HDATF
简体中文
开始交流
菜单

论文档案库

按编号从大到小列出全部收录论文及摘要。

收录论文 1,430资料更新日期 Research Monitoring基准测试
  1. 331Game TheoryarXiv:2606.06480

    DNQ: Deep Nash Q-Network for Partially Observable n-Player Games

    • It trains bidding agents with a solver-in-the-loop equilibrium supervision framework.
    • A shared critic predicts a pairwise payoff matrix or an N-player payoff tensor to improve training efficiency.
    • The pairwise formulation greatly reduces compute cost compared with the exact N-player formulation.
    • The exact method becomes computationally infeasible as the number of agents increases.
    • It empirically shows the tradeoff between strategic fidelity and scalability.
    原文摘要PDF
  2. 330LLMs / NLParXiv:2606.06481

    Operation-Guided Progressive Human-to-AI Text Transformation Benchmark for Multi-Granularity AI-Text Detection

    • The paper proposes OpAI-Bench, a benchmark that models the nine-stage progressive human-to-AI co-editing process.
    • It evaluates AI-text detectors at four granularities: document, sentence, token, and span.
    • It covers five representative AI editing operations across four domains.
    • It finds non-monotonic patterns in which intermediate versions of mixed-authorship text are harder to detect than either fully human text or high-ratio AI text.
    • AI detectability depends on the editing ratio, operation type, domain, and cumulative history.
    原文摘要PDF
  3. 329Machine LearningarXiv:2606.06486

    Regret Minimization with Adaptive Opponents in Repeated Games

    • It introduces RP-Regret, a new metric for repeated games with adaptive opponents.
    • RP-Regret is fundamentally a nonconvex problem, and the paper proposes three algorithms to address it.
    • When all players minimize RP-Regret, they can learn a subgame-perfect equilibrium.
    • On the Stag-Hunt game, the methods converge to more cooperative and higher-utility solutions than existing approaches.
    • It identifies necessary conditions on the memory of comparison strategies and opponent strategies.
    原文摘要PDF
  4. 328RoboticsarXiv:2606.06491

    TempoVLA: Learning Speed-Controllable Vision-Language-Action Policies

    • Existing VLA models learn only fixed speeds, but TempoVLA controls execution speed with an explicit speed condition.
    • VSTA augmentation preserves motion semantics while retiming demonstrations at arbitrary speeds.
    • It achieves bidirectional flexible speed control in both simulation and the real world.
    • Combined with a large multimodal model, it enables risk-aware dynamic speed control.
    • VSTA also improves the base 1x-speed performance by increasing data utilization.
    原文摘要PDF
  5. 327Software EngineeringarXiv:2606.06492

    Code2LoRA: Hypernetwork-Generated Adapters for Code Language Models under Software Evolution

    • A hypernetwork generates repository-specific LoRA adapters to inject repository knowledge at inference time without token overhead.
    • It supports two modes, Code2LoRA-Static for static snapshots and Code2LoRA-Evo for evolving codebases.
    • Code2LoRA-Evo adapts to commit-level diffs with a GRU hidden state.
    • It introduces RepoPeftBench, a new benchmark built from 604 Python repositories.
    • It comes close to the repository-specific LoRA upper bound with 66.2 percent exact match on the static track and 60.3 percent exact match on the evolving track.
    原文摘要PDF
  6. 326RoboticsarXiv:2606.06493

    HANDOFF: Humanoid Agentic Task-Space Whole-Body Control via Distilled Complementary Teachers

    • We propose a concise, modular interface between task planning and whole-body control.
    • We distill three experts, whole-body motion tracking, locomotion, and fall recovery, into a mixture of experts.
    • We achieve SOTA speed tracking and a wide manipulation workspace on Unitree G1 hardware.
    • A VLM-based agent planner enables control through natural-language commands.
    • It performs a wide range of tasks without task-specific data or controller fine-tuning.
    原文摘要PDF
  7. 325Machine LearningarXiv:2606.06494

    TailLoR: Protecting Principal Components in Parameter-Efficient Continual Learning

    • It designs low-rank updates by using the singular-value basis, U and V, of the pretrained weights as a fixed reference frame.
    • A soft spectral penalty suppresses updates along principal-component directions to reduce interference between tasks.
    • Fine-grained adaptation is routed into the long-tail spectral coordinates that have higher flexibility.
    • It improves continual-learning performance for PEFT methods based on spectral decomposition.
    原文摘要PDF
  8. 324LLMs / NLParXiv:2606.05122

    Self-Evaluation Is Already There: Eliciting Latent Judge Calibration in Base LLMs with Minimal Data

    • This paper finds that the capability is generally present even before any targeted training. With only a few-shot prompt, the base model already predicts multi-attribute quality scores from an external judge on open-ended responses across three benchmarks far above chance.
    • From just 160 unique examples, about 31 times fewer than the RL baseline, SEE improves held-out calibration across three benchmarks while preserving answer quality.
    • These results recast judge-aligned self-evaluation as a problem of extraction rather than acquisition.
    原文摘要PDF
  9. 323Computer VisionarXiv:2606.05124

    Geometry Gaussians: Decoupling Appearance and Geometry in Gaussian Splatting

    • Since the success of 3D Gaussian Splatting (3DGS) for novel view synthesis, many works have explored how to use it for geometric surface representations as well.
    • This study shows that base 3DGS is inherently ill-suited to represent texture and geometry at the same time by training with complete ground-truth texture and geometry.
    • The authors also propose a simple solution that applies a single additional geometric opacity parameter to each splat and combines it with an optional transparency curation optimization pipeline.
    原文摘要PDF
  10. 322Machine LearningarXiv:2606.05129

    Preserving Data Privacy in Learning Causal Structure with Fully Homomorphic Encryption

    • This paper proposes a fully homomorphic encryption (FHE)-based method that keeps data encrypted during transmission and computation by performing calculations on ciphertexts.
    • It also shows that the method can be portably extended to support differential privacy, making it easy to scale beyond FHE.
    • Experimental results show that the method achieves high consistency and comparable causal structures to the plaintext version on the tested datasets.
    原文摘要PDF
  11. 321Machine LearningarXiv:2606.05130

    Towards Efficient and Evidence-grounded Mobility Prediction with LLM-Driven Agent

    • Supervised sequence models achieve high accuracy, but they require task-specific training and provide little transparency at the decision level.
    • This paper proposes a training-free LLM-driven agent framework that formulates next-location prediction as adaptive, evidence-controlled decision making.
    • Across three mobility datasets, AgentMob achieves the strongest overall performance among training-free LLM-based methods, and GPT-5.4 reaches 71.42 percent Acc@1 on BW, 33.14 percent on YJMob100K, and 33.50 percent on Shanghai ISP.
    原文摘要PDF
  13. 319LLMs / NLParXiv:2606.05134

    Activation-Based Active Learning for In-Context Learning: Challenges and Insights

    • However, this paper reports a negative result: MLP outputs, viewed through the lens of large activations or the first four moments, do not correlate with example quality or task performance.
    • Specifically, across all tasks and models tested, the absolute Spearman correlation coefficient reached only 0.33 at most, showing that activation-based sampling should not be used for in-context learning.
    • The authors hypothesize that this may be due to superposition, the phenomenon where models represent more features than they have dimensions, which suggests methods such as Sparse Autoencoders (SAEs) may be a promising direction.
    原文摘要PDF
  14. 318Machine LearningarXiv:2606.05138

    Generating Financial Time Series by Matching Random Convolutional Features

    • To mitigate this, recent approaches train generators to minimize the mismatch between non-learned feature representations of real and generated time series.
    • This paper shows that a generator trained by matching random SOCK features consistently outperforms signature-based and diffusion-based baselines across multiple small-sample financial datasets.
    • The authors further demonstrate SOCK's expressiveness on two-sample hypothesis testing and time-series classification tasks, where SOCK matches or surpasses existing unsupervised feature maps.
    原文摘要PDF
  15. 317Machine LearningarXiv:2606.05139

    BBOmix: A Tabular Benchmark for Hyperparameter Optimization of Unsupervised Biological Representation Learning

    • Deep unsupervised learning architectures, especially autoencoders, are increasingly used in this field for dimensionality reduction and representation learning.
    • To make large-scale unsupervised HPO research broadly accessible, the paper introduces BBOmix, the first open-source tabular benchmark for unsupervised representation learning on real biological data.
    • The benchmark includes 105,000 evaluations across four AE architectures and seven multi-omics modalities in the TCGA and SCHC datasets.
    原文摘要PDF
  16. 316Computer VisionarXiv:2606.05142

    GeM-NR: Geometry-Aware Multi-View Editing for Nonrigid Scene Changes

    • The paper demonstrates that the method can handle edits with large changes in geometry and appearance, which existing methods struggle with.
    • The authors perform extensive evaluations showing that the method improves consistency across a range of editing tasks, including generating a 3D representation of the edited scene.
    • Both quantitative and qualitative results indicate state-of-the-art performance in edit quality as well as geometric and photometric consistency across multiple viewpoints.
    原文摘要PDF
  17. 315RoboticsarXiv:2606.05143

    HORIZON: Recoverability-Governed Curriculum for Physical-Domain Scaling

    • In on-policy learning, a new dynamic is useful only when it stays close enough to the current policy to generate corrective on-policy data instead of collapsing rollouts into irreversible failure.
    • Using quadruped locomotion as a physically challenging benchmark for embodied generalization, the paper introduces HORIZON, a checkpoint-based frontier curriculum that expands the physical domain only within the recoverability boundary of the current policy.
    • Taken together, the results frame physical-domain generalization as a continual-growth problem for embodied control and make recoverability the organizing principle for on-policy scaling.
    原文摘要PDF
  18. 314Machine LearningarXiv:2606.05145

    Failed Reasoning Traces Tell You What Is Fixable (But Not by Reading Them)

    • The paper argues that this throws away an important signal, because some failures come from unlucky sampling and can be helped by more rollouts, while other failures are structural and resist resampling regardless of budget.
    • The authors propose that failed reasoning traces encode recoverability structure, meaning an inference-time signature of which test-time intervention can rescue a given failure.
    • Three task-level trajectory features derived from the structure of available interventions recover this structure from the distributional signature of failed rollouts rather than from the text of the rollouts themselves.
    原文摘要PDF
  19. 313Computer VisionarXiv:2606.05149

    An Open-Source Two-Stage Computer Vision Pipeline for Fine-Grained Vehicle Classification using Vision Transformers

    • Evaluated on 3,805 annotated overtaking events collected from bike lanes in Ann Arbor, Michigan, the pipeline achieved 0.94 accuracy, and class-wise F1 scores ranged from 0.91 for minivans to 0.97 for SUVs.
    • Three of the four well-represented categories retained F1 scores above 0.90 even under domain shift.
    • The complete pipeline, including the inference script, training code, evaluation utilities, and model weights, is released as open-source software to support reproducibility and reuse across roadside video archives and bicycle safety research.
    原文摘要PDF
  20. 312Neural NetworksarXiv:2606.05150

    Multi-Column RBF Neural Network Using Adaptive and Non-Adaptive Particle Swarm Optimization

    • Error correction, the state-of-the-art gradient-based learning method, chooses the optimal hidden units to improve accuracy.
    • Both ErrCor and PSO demonstrate improved results and competitive convergence.
    • Inspired by the success of MCRN, the authors propose two new approaches for improving PSO performance: multi-column RBFN with PSO, or MC-PSO, and multi-column RBFN with APSO, or MC-APSO.
    原文摘要PDF
  21. 311Machine LearningarXiv:2606.05152

    Reinforcement Learning from Rich Feedback with Distributional DAgger

    • The paper shows that RL methods using existing self-distillation objectives based on reverse KL or Jensen-Shannon do not guarantee monotonic policy improvement, meaning the update can increase the probability of worse actions even when the expert has higher reward.
    • In contrast, it shows that forward cross-entropy allows monotonic policy improvement and provides regret guarantees.
    • It also shows that this objective improves Pass@N by optimizing a lower bound on the teacher's weighted success likelihood.
    原文摘要PDF
  22. 310LLMs / NLParXiv:2606.05158

    Streaming Communication in Multi-Agent Reasoning

    • The paper introduces StreamMA, a multi-agent reasoning system that reduces latency by streaming each reasoning step to downstream agents as soon as it is generated, thereby pipelining neighboring agents.
    • Across eight reasoning benchmarks spanning math, science, and code, two state-of-the-art LLMs, Claude Opus 4.6 and GPT-5.4, and three topologies, Chain, Tree, and Graph, StreamMA outperforms both baselines.
    • Beyond these contributions, the authors discover a 'step-level scaling law': increasing the number of steps per agent consistently improves both effectiveness and efficiency, and this is a new scaling dimension that is orthogonal to and composable with agent-count scaling.
    原文摘要PDF
  23. 309RoboticsarXiv:2606.05159

    X4Val: Learning Neural Surrogates for Variance-Reduced Policy Evaluation

    • It motivates an evaluation method that uses heterogeneous data sources, including simulation, historical policy logs, and data collected from related platforms or environments.
    • The paper provides both theoretical analysis and empirical evaluation for autonomous driving and real-world robot manipulation tasks, where X4Val achieves up to a 38.4% reduction in variance and shows consistent gains over strong baselines in these domains.
    • These results show that using unpaired heterogeneous data can substantially improve the sample efficiency of rigorous robot system verification.
    原文摘要PDF
  24. 308RoboticsarXiv:2606.05160

    GRAIL: Generating Humanoid Loco-Manipulation from 3D Assets and Video Priors

    • This privileged setting better conditions 4D reconstruction, allowing model-based object tracking, human motion estimation, and interaction-recognition optimization to reconstruct metric 4D human-object interaction (HOI) trajectories while reducing depth ambiguity and shape mismatch.
    • The authors retarget the reconstructed motions to a humanoid robot and learn complementary general-purpose task trackers: an object-aware latent adapter for manipulation and a scene-aware tracker for terrain traversal.
    • Using only data generated by GRAIL, they learn egocentric visual policies through a sim-to-real pipeline and deploy them on the Unitree G1 humanoid, achieving real-world success rates of 84% on diverse object picking tasks and 90% on stair climbing.
    原文摘要PDF
  25. 307LLMs / NLParXiv:2606.05161

    Beyond Text Following: Repairable Arbitration Reversals in Audio-Language Models

    • The paper examines this question using an audio-fixed counterfactual method that removes only conflicting text and measures the resulting change in model preferences.
    • Across five ALMs and four conflict tasks, sign flips appeared in 64.1 percent of conflict samples. In other words, the same audio-only branch preferred the audio-supported answer, while the combined branch preferred the text-supported answer.
    • Under a strict 5 percentage point fidelity-drop budget, GACL improves nAUC by 17.8 points over the best contrastive-learning baseline and transfers to vision-text mediation without retraining, with up to a 40.5 percentage point gain.
    原文摘要PDF
  26. 306Computer VisionarXiv:2606.05162

    Controllable Dynamic 3D Shape Generation via 3D Trajectories and Text

    • To address this, the paper adopts 3D trajectories as controllable spatial guides that specify the exact path selected points should follow.
    • To robustly handle trajectory inputs with arbitrary configurations, from dense to sparse and unevenly distributed, we further propose shape-aware trajectory embeddings that map the input trajectory set into a shape-aware token set covering the entire object.
    • Quantitative and qualitative evaluations, along with user studies, show that this approach follows prompts more faithfully while preserving motion quality and generating more expressive motions.
    原文摘要PDF
  27. 305Machine LearningarXiv:2606.05165

    STRIDE: Training Data Attribution via Sparse Recovery from Subset Perturbations

    • The paper proposes a shift in perspective: instead of estimating parameter changes, it models the functional effect of training data in activation space.
    • The authors introduce STRIDE, which stands for Steering-based Training Data Influence Decomposition, a framework that formulates TDA as a sparse recovery problem in the spirit of compressed sensing.
    • STRIDE achieves SOTA on LLM pretraining attribution while being one order of magnitude, specifically 13x, faster than existing methods.
    原文摘要PDF
  28. 304LLMs / NLParXiv:2606.02545

    Transferable Self-Harm Surveillance from Emergency Department Triage Notes Using an Evidence-Augmented Machine Learning Approach

    • The approach achieved AUPRC scores of 0.887 ± 0.016 and 0.884 ± 0.012 on internal and external validation, respectively.
    • Prospectively, it achieved an AUPRC of 0.881 ± 0.008 at the development institution and 0.879 ± 0.012 and 0.816 ± 0.015 at two external institutions without site-specific retraining.
    • Its key advantage is that it can identify major self-harm methods with 95 percent accuracy, supporting more granular surveillance beyond binary classification.
    原文摘要PDF
  29. 303LLMs / NLParXiv:2606.02548

    SN-WER: Script-Normalized WER for Multi-Script Indic ASR Evaluation

    • Before computing WER, the paper proposes Script-Normalized WER (SN-WER), a training-free, evaluation-only metric that transliterates both reference and hypothesis text into the standard script for each language.
    • In controlled stress tests, artificially induced WER inflation from romanization was reduced by 67%, and in lexical substitution control experiments sensitivity to semantic errors remained almost unchanged, with Delta SN-WER to Delta WER at about 1.09.
    • SN-WER is robust to transliteration choices and normalization changes, and it showed a low token collision rate of under 0.1% in the Indian language settings evaluated.
    原文摘要PDF
  30. 302RoboticsarXiv:2606.02551

    AFUN: Towards an Affordance Foundation Model for Functionality Understanding

    • Existing methods usually cover only part of this task; they either localize task-relevant regions without specifying executable actions, or they predict actions but remain limited in scalability.
    • This paper presents ourmodel, a step toward an affordance foundation model for functionality understanding.
    • From a single RGB-D observation and a natural-language task description, ourmodel predicts a task-conditioned functionality mask that indicates where to interact and a 3D post-contact action curve that indicates how to interact.
    原文摘要PDF
  31. 301Computer VisionarXiv:2606.02552

    Modeling Depth Ambiguity: A Mixture-Density Representation for Flying-Point-Free Depth Estimation

    • The authors trace this artifact to a standard modeling choice, namely assigning a single depth hypothesis to each pixel.
    • They address this with MDA, a mixture-density representation that allows multiple depth hypotheses and their associated probabilities to be predicted for each pixel.
    • Across different backbones, MDA substantially improves boundary reconstruction and removes most flying-point artifacts even under severe input blur, while adding only negligible runtime overhead.
    原文摘要PDF
  32. 300Computer VisionarXiv:2606.02553

    LongLive-RAG: A General Retrieval-Augmented Framework for Long Video Generation

    • To make retrieval more discriminative, the authors introduce Window Temporal Delta Loss, which suppresses redundant local similarity and encourages embeddings to capture meaningful temporal change.
    • Experiments across several AR backbones and generation lengths show improved long-form video quality and the best average VBench-Long ranking.
    • As far as the authors know, LongLive-RAG is the first open-ended AR long-video generation method to formalize its self-generated latent history as content-addressable retrieval memory.
    原文摘要PDF
  33. 299LLMs / NLParXiv:2606.02556

    HERO'S JOURNEY: Testing Complex Rule Induction with Text Games

    • The paper introduces HERO'S JOURNEY, a benchmark for rule induction in goal-directed episodic tasks, where agents must infer hidden rules from demonstrations and act accordingly through multi-step execution.
    • Evaluation of state-of-the-art LLMs found that the models show evidence of rule induction, but the ability is limited and uneven across tasks.
    • Induction-specific steering methods improve performance on attribute tasks but do not reliably improve procedural tasks, suggesting that the gap in procedural induction remains an open problem.
    原文摘要PDF
  34. 298LLMs / NLParXiv:2606.02559

    From Layers to Submodules: Rethinking Granularity in Replacement-Based LLM Compression

    • Existing replacement-based methods share two design constraints: layer-level granularity and continuous selection.
    • Building on this intuition, the authors introduce SubFit, or submodule-level fitted residual replacement, which compresses LLMs at the submodule level. Attention and feed-forward submodules are selected non-contiguously, and each receives its own lightweight fitted residual bypass path.
    • SubFit operates after training and requires only calibration data.
    原文摘要PDF
  35. 297RoboticsarXiv:2606.02562

    Permissive Safety Through Trusted Inference: Verifiable Belief-Space Neural Safety Filters for Assured Interactive Robotics

    • The paper proposes an algorithmic approach that explicitly considers the reliability of a robot's runtime reasoning module and uses conformal prediction to certify high-probability safety for BeliefSF.
    • The method leverages the structure of belief-space safety filtering by concentrating verification on the regions where reasoning is expected to be trustworthy.
    • On a simulated human-vehicle interaction benchmark, the approach validates much more permissive belief-space safety filters than standard conformal prediction baselines.
    原文摘要PDF
  36. 296Machine LearningarXiv:2606.02563

    IntraShuffler: A Privacy Preserving Framework for Heterogeneous DP Federated Learning

    • Heterogeneous differential privacy (HDP) in federated learning allows clients to choose individual privacy budgets and supports budget-aware aggregation.
    • Experiments across four different datasets show that IntraShuffler reduces gradient reconstruction risk by more than 60% and lowers surrogate inference accuracy from 0.78 to 0.33, while maintaining similar model utility across multiple FL aggregation rules.
    原文摘要PDF
  37. 295Computer VisionarXiv:2606.02564

    VLMs are Good Teachers for Video Reasoning via Adaptive Test-Time Optimization

    • State-of-the-art VGM delivers strong visual quality, but it often struggles to understand and follow task-specific rules, which leads to logical failures across diverse reasoning scenarios.
    • VLMs are difficult as solvers, but they have strong perceptual abilities that allow them to evaluate whether process constraints are satisfied and whether the final goal is achieved.
    • These results show that integrating VLMs as test-time teachers offers a promising paradigm for achieving generalizable video reasoning.
    原文摘要PDF
  38. 294Computer VisionarXiv:2606.02565

    Policy-based Foveated Imaging and Perception

    • Existing acquisition strategies, such as spatial or temporal downsampling, solve this problem by irreversibly discarding information before task relevance can even be assessed.
    • The paper introduces a real-time, predictive, task-aware foveated imaging system that operates directly at image acquisition time.
    • Extensive simulation across multiple perception tasks shows that this approach achieves high task performance under strict pixel budgets and substantially outperforms related baselines operating at the same bandwidth.
    原文摘要PDF
  39. 293AI / GeneralarXiv:2606.02568

    ClinEnv: An Interactive Multi-Stage Long Horizon EHR Environment for Agents

    • Static benchmarks cannot validate this, and existing interactive medical benchmarks each compromise on at least one dimension.
    • This paper presents ClinEnv, an interactive benchmark that evaluates LLMs as attending physicians on real inpatient cases under a paradigm called Longitudinal Inpatient Simulation.
    • Each case is automatically organized into a sequence of ordered decision steps, and at every step the model must actively query four specialized agents before finalizing medications, procedures, or diagnoses.
    原文摘要PDF
  40. 292Computer VisionarXiv:2606.02569

    AdaCodec: A Predictive Visual Code for Video MLLMs

    • However, existing video multimodal large language models, or video MLLMs, usually encode each sampled frame independently as an RGB image, which causes visual tokens to repeat content that already appears in earlier frames.
    • Across 11 benchmarks, AdaCodec outperforms the frame-wise RGB baseline of Qwen3-VL-8B under the same visual-token budget.
    • Even with one-seventh of the budget, AdaCodec with 32K tokens beats the 224K baseline on every long-form video benchmark, and on five general video benchmarks it raises the average score while reducing time to first token from 9.26 seconds to 1.62 seconds.
    原文摘要PDF
  41. 291Computer VisionarXiv:2606.02572

    VISReg: Variance-Invariance-Sketching Regularization for JEPA training

    • Sketching-based methods such as SIGReg address this by aligning embeddings to an isotropic Gaussian, but they lack flexibility and suffer from vanishing gradients under collapse.
    • VISReg scales linearly, outperforms existing regularization methods on low-quality datasets, and shows robustness in long-tail and low-rank settings.
    • Pretrained on ImageNet-1K, VISReg achieves SOTA performance on out-of-distribution datasets.
    原文摘要PDF
  42. 290Computer VisionarXiv:2606.02573

    HumanNOVA: Photorealistic, Universal and Rapid 3D Human Avatar Modeling from a Single Image

    • Achieving both photorealism and generalization is difficult because diverse, high-quality 3D human data are scarce.
    • These two strategies allow scaling to as many as 100,000 assets, greatly increasing both the quantity and diversity of data needed to train robust models.
    • Extensive experiments on multiple benchmarks quantitatively and qualitatively demonstrate the superiority of this approach and its robustness under diverse input-image conditions.
    原文摘要PDF
  43. 289Computer VisionarXiv:2606.02575

    From Zero to Hero: Training-Free Custom Concept Spawning in World Models

    • The paper introduces a training-free method for concept spawning called SPAWN, which stands for Swapping Pinned Anchor with Windowed iNjection.
    • SPAWN exploits a structural property of image-to-video backbones: the first slot in context memory is pinned to the reference frame and serves as the anchor for every generated chunk.
    • Experiments show that SPAWN integrates concepts with consistent lighting, scale, and perspective while preserving identity and temporal consistency, demonstrating that controllable concept spawning is possible in existing autoregressive world models without any training.
    原文摘要PDF
  44. 288Computer VisionarXiv:2606.02576

    ProtoAda: Prototype-Guided Adaptive Adapter Expansion and Geometric Consolidation for Multimodal Continual Instruction Tuning

    • Multimodal large language models achieve strong performance through instruction tuning, but real-world deployment requires them to keep acquiring new vision-language abilities, so multimodal continual instruction tuning is essential.
    • To address this problem, the paper proposes ProtoAda, a prototype-guided adaptive tuning framework.
    • Extensive experiments across multiple benchmarks show that ProtoAda achieves superior performance, especially on tasks where answer structure is easily damaged by sequential tuning.
    原文摘要PDF
  45. 287RoboticsarXiv:2606.02577

    RoboDream: Compositional World Models for Scalable Robot Data Synthesis

    • This paper proposes a generalizable body-centric world model that achieves scalable data generation by synthesizing realistic demonstrations in new objects, new scenes, and new viewpoints.
    • The approach anchors generation to rendered robot motions while conditioning on explicit scene and object priors, effectively separating trajectory execution from environment synthesis.
    • Real-world experiments show that the generated data consistently improves downstream policy performance across diverse manipulation tasks and substantially reduces the need for real-world data.
    原文摘要PDF
  46. 286Computer VisionarXiv:2606.02578

    Mitigating Perceptual Judgment Bias in Multimodal LLM-as-a-Judge via Perceptual Perturbation and Reward Modeling

    • To address this issue, the paper introduces the Perceptually Perturbed Judgment Dataset, which constructs minimally edited counterfactual responses that separate perceptual errors and enable verifiable supervision.
    • Experiments across multiple MLLM-as-a-Judge benchmarks show that this approach substantially improves perceptual fidelity, ranking consistency, and alignment with human judgments.
    • The result establishes a scalable and generalizable path toward training multimodal judges that are perceptually grounded, interpretable, and robust to vision reasoning conflicts.
    原文摘要PDF
  47. 285Computer VisionarXiv:2606.02580

    Thinking in Blender: Staged Executable Inverse Graphics with Vision-Language Models

    • The paper introduces Staged Executable Inverse Graphics, or SEIG, an agent framework for reconstructing 3D scenes from a single image by progressively refining scene factors, including geometry, materials, composition, and lighting, directly in executable Blender code space.
    • The experiments show that staged reconstruction substantially improves reconstruction fidelity and highlights the importance of task decomposition in executable inverse graphics with a general-purpose VLM.
    • It finally demonstrates a variety of downstream applications enabled by the reconstructed and editable Blender scenes.
    原文摘要PDF
  48. 284Computer VisionarXiv:2605.31598

    Linear Scaling Video VLMs for Long Video Understanding

    • Existing efficiency methods improve scalability, but they often lose accuracy relative to full self-attention, for example through aggressive frame or token dropping or coarse attention approximations.
    • These results suggest practical progress toward scalable long-video understanding.
    • Video vision-language models, or VLMs, are increasingly used in long-horizon and streaming settings, but most video encoders still rely on spatiotemporal self-attention, so compute and latency grow quadratically with the number of frames.
    原文摘要PDF
  49. 283Machine LearningarXiv:2605.29548

    Why Larger Models Learn More: Effects of Capacity, Interference, and Rare-Task Retention

    • The result points to data-induced competition for resources, or neurons.
    • The paper develops a simple phenomenological argument that power-law scaling already suggests larger models can learn parts of the data distribution that smaller models cannot, even with infinite training data.
    • To test this claim and identify its cause, the authors study model scaling in a synthetic setting made of a mixture of tasks with monotonic scaling curves.
    原文摘要PDF
  50. 282Machine LearningarXiv:2605.30865

    GlucoFM: A Dual-Stream Foundation Model for Continuous Glucose Monitoring

    • Across four diverse cohorts and seven clinical prediction tasks, GlucoFM achieves the strongest subject-separation linear probing performance among evaluated baselines and improves average PR-AUC by 4.1 points over the best CGM-specific foundation model.
    • The gains are most pronounced on core metabolic outcomes, where it leads PR-AUC on all diabetes risk and beta-cell dysfunction tasks and on 3 of 4 insulin resistance tasks.
    原文摘要PDF