跳到正文
研踪PaperTrail

论文库

检索计算机领域论文,按研究方向、作者或关键词筛选。

部分采集异常
正在读取账号与研究记录…

全站更新概况

+1352新增论文
+1083其中 AI 相关
+143已有论文修订
+94新增 AI 导读

arXiv 最近完成扫描:2026-11-19 06:59 · 全站最近入库:2026-11-19 15:04各信源进度 →

统计口径与数据新鲜度

统计 2026-11-18 15:26 — 2026-11-19 15:26(北京时间) · 点击数字查看对应简报;AI 相关包含在新增论文内。

最近完成检查:2026-11-19 06:59最近发现新增版本:2026-11-19 15:04库内论文最新更新时间:2026-11-19 02:59

每分钟刷新已保存的数据。空检查不代表采集异常;新收录可能包含历史论文。AI 导读按新增版本计数。

27,409 已收录论文22,787 AI 相关8,342 有修订112 HF 社区发现

浏览论文库

时间范围跟随排序:版本更新、本站收录或首次发表。最新收录可能包含历史论文。部分信源异常,查看状态 →

找到 4,873 篇论文 · 按版本更新时间排序上轮扫描起点 2026-11-19 06:57(北京时间)
新论文 · v1arXiv:2611.24940

$Z$-sight: Unifying Vision-Language-Action and Latent World Modeling

Junjie Ye, Yiqi Zhao, Muchen Xu, Celina Shiyu Wang, Hao Dong, Ruohai Ge, Rong Xue, Benjamin J Kolligs 等 15 位作者

原文摘要

We introduce $Z$-sight, a unified robot foundation model that couples vision-language-action (VLA) and generative latent world modeling. Video-based world action models (WAMs) show the value of joint future-and-action generation for control, yet inherit the heavy cost of video generators. Recent latent-space policies avoid pixel generation, yet typically align intermediate features or rely on frozen target encoders, decoupling future prediction from the VLM's native feature space. $Z$-sight instead predicts the future directly in the visual token space in which its VLM perceives the present. Coupling a foresight expert and a continuous action expert with the VLM in a mixture-of-transformers, $Z$-sight jointly denoises future spatial tokens and actions at inference without pixel reconstruction or external encoders. Trained on human video and robot trajectories, $Z$-sight achieves 81.9% success on SIMPLE (+9.6 points over the strongest baseline) and 73.3% across six Unitree G1 tasks (+20.0 points) at 110 ms per chunk, while remaining competitive on tabletop benchmarks.

cs.RO · cs.CV

新论文 · v1arXiv:2611.24938

Phoenix: Learning Scene Structure Representations for Generalizable World-Action Models

Yupeng Zheng, Chaoyue Li, Linbo Wang, Xiang Li, Yuhang Zheng, Xiaotong Su, Songen Gu, Yichen Zhang 等 18 位作者

原文摘要

World-action models jointly predict future observations and actions, yet reconstruction-oriented visual representations do not explicitly emphasize the spatial relationships and object-state changes needed for manipulation. We introduce Phoenix, which learns representations of scene structure for world-action modeling. Specifically, Phoenix Encoder enriches DINOv3's general visual semantics through dual-teacher distillation of geometric and spatiotemporal semantic knowledge. Next, CS-VAE causally aggregates and compresses Phoenix Encoder features into generative latents while preserving scene information and spatial layout. Finally, Phoenix jointly predicts these latents, complementary appearance latents, and robot actions, linking scene evolution with action generation. To study pretraining data efficiency, we curate approximately 39,000 hours of egocentric and robot manipulation data for representation learning and use the robot demonstration subset for WAM pretraining. Following benchmark-specific post-training, Phoenix achieves state-of-the-art performance across simulation and real-robot benchmarks spanning single-arm, bimanual, and dexterous manipulation. It demonstrates strong robustness to perturbations, achieving success rates of 94.8% on LIBERO-Plus and 72.4% on RoboTwin 2.0 C2R, alongside strong cross-embodiment transfer and real-world manipulation performance on Bench2Dex and RoboDojo-RealWorld, respectively. Pretraining-scale analysis shows larger performance gains from data scaling than a video-VAE-only baseline, indicating stronger scaling benefits and more effective data utilization. Moreover, Phoenix maintains high success rates when adapted with limited downstream demonstrations, demonstrating strong few-shot adaptation. Our code, pretraining data, and models are available on the project page to help the community reproduce our results.

cs.RO

版本更新 · v2arXiv:2611.15789

VibePrevisBench: Evaluating Agentic Coding for 4D Scene Generation in Blender

Yuzhe Chen, Muxi Diao, Lele Yang, Wenbin Xu, Xuyang Liu, Qile Su, Changyi Liu, Jiankang Chen 等 21 位作者

原文摘要

Controllable real-world video generation models can condition on a rendered graybox proxy video, which fixes layout, motion and camera. Coding agents can build such videos by generating a 4D scene in Blender, yet no benchmark checks both these rendered proxy videos and the complex, multi-shot scenes behind them. We introduce VibePrevisBench, a benchmark for 4D scene generation: an agent turns a brief into an executable 4D scene and renders its proxy video. Its 242 tasks span four input kinds, including scripts, photos, continuations, and reference videos. Each delivery is scored along two evaluation dimensions. Rule-based Scene Check re-runs the code and compares the exported 4D scene with hidden ground truth facts in nine dimensions. VLM-based Screen Check asks a VLM judge whether facts that pass the rules are recognizable in the proxy video. We analyze the evaluation results and trajectories of 11 frontier coding agents and identify three key observations. First, multimodal understanding separates the agents, especially in understanding space and camera motion. Second, all agents seem to already master basic camera work and staging, and fail at challenging requests. Third, agents inspect their own renders but fail to correct their errors. Agents thus need better visual understanding of space and camera, explicit checks for challenging requests, and item-by-item verification of renders against the brief.

cs.CV · cs.AI

新论文 · v1arXiv:2611.24926

Grana: Compositional Reward Modeling for Long-Horizon Robot Manipulation

Zi-ang Cao, Kaiyuan Zheng, Yinzhen Xu, Haonan Duan, Jiasheng Gu, Fengyuan Hu, Nan Zhu, Kun Lei 等 17 位作者

原文摘要

Robot reward modeling faces two challenges: generalizing across tasks and accurately tracking progress in long-horizon manipulation. General-purpose reward models demonstrate broad task transfer, but transfer from short-task pretraining to long-horizon tasks remains underexplored. Long-horizon progress models often rely on task-specific subtask classifiers or fixed action vocabularies, limiting the reuse of learned progress knowledge when adapting to new tasks. To address these challenges, we introduce Grana, a vision-language robot reward model that learns to identify unfinished subtasks and estimate their continuous progress using open-vocabulary subtask descriptions. We train a language-conditioned model to estimate progress in short-horizon tasks and individual subtasks of long-horizon tasks. For long-horizon tasks, we combine the selected subtask's predicted progress with credit for preceding completed subtasks to estimate whole-task progress. General progress prediction requires training the model across diverse tasks with progress supervision that accounts for imperfect execution. Elapsed-time proxies can misrepresent progress when relative subtask durations vary, or failures reverse local progress. We therefore introduce an annotation protocol that converts sparse subtask boundaries and action-quality annotations into dense progress targets, without requiring annotators to assign numerical progress scores. Pretraining with this supervision on 200 diverse, mostly short-horizon manipulation tasks enables zero-shot reward prediction on unseen tasks and improves progress prediction after adaptation to multi-minute, long-horizon tasks. Experiments demonstrate (i) generalizable reward prediction, including adaptation to five long-horizon tasks, (ii) benefits from pretraining and composing subtask progress estimates, and (iii) improved downstream policy performance.

cs.RO

新论文 · v1arXiv:2611.24924

Agentic Harness Engineering Beta: Requirement-Centered Evolution of Agent Harnesses

Jiahang Lin, Jiahao Chen, Yiding Wang, Shichun Liu, Chengjun Pan, Shihan Dou, Honglin Guo, Yuyang Hu 等 14 位作者

原文摘要

Real-world harness engineering is driven by explicit and implicit user requirements, which guide development and validation. However, existing harness evolution methods often overlook implicit requirements and interaction needs, while using a single evolution agent to modify the harness and benchmark scores to select changes. This process often produces narrow patches and leads to benchmark-specific local optima. We introduce Agentic Harness Engineering Beta (AHE-Beta), a requirement-centered loop that brings real-world engineering practices into harness evolution. The loop consists of three steps: (1) User-driven evaluation lets simulated users raise their requirements as issues while they work with the harness over multiple turns. (2) Team-based modification implements those requirements as generalizable changes rather than task-specific patches. (3) Requirement-driven selection merges a change only when it resolves a verified requirement and passes review. Empirically, on held-out tasks, AHE-Beta raises pass@1 from 21.9% to 33.3% while using 40.8% fewer tokens than Base Harness, and achieves higher pass@1 than Claude Code, Meta-Harness, and AHE. In a blind human A/B test, real users prefer AHE-Beta in 31 of 48 valid pairs and rate it higher on all five dimensions, matching the directional preference of the user simulator. The human A/B results show that AHE-Beta achieves its goal of aligning harness evolution with real-world user requirements. Without further evolution, the gains transfer to GLM-5.3 and DeepSeek-V4.1-Flash with pass@1 improvements of 15.8% and 5.3% over Base Harness. Together, these results show that requirement-centered harness evolution is a more general method for improving agent capabilities.

cs.SE · cs.AI

版本更新 · v2arXiv:2611.22464

MAVE: Reconstructing Editable 3D Blockouts from Film Shots

Yifan Chang, Jian Yang, Jiangnan Shao, Xiaofeng Wang, Zhaoxin Fan, Renjing Xu, Ming Lu, Xingang Wang 等 9 位作者

原文摘要

Animated 3D blockouts support previsualization and provide structural conditions for video generation, but manually reconstructing them from reference videos is costly. This limits both their use by creators and the production of paired training data. We introduce MAVEBench, a benchmark of 50 film shots and 416 keyframes for evaluating camera motion, spatial layout, and character blocking. We propose MAVE, a video-to-blockout framework combining Metric observation, Agentic modeling, and evidence-based VErification. It first recovers cameras, human motion, and geometric evidence in a shared metric coordinate system. An agent then constructs an editable environment under constraints from these observations, while programmatic checks report inconsistencies to guide refinement. The outputs are editable Blender scenes and frame-aligned blockout videos. On MAVEBench, MAVE improves overall scores by 27.5-41.1 points across three agents over their corresponding vanilla baselines. With DeepSeek-1017, it achieves a score of 73.2 at an average API cost of only USD 0.054 per shot, outperforming the strongest vanilla baseline while reducing total reconstruction time by 53%. We also construct MAVEData-3K, a dataset of 3,000 blockout-video pairs for model training, and will release it publicly. Creators can change viewpoints and character paths, then render the edited scenes to condition video generation. Code is available at https://github.com/yfChang-cv/MAVE.

cs.CV · cs.AI

新论文 · v1arXiv:2611.24891

Cap-and-Track: Optimizing LLM Inference Request Dispatch Across Shifting Capacity

Benjamin Novotny, Doug Szeto, Chang Wang, Shicheng Chen, George Zhihua Yuan, Bob Kamma, Michael Ge, John Xin Zhou 等 9 位作者

原文摘要

A fixed dispatch rate for LLM inference becomes stale as serving fleets resize and request sizes change. Generative output length is unknown at admission, and shared provider pools can lose capacity. We present Cap-and-Track, a design for asynchronous inference across provider capacity and internal GPUs. Meta's traffic shaper implements Cap-and-Track for post and live-video understanding. Refusals signal the backend's current admission ceiling. Durable priority queues hold excess work outside execution, including refused requests for retry. Cap is a fail-fast admission check at the saturated resource that refuses work before a long wait. Using only those refusals, the Track dispatcher probes upward and backs off. It follows the ceiling without output-length prediction or a fleet-level rate to retune. In a concurrent same-traffic production A/B, the shaper ran per-replica Cap plus Track. It served a classifier on 43% fewer H100s than its queued, fixed-rate precursor, though lower-priority work waited longer at backfill peaks. In steady-capacity real-engine replay, Track ran within 1% of the goodput of a fixed rate tuned above measured capacity, with about a quarter of that rate's refusal share. The shaper carries more than 300 million online requests and hundreds of terabytes of media a day. During a backfill, it handled 1.1 billion requests in its busiest 24 hours. Lessons include keeping queues inside inference engines just long enough to fill batches, steering dispatch by capacity refusals alone, keeping dispatch stateful, and sizing fleets by how long low-priority work can wait.

cs.LG

新论文 · v1arXiv:2611.24867

Choosing What to Observe: Budget-Paired Counterfactual Audits for Credible Evaluation of Visual Agents

Ruiqi Liu, Jinming Li, Mo Wang, Yang Lu, Yanlan Kang

原文摘要

Visual agents that read dense documents must ration a finite observation budget across candidate regions, yet answer accuracy conflates where an agent looks, how it adapts to the reads that remain, and how it an?swers once observation stops. We introduce budget-paired counterfactual audits, an evaluation protocol that holds public evidence and region bindings fixed while perturbing only the remaining budget or de?clared task parameters, and credits every acquired history with exact continuation values. The resulting accounting decomposes lost expected accuracy into stepwise acquisition loss and terminal decision loss, separating planning from fixed-role heuristics. Across forty paired routing anchors, five open-weight

cs.AI

新论文 · v1arXiv:2611.24864

NereidVLA: Learning to Understand and Navigate the Underwater World

Chenjie Zhao, Zhiliang Zhu, Quan Chen, Yuquan Xue, Xingxue Shi, Puyue Wang, Zhiqiang Zhao, Kangye Zheng 等 11 位作者

原文摘要

Underwater navigation requires translating visual spatial cues into motion, yet marine vision-language datasets rarely include paired robot trajectories. We present NereidVLA, a vision-language-action model that transfers marine spatial knowledge to language-conditioned 3D navigation. We first train Sea2Know on approximately two million marine examples to learn target grounding, spatial relations, and scene structure without paired robot-motion supervision. We then adapt it using expert trajectories executed under simulated underwater vehicle dynamics. Given RGB histories and a language instruction, NereidVLA predicts discrete motion tokens that decode into local 3D trajectories for closed-loop tracking. Evaluation combines an independent six-dimensional underwater navigation understanding benchmark with offline prediction and closed-loop navigation. Under matched navigation training and execution settings, marine spatial pretraining increases simulation success from 72.4% to 84.4%. Without fine-tuning on the real test environment, NereidVLA achieves 78.75% success in laboratory-pool robot trials, compared with 47.50% for InternVLA-N1. These results show that marine spatial supervision can support downstream motion learning and physical underwater navigation.

cs.CV · cs.RO

新论文 · v1arXiv:2611.24846

Atomic MathReasoner: Instance-Level Composition of Atomic Reasoning Skills for Mathematical LLMs

Jiashu Li, Tianyu Gao, Qinglong Liu, Jinping Wang, Ye Wei, Zhiqiang Gao

原文摘要

Most existing approaches enhance mathematical reasoning in large language models (LLMs) as a monolithic capability, despite the fact that different mathematical problems require different reasoning skills. Recent work has explored a more fine-grained perspective by decomposing mathematical reasoning into atomic skills and developing specialized modules for individual capabilities. However, how to effectively utilize these atomic abilities remains underexplored. To address this challenge, we propose Atomic MathReasoner (AMR), an instance-level framework for composing atomic reasoning skills. AMR first trains a library of capability-specific LoRA experts and performs gradient-free optimization over held-out validation subsets to discover a diverse set of low-loss composition prototypes. It evaluates the preference of each prototype using the token-level negative log-likelihood of reference solutions, filters ambiguous assignments through a confidence margin, and distills the resulting pseudo-labels into a lightweight router. During inference, the router selects an appropriate composition prototype for each problem, enabling problem-specific utilization of atomic skills without costly online optimization. Experiments on the AtomicMath benchmark across general and mathematics-specialized LLMs demonstrate consistent improvements. On Qwen2.5-Math-Instruct-7B, AMR increases average field accuracy from 64.4% to 68.1% at Level 1 and from 59.3% to 63.7% at Level 2, while improving average logical reasoning performance from 39.4% to 46.8%. These results demonstrate that effectively composing specialized atomic skills provides a promising approach to strengthening mathematical reasoning in LLMs.

cs.LG

版本更新 · v2arXiv:2503.01935

MultiAgentBench: Evaluating the Collaboration and Competition of LLM agents

Kunlun Zhu, Hongyi Du, Zhaochen Hong, Xiaocheng Yang, Shuyi Guo, Zhe Wang, Zhenhailong Wang, Cheng Qian 等 11 位作者

原文摘要

Large Language Models (LLMs) have shown remarkable capabilities as autonomous agents, yet existing benchmarks either focus on single-agent tasks or are confined to narrow domains, failing to capture the dynamics of multi-agent coordination and competition. In this paper, we introduce MultiAgentBench, a comprehensive benchmark designed to evaluate LLM-based multi-agent systems across diverse, interactive scenarios. Our framework measures not only task completion but also the quality of collaboration and competition using novel, milestone-based key performance indicators. Moreover, we evaluate various coordination protocols (including star, chain, tree, and graph topologies) and innovative strategies such as group discussion and cognitive planning. Notably, gpt-4o-mini reaches the average highest task score, graph structure performs the best among coordination protocols in the research scenario, and cognitive planning improves milestone achievement rates by 3%. Code and datasets are public available at https://github.com/MultiagentBench/MARBLE.

cs.MA · cs.AI · cs.CL · cs.CY

新论文 · v1arXiv:2611.24816

RoboMorpher: Scaling Robot Morphology Representations for Cross-Embodiment Action Prediction

Wenwei Lin, Yan Qin, Weiliang Deng, Zixuan Li, Kaixuan Wang, Zanxin Chen, Yue Chen, Tianshuo Yang 等 15 位作者

原文摘要

Learning shared action representations is central to cross-embodiment manipulation. However, identical joint values can produce different spatial configurations across robots, limiting representations learned from joint values alone. We introduce RoboMorpher, a morphology-conditioned action tokenizer that jointly reconstructs joint angles and link poses to associate control coordinates with their kinematic effects. MorphGen supplies scalable kinematic supervision by generating serial-arm descriptions from survey-derived geometric priors. We pretrain RoboMorpher on 40 million morphology-state pairs and integrate the frozen tokenizer into a vision-language-action policy for discrete prediction and continuous refinement. We also introduce EmbodiBench, a ten-task benchmark with 30 generated URDFs per task. With downstream URDFs excluded from tokenizer pretraining, the policy achieves 56.40% success on RoboTwin 2.0 and 23.27% on EmbodiBench, exceeding continuous joint-space control by 9.73 and 15.34 percentage points. Removing link-pose supervision reduces EmbodiBench success by 17.94 percentage points, supporting morphology-aware geometric supervision for joint-action representation learning.

cs.RO

新论文 · v1arXiv:2611.24809

On-Policy Morphing into Hybrid Attention Models

Disen Lan, Xuanda Wang, Jianbin Zheng, Huafeng Kuang, Wei Tang, Yu Cheng

原文摘要

Hybrid attention models combine full and linear attention to achieve a better quality--efficiency trade-off. Converting pretrained Transformers into such models requires identifying which layers should retain full attention under a prescribed budget. We identify two off-policy mismatches in FlashMorph's layer selection: (i) continuous branch interpolation and teacher-forced hidden states differ from discrete hybrid computation; and (ii) frozen model weights fail to account for parameter adaptation during subsequent recovery. We introduce FlashMorph2, an on-policy layer-selection method that (I) evaluates sampled discrete hybrid architectures end to end using their own hidden states, and (II) jointly optimizes layer-selection scores and model weights under the recovery objective to reflect post-adaptation performance. Experiments across Qwen3 model scales and linear-attention variants show that FlashMorph2 consistently outperforms existing layer-selection methods, improving average scores over FlashMorph by $+2.42$ points on RULER, $+1.08$ on LongBench, and $+0.76$ on recall-intensive tasks, while completing layer selection in only $0.72$--$3.57$ GPU hours across 0.6B--8B models, demonstrating high efficiency and scalability.

cs.CL

新论文 · v1arXiv:2611.24789

HINT-WAM: Hindsight-Informed Training for World Action Models

Phuong Tuan Dat, Qi Li, Xinchao Wang

原文摘要

World action models (WAMs) have shown promising generalization and data efficiency in robot manipulation by coupling action generation with predictions of future visual dynamics. Translating these capabilities into reliable task execution, however, remains a challenge for existing post-training methods. Offline supervised fine-tuning provides dense supervision on expert trajectories, but its flow-matching action expert is trained only on noised demonstration actions and never on the intermediate states produced by its own sampler; task-specific adaptation can also compromise pretrained capabilities. On-policy reinforcement learning trains on the policy's own behavior, yet sparse task rewards make exploration and credit assignment difficult, while dense reward shaping still provides scalar evaluations rather than direct action targets. We introduce \textbf{HINT-WAM}, the first on-policy self-distillation framework that supervises the action expert on its own sampling trajectories using hindsight, the future video of the demonstration, as privileged information available to the WAM itself. For each demonstration observation, the student generates its action through its own denoising trajectory, and at the states along this trajectory the same model acts as a teacher conditioned additionally on this hindsight, providing a conditional action distribution as a dense training target, while the student sees only deployment-time inputs and needs no environmental rewards or simulator interaction. To transfer this supervision to a flow-matching action expert, HINT-WAM distills the teacher into the student with an energy score, a distribution-level objective that matches the teacher's action distribution from the student's own samples rather than regressing its velocity field. The project web is available in https://tuandattt.github.io/hintwam/.

cs.RO

新论文 · v1arXiv:2611.24764

Horizontal Flow Matching for Single-Cell Spatiotemporal Dynamics

Zihan Wang, Jianzhe Li, Zhenyi Zhang, Peijie Zhou

原文摘要

Learning the spatiotemporal dynamics of cells from spatial transcriptomics requires connecting snapshots of different cells acquired in different coordinate frames. A rotation or translation of a section can appear as cell movement and distort the inferred correspondence. We introduce horizontal flow matching, which learns changes in tissue organization, molecular state, and cell abundance up to shared rigid motion. The dynamics follow a continuity--reaction equation whose action assigns no cost to shared rigid motion, and the resulting horizontal projector removes this component from spatial prediction errors during learning and from predicted velocities during generation. An equivalent endpoint problem jointly estimates correspondence and relative pose in Wasserstein--Fisher--Rao geometry and yields analytic paths whose velocity and growth serve as regression targets. Regressing these targets recovers the horizontal dynamics under ideal population conditions. Our model, HORIZON, uses an equivariant transformer to predict velocity and growth from the current tissue, so that a tissue can be generated continuously from a single initial snapshot. On a simulated tissue observed in perturbed coordinate frames, alignment improves reconstruction and projection reduces rigid drift. On four spatial transcriptomics time courses, HORIZON reconstructs tissue geometry, gene expression, and cell abundance more accurately than existing spatial and unbalanced trajectory methods, and interpolates held-out developmental stages.

cs.LG · cs.AI · q-bio.QM

新论文 · v1arXiv:2611.24763

Elysium: Simulation-Free Unbalanced Schrödinger Bridges

Zihan Wang, Jianzhe Li, Zhenyi Zhang, Peijie Zhou

原文摘要

Reconstructing cellular dynamics from snapshots is challenging because the destructive nature of single-cell measurements makes tracking the same cells over time impossible. Cellular dynamics are inherently stochastic and involve changes in population size through proliferation and apoptosis. Growth-based methods represent population change as a rate on the density, which does not assign exits and entries to individual cells. Unbalanced Schrödinger bridges model individual events, but neural solvers for killed diffusion bridges simulate trajectories during training. We introduce Elysium, a simulation-free solver that jointly learns state transition and mass change. Using a killed Brownian reference and its time reversal, Elysium constructs direct regression objectives from an optimal endpoint coupling and conditional bridge distributions, without simulating trajectories during training. We show that the minimizers of the expected training objectives recover the optimal unbalanced Schrödinger bridge. Empirically, Elysium reconstructs state transitions, population size, and event distributions on simulated benchmarks with known killed-diffusion ground truth, outperforming existing deterministic and bridge-based solvers. On single-cell time courses, it interpolates cell state distributions at unobserved time points, tracks population expansion and contraction, and captures biphasic drug-response dynamics.

cs.LG · cs.AI · q-bio.QM

版本更新 · v2arXiv:2606.27866

FlexMoE: Learning One-for-All Multi-Budget Expert Pruning for Flexible MoE Deployment

Fan Mo, Yuxuan Han, Geng Zhang, Wangbo Zhao, Yang You

原文摘要

Mixture-of-Experts (MoE) language models scale model capacity through sparse expert activation, yet deployment still requires storing and serving all experts. Moreover, deployment budgets can vary substantially across devices, users, and workloads. Existing MoE compression methods typically optimize a separate compressed model for each target budget. We introduce FlexMoE, a post-training one-for-all framework that converts a large pretrained MoE model into a multi-budget family of smaller deployable subnetworks. FlexMoE first ranks expert FFN channels by importance to construct a sliceable weight space shared across budgets, and then learns expert-wise discrete retention actions under gradually increasing cost pressure, allowing a single action-training run to produce reliable pruning masks across budgets. Moreover, only a single mid-budget fine-tune can further recover degraded performance and yields a shared recovered model that transfers to other unseen budgets, avoiding repeated per-budget recovery and separated checkpoint storage. Across tested MoE LLMs with diverse architectures and scales, FlexMoE consistently outperforms strong MoE compression baselines. In particular, on Qwen2-57B-A14B, it retains ~99.8% of base performance while pruning 50% of routed-expert parameters even without recovery fine-tuning. For deployment, pruned subnetworks deliver real memory reduction and end-to-end throughput gains, while our kernel co-design enables runtime budget switching without checkpoint reloading and delivers substantial inference acceleration.

cs.LG

新论文 · v1arXiv:2611.24738

Jeveration: System One Models Now Generate

Yuhe Liu, Qi Li, Shihua Zhang, Xinchao Wang

原文摘要

System One models such as Jev return decisions through fast, parallel prediction. In this paper, we introduce Jeveration, showing that the same decision paradigm can also generate high-quality images in a single forward pass. Jeveration builds on a frozen vision-language model, repurposing its visual-semantic knowledge for holistic image generation. The key challenge is to maintain global consistency. A single prompt admits many valid images, yet parallel predictions must converge on a coherent whole. We therefore condition all positions on a shared random state to coordinate parallel predictions, so that each draw yields a different but coherent image. Since the model tends to ignore this state under direct image supervision, we first learn a distribution over discrete image codes to guide single-pass generation. Beyond text-to-image generation, we further unify image-to-image generation and dense prediction within the same architecture, so that one checkpoint supports all three tasks, each in one forward pass. Experiments show that Jeveration generates a 512x512 image in 78 ms, up to 122x faster than existing unified models, while maintaining competitive image quality and dense prediction performance.

cs.CV

版本更新 · v2arXiv:2605.24173

Extracting Training Data from Diffusion Language Models via Infilling

Yihan Wang, N. Asokan

原文摘要

Memorization in large language models has been studied almost exclusively through prefix-conditioned extraction, a natural choice for autoregressive models. However, diffusion language models (DLMs) can denoise masked tokens at arbitrary positions. Thus, prefix-only probing reveals only one facet of memorization in DLMs and significantly underestimates the risk of training-data extraction. In order to realistically model extractability of training data in DLMs, we introduce infilling extraction, a data-extraction protocol parameterized by an arbitrary binary mask that subsumes prefix-only probing and accounts for the bidirectional inductive bias of DLMs. Instantiating it on LLaDA-8B and Dream-7B across five extraction modes, three training pipelines, and three corpora covering verbatim and partial leakage, we find that mask geometry governs extractability: edge-conditioned masks extract up to three times more verbatim sequences than prefix-conditioned ones, and bidirectional access opens channels inaccessible in autoregressive models. In particular, an adversary who obtains a PII-redacted copy of the training data can use targeted infilling to raise PII recall on fine-tuned DLMs by 2.5-4.5x over prefix-conditioned extraction, substantially narrowing the gap to a scale-matched autoregressive model and, for LLaDA on email addresses, exceeding it under greedy prefix-conditioned extraction. Tunable parameters for decoding measurably affect extraction performance, while a follow-up supervised fine-tuning stage does not eliminate the prior memorization.

cs.CL · cs.AI · cs.CR · cs.LG

新论文 · v1arXiv:2611.24731

HandWorldBench: A Unified Benchmark for Hand-Centric World Models

Zhaolu Wang, Jiacheng Xu, Lihong Wang, Shuqing Wang, Jianjie Fang, Ziyou Wang, Xvyuan Liu, Rongze Tang 等 15 位作者

原文摘要

Recent advances in world models are opening new possibilities for fine-grained embodied tasks, with hand-object interaction offering a particularly demanding setting. Modeling these interactions involves capturing how hand actions change objects and how the resulting states persist over time. Yet systematic evaluation tailored to these hand-centric capabilities remains underdeveloped. To address this gap, we introduce HandWorldBench, a unified benchmark that systematically evaluates hand-centric world models across human hands, simulated dexterous hands, and real dexterous hands along six dimensions: visual quality, hand fidelity, appearance consistency, instruction following, physical consistency, and temporal coherence. To construct the benchmark, we develop a data curation pipeline that extracts hand-object interaction clips and screens them for quality, yielding 20,500 high-quality egocentric clips from over 600 hours of footage. We then manually select 500 clips covering seven task families for evaluation. Benchmarking eleven models under a unified protocol reveals that stronger hand-pose control does not consistently translate into coherent object-state evolution and persistence or robust cross-embodiment performance, with several specialized hand-control models underperforming general-purpose text/image-to-video generators on dexterous-hand tasks.

cs.CV

版本更新 · v3arXiv:2210.01293

ThinkSum: Probabilistic reasoning over sets using large language models

Batu Ozturkler, Esmeralda S. Whitammer, Zhen Wang, Nebojsa Jojic

原文摘要

Large language models (LLMs) have a substantial capacity for high-level analogical reasoning: reproducing patterns in linear text that occur in their training data (zero-shot evaluation) or in the provided context (few-shot in-context learning). However, recent studies show that even the more advanced LLMs fail in scenarios that require reasoning over multiple objects or facts and making sequences of logical deductions. We propose a two-stage probabilistic inference paradigm, ThinkSum, which reasons over sets of objects or facts in a structured manner. In the first stage (Think - retrieval of associations), a LLM is queried in parallel over a set of phrases extracted from the prompt or an auxiliary model call. In the second stage (Sum - probabilistic inference or reasoning), the results of these queries are aggregated to make the final prediction. We demonstrate the possibilities and advantages of ThinkSum on the BIG-bench suite of LLM evaluation tasks, achieving improvements over the state of the art using GPT-family models on thirteen difficult tasks, often with far smaller model variants. We also compare and contrast ThinkSum with other proposed modifications to direct prompting of LLMs, such as variants of chain-of-thought prompting. Our results suggest that because the probabilistic inference in ThinkSum is performed outside of calls to the LLM, ThinkSum is less sensitive to prompt design, yields more interpretable predictions, and can be flexibly combined with latent variable models to extract structured knowledge from LLMs. Overall, our proposed paradigm represents a promising approach for enhancing the reasoning capabilities of LLMs.

cs.CL

新论文 · v1arXiv:2611.24723

Multi-Band Galaxy Novel-View Synthesis with Radiative Gaussian Fields

Yuzhe Wu, Jialin Zhu, Doğa Yılmaz, Daisuke Kawata, He Wang

原文摘要

We propose a novel view synthesis tool for galaxies based on Gaussian Splatting (GS). Unlike RGB data, where GS has achieved success, galaxy observations consist of multi-band signals from a massive number of bodies, contain complex patterns of cross-band interactions, and capture light propagation in highly heterogeneous media. As a result, the underlying physical process poses great challenges in model representation and optimization, which have not been investigated in existing research. To this end, we propose a physically inspired Gaussian Splatting method, GalaxyGS, which is equipped with physically meaningful components that capture both the energy emission mechanism of different astronomical bodies and the light propagation between them. In the absence of directly comparable baseline methods, we adopt a wide range of possible alternative approaches. Comparisons conclusively demonstrate that GalaxyGS is better in accuracy and physical fidelity across datasets and metrics.

cs.CV

版本更新 · v3arXiv:2110.08294

Coherence boosting: When your pretrained language model is not paying enough attention

Esmeralda S. Whitammer, Zhen Wang, Nebojsa Jojic

原文摘要

Long-range semantic coherence remains a challenge in automatic language generation and understanding. We demonstrate that large language models have insufficiently learned the effect of distant words on next-token prediction. We present coherence boosting, an inference procedure that increases a LM's focus on a long context. We show the benefits of coherence boosting with pretrained models by distributional analyses of generated ordinary text and dialog responses. It is also found that coherence boosting with state-of-the-art models for various zero-shot NLP tasks yields performance gains with no additional training.

cs.CL

新论文 · v1arXiv:2611.24706

Can Omni-modal LLMs Understand Spoken Intentions?

Zijie Xin, Tianyi Wang, Daoze Zhang, Jie Yang, Changli Tang, Qijie Wei, Fengyun Rao, Xirong Li

原文摘要

Omni-modal large language models (omni-modal LLMs) can respond to spoken requests directly through voice and video calls. Such requests rely on the scene shared by the user and the assistant, so the same utterance may call for different help or for no response. Most audio-visual benchmarks provide questions as separate text input, and interactive benchmarks mostly score a response as a whole. It thus remains unclear whether a model recognizes when the user is asking for help, and whether its reply meets the user's need once the request is recognized. We propose O-IN-Bench to evaluate how well omni-modal LLMs understand spoken intentions. A model watches a complete video with native speech and outputs one text response, or [SILENCE] when no response is needed. O-IN-Bench contains 1,000 Chinese and English samples in five categories and 16 subcategories. Contrastive pairs combine the same request with different scene conditions to test whether the response changes with the context. A sample-specific rubric judges language, intent, and content separately, and O-IN-Engine verifies each rubric against the video. Across 30 omni-modal model configurations, the best overall score is 59.6, achieved by Gemini3.7-Flash. Gemini3.7-Flash passes 74.4% of the intent judgments but only 42.6% of the content judgments. When the speech is a conversation between peers rather than a request to the assistant, even strong models often keep responding. When a reply must satisfy several people, the best score is only 25.9. These results indicate that recognizing a request and meeting its requirements are separate abilities, and O-IN-Bench measures each of them.

cs.CV

版本更新 · v2arXiv:2603.07432

How Far Does Online RL Generalize in Mobile GUI Agents? A Controlled Study with AndroidWorld-Generalization

Li Gu, Zihuan Jiang, Zhixiang Chi, Huan Liu, Ziqiang Wang, Yuanhao Yu, Linqiang Guo, Yang Wang

原文摘要

Online reinforcement learning (RL) has advanced mobile GUI agents on AndroidWorld to over 80\% success, but these systems train and evaluate in the same environment, and their training tasks are not released for auditing. How far RL generalizes beyond its training contexts therefore remains unmeasured. We formalize the mobile environment as a Contextual MDP and instantiate a ladder of three progressively distant context shifts as AndroidWorld-Generalization: Unseen Instance, Unseen Template, and Unseen Application, with separation verified manually. In addition, we provide an open RL training system that makes studies at this scale feasible. Applying the identical RL algorithm to two priors of different origin and competence, UI-TARS-7B-SFT at 23.0\% and Qwen3-VL-8B at 52.6\%, we find that RL improves both in distribution but that gains fall monotonically up the ladder (+17.5/+11.1/+4.2 and +10.3/+7.4/+0), with training performance rising throughout while evaluation performance saturates. The decay is algorithm-agnostic, reproduced under PPO, and RL contributes less as the prior strengthens. Few-shot adaptation in the target application recovers +22.9, motivating test-time adaptation as a complement to further RL training. We release the benchmark, training system, recipes, and model weights.

cs.CV · cs.CL · cs.HC · cs.LG

新论文 · v1arXiv:2611.24654

CurvDelta: Low-Curvature Delta Modeling For Surgical Learning and Unlearning

Zeguan Xiao, Zhuocheng Liang, Yun Chen, Yong Wang, Guanhua Chen, Pinyan Lu

原文摘要

A central challenge in adapting large language models (LLMs) is capability preservation: an adaptation can achieve its target while degrading general capabilities. We introduce CurvDelta, a scalable framework that incorporates the preservation constraint into the parameterization of the model update. By analyzing the local geometry of the preservation constraint around the pre-adaptation model, CurvDelta constructs a low-curvature space in which parameter updates leave the preservation loss nearly invariant. We make its construction practical for LLMs through Kronecker-factored approximate curvature (K-FAC) and a factorized projection. We also introduce CurvLoRA, a low-rank variant within the same preservation geometry, designed for more efficient adaptation. Both methods apply to preservation-constrained adaptation tasks such as capability learning and selective unlearning. We evaluate our methods on mathematical reasoning, code generation, and knowledge unlearning. For mathematical reasoning and code generation, our approaches achieve competitive downstream performance while better preserving general capabilities, offering a favorable acquisition--retention trade-off relative to the baselines. For knowledge unlearning, our main results on TOFU and MUSE show a better balance between targeted forgetting and knowledge retention than the baselines, with retention performance close to the from-scratch retraining reference.

cs.CL

新论文 · v1arXiv:2611.24644

Equal Steps, Unequal Curvature: Where Muon's Update Goes Wrong

Biswa Sengupta, Jinhua Wang, Léo Maxime Brunswic

原文摘要

Muon updates each hidden weight matrix along the polar factor of its momentum, the steepest-descent direction under the spectral norm. That choice fixes two allocations without consulting the loss: every singular direction of the momentum receives the same step, and in a non-square matrix each neuron's step is set by its leverage score. Whether they are good is a question about curvature, which steepest-descent arguments do not answer; we measure it. Using held-out gradients and Hessian-vector products, we decompose Muon's step by singular-value band and by neuron group and compute each group's second-order optimal multiplier. In a 124M-parameter decoder trained on FineWeb-Edu, the multipliers rise monotonically from the top singular directions to the tail: Muon over-steps the high-curvature top of its spectrum. We prove in Lean when such a diagnostic licenses an intervention at Muon's own update norm. On the measured curvature, the criterion predicts, and held-out losses confirm, that moving step length off the top helps mid-training and early in the cooldown, but not at its end. We then test interventions under per-step norm matching, tuned baselines and paired seeds. Orthogonalizing Q, K and V separately rather than as one fused matrix lowers the loss by 0.0250 nats and accounts for all of the gain of block-wise variants. A matmul-only, rotation-equivariant spectral tilt $σ^{-1/4}$ applied in the cooldown lowers the loss of this stronger baseline by a further 0.0120 nats, additively with NorMuon (0.0158 nats together, three seeds; the baseline needs about 4.3% more tokens to match), and by 0.0188 nats at 355M parameters (three seeds). Energy-matched Schatten-1 weight decay adds a further 0.0052 nats. Geometrically principled changes that do not act on the measured miscalibration, namely exact polar factors, steepest-descent input whitening and top-heavy decay, do not help.

cs.LG

新论文 · v1arXiv:2611.24643

AggCLIP: Aggregating CLIP for Zero-Shot Multimodal Anomaly Detection

Zhi Li, Hanyu Wang, Fei Shen, Zhengtao Zhang

原文摘要

Zero-Shot Multimodal Anomaly Detection (ZSMAD) aims to identify both appearance and geometric defects in novel categories without relying on target-specific training data, making it especially valuable in real-world scenarios constrained by data scarcity, privacy concerns, or high annotation costs. While recent approaches successfully adapt vision-language models like CLIP to achieve this open-vocabulary generalization, they still face significant challenges: (i) severe geometric information loss during 3D-to-2D point cloud rendering; (ii) spatially noisy 2D feature maps produced by raw CLIP patch embeddings, compounded by shallow late-stage modality fusion; and (iii) severe cross-modal semantic misalignment due to decoupled processing streams. To address these limitations, we propose AggCLIP, a novel framework that adapts frozen CLIP representations for ZSMAD by explicitly incorporating local geometric structure and visual spatial context. Specifically, to accurately localize fine-grained defects, we design a Geometry-Aware Spatial Aggregation (GSA) module that refines back-projected 3D point features via density-adaptive neighborhood weighting, alongside a Context-Aware Spatial Aggregation (CSA) module that smooths multi-scale 2D patch embeddings using a lightweight trainable adapter. Furthermore, to prevent semantic drift between modalities, we introduce a Dual-Stream Prompt Contrastive (DSPC) loss, which aligns the abstract concepts of normality and abnormality across the point cloud and RGB streams. Extensive experiments demonstrate that the proposed AggCLIP framework achieves state-of-the-art performance on the MVTec 3D-AD and Eyecandies benchmarks under both one-vs-rest and cross-dataset zero-shot settings, notably surpassing the strongest competitive methods by up to 2.9% in pixel-level AUPRO.

cs.CV

新论文 · v1arXiv:2611.24633

H3Remover: Reinforcing Video Object Removal without Paired Data

Kangjie Chen, Xiangyu Li, Jinhao Deng, Dongbin Zhang, Chaoda Zheng, Shijia Chen, Yuqi Zhang, Choo Sin Wai 等 13 位作者

原文摘要

Pretrained video editors possess rich removal and completion priors, yet may retain targets or alter unrelated content. We present H3Remover, which reinforces MiniMax-H3 Base Ref2VA for video object removal without paired object-free training targets. To discourage superficial edits such as recoloring, a target-removal reward favors calibrated depth changes consistent with revealing farther content. To limit unintended scene changes, a background-geometry reward encourages preservation of relative depth ordering. To preserve background appearance without retaining unwanted shadows or reflections, an appearance reward constrains RGB changes in source-derived protected regions while exempting targets and estimated effects. Frozen HPSv3 complements these constraints with perceptual-quality feedback for synthesized content. Evaluation regions remain fixed across candidates, and DiffusionART optimizes LoRA adapters with the backbone frozen. Experiments show substantial reconstruction gains over original H3, with qualitative comparisons demonstrating more effective object removal and better preservation of surrounding scene content.

cs.CV · cs.AI

版本更新 · v2arXiv:2603.17712

AERR-Nav: Adaptive Exploration-Recovery-Reminiscing Strategy for Zero-Shot Object Navigation

Jingzhi Huang, Junkai Huang, Haoyang Yang, Haoang Li, Yi Wang

原文摘要

Zero-Shot Object Navigation (ZSON) in unknown multi-floor environments presents a significant challenge. Recent methods, mostly based on semantic value greedy waypoint selection, spatial topology-enhanced memory, and Multimodal Large Language Model (MLLM) as a decision-making framework, have led to improvements. However, these architectures struggle to balance exploration and exploitation for ZSON when encountering unseen environments, especially in multi-floor settings, such as robots getting stuck at narrow intersections, endlessly wandering, or failing to find stair entrances. To overcome these challenges, we propose AERR-Nav, a Zero-Shot Object Navigation framework that dynamically adjusts its state based on the robot's environment. Specifically, AERR-Nav has the following two key advantages: (1) An Adaptive Exploration-Recovery-Reminiscing Strategy, enables robots to dynamically transition between three states, facilitating specialized responses to diverse navigation scenarios. (2) An Adaptive Exploration State featuring Fast and Slow-Thinking modes helps robots better balance exploration, exploitation, and higher-level reasoning based on evolving environmental information. Extensive experiments on the HM3D and MP3D benchmarks demonstrate that our AERR-Nav achieves state-of-the-art performance among zero-shot methods. Comprehensive ablation studies further validate the efficacy of our proposed strategy and modules.

cs.RO · cs.CV · cs.LG