HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 394Computer VisionarXiv:2606.09967

    ABot-Earth 0.5: Generative 3D Earth Model

    • It is the first large generative model to use 3D Gaussian Splatting (3DGS) directly and synthesize 3D scenes from satellite images alone.
    • It generates scenes in less than 10 minutes per km2, which is dramatically more efficient than photogrammetry pipelines that take months or years.
    • It combines satellite, aerial, and urban data sources, and it uses the ABot-3DGS reconstruction engine to build high-quality training data.
    • Its native multi-LOD hierarchy streams trillions of Gaussian primitives in real time, from globe level down to street level on web maps.
    • It covers 300 cities across 190 countries and scores better than Google Earth on aesthetics, although its precise geometry and texture are still weaker.
    • It offers an editable platform through a hybrid approach that combines high-quality landmark reconstruction with synthesized backgrounds.
    AbstractPDF
  2. 393Machine LearningarXiv:2606.09821

    Rethinking the Divergence Regularization in LLM RL

    • It replaces DPPO's hard mask with an advantage-weighted l2_2 regularization term that guarantees smooth gradients.
    • It preserves the geometry of the binary-TV trust region, combining divergence-based stability with SPO-style smoothness.
    • Inside the boundary, it keeps the policy gradient, and beyond the boundary it provides corrective gradients that pull the policy toward the behavior policy.
    • The method was cross-validated on Qwen3-4B and Qwen3-30B-A3B, a MoE model, under BF16 and FP8 settings, and it consistently outperformed GRPO, DPPO, and SPO.
    • An ablation confirmed that absolute advantage weighting by |A_t| stabilizes the trust-region boundary independently of the reward scale.
    • Compared with other divergence regularizers such as KL, TV, and chi2, DRPO performs better, which suggests that gradient shape matters more than the exact form of the objective.
    AbstractPDF
  3. 392LLMs / NLParXiv:2606.11926

    Toward Generalist Autonomous Research via Hypothesis-Tree Refinement

    • Autonomous Optimization, or AO, is defined as a task in which an initial output is improved through repeated experiments without step-by-step human supervision.
    • The hypothesis tree implements cumulative research state by binding each node to a hypothesis, a version of the output, experimental evidence, and refined insight.
    • It separates the coordinator, which handles the global strategy, from the executor, which runs individual experiments in isolated worktrees.
    • Across six real research tasks, including model training, harness work, and data synthesis, it reports more than 2.5x average relative gains over Codex and Claude Code.
    • On MLE-Bench Lite, it achieves 86.36 percent Any Medal with GPT-5.5, and ablations confirm that HTR and insight propagation are key.
    • It demonstrates generalization by transferring the BrowseComp optimization harness to a previously unseen search-agent task.
    AbstractPDF
  4. 391LLMs / NLParXiv:2606.09659

    End-to-End Context Compression at Scale

    • It trains a 0.6B encoder and a 4B decoder based on Qwen3-4B on more than 350B tokens at compression ratios of 1:4, 1:8, and 1:16.
    • A large architecture search optimizes the encoder-decoder design, including pooling, adapter structure, and training stage design.
    • It achieves 8.8x faster TTFT on RULER 4K and 5.2x faster TTFT on LongBench 64K while keeping accuracy ahead.
    • It proposes an agentic harness in which the agent reads the full compressed context at once and uses the EXPAND tool to open only the chunks it needs.
    • It shows near-linear memory scaling up to 1M tokens and is compatible with standard inference engines such as vLLM and SGLang.
    AbstractPDF
  5. 390Machine LearningarXiv:2606.11722

    ICA Lens: Interpreting Language Models Without Training Another Dictionary

    • With GPU-parallel FastICA, row normalization, and a p95-LIM convergence criterion, it increases accepted layers in GPT-2 Small by 400 percent and reduces iterations by 21.5 percent.
    • ICA directions have significantly higher non-Gaussianity than random projections, and deeper layers show increasing context dependence, confirming the expected pattern of effective receptive field growth.
    • It is competitive with open SAE models on SAEBench Sparse Probing, and it outperforms SAE under small to medium budgets on Targeted Probe Perturbation.
    • ICA components capture a wide range of linguistic structures, including lexical form, local syntax, discourse context, semantic ambiguity, and long-range repetition.
    • It proposes ICA as a complementary first lens rather than a replacement for SAE and positions it as a cost-effective exploratory tool.
    AbstractPDF
  6. 389AgentsarXiv:2604.11548

    SemaClaw: A Step Towards General-Purpose Personal AI Agents through Harness Engineering

    • The paper formalizes a paradigm shift in AI engineering from prompt and context engineering to harness engineering. As model capabilities converge, the harness layer becomes the main site of architectural differentiation.
    • DAG Teams proposes a two-phase hybrid orchestration approach in which the LLM dynamically generates a task graph (DAG) and a deterministic scheduler executes it, achieving both flexibility and execution traceability.
    • PermissionBridge is a behavioral safety system that performs runtime authorization checkpoints before high-risk actions such as file edits, external API calls, and code execution. It is a native runtime primitive, not a tool-level wrapper.
    • We introduce a hierarchical memory architecture for three-tier context management, consisting of compressed working memory, retrieval-based external memory, and SOUL.md anchor persona partitions.
    • Agentic Wiki externalizes task-derived knowledge into a Markdown-file-based user-owned knowledge base, creating a personal knowledge infrastructure that future agent sessions can retrieve and use.
    • In LangChain experiments, keeping the model fixed and improving only the harness increased task completion from 52.8% to 66.5%, a gain of 13.7 percentage points, showing that harness design is a major determinant of agent performance.
    AbstractPDF
  7. 388RoboticsarXiv:2606.09215

    MotionWAM: Towards Foundation World Action Models for Real-Time Humanoid Loco-Manipulation

    • It uses intermediate denoising features from a video DiT as policy conditioning, which makes real-time execution possible without generating full videos.
    • Instead of splitting the upper and lower body, it controls the whole body in a single latent space using a unified motion latent variable.
    • Its three-stage training pipeline consists of 2,136 hours of egocentric video pretraining, cross-embodiment action post-training, and whole-body fine-tuning.
    • On the Unitree G1, it achieves a 76.1 percent average success rate across 9 tasks, which is more than 32 percentage points above the strongest VLA baseline at 43.9 percent.
    • It enables task-driven foot interaction, such as kicking a ball or pressing a pedal, which split policies cannot perform.
    • It runs in real time at 4.9 Hz on an NVIDIA A100, about seven times faster than Cosmos Policy at 0.7 Hz.
    AbstractPDF
  8. 387Machine LearningarXiv:2606.11087

    Test-Time Gradient Guidance of Flow Policies in Reinforcement Learning

    • It proposes QGF, or Q-Guided Flow, an algorithm that performs policy optimization only at test time.
    • A single Euler-step approximation uses the critic gradient of the clean action during denoising, so backpropagation through time is not needed.
    • It avoids the instability of training-time actor-critic methods and scales favorably as model size increases.
    • It achieves better performance than Best-of-N with far fewer FLOPs.
    • It shows consistent gains over existing methods on difficult goal-conditioned RL tasks.
    • When it uses a QAM critic, it extracts policies better than QAM itself.
    AbstractPDF
  9. 386RoboticsarXiv:2606.08737

    Dream-Tac: A Unified Tactile World Action Model for Contact-Rich Robot Manipulation

    • It is the first unified Tactile World Action Model that jointly predicts vision and touch.
    • Contact-gated fusion is a gate mechanism that emphasizes tactile signals only when the amount of touch change is large.
    • A contact-aware attention bias controls cross-modal interaction among vision, touch, and action.
    • FlashAttention-based gate optimization speeds training by 2.94x, and diffusion-step caching speeds inference by 1.8x.
    • It improves average accuracy by 31.7% across six contact-intensive tasks and exceeds Cosmos Policy by 31.6%.
    • It is robust under out-of-distribution changes in table height, space, object identity, and background.
    AbstractPDF
  10. 385RoboticsarXiv:2606.09827

    MemoryVLA++: Temporal Modeling via Memory and Imagination in Vision-Language-Action Models

    • We introduce a cognition-inspired full-temporal model that combines working memory, episodic memory, and future imagination through a world model.
    • The Perceptual-Cognitive Memory Bank stores low-level details and high-level semantics together, and updates memory when redundancy is detected.
    • A latent-space world model imagines future states through partial denoising only, without requiring pixel-level prediction.
    • We achieve 98.4% on Libero and 74.0% on SimplerEnv, with real-robot gains of +9% for general capability, +26% for memory, and +28% for imagination.
    • It is especially strong on long-horizon tasks, with 4.29 on Calvin and 44.4% on Mikasa-Robo.
    • It is ready for real-time deployment, with 66.4 Hz inference on an RTX 4090.
    AbstractPDF
  11. 384RoboticsarXiv:2606.10340

    OMG: Omni-Modal Motion Generation for Generalist Humanoid Control

    • A hierarchical structure with a motion-generation brain and a motion-tracking cerebellum generalizes whole-body humanoid control.
    • OMG-Data is a 1,174.66-hour omni-modal humanoid motion corpus containing language, audio, and human motion.
    • OMG-DiT is a diffusion transformer conditioned on language, audio, and human references, and it can be extended to new modalities.
    • It achieves an FID of 6.03 on text-to-motion with OMG-XL, which is a large improvement over prior methods.
    • It supports zero-shot compositionality, so unseen combinations of language and audio can be combined at inference time.
    • Performance improves consistently as the model scales from B to L to XL.
    AbstractPDF
  12. 383Computer VisionarXiv:2606.11187

    Next Forcing: Causal World Modeling with Multi-Chunk Prediction

    • We solve the myopic supervision problem in autoregressive video generation with multi-chunk prediction (MCP).
    • A causal MCP chain lets near-future predictions inform far-future predictions and provides dense supervision to the main model.
    • We achieve SOTA on RoboTwin, with 94.1% on Clean and 93.5% on Random.
    • At 50 fps, it delivers a 93.1% relative improvement over LingBot-VA and a 2.3x training speedup.
    • General video pretraining also reduces FVD by more than 50%, demonstrating generality beyond robot-specific data.
    • Keeping the MCP module at inference time doubles inference speed while preserving accuracy.
    AbstractPDF
  13. 382Computer VisionarXiv:2606.10651

    Kwai Keye-VL-2.0 Technical Report

    • It is a 30B MoE model with 3B active parameters, designed as an efficiently deployable multimodal foundation model.
    • It applies DeepSeek Sparse Attention to a GQA-based architecture for the first time, enabling 256K lossless context.
    • Cross-Modal MOPD integrates video, code, tools, and search into a single MoE backbone.
    • Context-RL and Video-RL reduce visual hallucination and stabilize long-horizon decision making.
    • It achieves state-of-the-art results on TimeLens, with 58.5 mIoU on ActivityNet, and on LongVideoBench, with 74.1.
    • It is also strong on agent benchmarks such as LiveCodeBench v6 at 64.2 and OJBench at 71.5.
    AbstractPDF
  14. 381AI / GeneralarXiv:2606.08432

    Trajectory-Refined Distillation

    • It is the first to formalize the prefix-failure problem in OPD, showing that token-level improvements cannot solve it at the root.
    • TRD performs distillation after refining trajectories within the on-policy support region through trajectory-level refinement.
    • It can be applied to both OPD and OPSD self-distillation and yields consistent gains across the Qwen3 model family.
    • On AMOBench, Qwen3-8B shows roughly 50 percent relative improvement in Pass@16, and the refined trajectories are compressed by about 9 times.
    • This trajectory-level approach outperforms prior token-level loss-improvement methods such as clipping, reweighting, and top-K selection.
    AbstractPDF
  15. 380LLMs / NLParXiv:2606.09498

    Self-Harness: Harnesses That Improve Themselves

    • It proposes Self-Harness, a paradigm in which the agent improves its own harness through weakness mining, harness proposal, and proposal validation.
    • Across three models, MiniMax M2.5, Qwen3.5-35B-A3B, and GLM-5, the held-out pass rates improve from 40.5% to 61.9%, from 23.8% to 38.1%, and from 42.9% to 57.1%, respectively.
    • Different harness changes are produced for different models, so the updates are model-specific rather than generic instruction additions.
    • A regression-test-based acceptance rule prevents overfitting, and the edits remain small and auditable.
    • The work is directly related to existing harness systems such as SemaClaw and OpenClaw.
    • It is a meaningful step forward in the self-improvement research line for agents.
    AbstractPDF
  16. 379Computer VisionarXiv:2606.09056

    MilliVid: Hierarchical Latents for Long-Range Consistency in Video Generation

    • It uses a hierarchical autoencoder design with different token counts per frame at different levels.
    • The coarsest level, with 4 tokens per frame, captures scene structure and object permanence.
    • Coarse-to-fine rollout maximizes temporal coverage within a fixed sequence length.
    • Learned hierarchical tokenization provides better compression quality than simple downsampling.
    • It trains the model to predict far-future frames, encouraging long-range context dependence.
    • It shows a significant improvement in consistency over FramePack on long Minecraft videos.
    AbstractPDF
  17. 378AI / GeneralarXiv:2606.06741

    OpenSkill: Open-World Self-Evolution for LLM Agents

    • Open-world self-evolution is defined as the process of building skills and verification signals at the same time from task prompts alone.
    • In the open world, grounding knowledge and verification anchors are acquired independently.
    • Self-built virtual tasks are used for skill refinement, with no target supervision at all.
    • On SkillsBench, it improves by 8.9 percent over the previous strongest baseline.
    • The skills transfer across models, so the same effect applies even to weaker models.
    • The self-built verifier covers 88.9 percent of the intent represented in the ground truth.
    AbstractPDF
  18. 377Software EngineeringarXiv:2606.07297

    SWE-Explore: Benchmarking How Coding Agents Explore Repositories

    • A new benchmark that evaluates repository exploration independently from patch generation.
    • It derives trajectory-grounded, line-level ground truth from successful agent trajectories.
    • It covers diverse coding settings with 848 issues, 10 languages, and 203 repositories.
    • Agent exploration forms a clear top tier compared with classical retrieval.
    • File-level hit rates are strong, but core line recall remains only 0.14 to 0.19, making it the bottleneck.
    • Missing key context is more damaging to patch success than noisy context.
    AbstractPDF
  19. 376AI / GeneralarXiv:2606.08702

    ConMem: Structured Memory-Guided Adaptation in Training-Free Multi-Agent Systems

    • It distills past trajectories into typed signed memory cards to create reusable strategic units.
    • It organizes the cards as typed graphs with supports, satisfies, conflicts, and constrains relations.
    • It uses needs-aware retrieval to search for cards that match the current task.
    • It resolves conflicts and recovers dependencies through the relation graph to produce consistent guidance.
    • Context-budget control prunes more than 50% of candidate cards and reduces planning overhead by 80%.
    • It is training-free and does not require changes to the host model weights or orchestration.
    AbstractPDF
  20. 375LLMs / NLParXiv:2606.07502

    Your UnEmbedding Matrix is Secretly a Feature Lens for Text Embeddings

    • It discovers representation collapse in LLM embeddings, where they align with high-frequency meaningless tokens.
    • It identifies the edge-spectrum subspace of the unembedding matrix as the cause.
    • EmbedFilter removes that subspace with a simple linear transform and does not require additional training.
    • It improves MTEB performance by up to 14.1% and works across different LLM backbones.
    • It also saves storage and speeds up retrieval through distance-preserving dimensionality reduction.
    • It outperforms existing correction methods such as BERT-whitening in unsupervised settings.
    AbstractPDF
  21. 374RoboticsarXiv:2606.09811

    AHA-WAM:Asynchronous Horizon-Adaptive World-Action Modeling with Observation-Guided Context Routing

    • The design separates Video DiT as a low-frequency planner and Action DiT as a high-frequency executor through an asynchronous decomposition.
    • Rolling KV memory preserves the planner's temporal context over time.
    • OVCR dynamically routes and updates the planner context based on the latest observations.
    • Horizon-adaptive offset training learns arbitrary phase relationships between planner and executor.
    • The method reaches SOTA-level performance, with 92.80 percent on RoboTwin and 78.3 percent success in the real world.
    • It achieves control frequencies up to 56.9 Hz, which is a 10.82x speedup over Fast-WAM.
    AbstractPDF
  22. 373Machine LearningarXiv:2606.09079

    FlashMemory-DeepSeek-V4: Lightning Index Ultra-Long Context via Lookahead Sparse Attention

    • LSA is a paradigm that predicts future context demand and loads only the necessary KV chunks onto the GPU.
    • It decouples the indexer from the backbone, making it possible to train on a single H20 GPU in one hour.
    • It replaces fixed Top-k with dynamic chunk recall based on a sigmoid threshold.
    • It improves accuracy by 0.6% while using only 13.5% KV cache on average.
    • At 500K length, it cuts memory use by 90% and still improves accuracy by 1.9%.
    • It has a clear limitation: performance drops sharply on the dense-search MRCR benchmark, from 76% to 48%.
    AbstractPDF
  23. 372AI / GeneralarXiv:2606.09669

    SpatialWorld: Benchmarking Interactive Spatial Reasoning of Multimodal Agents in Real-World Tasks

    • Design a simulator-agnostic evaluation protocol that integrates eight simulators.
    • Build a dataset of 760 tasks across six domains, including household chores, travel, social cooperation, and digital games.
    • Provide an MLLM-native interface with first-person vision-only observations and text-based actions.
    • Space reasoning remains a major challenge, with GPT-5 at 17.4% and Qwen-3.5 at 14.1%.
    • Domain-specific leaders differ: GPT-5 leads on household chores, while Gemini-3.1-Pro leads on games.
    • It finds a mismatch between task success rate and execution efficiency.
    AbstractPDF
  24. 371Computer VisionarXiv:2606.09828

    Latent Spatial Memory for Video World Models

    • We build 3D spatial memory directly in latent space, eliminating the cost of round-tripping through pixel space.
    • Depth-guided back-projection loads latent tokens into 3D world space.
    • Occlusion-aware latent-resolution projection is used to obtain the target view.
    • We exclude dynamic objects and the sky from memory updates to improve long-term stability.
    • Compared with an RGB cache, the system runs 10.57 times faster end to end and uses 55 times less GPU memory.
    • Two-stage training enables stable convergence of the side branch and LoRA.
    AbstractPDF
  25. 370RoboticsarXiv:2606.06556

    Robots Need More than VLA and World Models

    • It redefines VLA models as one layer in the physical intelligence stack and argues that upstream and downstream grounding mechanisms are essential.
    • It advocates a shift to grounding-centered pipelines that convert unstructured physical experience into robot-usable supervision.
    • It proposes four essential interfaces: a physical data engine, embodiment retargeting, physics-based world models, and task-conditioned reward grounding.
    • It presents a vision for self-improving systems that use success, failure, and human correction as structured supervision in the deployment loop.
    • It revisits state-of-the-art research in robot-native data, video-based methods, simulation, and world models from the perspective of grounding bottlenecks.
    AbstractPDF
  26. 369Computer VisionarXiv:2606.07433

    Watch, Remember, Reason: Human-View Video Understanding with MLLMs

    • It formalizes video understanding as the functional process of observing, remembering, and reasoning, and proposes an integrated formulation for perceptual representations, memory state, reasoning traces, and final prediction.
    • It systematically organizes the viewing domain into four subareas: fine-grained perception, holistic perception, audiovisual perception, and efficient processing.
    • In memory, it distinguishes offline memory from streaming memory to analyze the tension between redundancy and evidence sparsity in long videos.
    • In reasoning, it integrates and classifies different paradigms such as agentic versus non-agentic and text-only versus video-involved reasoning.
    • It covers five application domains, including egocentric, sports, instructional, medical, and narrative, along with major training datasets and evaluation benchmarks.
    AbstractPDF
  27. 368Machine LearningarXiv:2606.06624

    Principles and Practice of Deep Representation Learning: or a Mathematical Theory of Memory

    • It presents data compression as a unifying principle behind all representation learning methods, including PCA, denoising, score matching, and information gain maximization.
    • It mathematically interprets ResNet, CNN, and Transformer architectures as iterative optimization algorithms and clarifies the principles behind architecture design.
    • It proves that the encoder-decoder structure is a necessary design for ensuring the accuracy and consistency of representations through a closed-loop transcription framework.
    • In v2.0, it adds the theoretical justification for contrastive learning and DINO, derives a causal CRATE transformer, and expands the relationship between distribution learning and representation learning.
    • It demonstrates the practical value of these theoretical principles step by step in large-scale real-world applications involving visual, motion, and text data.
    AbstractPDF
  28. 367Computer VisionarXiv:2606.06476

    Thinking with Imagination: Agentic Visual Spatial Reasoning with World Simulators

    • It formulates spatial reasoning as an interactive evidence-gathering problem, so the VLM actively requests and uses imagined visual evidence.
    • It applies view-consistency tuning to Astra-WM to address the lack of spatial consistency in existing image generation models.
    • A two-stage RL curriculum teaches valid simulator calls in stage 1 and encourages selective imagination in stage 2.
    • It shows that simply connecting a simulator to Qwen3-VL degrades performance, but meaningful gains appear after RL training.
    • It shows that useful imagination requires both a high-quality world simulator and an agent policy that decides when, where, and how to imagine.
    AbstractPDF
  29. 366Computer VisionarXiv:2606.07100

    LARA: Latent Action Representation Alignment for Vision-Language-Action Models

    • We overcome the limitations of existing separate training pipelines by jointly optimizing LAM and VLA with a bidirectional representation-alignment loss.
    • LAM is grounded in real action trajectories, which suppresses learning spurious visual variations such as background and lighting.
    • VLA is regularized by LAM's forward-dynamics prediction, reducing hallucinated trajectories that are kinematically feasible but functionally wrong.
    • We find that inserting alignment at DiT layer L-2 is optimal, and that alignment in deeper layers is effective.
    • We achieve consistent performance gains on three simulation benchmarks and in a real robot manipulation environment.
    AbstractPDF
  30. 365LLMs / NLParXiv:2606.07229

    MMAE: A Massive Multitask Audio Editing Benchmark

    • It establishes a unified taxonomy for audio editing for the first time, spanning seven modalities, six complexity levels, and eight operation types.
    • It introduces an evaluation paradigm that decomposes free-form editing tasks into a verifiable multidimensional rubric, which enables objective and interpretable assessment.
    • It builds the benchmark from 2,000 high-quality samples and 17,741 evaluation rubrics selected through a human-agent collaboration pipeline.
    • Evaluation of five state-of-the-art models shows an EMR below 5 percent and 0 percent on complex mixed modalities, which exposes the fundamental limitations of current systems.
    • It confirms that introducing an external agent planner does not consistently help and that both understanding and generation remain bottlenecks.
    AbstractPDF
  31. 364AI / GeneralarXiv:2606.07080

    dots.tts Technical Report

    • It trains a high-resolution AudioVAE that encodes 48 kHz speech into a 25 Hz, 128-dimensional latent space and adds semantic structure through WavLM alignment loss.
    • It uses a three-module split of semantic encoder, LLM, and AR flow-matching head to reduce long-range error accumulation.
    • It introduces SOAR self-corrective alignment, which exposes the flow-matching head to its own inference errors without a reward model.
    • CFG-aware MeanFlow distillation enables 2 to 4 NFE inference steps and achieves 0.231 RTF and 54 ms first-packet latency on H800.
    • It reports open-source SOTA on Seed-TTS-Eval with WER 2.92 and SIM 79.2, and releases the full system under Apache 2.0.
    AbstractPDF
  32. 363Machine LearningarXiv:2606.07082

    On the Geometry of On-Policy Distillation

    • OPD lies in a relaxed off-principal regime between SFT, which aligns principal components, and RLVR, which localizes non-principal components.
    • The authors discover subspace locking, in which updates lock into a narrow low-dimensional channel soon after training begins.
    • Under a rank-16 projection constraint, OPD preserves performance while SFT degrades sharply, which shows that the locked subspace is functionally sufficient.
    • Token sparsification and off-policy rollouts preserve the rank trajectory, but objective mixing changes it, which confirms that objective composition is the key control axis.
    • Controlled experiments based on Qwen3-8B use multidimensional diagnostic metrics such as update sparsity, spectral drift, and principal-component rotation.
    AbstractPDF
  33. 362AI / GeneralarXiv:2606.05661

    Continual Learning Bench: Evaluating Frontier AI Systems in Real-World Stateful Environments

    • It is a set of high-quality, expert-validated tasks from six real-world domains, and each task contains latent structure that cannot be solved through pretraining alone.
    • It proposes an evaluation methodology that introduces a gain metric to separate base model capability from experience-based learning.
    • Simple in-context learning (ICL) outperforms dedicated memory systems such as Mem0 and ACE on most tasks.
    • Even the best system reaches only 25.4% normalized gain, suggesting that the continual learning ability of current frontier models is still limited.
    • Memory modules are frequently observed to hurt performance by producing false generalizations and stale beliefs.
    • In terms of cost efficiency, ICL is also superior, and expensive systems fail to deliver proportional gains.
    AbstractPDF
  34. 361Information RetrievalarXiv:2606.06260

    OneReason Technical Report

    • The paper analyzes the cause of the problem in generative recommendation, where thinking mode fails to outperform non-thinking mode, along two axes: perception and cognition.
    • In pretraining, it obtains strong item-level token perception through an itemic tokenizer and four-granularity training.
    • It uses a three-level cognition-enhanced chain-of-thought, or CoT, SFT scheme that goes from R0, which stands for perception, to R1, derivation, to R2, evolution, and finally R3, recommendation.
    • The authors use a specialize-then-unify RL strategy that first specializes by task and then unifies the models to strengthen thinking ability.
    • On the Kuaishou production multi-domain benchmark, the thinking mode consistently outperforms the non-thinking mode.
    • They find a transfer effect in which CoT supervision also improves non-thinking inference.
    • OneReason-8B and 0.8B models are planned for open sourcing.
    AbstractPDF
  35. 360RoboticsarXiv:2606.06155

    AffordanceVLA: A Vision-Language-Action Model Empowering Action Generation through Affordance-Aware Understanding

    • It introduces a structured affordance intermediate representation to close the gap between the semantic space of VLMs and the 3D physical action space.
    • It models the task progressively through Which2Act, a visual latent prediction stage, Where2Act, a 2D affordance map stage, and How2Act, a 3D geometry stage.
    • Its MoT architecture consists of three expert modules for understanding, affordance generation, and action.
    • It addresses label sparsity with a three-stage progressive training scheme and an automatic affordance annotation pipeline.
    • On real robots, it achieves 88.3 percent on basic tasks and 82.9 percent on complex tasks, which is 38.1 percentage points better than Pi0 on complex tasks.
    • The authors hypothesize that affordances serve as a semantic anchor that preserves the visual-language capabilities of the VLM backbone.
    AbstractPDF
  36. 359RoboticsarXiv:2606.06194

    ActiveMimic: Egocentric Video Pretraining with Active Perception

    • Camera motion in egocentric video is reinterpreted as viewpoint action rather than noise.
    • VGGT plus SAM-3D-Body reconstruct camera and wrist trajectories from a single RGB input and build a 27-dimensional unified action representation.
    • The model jointly learns active perception and manipulation with flow matching on the large-scale Ego4D dataset.
    • On four real-robot tasks, it matches or exceeds pi0 on evaluations such as Restocking at 90.1 percent and Finding at 91.7 percent.
    • The results show that active perception capability originates from egocentric pretraining rather than from robot fine-tuning.
    • Layer-level analysis confirms that camera-motion supervision promotes representation transfer from humans to robots.
    AbstractPDF
  37. 358Computer VisionarXiv:2606.03746

    Qwen-Image-Flash: Beyond Objective Design

    • It trains a 4-NFE student model distilled from Qwen-Image-2.0 using distribution matching distillation, or DMD.
    • A counterintuitive data finding emerges: increasing diversity hurts performance, while consistent single-category data transfers better.
    • Step-wise multi-teacher guidance combines complementary strengths across teacher models and stabilizes training.
    • It achieves the best overall performance at a balanced T2I and Edit ratio, and edit supervision also improves generation quality.
    • Qwen-Image-Flash produces high-quality results with only 4 NFE in scenarios such as poster generation.
    AbstractPDF
  38. 357Computer VisionarXiv:2606.05769

    Imagine Before You Predict: Interleaved Latent Visual Reasoning for Video Event Prediction

    • We propose an interleaved latent visual reasoning architecture that alternates between language tokens and continuous latent visual tokens.
    • Using visual-gain data curation, we built 50K high-quality training examples in which future visual hints genuinely help prediction.
    • LA-DAPO combines outcome-contrastive reward and temporal-diversity reward as a latent-awareness RL objective.
    • On FutureBench, Qwen3-VL-8B improves from 61.0 to 85.4, a 10.4-point gain over the previous best, Video-CoE.
    • On TwiFF-Bench, the average score rises from 2.44 to 3.04, and latent span usage increases adaptively as reasoning depth grows.
    • With the same data, text-only SFT reaches only 65.0, whereas interleaved latent SFT reaches 73.2, showing that the gain does not come from extra supervision.
    AbstractPDF
  39. 356AI / GeneralarXiv:2606.05818

    Benchmarks in Leipzig

    • The benchmark contains 100 research-level mathematics problems posed by 49 mathematicians, 35 of whom participated directly in the workshop, spanning 25 subfields including algebraic geometry, combinatorics, and representation theory.
    • It uses a three-stage evaluation pipeline: Stage 1 with 5 models and 1 run each, Stage 2 with 3 models and 20 runs each, and Stage 3 with 2 models and 3 heavy-thinking runs each.
    • The number of unsolved problems drops sharply from 41 to 16 to 2, suggesting that benchmark-style practice has reached the limits of the best-performing models.
    • GPT-5.5 solves 44 problems on a single attempt, the best result, while performance varies dramatically across models, with Grok 4.3 solving only 6 problems.
    • Even repeated runs of the same model show very large performance variance, indicating a deterministic limit in mathematical reasoning.
    • It is a public benchmark based on the ScienceBench platform, and the LLM review process found and removed 3 erroneous problems.
    AbstractPDF
  40. 355Machine LearningarXiv:2606.05254

    Flash-WAM: Modality-Aware Distillation for World Action Models

    • It identifies the structural reason why prior consistency distillation fails in video-action joint diffusion: gradient signals vanish because of an asymmetric noise schedule.
    • It selects modality-specific consistency functions, using linear-gradient scaling for actions and variance-preserving parameterization for video.
    • It delivers a 23 times speedup, reducing per-chunk latency from 8.1 seconds to 348 milliseconds and enabling real-time control.
    • It preserves an 85.5 percent success rate on RoboTwin 2.0 and 95.7 percent on LIBERO under the 1v/2a setting.
    • It achieves a 60 percent success rate on a real-world Unitree G1 humanoid, which is a substantial improvement over the naive LCM baseline at 24 percent.
    • A structural analysis of the consistency function family theoretically explains the limits of gradient scaling.
    AbstractPDF
  41. 354AI / GeneralarXiv:2606.03883

    Reasoning Structure of Large Language Models

    • We build a scalable benchmark based on 21 two-dimensional grid logic puzzles, enabling controlled scaling studies across four difficulty levels.
    • We develop a pipeline that automatically converts unstructured text reasoning traces into atomic claim-dependency graphs.
    • We propose the reasoning efficiency metric eta, a topology-based measure that distinguishes focused reasoning from diffuse search.
    • We empirically show that token count is a poor proxy for reasoning quality, with a correlation of r = -0.05 with eta, and that extra tokens are mostly consumed by verification overhead.
    • Analysis of open-source reasoning models shows that the highest difficulty level is barely solved by any model, even with large token budgets, confirming the limits of simple compute scaling.
    • Early errors are strongly associated with inefficient traces, with r = 0.28, and eta increases when the time of error occurrence is later.
    AbstractPDF
  42. 353Machine LearningarXiv:2606.04662

    Why Muon Outperforms Adam: A Curvature Perspective

    • A second-order Taylor expansion decomposes the loss reduction of Muon and Adam. The first-order gains are similar, but the difference appears in the second-order curvature penalty.
    • The update norms are similar, but Muon has lower NDS (normalized direction sharpness), so its curvature cost is smaller.
    • Controlled experiments with Zipf-PCFG show that data imbalance amplifies Muon's NDS advantage.
    • In the middle and later stages of training, Muon's lower NDS comes mainly from reduced within-layer curvature.
    • On heterogeneous-curvature second-order problems, Muon allocates update energy evenly across high- and low-curvature directions.
    • It is rigorously proven that when curvature heterogeneity is sufficiently large, Muon achieves a lower loss than GD.
    AbstractPDF
  43. 352RoboticsarXiv:2606.05979

    World-Language-Action Model for Unified World Modeling, Language Reasoning, and Action Synthesis

    • It unifies WAM's future state prediction and VLA's language reasoning in a single autoregressive transformer.
    • It uses a dual structure based on meta-queries, with a World Expert for visual prediction and an Action Expert for action generation.
    • The World Expert can be disabled at inference time, which enables real-time control with 40 ms inference.
    • Test-time scaling further improves robot control accuracy.
    • It achieves state-of-the-art results on RoboTwin2.0 Clean at 92.94% and RMBench at 56.5%.
    • It also learns new tasks from cross-embodiment videos without action annotations, which points to a scalable direction for robot learning.
    AbstractPDF
  44. 351RoboticsarXiv:2606.01955

    WALL-WM: Carving World Action Modeling at the Event Joints

    • It shifts the paradigm from chunk-centric optimization to event-grounded optimization, aligning natural boundaries across language, video, and action.
    • It builds a data ecosystem made up of event captions and cluster-balanced sampling.
    • It uses dual inference modes: Event mode for variable-length event execution and Unified mode for fixed chunks with Staircase Decoding.
    • It provides a scalable training recipe built on large-scale pretraining infrastructure with the Muon optimizer.
    • It demonstrates strong generalization across language, scene, and task settings.
    • It outperforms the existing WAM on both physical video prediction and executable control.
    AbstractPDF
  45. 350Machine LearningarXiv:2606.06021

    OPRD: On-Policy Representation Distillation

    • It moves distillation from output-space matching to hidden-state-space matching, overcoming the structural limitations of all prior OPD variants.
    • It removes the variance problem of Monte Carlo KL estimation for large-vocabulary models such as Qwen with about 150K vocabulary by using deterministic MSE instead.
    • The teacher's layer-wise and position-wise intermediate representations, which LM-head projection would discard, are transferred directly to the student.
    • It fully closes the student-teacher gap on three competitive math benchmarks, AIME 2024, AIME 2025, and AIMO.
    • It achieves 1.44x faster training than top-k OPD on the same hardware and uses up to 54 percent less actor-update memory.
    • Its memory and distributed advantages are especially strong in expanded scenarios such as multi-model RL merging and on-policy self-distillation, or OPSD.
    AbstractPDF
  46. 349Machine LearningarXiv:2606.06096

    OrderGrad: Optimizing Beyond the Mean with Order-Statistic Policy Gradient Estimation

    • We propose an unbiased gradient estimator for order-statistic objective functions, and by changing only the rank weights, it can express many distributional objectives such as VaR, CVaR, trimmed mean, and best-of-K.
    • It is a plug-and-play design that can be applied to existing likelihood-ratio and reparameterization updates in one line of code by reformulating them as reward transformations.
    • Top-M@K shows consistent gains over the existing Max@K, which is the special case max@K with m=1, on both pass@1 and pass@256.
    • Combining a correct-answer reward for Top-M and a length penalty for Bottom-M shortens responses far more effectively than the simple scalarization used by GRPO.
    • Choosing K, the target size, and m, the number of top ranks, controls the exploration-exploitation tradeoff, with larger K improving pass@256 while reducing pass@1.
    • It has effectively negligible overhead, with an O(N log N) sorting cost.
    AbstractPDF
  47. 348LLMs / NLParXiv:2606.06447

    Latent Reasoning with Normalizing Flows

    • We place a normalizing flow inside the LLM backbone to provide exact likelihoods for continuous CoT while preserving left-to-right decoding and KV-cache support.
    • Continuous thought positions are handled by the NF head, while text positions are handled by the LM head, and both are processed jointly in the same causal stream.
    • The system supports both supervised likelihood training and policy-gradient optimization through a unified interface.
    • Even with Gaussian perturbations in latent space, pass@1 remains nearly unchanged, showing that the learned continuous thought space is locally smooth and functionally robust.
    • On code-generation benchmarks, it improves accuracy while reducing inference cost compared with explicit CoT and prior latent reasoning methods such as LaDiR.
    AbstractPDF
  48. 347Computer VisionarXiv:2606.02800

    Cosmos 3: Omnimodal World Models for Physical AI

    • It is an omnimodal design based on a single mixture-of-transformers architecture that flexibly supports input and output across language, image, video, audio, and action.
    • A dual-tower layer structure and dual-stream joint attention effectively fuse multimodal information.
    • It serves both as a reasoner and a generator, and is released in Super and Nano scales.
    • It uses large synthetic datasets, such as SDG-PhyxSim, SDG-RobotSim, and SDG-DriveSim, for training.
    • According to Artificial Analysis, it was rated the best open-source text-to-image and image-to-video model, and the top policy model in RoboArena.
    • It fully releases code, checkpoints, datasets, and evaluation benchmarks under the Linux Foundation OpenMDW-1.1 license.
    AbstractPDF
  49. 346Machine LearningarXiv:2606.06080

    On Advantage Estimates for Max@K Policy Gradients

    • The paper mathematically proves that the existing PKPO advantage estimator is unbiased for policy gradients but not centered, and addresses this with a Leave-Two-Out (L2O) baseline.
    • It derives the canonical advantage u_i - v_i for finite-batch max@K and unifies existing pass@K and max@K estimators under a single framework.
    • It is efficient with a vectorized implementation that has O(B^2) time complexity and integrates naturally into group-based RL pipelines.
    • On Llama-3.2-3B-Instruct, it reduces gradient variance by 77.4 percent relative to PKPO and improves pass@256 by 5.2 percent on Qwen2.5-Math-7B.
    • It shows consistent gains over strong post-training baselines for LLMs, including GRPO and Entropy-Adv.
    AbstractPDF
  50. 345AI / GeneralarXiv:2606.05405

    Agents' Last Exam

    • It is a large benchmark with 55 subdomains, 13 industry clusters, and 1,490 task instances based on the O*NET and SOC 2018 occupational taxonomy.
    • Tasks are collected from projects already completed by working professionals and are evaluated automatically with deterministic scripts and structured rubrics, removing dependence on human raters.
    • The tasks are classified into three tiers by difficulty, Near-Term with 59 tasks, Full-Spectrum with 55, and Last-Exam with 36, to support agent evaluation at different levels.
    • Failure analysis shows that lack of domain knowledge, meaning understanding and approach, accounts for about 75 percent of all failures, so knowledge is the main bottleneck rather than execution ability.
    • Model choice creates about three times larger performance differences than agent harness choice, and using more resources does not always lead to better performance.
    • Even the strongest configuration, Codex plus GPT-5.5, reaches only a 2.6 percent average pass rate, which suggests that the benchmark is still far from saturated.
    AbstractPDF