HDATF
English
Start a conversation
Menu

Paper Archive

Every archived paper with its summary, highest number first.

Papers 1,193Archive updated BenchmarksRSS feed
  1. 745Computer VisionarXiv:2607.26037

    Wonder: Video World Model Done Better

    • It converts camera motion into rendered visual cues rather than abstract data, which prevents control accuracy from collapsing after distillation.
    • A sparse full-fidelity memory preserves every past frame while keeping attention cost constant.
    • A mixture-of-students over time and camera-aware adversarial regularization preserve both visual quality and control accuracy during distillation.
    • It achieves meaningful improvements over the strongest baseline on rotation position error, going from 0.0174 to 0.0132 for translation and from 0.1155 to 0.0784 for rotation.
    • It also supports video input, so it can re-shoot existing videos by changing the camera path in real time.
    AbstractPDF
  2. 744Machine LearningarXiv:2607.24645

    Sparse Autoencoders Encode Both Concepts and Functions: The Downstream Geometry of Feature Effects

    • It measures the geometry of how output logits change when a feature is removed.
    • Most features do not provide one stable control direction.
    • Value-like features more often show low-dimensional structure, but their effects are spread across multiple directions.
    • Pointer-like features produce mostly scattered effects.
    • The paper compares ReLU, TopK, and Matryoshka Batch TopK on Gemma-2-2B.
    • Even when a feature is interpretable and causal, that does not guarantee stable control.
    AbstractPDF
  3. 743AI / GeneralarXiv:2607.24280

    From Proprietary to Open-Source: Bridging the Distribution Gap via Multi-Agent Protocol Distillation in Agentic Search

    • It uses commercial models only as offline teachers, so inference cost is zero.
    • The structured JSON convention created by the multi-agent system links teacher and student.
    • It strips away surface style to reduce stylistic contamination and hallucination.
    • It achieves average success rates of 39.4% with Qwen3-1.7B and 44.4% with Qwen3-4B.
    • Claude, GPT, and Gemini teachers are all within 2 points of each other, so API swapping is easy.
    • Convention synthesis is a one-time cost of about $1,454.
    AbstractPDF
  4. 742LLMs / NLParXiv:2607.24223

    A New Role for Relevance: Guiding Corpus Interaction in Agentic Search

    • It redefines relevance as the order in which grep explores the corpus, rather than as a top-k document selection problem.
    • Document ranking provides global relevance, while match reranking provides local relevance.
    • It seeds the retrieval start points with question-matched passages so that decisive evidence is found earlier.
    • On GPT-5.4-mini, it reaches 84 percent accuracy with only 23.9 tool calls, which is far fewer calls than competing methods.
    • It retains 79 percent accuracy even when scaled to one million documents, showing only a mild performance drop.
    AbstractPDF
  5. 741Information RetrievalarXiv:2607.24617

    LaRec: Unleashing LLM-based Latent Reasoning for Generative Recommendation

    • It greatly reduces response latency by reasoning in latent space instead of through long text reasoning.
    • During latent pretraining, step-level alignment and process direction alignment provide rich supervisory signals to intermediate reasoning steps.
    • A user-specific Gaussian mixture distribution is used to sample the starting point of reasoning within the relevant latent space.
    • GRPO is combined with a sparse-hit reward and a dense semantic-similarity reward to align reasoning with the recommendation objective.
    • The latency of latent reasoning is close to no-reasoning and much lower than chain-of-thought.
    • An online A/B test on a real advertising platform records improvements such as a 2.93% conversion rate.
    AbstractPDF
  6. 740RoboticsarXiv:2607.23969

    LeapBot-WA: World-Anchor Action Models via Predictive Latent Alignments

    • It shifts from pixel-level video reconstruction to predicting abstract physical changes in latent space.
    • It uses a co-embedding prediction structure as a world anchor, focusing on state changes rather than visual appearance.
    • An isotropic semantic autoencoder bridges the mismatch between the predicted features and the diffusion prior.
    • The heavy prediction branch is used only during training and is removed at inference, so it runs without extra cost.
    • It achieves state-of-the-art results on RoboTwin 2.0 at 91.46 percent and on LIBERO at 97.3 percent without large-scale trajectory pretraining.
    • Isotropic regularization raises latent rank from 38.1 to 92.3 and improves success rate to 71.3 percent.
    AbstractPDF
  7. 739RoboticsarXiv:2607.24159

    DeVA: Decoupled Video-Action Model with physical guidance for robot policy learning

    • The method separates video prediction and action prediction into different experts so that policy learning becomes more manageable.
    • It passes representations from multiple layers of the video backbone to the action expert through cross-attention and bridge tokens.
    • It injects auxiliary supervision for object-manipulation regions and relative depth, which strengthens physical information.
    • With 50 demonstrations per RoboCasa task, it reaches a 72.0 percent success rate and improves by up to 22 points over the same budget.
    • On three real-world dual-arm robot tasks, it averages 74 percent and outperforms GR00T-N1.6 at 48 percent and Cosmos Policy at 34 percent.
    • It converges faster with up to 20 times fewer training examples than Cosmos-Policy.
    AbstractPDF
  8. 738ResearcharXiv:2607.24528

    On Feige's conjecture

    • It constructed a short proof by connecting Vlassis and Thomas's distribution-free p-value result to the random-variable-sum problem in Feige's conjecture.
    • It showed that the probability that the sum of n independent nonnegative random variables, each with expectation 1, is less than n+1 is at least (n/(n+1))^n.
    • Because this lower bound is at least 1/e, it establishes Feige's conjecture, posed in 2004.
    • It shows how resolving Gaffke's conjecture in statistics can lead to a proof of a probability inequality, and GPT-5.6 Sol assisted the proof search.
    • However, the central theorem focuses on Feige's conjecture with threshold n+1, rather than resolving the full range of more general small-deviation parameters.
    AbstractPDF
  9. 737ResearcharXiv:2607.23980

    Sharp small-deviation inequalities for sums of independent nonnegative random variables

    • It combined the exact Dirichlet calibration theorem of Vlassis and Thomas with results from convex geometry, including Grünbaum's centroid theorem.
    • Treating 0<δ<1 and δ≥1 separately, it proved explicit lower bounds on small-deviation probabilities for sums of independent nonnegative random variables.
    • For δ≥1, the lower bound is optimal for every n and is at least 1/e, settling Feige's conjecture.
    • The result provides a sharp reference for analyzing the lower tails of sums of independent nonnegative random variables.
    • However, optimality for every n is established only for δ≥1, and the lower bound in the range 0<δ<1 has not been proved equally sharp.
    AbstractPDF
  10. 736RoboticsarXiv:2607.24485

    τ: Learning Touch-Augmented Vision-Language-Action Models from Future Visual Supervision

    • It learns spatiotemporal tactile representations using a self-supervised objective that predicts future visual features from current touch and subsequent actions.
    • It combines the learned tactile representations with pretrained visual and language features to generate robot actions, while using the prediction branch only during training.
    • It outperformed vision-only VLAs and existing tactile policies on four contact-intensive manipulation tasks, including plug and USB insertion.
    • Removing the additional branch at inference allows it to improve contact-critical robot manipulation without increasing deployment cost.
    • Evaluation focuses on TacAura's four representative tasks and limited task-specific data, leaving validation across a wider variety of robots and contact situations outstanding.
    AbstractPDF
  11. 735AI / GeneralarXiv:2607.24459

    From Execution to Capability: Scientific Experience Consolidation via Procedural Knowledge Synthesis

    • It defines the problem of turning validated execution experience into reusable procedural knowledge and persistent model capability.
    • It extracts procedures by contrasting successes and failures, then keeps only the ones that pass development validation gates.
    • It increases training data by synthesizing new questions from failures even when no ground truth is available.
    • It empirically shows the gap between abstraction and execution, where weaker models cannot use the abstract procedures directly.
    • A stronger model concretizes the procedures into executable code and turns them into standard supervised training data.
    • The 9B student improves by 5.62 points on substeps and 11.25 points on full problems even without the original procedures.
    AbstractPDF
  12. 734Computer VisionarXiv:2607.23855

    OmniVAE: An Audio-Video VAE with Cross-Modal Alignment for Joint Generation

    • Training audio VAE and video VAE separately causes their latent spaces to diverge, making synchronization difficult.
    • OmniVAE is an audio-video VAE that finely aligns the two spaces through joint training.
    • It matches temporal semantics with segment-level contrastive learning.
    • It distills semantic features from a pretrained encoder to improve the trainability of each latent space.
    • Both objectives are used only during training, so inference cost does not increase.
    • On Verse-Bench, both synchronization, DeSync dropping from 0.884 to 0.570, and audio quality improve.
    AbstractPDF
  13. 733LLMs / NLParXiv:2607.24720

    The Physics of Multi-Turn Long-Horizon Planning: From Pre-training to Post-training via Single- and Multi-Teacher On-Policy Agentic Distillation

    • Making the world model explicit as a chain of thought greatly improves long-horizon generalization, from 93.1 versus 0.3.
    • Atomic skills alone do not generalize compositionally, but adding only 5% long-horizon data causes a sharp jump.
    • Mixing in bad trajectories causes error accumulation and collapses long-horizon performance.
    • On-policy distillation has a wider effective range than GRPO under long-horizon and low-quality conditions.
    • Multi-teacher distillation converges on shared planning patterns, while conflicting patterns cause severe forgetting.
    • The experiments use full-parameter supervised learning based on Qwen2.5-100M.
    AbstractPDF
  14. 732Computer VisionarXiv:2607.24591

    CameraAnything: Refilming Videos with Arbitrary Camera Control

    • It jointly controls extrinsic camera parameters such as position and orientation, and intrinsic parameters such as focal length and resolution.
    • The method combines per-pixel ray representations and resolution-aware 3D positional encodings with self-attention.
    • It handles original-resolution editing in a single generation pass without cropping or outpainting.
    • It creates paired training data with a synthesis pipeline based on synchronized multi-camera recordings.
    • Its real-world rotation error is 2.76 degrees, which is lower than 5.12 degrees for ReCamMaster and 13.99 degrees for TrajectoryCrafter.
    • It generalizes to real videos even when trained only on synthetic data.
    AbstractPDF
  15. 731Computer VisionarXiv:2607.24027

    Sol-Attn: Accelerating Video Generation Inference via On-the-Fly Attention Sparsification

    • It speeds up a pretrained model without training by sparsifying only the attention.
    • Block proxy scores are modeled as Gaussian, so blocks are selected using thresholds based on the mean and standard deviation.
    • It does not build a proxy-score map in memory, which reduces routing cost.
    • It reuses discarded block scores for approximate correction, so accuracy loss stays small even under aggressive sparsification.
    • It achieves end-to-end speedups of 2.1x for video generation and 2.3x for editing, without degrading quality.
    • When integrated with Sol-Engine, it reaches up to 5x acceleration for video generation.
    AbstractPDF
  16. 730Computer VisionarXiv:2607.24731

    Rethinking Classifier-Free Guidance in On-Policy Diffusion Distillation

    • OPD, which only matches CFG synthesis, fails to capture branch-level stagnation, so errors cancel out.
    • Privileged information in the teacher's negative branch causes the negative-branch asymmetry, or NBA, failure.
    • PDM separately constrains the positive prediction and the CFG direction, meaning positive minus negative, to prevent cancellation.
    • When the guidance scale changes, the baseline method breaks, but PDM remains stable.
    • By adjusting direction weights and the number of supervised states, it finds the best trade-off between accuracy and efficiency.
    • It improves control fidelity in pose, depth, and dense versus sparse scribble video control.
    AbstractPDF
  17. 729Computer VisionarXiv:2607.24241

    FilmBench: A Film-Grade Benchmark for Cinematic Video Generation

    • It constructed 1,169 prompts reverse-engineered by professional directors from scenes in award-winning films, of which 1,056 are multi-shot.
    • It evaluates cinematographic language along 3 axes, 12 elements, 35 detailed metrics for T2V, and 3 metrics specific to R2V.
    • The FilmOps automatic evaluator reproduced human rankings of 9 T2V and 7 R2V models with Spearman correlations of 0.95 and 0.96, respectively.
    • It enables comparison of dynamic aesthetics and multi-shot continuity, which existing web-video-centered evaluations miss, against filmmaking standards.
    • However, because the benchmark is tailored to film and professional video production, it does not directly represent performance on short vertical videos or general user-generated videos.
    AbstractPDF
  18. 728RoboticsarXiv:2607.24744

    Data Pyramid for Embodied Manipulation

    • It organizes embodied-manipulation data into five sources: real robots, UMI-based collection, first-person and third-person video, simulation, and general vision-language data.
    • It situates each source within the tension between scalability and robot alignment, comparing quality, diversity, reusability, and physical fidelity.
    • It analyzes how embodied brains, VLAs, and world-action models select and mix multiple data sources during pretraining.
    • This classification can serve as a reference framework for robotics researchers designing data mixtures around target capabilities and collection costs.
    • A limitation is that six challenges remain unresolved, including large-scale tactile data, failure and recovery records, and action alignment across different robots.
    AbstractPDF
  19. 727LLMs / NLParXiv:2607.24653

    Kimi K3: Open Frontier Intelligence

    • Kimi K3 is a mixture-of-experts model that activates 104 billion of its 2.8 trillion total parameters and supports visual input and a 1-million-token context.
    • It combines Kimi Delta Attention, Attention Residuals, and Stable LatentMoE, using 16 of 896 experts for each token.
    • Reinforcement learning across general reasoning, agentic work, coding, and multiple reasoning-effort levels strengthened long-horizon work and compositional generalization.
    • Its overall scaling efficiency is approximately 2.5 times higher than Kimi K2's, and all weights are released for use in frontier-level model research and deployment.
    • Although it was stronger than other open and commercial models in the evaluation suite, it still lagged behind top closed models such as Claude Fable 5 and GPT-5.6 Sol.
    AbstractPDF
  20. 726Computer VisionarXiv:2607.22148

    dRAE: Representation Autoencoder with Hyper-Spherical Codes

    • It clearly diagnoses the cause of VQ codebook collapse as a geometric mismatch between Euclidean distance and the geometry of high-dimensional representation spaces.
    • The best choice is a combination of angle-based, meaning cosine similarity, code assignment and Euclidean commitment loss, while using angles for everything hurts performance.
    • It keeps 100 percent codebook utilization even when the codebook scales to 131,072 entries, and it shows faster convergence and more stable training than VQ and IBQ.
    • On class-conditional ImageNet generation, it reaches gFID 4.45 with HSQ 65536, which is a large improvement over VQ, and it also performs well on text-to-image.
    • With only 12 million image-text pairs, it reaches GenEval 0.63 and DPG-Bench 80.58, which shows an efficient generation pipeline.
    AbstractPDF
  21. 725AI / GeneralarXiv:2607.22319

    Towards Trustworthy and Cost-Efficient Data Integration: From Naïve RAG to Agentic RAG

    • It diagnoses the three major problems in LLM-based data integration, namely hallucination, label dependence, and cost, as stemming from a lack of knowledge grounding.
    • It lays out an evolution path from Naive RAG to GraphRAG and KG-RAG and then to Agentic RAG.
    • In Agentic RAG, it proposes a structure in which five agents, Planner, Retriever, Reasoner, Decision, and Evaluator, collaborate.
    • It identifies major research challenges such as memory-retrieval knowledge conflict, batch retrieval noise, and agent initialization.
    • It emphasizes that a new benchmark is needed to evaluate reliability through evidence-based reasoning, cost efficiency, and autonomy at the same time.
    AbstractPDF
  22. 724RoboticsarXiv:2607.22530

    ViTacWorld: Scaling Visuo-Tactile World Models for Contact-Rich Robot Manipulation

    • It proposes the first action-conditioned visual and tactile world model for robots and builds a training pipeline that unifies simulation and real-world data.
    • By modeling touch as an additional view, it uses view-aware conditioning and cross-view attention to preserve visual-tactile consistency.
    • Adding generated rollouts to policy training improves the average success rate of the pi0.5 plus tactile policy from 42.5 percent to 67.5 percent, a gain of 25 points.
    • Pretraining and simulation data greatly improve generation quality across PSNR, SSIM, and LPIPS.
    • It can also be used as a policy evaluation tool to predict visual and tactile outcomes for a given action sequence.
    AbstractPDF
  23. 723RoboticsarXiv:2607.22535

    Robot-Factored World Models via Robot Rendering

    • It converts actions into nominal trajectories for a robot controller and renders them as input, passing action information without leaking future states.
    • By adding end-effector depth and scene depth, it enables contact and occlusion judgments that are impossible on the 2D image plane.
    • Across both SVD and Wan backbones, it improves PSNR, SSIM, and LPIPS over vector-conditioned baselines.
    • A new robot, xArm6 plus Inspire F1, not used in training can be applied zero-shot simply by changing the rendering.
    • Rendering human manipulation videos after retargeting the hand motions to the robot makes robot manipulation videos possible.
    AbstractPDF
  24. 722Computer VisionarXiv:2607.22531

    Twins: Learn to Predict Unified Representations with Focal Loss

    • It builds unified representations by simply concatenating ViT semantic features and VAE detail features channel by channel, with fixed token length.
    • It finds that when training a DiT, the ViT branch learns well but the VAE branch does not, and identifies three causes: frequency, dimensionality, and conditional dependence.
    • Applying focal loss to flow matching balances the VAE channels by weighting their errors, improving ImageNet gFID by up to 10.57 over MSE.
    • Compared with a single encoder such as SigLIP2, it preserves understanding performance while substantially improving reconstruction quality, as measured by PSNR.
    • With classifier-free guidance, it achieves an FID of 1.59, which is a solid generation quality result.
    AbstractPDF
  25. 721LLMs / NLParXiv:2607.22043

    Scaling Native Multimodal Pre-Training From Scratch

    • The scaling law for language objectives is almost unaffected by data composition, but multimodal objectives are very sensitive to the data ratio.
    • Text-heavy data is more efficient for larger models, so the optimal allocation of resources shifts toward model capacity.
    • It derives, in formula form, an efficiency frontier that specifies the optimal combination of model size, token count, and data ratio.
    • It confirms a cross-modal transfer effect in which native multimodal training also improves pure text-space reasoning ability.
    • It observes the natural emergence of multimodal in-context learning in a 3B-parameter model.
    AbstractPDF
  26. 720Computer VisionarXiv:2607.22534

    SM4RT: Learning Structured Motion Geometry for 4D Reconstruction

    • It proposes Structure-of-Motion, or SoM, a representation that expresses video motion as a combination of rigid motion bases rather than individual points.
    • It infers 3D geometry, motion, and scene motion structure in a single forward pass.
    • On the Kubric benchmark, it achieves state-of-the-art tracking accuracy, APD 90.62, and shows strong structure preservation on the deformation metric VM.
    • The DA3 backbone outperforms VGGT by a wide margin, and simple linear fusion is the most effective feature fusion method.
    • It is efficient at inference, taking 2.86 seconds per sample and using only 15.72 GB of GPU memory.
    AbstractPDF
  27. 719LLMs / NLParXiv:2607.22529

    Skill Self-Play: Pushing the Frontier of LLM Capability with Co-Evolving Skills

    • It uses reusable skills as the main carrier so that task diversity and verification reliability are both preserved.
    • A proposer, a solver, and a skill controller all evolve together in the reinforcement-learning loop.
    • The skill library creates about 20 new skills per iteration, updates existing skills, and drops useless ones.
    • With Qwen3-4B, it gains up to 42.9 points on tool use, BFCL, and 12.0 points on logical reasoning, ZebraLogic.
    • Dynamic skill evolution is the key, and it adds another 2.6 points in overall accuracy compared with self-learning without skills.
    • Even models that start out misaligned show a large turnaround effect.
    AbstractPDF
  28. 718Computer VisionarXiv:2607.21580

    GraphVid: Interactive Graph-Controllable Video Generation

    • It proposes the first framework for generating video conditioned on a scene graph, where objects are nodes and interactions are edges.
    • The Edge-Aware Graph Reasoning module performs message passing while reflecting the directional properties of edges.
    • It builds a new GraphVid-Bench dataset with about 27K interaction-centered videos.
    • It reduces FID by 39.9%, reduces FVD by 37.6%, and improves PSNR from 9.87 to 15.98 compared with Motion-I2V.
    • With only 0.6B trainable parameters, it is far more efficient than existing methods that use billions of parameters.
    • Because the graph representation is backbone-independent, it works with both LTX (2B) and Wan 2.2 (5B).
    AbstractPDF
  29. 717ResearcharXiv:2607.21367

    A solution to 2-copy distillability of Werner states

    • It completely resolves the 2-copy distillability problem for Werner states in arbitrary dimensions.
    • It proves that the 1-copy and 2-copy distillation thresholds are exactly the same, showing that the second copy provides no advantage.
    • It uses geometric estimates on symmetric and antisymmetric tensor subspaces as the key tool.
    • It resolves the first finite-copy step of the broader open problem of NPT-state distillability.
    • At the boundary case α = -1/2, the proof combines block-operator inequalities with variational methods.
    AbstractPDF
  30. 716LLMs / NLParXiv:2607.21051

    Sample-Efficient Learning from Agent Experience

    • The paper defines the problem of Experience Distillation, which distills an agent's interaction experience into weights, and proposes a solution.
    • It removes extra environment use by using a one-step branch rollout that resamples only the teacher's next action at each point in the collected trajectory.
    • It preserves 64.8 percent of the in-context learning effect, which greatly outperforms standard SFT at 3.8 percent recovery.
    • It matches reinforcement-learning baselines while using more than 9.6 times fewer environment samples.
    • The experimental details, including experience preprocessing, strengthened teacher inference, and branch packing, are validated empirically.
    AbstractPDF
  31. 715Machine LearningarXiv:2607.21427

    Context-weighted Discrete Flow Matching

    • It confirms experimentally that the more filled neighbors a token has, the lower its prediction entropy becomes.
    • It proposes a context-weighted sampler that reflects local context density in sampling without any extra training.
    • A Scaled Cross-Entropy loss reweights the training signal and lowers OpenWebText generation perplexity by 63%.
    • On molecule generation with QM9, it increases the number of valid molecules by 2.8x and the number of novel molecules by 1.9x.
    • It keeps quality similar to semi-autoregressive block-diffusion methods while preserving the flexibility of order-free generation.
    AbstractPDF
  32. 714AI / GeneralarXiv:2607.21503

    Agentic Context Management: Solving Agent Memory and Cost by Treating Them as Lifecycle and Architecture Problems

    • It redefines agent context management as a 5-stage lifecycle, design, extraction, scope, prediction, and compression, instead of treating it as simple storage and retrieval.
    • It proves mathematically that unlimited context accumulation increases token cost as O(n²), while naive summarization causes sharp drops in accuracy.
    • Its scope-aware design, which respects organizational layers such as user, customer, and client, simultaneously supports privacy isolation and organizational knowledge accumulation.
    • The reference implementation, Maximem Synap, achieves 92 percent on LongMemEval and 93.2 percent on LoCoMo.
    • It contrasts itself with commercial memory systems such as Mem0, Zep, and Cognee, which mostly focus on extraction and storage and do not cover the full 5-stage lifecycle.
    AbstractPDF
  33. 713ResearcharXiv:2607.21554

    Strategic Plan for Neutral Atom Quantum Computation

    • It presents a comprehensive roadmap covering all areas of neutral-atom quantum computing, including hardware, error correction, compilation, and networking.
    • It confirms scaling trends of about 1.8x more qubits per year and about 40% lower two-qubit gate error per year.
    • It emphasizes integrated photonics as the key technical path for scaling to large systems.
    • It explains how new error-correcting codes such as qLDPC can greatly reduce resource requirements compared with surface codes.
    • It summarizes the possibility and technical challenges of distributed quantum computing by networking multiple quantum processors.
    AbstractPDF
  34. 712Computer VisionarXiv:2607.21576

    Self-Supervised Learning of Structured Dynamics from Videos

    • The method proposes SDM, which structures video motion into two tokens, one for camera motion as the primary token and one for object motion as the residual token.
    • A lightweight model is placed on top of a frozen DINOv2 image backbone, which lets it learn dynamics representations without video pretraining.
    • It creates a new benchmark called ProbeMotion and measures performance systematically on synthetic and real videos.
    • With only weak supervision, it reaches performance similar to or better than VGGT trained on 17 datasets.
    • The primary token adapts automatically so that it captures the camera in camera-motion scenes and objects in static-camera scenes.
    AbstractPDF
  35. 711Machine LearningarXiv:2607.21585

    Expanding Flow Maps

    • It is the first to propose a new generative model structure called Expansion Flow, or EFlow, which increases dimensionality during generation.
    • It alternates expand and transport operations to generate high-quality samples in only one or two steps.
    • For molecular graph generation, it reduces the one-step FCD error to 0.44 versus 2.14 for the prior CFM method.
    • It also outperforms fixed-length models on language generation, which shows that variable-length generation does not hurt quality.
    • It demonstrates that the same framework works in both continuous domains, such as molecular structure, and discrete domains, such as graphs and language.
    AbstractPDF
  36. 710Computer VisionarXiv:2607.21072

    Show, Don't Tell: Evaluating Spatial Cognition in Generative Pixels Rather Than LLM Text

    • It proposes the ProVisE framework, which makes image generation models express spatial answers directly in pixel space.
    • It builds the SpatialGen-Bench diagnostic benchmark with 470 samples and 14 subtasks.
    • An agent builder automatically constructs and validates an evaluation protocol for a new benchmark.
    • Image generation models are competitive on intuitive visual-expression tasks.
    • Eighty-eight percent of failures come from errors in spatial reasoning itself, not from the visual communication channel.
    AbstractPDF
  37. 709RoboticsarXiv:2607.21582

    Scale Up Strategically: Learning Compositional Generalization via Bias-Aware Evaluation and Data Collection for Robotic Manipulation

    • It builds a framework that decomposes language instructions into five factors, color, verb, object, size, and space, to measure bias.
    • It finds a consistent bias structure across six base policies in the order color, object, spatial, verb, and size.
    • It proposes two quantitative metrics, FDR for pairwise bias across factors and FDH for the full ranking.
    • The bias diagnostics show that focusing data collection on weak factors improves generalization in real robots.
    • It achieves better performance than the previous full dataset even when using only half of the demonstrations.
    AbstractPDF
  38. 708ResearcharXiv:2607.21509

    Aspects of Closed Matricial Worlds

    • It analyzes the Hilbert space of two-dimensional quantum gravity with a positive cosmological constant using a matrix model.
    • It treats both the Wheeler-DeWitt equation and the path-integral perspective, and discusses their differences.
    • It explicitly computes the corrections that nontrivial topology contributes to the path-integral wavefunction.
    • It presents explicit results up to genus 6 and shows an analogy with topological gravity in the large-volume limit.
    • It names the double-scaling limit of the pure-gravity matrix model a 'closed matricial world'.
    AbstractPDF
  39. 707AI / GeneralarXiv:2607.21557

    OpenForgeRL: Train Harness-native Agents in Any Environment

    • It records harness model calls as proxies so standard RL code, such as veRL, can train on them.
    • Each rollout is isolated in a Kubernetes container, allowing complex environments to run safely.
    • Training and deployment share the same harness, eliminating train-deploy mismatch.
    • OpenForge-Claw reaches 55.9 pass@3 on ClawEval, and OpenForge-GUI reaches 37.7 on OSWorld.
    • RL improves self-verification and tool coverage, but error recovery remains weak.
    AbstractPDF
  40. 706Computer VisionarXiv:2607.21595

    3D-Aware VLMs with Implicit and Explicit Geometries

    • We inject both implicit geometry tokens (IGT) and explicit geometry tokens (EGT) into the VLM at the same time.
    • EGT encodes detailed 3D information, such as depth maps and point clouds, into lightweight embeddings.
    • A 3D-aware adapter effectively fuses 2D vision, implicit 3D, and explicit 3D information.
    • It uses only RGB video as input, so no additional 3D sensors are needed.
    • It achieves SOTA on four tasks: 3D video detection, 3D visual grounding, 3D dense captioning, and spatial reasoning.
    AbstractPDF
  41. 705Computer VisionarXiv:2607.21556

    Visual Contrastive Self-Distillation

    • It learns visual dependence by comparing the difference in token probabilities between the original image and an image with the content removed.
    • It needs no external teacher, privileged ground truth, visual evidence signal, or reasoning trajectory.
    • It adds no inference-time cost, because the EMA teacher is evaluated twice only during training.
    • On Qwen3-VL, it improves by +4.77 percentage points at 2B, +1.86 points at 4B, and +3.75 points at 8B.
    • It consistently outperforms previous self-distillation methods across 7 vision-language benchmarks.
    AbstractPDF
  42. 704RoboticsarXiv:2607.21588

    AXIS: A Growable Community-Driven Data Engine for Scalable Robot Manipulation

    • Browser-based MuJoCo-WASM teleoperation lets the system collect demonstration data without specialized equipment.
    • It builds a scalable community dataset with 207 tasks and more than 50K trajectories.
    • It integrates automatic success verification, quality filtering, trajectory smoothing, and IsaacSim augmentation into one pipeline.
    • Fine-tuning a pi0.5 model on AXIS data improves success rate by 5.8 percent and beats RoboCasa365 by 37.3 percent.
    • Performance improves consistently as the amount of data grows.
    AbstractPDF
  43. 703Computer VisionarXiv:2607.21594

    Streaming Multi-Agent Autoregressive Diffusion Model with World State Registers

    • It uses a world-state register, which is a learnable token, to explicitly manage shared state across multiple agents.
    • The register is grounded with three supervision signals: agent state, bird's-eye view, and scene text.
    • A Mixture-of-Transformers structure reduces interference between state updates and frame generation.
    • The Self-Forcing paradigm is used to update the register dynamically during long-horizon rollouts.
    • In two-agent Minecraft experiments, both logical consistency and generation quality improve.
    AbstractPDF
  44. 702LLMs / NLParXiv:2607.20911

    Tencent WorkBuddy Bench: A Multi-Domain Coding-Agent Benchmark with Contamination-Resistant Task Construction

    • It reverse-engineers real commit, pull request, and security scenarios to create tasks that cannot be found through web search.
    • The benchmark contains 260 tasks across four domains: 80 code, 70 web, 50 office, and 60 security tasks.
    • Each task is rewritten as a everyday, conversational persona request to prevent contamination.
    • Claude Opus 4.8 ranks first with an overall score of 75.0 percent, and GLM-5.2 ranks second with 72.9 percent.
    • The authors release the full code, environment images, and evaluation tools so anyone can reproduce and audit the benchmark.
    AbstractPDF
  45. 701Machine LearningarXiv:2607.21366

    Hilbert Operator for Progressive Encoding (HOPE): A Mathematical Framework for Deconstructing Learned Representations in Deep Networks

    • We model neurons as rank-1 operators in a Hilbert space, unifying pruning and merging under a single mathematical criterion.
    • It works without data or hyperparameters by using batch-normalization statistics.
    • Macro block eviction makes it possible to evaluate the entire residual connection path.
    • In DEFT fine-tuning, we propose merging redundant neurons to make room in the parameter space for new tasks.
    • The entire framework is mathematically rigorous, and concept-proving experiments validate it.
    AbstractPDF
  46. 700AI / GeneralarXiv:2607.21461

    AREX: Towards a Recursively Self-Improving Agent for Deep Research

    • It exploits the asymmetry between discovery and verification, where finding an answer is hard but checking it is easy, and uses verification as the starting point for the next research round.
    • A self-contained context compression tool keeps only the essentials from long conversations, making long-horizon research possible.
    • With a 4B model, it achieves 82.5 on BrowseComp and 52.4 on HLE, far outperforming similarly sized competing models.
    • It teaches search, tool use, and verification step by step through agent mid-training and long-horizon reinforcement learning.
    • It reduces the credit-assignment problem of sparse rewards by weighting the moments when key evidence is obtained.
    AbstractPDF
  47. 699Computer VisionarXiv:2607.21553

    SANA-Video 2.0: Hybrid Linear Attention with Attention Residuals for Efficient Video Generation

    • It mixes linear attention and softmax attention in a 3-to-1 ratio so that it can achieve both speed and quality.
    • The AttnRes technique improves information quality in deep layers by about 12 percent.
    • A single H100 GPU can generate a 720p video in 13 seconds, which is 120 times faster than Wan 2.2.
    • It directly learns the optimal balance point by showing that 25 percent softmax attention is the best trade-off between quality and efficiency.
    • Kernel optimization through Sol-Engine delivers an additional 3.58 times speedup.
    AbstractPDF
  48. 698LLMs / NLParXiv:2607.20062

    Solar Open 2 Technical Report

    • A 250B-A15B MoE structure scales up from Solar Open 1 (102B) while routing only 2.3% of the parameters.
    • Hybrid attention, 75% linear and 25% softmax, plus NoPE, supports a 1M-token context.
    • The Korean tokenizer is 24% more efficient than global models, meaning it represents the same Korean text with fewer tokens.
    • It ranks first among comparable open models and commercial APIs on average Korean benchmarks, and is nearly on par with DeepSeek-V4-Pro (1.6T) on Ko-GDPval.
    • It integrates 12 domain experts into a single model through multi-teacher on-policy distillation (MOPD).
    AbstractPDF
  49. 697ResearcharXiv:2607.19918

    Scalable Keyword Spotting via Modular Network Expansion

    • It freezes the entire base network, including batch-normalization statistics, so existing keyword performance is preserved at the pixel level.
    • With only 10K additional parameters, 6.7% of the 150K base, it improves FRR for new keywords from 6.46% to 4.37%.
    • It achieves better performance with fewer MACs, 16.34M versus 18.45M and 20.52M, than Adapter or LoRA.
    • A core-first decision rule preserves the exact behavior of the existing keyword threshold.
    • It can scale using only new keyword data, without access to the original training data.
    AbstractPDF
  50. 696Computer VisionarXiv:2607.20417

    ATSplat: Compact Feed-forward 3D Gaussian Splatting with Adaptive Token Expansion

    • An adaptive token expansion, or ATE, module automatically finds regions with large rendering error and adds Gaussians there.
    • Compared with dense baselines, it reduces the number of Gaussians by more than 5.7x while still reaching state-of-the-art quality.
    • It reconstructs 12 images at 512×960 in under 1 second and renders at 1,136 FPS.
    • Its 3D anchor-offset design decouples Gaussian placement from the input pixel grid.
    • Uncertainty predictions align well with the actual rendering errors, as shown visually.
    AbstractPDF