post-Training Without Rubric Verifiers

    • Its key idea is token-level self-distillation from a rubric-conditioned teacher to an unconditioned student.
    • It uses the condition enhancement effect, where rubric-only information boosts base model performance by 30 to 45 percentage points.
    • Compared with GRPO, it achieves medical gains of +6.1 versus +5.9 percentage points and science gains of +4.9 versus +4.5 percentage points, while making zero judge calls.
    • On Qwen-2.5-7B, it roughly halves response length compared with GRPO and avoids verbosity drift.
    • It performs on-policy student rollout distillation with clipped Jensen-Shannon divergence.
    • Because GRPO with a stronger judge can outperform RGSD in some settings, the two methods are complementary.
    AbstractPDF
  40. 405Machine LearningarXiv:2606.13473

    MaxProof: Scaling Mathematical Proof with Generative-Verifier RL and Population-Level Test-Time Scaling

    • The paper trains three specialized capabilities for proof generation, verification, and repair separately and then merges them into a single M3 model.
    • Its defense-in-depth verifier minimizes false positives using bad-case filtering, normalization, multiple judgments, and pessimistic minimum aggregation.
    • MaxProof test-time scaling uses 32 initial candidates and up to 10 rounds of PATCH and REWRITE refinement.
    • It achieves 27 to 35 points, a gain of 8, on IMO 2025, and 26 to 36 points, a gain of 10, on USAMO 2026, exceeding the gold-medal threshold in both contests.
    • It documents four reward-hacking patterns discovered in the M2 cycle and the corresponding defenses.
    • It is cost-efficient: using GLM-5.1, it sets a new SOTA on 26-circle packing for less than 11 dollars in API cost.
    AbstractPDF
  41. 404AI / GeneralarXiv:2606.12683

    From AGI to ASI

    • It defines ASI as a system that is broadly superior to large groups of human experts and quantifies it with Legg-Hutter score.
    • It outlines four transition paths: scaling, paradigm shift, recursive self-improvement, and multi-agent collective intelligence.
    • Effective compute is growing at about 10x per year, driven by 1.5x hardware growth, 2.5x investment growth, and 3x algorithmic efficiency gains.
    • Main friction factors include the data wall, compute and energy costs, rising research difficulty, and orchestration cost.
    • Recursive self-improvement may enable explosive growth, but plateau is also possible.
    • The paper emphasizes that AI safety, governance, and verification research must come before the transition to ASI.
    AbstractPDF
  42. 403AI / GeneralarXiv:2606.13392

    MiniMax Sparse Attention

    • It proposes block sparse attention based on GQA, where a lightweight index branch independently selects Top-k blocks for each GQA group.
    • From-scratch training on 3T tokens in a 109B MoE model demonstrates downstream performance equivalent to GQA.
    • At 1M context length, it reduces attention FLOPs per token by 28.4x and speeds up prefill by 14.2x and decoding by 7.6x on H800.
    • It optimizes GPU tensor-core utilization with an exp-free TopK kernel and KV-outer sparse attention.
    • The index branch is trained with a KL alignment loss and separated from the main branch with stop-gradient.
    • It is publicly released as the commercial multimodal model MiniMax-M3 on HuggingFace.
    AbstractPDF
  43. 402Computer VisionarXiv:2606.11683

    Reason, Then Re-reason: Cross-view Revisiting Improves Spatial Reasoning

    • It reformulates spatial reasoning as a two-stage process of hypothesis and verification rather than a single turn.
    • It strategically synthesizes complementary oblique views using VGGT monocular 3D reconstruction.
    • It is a training-free framework that can be applied at inference time without modifying the MLLM architecture.
    • It improves the VSI-Bench average by 5.2 points; at the sample level, the positive flip rate is 71.6% versus a negative flip rate of 28.4%, a 2.52 to 1 ratio.
    • On an A100 GPU, it takes about 11 seconds per sample; reducing the VGGT input to 20 frames preserves a 2.8-point gain in about 4 seconds.
    • Oblique Sweep outperforms bird's-eye views because it exposes spatial information while avoiding viewpoint-distribution mismatch.
    AbstractPDF
  44. 401Computer VisionarXiv:2606.11289

    i1: A Simple and Fully Open Recipe for Strong Text-to-Image Models

    • A single strong text encoder, T5Gemma-2B, and a large adapter with two transformer blocks are more effective than multiple encoders.
    • A dual-stream DiT backbone, long skip connections, and QK-Norm form a strong backbone design, while AdaLN offers little benefit in T2I.
    • Training on long captions outperforms short captions, but prompt rewriting is needed at inference time.
    • Equal weighting, meaning the same number of images from each dataset, is a strong default for mixing curated datasets.
    • i1-3B uses only public data and is on average 29.5 percentage points better than the previous best fully open model across five benchmarks.
    • Larger models such as 17B HiDream-I1 and 12B FLUX.1[Dev] are also surpassed, confirming the value of design exploration.
    AbstractPDF
  45. 400Machine LearningarXiv:2606.12360

    Anatomy of Post-Training: Using Interpretability to Characterize Data and Shape the Learning Signal

    • Using SAE features and two-sample hypothesis tests, we identify latent concepts in preference data, including length, helpfulness, morality, and refusal.
    • We propose a data-centric framework that explains away the identified concepts to pre-remove undesirable learning signals.
    • Training on the Dolci dataset reveals reduced jailbreak robustness, meaning that safety is worse than with standard SFT.
    • We diagnose and mitigate excessive stylization and sycophancy, and the framework can amplify model personalities such as playful behavior and formality.
    • We present a unified lens for various post-training protocols, including DPO, SFT, and data filtering.
    • The approach is grounded in the theory that causally valid features change model probabilities in a predictable way.
    AbstractPDF
  46. 399Computer VisionarXiv:2606.13679

    InterleaveThinker: Reinforcing Agentic Interleaved Generation

    • A three-agent Planner-Gen-Critic structure addresses visual over-reliance and error accumulation across steps.
    • The Planner preserves the global goal by blocking intermediate feedback, and the Critic performs step-by-step evaluation and prompt refinement.
    • The authors build high-quality datasets for Interleave-Planner-SFT-80k, Critic-SFT-112k, and Critic-RL-13k.
    • A dual-reward strategy combining accuracy reward and step-level reward optimizes the whole trajectory through single-stage RL.
    • With FLUX.2-klein as the base model, it improves WISE from 0.47 to 0.73 and RISE from 13.3 to 28.9, which reaches the level of Nano Banana and GPT-5.
    • The image generator is replaceable, and the method works with both FLUX.2-klein and Qwen-Image-Edit.
    AbstractPDF
  47. 398Computer VisionarXiv:2606.12195

    InternVideo3: Agentify Foundation Models with Multimodal Contextual Reasoning

    • MCR is a closed-loop multimodal reasoning formulation that integrates observation, reasoning, tool actions, feedback, and memory into a shared evolving context.
    • M2LA, or Multimodal Multi-head Latent Attention, compresses the KV cache to make long multimodal rollouts efficient.
    • The staged training pipeline includes M2LA conversion, continued pretraining, short-to-long video SFT, verifiable-task RL, and on-policy distillation.
    • It shows substantial gains over InternVideo2.5-7B on long-video benchmarks such as Video-MME, MLVU, and EgoSchema.
    • It instantiates a video agent with retrieval and verification tools to demonstrate the practicality of recursive multimodal reasoning.
    • Because the open-weight ecosystem is evolving rapidly, the paper emphasizes the principled value of context efficiency and closed-loop reasoning over state-of-the-art claims.
    AbstractPDF
  48. 397Machine LearningarXiv:2606.11709

    RLCSD: Reinforcement Learning with Contrastive On-Policy Self-Distillation

    • It discovers privilege-induced style drift, where privileged conditioning concentrates the learning signal on style tokens such as Therefore and Wait.
    • Contrastive hints from correct and incorrect answers cancel style components and concentrate the signal on task tokens such as mathematical content.
    • On Qwen3-8B, it achieves 90.8 percent on AMC23, 77.5 percent on AIME24, and 69.7 percent on AIME25, outperforming GRPO and OPSD variants.
    • It addresses the training instability and response-length contraction problems of prior OPSD methods, including OPSD, SDPO, SRPO, and RLSD.
    • The contrastive principle is general, so it can be plugged into existing OPSD methods and extended to cross-model distillation.
    • K-marginalization and two-path loss aggregation are key contributors to preserving performance.
    AbstractPDF
  49. 396Machine LearningarXiv:2606.12370

    Breaking Entropy Bounds: Accelerating RL Training via MTP with Rejection Sampling

    • We found a clear negative linear relationship between MTP acceptance rate and policy entropy: as entropy rises, acceptance rate falls linearly.
    • We mathematically and empirically show that probabilistic rejection sampling is far more robust to entropy variation than greedy target-only sampling.
    • Instead of conventional CE/KL losses, we propose an end-to-end TV loss that directly optimizes rejection-sampling acceptance rate, improving it by about 10 percentage points.
    • A light MTP pretraining phase before RL is sufficient to keep acceptance rates stable throughout RL, eliminating the need for online MTP co-training.
    • Across mathematical reasoning, code generation, and agentic tasks, it reaches up to 95% acceptance and an additional 25% increase in reasoning throughput.
    • In asynchronous RL pipelines for Qwen3.5/3.6/3.7, it achieves up to 1.8x end-to-end acceleration.
    AbstractPDF
  50. 395RoboticsarXiv:2606.10040

    Efficient-WAM: A 1B-Parameter World-Action Model with Low-Cost Future Imagination

    • From WAN-2.2-5B, the authors built a 0.8B compact video expert through structural pruning and knowledge distillation, then combined it with a 0.2B action expert and MoT.
    • By lowering the future prediction resolution, they reduced the token count from 240 to 60 and cut latency further from 430 ms to 377 ms.
    • Asymmetric denoising with 2 video steps and 10 action steps achieved 139 ms latency with less than 1 percentage point of performance loss.
    • It achieved an 86.7 percent average success rate in RoboTwin 2.0 simulation and 66.25 percent success on the real Astribot S1 robot.
    • On a consumer RTX 4090, it achieved 98 ms per chunk, which is 32x faster than Motus at 3215 ms and better than pi0.5 at 113 ms.
    • The design principle of action-centric future imagination is that the policy needs only structural dynamics cues, not photo-like images.
    AbstractPDF