NVIDIA, :, Anis Ahmad, Jon Allen, Malik Aqeel Anwar, Jesse Archer, Maciej Bala, Piotr Barejko 等 90 位作者
We present a shared framework that adapts Cosmos world foundation models for real-time interactive world simulation in Physical AI. By streaming visual observations in response to actions and scene controls, the framework enables users and embodied agents to interact with evolving virtual environments. It supports both predicting the visual consequences of actions and generating observations from physical state maintained by an external simulator. We demonstrate this approach across camera navigation and robot manipulation, controllable video transfer, and multi-view autonomous vehicle (AV) simulation. These examples provide a shared foundation for developing interactive virtual environments for Physical AI training and evaluation. Our code and model checkpoints are available at https://github.com/nvidia/cosmos and https://huggingface.co/collections/nvidia/cosmos3-examples. The project website is available at https://research.nvidia.com/labs/cosmos-lab/cosmos-simulation.
Ximo Zhu, Ruiqi Liu, Yi Han, Zhengan Wu, Gengsheng Li, ZiAn Wang, Zhu Yuantian, Jin Jiang 等 9 位作者
Proprietary language models exhibit strong reasoning and coding capabilities, but their output distributions are inaccessible, making fine-grained distillation methods that rely on teacher logits difficult to apply directly to black-box models. We observe that rich textual feedback from a black-box teacher can be treated as privileged information (PI) and converted, through a PI-conditioned privileged teacher, into dense supervision signals on the student's own trajectories. However, what PI should be acquired from the teacher to most effectively support student learning remains unclear. PI strategies that differ in where they focus, what information they request, and how much detail they provide can have substantially different effects on the current student, and their relative effectiveness can further change as training progresses. Based on this observation, we propose Evolving Privileged Information Strategies for Black-Box On-Policy Distillation (EvoOPD), which treats PI acquisition itself as an optimization problem and maintains a PI Strategy Bank that evolves with the student state. EvoOPD uses the measured repair utility of PI on the current student to adapt the focus, information type, and granularity of teacher feedback, allowing the acquired information to continually match the student's learning needs. We systematically evaluate EvoOPD on mathematical reasoning and code generation with Qwen3-1.7B, 4B, and 8B students. EvoOPD achieves the best overall performance in all six model-task settings, improving over the strongest baseline in each setting by an average of 2.14 percentage points in mathematics and 1.30 points in code. Further analysis shows that the relative effectiveness of PI acquisition strategies changes with student state, making a fixed strategy difficult to sustain as the best choice throughout distillation.
Vision-Language-Action Models (VLAs) enable robots to perform manipulation tasks by leveraging Vision-Language Models (VLMs) to process visual observations and language instructions. However, the action prediction objective itself lacks fine-grained task-specific perceptual supervision, making it prone to relying on task-irrelevant visual shortcuts, leading attention away from target objects and key interaction areas, thereby limiting its operational performance and generalization ability. To address this, we propose Automatic Perception-Guided Attention Alignment for Robot Manipulation ($A^3$-VLA), a framework that derives task-specific visual guidance from the manipulation process itself. Specifically, $A^3$-VLA first generates heuristic candidate regions for target objects and the robot from interaction-induced visual changes, and refines them with a visual segmentation model offline. These regions serve as explicit perceptual guidance through two complementary objectives: an attention allocation loss that encourages sufficient focus on task-critical regions when language or robot state tokens query the visual input, and a contrastive learning loss that aligns the representations of candidate target regions with language instructions while separating them from task-irrelevant regions. Extensive experiments on simulated and real-world manipulation tasks demonstrate that $A^3$-VLA consistently improves multiple VLA backbones, with an average success rate gain of 6.5% - 25.2% across benchmarks.
Finding high-performing formulations is difficult when few have been measured and experimental budgets are limited. We propose KolTa, an active learning framework that selects batches to discover the top 5% of formulations at a target temperature under a fixed query budget. KolTa combines a three-stage predictor of component contributions, role-conditioned interactions, and anchored temperature response with task-specific soft physics constraints. Its Physics-Residual-Calibrated Upper Confidence Bound (PRC-UCB) acquisition rule combines predicted utility, ensemble disagreement, and a target-temperature physics residual, with nonnegative exploration weights fitted to out-of-fold signed errors. We evaluate the framework on three measured tasks and one simulated task covering binary-mixture viscosity, drug solubility, ionic-liquid viscosity, and electrolyte conductivity. KolTa ranks first or second in mean area under the learning curve (AULC) on every task and outperforms seven of the eight comparator strategies in paired tests, winning 75 of 78 paired comparisons across tasks and seeds against those seven strategies. Across the three measured tasks, its largest gap to the best mean AULC is 0.032, compared with at least 0.056 for every comparator. On Drug, reaching 80% average recovery of the remaining targets across drugs requires an estimated median of 348 additional formulation queries, about 18% fewer than the fastest comparator. Researchers can adapt the framework to diverse formulation problems through component roles and physical constraints.
Xi Xiao, Yunbei Zhang, Lin Zhao, Janet Wang, Xinru Jiang, Yanshu Li, Chenrui Ma, Pan Wang 等 10 位作者
Segment Anything Model 3 (SAM 3) can segment objects from text prompts and track them across video, but its memory and computation costs make deployment difficult. Post-training quantization (PTQ) reduces these costs by using fewer bits. However, a quantized model can select the wrong object or lose track of its target across frames. We call this prompt-alignment drift. Controlled tests show that perturbing low-variance feature directions strongly affects target selection. We propose Semantics-Preserving Quantization (SPQ), which gives more calibration weight to concept-sensitive parameters, preserves residual features, and corrects the mean error in video memory. Fixed text concepts guide calibration without task annotations. All three components operate offline and add no inference operators. With a W4A8 image encoder, SPQ improves RefCOCO Pr@0.5 from 42.1 to 72.5 over PTQ4SAM while retaining 54.8 COCO mask AP. Experiments on LVIS, RefCOCO variants, and SA-V show consistent gains. A CLIP study also shows improved image-text retrieval.
Omni models answer questions about videos by jointly processing sampled frames and audio in one turn, with progress driven largely by scaling model size and omni training data. We ask whether a model needs to hear the audio directly at all, and present SleuthVLM, an agent harness built on a VLM that never hears the audio directly: speech will be transcribed into text, and the VLM will sleuth the video on its own without any predefined workflow. On four omni benchmarks with videos up to 90 minutes, SleuthVLM achieves the best average accuracy: \textbf{75.0%}, 2.6 points above Gemini 3.8 Flash, and it also tops the Video-MME-v2 leaderboard with \textbf{77.8%}; with an open-source 27B VLM, it beats all open-source omni models. We draw three findings: SleuthVLM beats a model that glances; hearing is not the bottleneck, agentic ability is; omni understanding needs better agentic ability, not just more omni data for modality alignment. So we provide a simple but effective starting point for doing omni video understanding. Omni understanding needs better agents, not just bigger models that hear.
Automatic modulation classification (AMC) underpins spectrum sensing and non-cooperative signal identification in cognitive radio, yet the two standard CPU-feasible approaches to it, a compact convolutional neural network (CNN) on raw in-phase and quadrature (I/Q) samples and a classical higher-order-cumulant feature classifier, are rarely compared against each other conditioned on signal-to-noise ratio (SNR). We show that they are, in fact, complementary: the CNN is markedly more robust at low SNR, where hand-crafted statistics collapse toward chance, while the cumulant classifier is more accurate at high SNR, where higher-order statistics cleanly separate dense constellations. We exploit this complementarity with FuseAMC, a lightweight decision-fusion model that trains a logistic-regression meta-combiner over the concatenated class-posteriors of an approximately 198k-parameter residual CNN and a gradient-boosted-tree cumulant classifier, on a validation split disjoint from test. On RadioML 2016.10a (5 seeds, 44k-frame held-out test set), FuseAMC reaches 55.8% overall accuracy, improving on the best cumulant-feature baseline by 5.1 points and on the CNN alone by 3.2 points (paired-bootstrap 95% confidence-interval lower bounds of 5.0 and 3.1 points, both positive), while matching the best classical baseline in its own strong high-SNR regime and the CNN in its strong low-SNR regime. With about 0.2 million parameters, a model footprint of a few megabytes, and roughly 3 MFLOPs per frame, the whole pipeline trains and runs on one CPU core, targeting general-purpose and embedded processors without hardware acceleration.
Hei Yi Mak, Hoang Le, Mehran Taghian Jazi, Yunke Peng, Huan Lin, Zhengzhi Tang, Zhexu Liu, Yao Wang 等 11 位作者
In this work, we study end-to-end FP4 reinforcement learning (RL) post-training for Mixture-of-Experts (MoE) large language models. Recent work on FP4 RL has introduced effective recovery recipes that restore accuracy under aggressive low-precision quantization for dense models, but we find that the same approach yields no improvement over the naive FP4 baseline in the MoE setting. We conduct an in-depth study to understand why. Our analysis identifies the root cause in the routing mechanism unique to MoE: because the router makes discrete expert-selection decisions, training-inference policy alignment becomes acutely sensitive to whether the same recovery technique is applied symmetrically on the training and inference sides. When an asymmetric recipe is used-including the dense-model approach, which corrects only the rollout-training-inference alignment deteriorates sharply, as measured by the Pearson correlation coefficient between training and inference probabilities, and router selection matching rates. Building on this finding, we propose MoEResQ, a quantization-recovery recipe for MoE RL post-training that enforces symmetry between the training- and inference-side corrections. MoEResQ restores training-inference alignment and substantially closes the accuracy gap to full precision, establishing symmetric recovery as the key principle in FP4 RL for MoE models.
Temporal-difference (TD) learning underlies many prediction and control algorithms, but can diverge when combined with off-policy learning, bootstrapping, and function approximation. TD with Regularized Corrections (TDRC) is theoretically sound and practically simple, but existing convergence theory covers only the one-step case under i.i.d. sampling, while the Huber Bellman-error formulation has not been extended to TDRC or eligibility traces. In this work, we establish convergence of constrained TDRC($λ$) with diminishing stepsizes and non-i.i.d. off-policy data, for state-dependent and history-dependent $λ$-parameters. We also propose TDRC($λ$)-Huber, which preserves the TDRC($λ$) fixed point while reducing sensitivity to large Bellman errors. For both algorithms, we apply an ODE-based weak-convergence method from stochastic approximation theory. For the nonlinear mean dynamics of TDRC($λ$)-Huber, we give an integral small-gain condition that guarantees uniqueness and global asymptotic stability of the target equilibrium. We then establish convergence of the constrained algorithm, derive computable sufficient conditions for convergence, and identify a regularization-based convergence mechanism shared by linear TDRC($λ$) and TDRC($λ$)-Huber. When the mean TD matrix is positive stable, both constrained algorithms converge under sufficiently strong regularization without requiring positive definiteness.
Qingqi Song, Yuzhou Chen, Lingqian Hu, Shenhao Wang
Various land-use models characterize land-use patterns and facilitate scenario planning. However, existing models are largely predictive, forecasting transitions based on numeric data without synthesizing new land-use maps from simple text prompts. Meanwhile, generative AI in urban systems remains unimodal, weakly controllable, and evaluated mostly by generic image quality rather than domain-specific metrics. To address these gaps, we introduce DualControlNet, a generative land-use model that jointly synthesizes high-resolution land-use maps and satellite imagery, addressing multimodality, controllability, and verification challenges. Trained on over 40,000 paired tiles from Orlando and Chicago, DualControlNet is evaluated using land-use, spatial consistency, and image quality metrics. Relative to unimodal baselines, DualControlNet improves land-use diversity consistency by 62.8% and building coverage ratio consistency by 16.3%, more than doubles cross-modal spatial consistency, and triples visual quality. Scenario-based experiments demonstrate that DualControlNet supports sustainable development through multidimensional controls, increasing density and diversity, and responding to text prompts like "high-density development." Overall, this work advances land-use modeling via a multimodal, controllable, and verifiable generative framework, lowering technical barriers and enabling intuitive exploration of urban futures.
Yu-Chao Wang, Yanhong Lu, Yingjie Victor Chen, Tim McGraw
In bespoke art commissions, laypeople know what they feel but lack the words to specify it: one participant wanted a laid-off truck driver depicted as "a ghost in his own machine" but left the medium, scale, and palette unsaid. We frame this as an articulation bottleneck at an under-served upstream stage: requirement discovery, which precedes any artist or image generator and forces the commissioner to constitute intent in the first place. We present MAIA (Multi-Agent Intent Articulation), a multi-agent system that scaffolds this stage through Socratic inquiry under a "Verification over Invention" rule, turning vague affect into a text-only brief of visual terms the user verifies. In a within-subjects study (N = 16), the full configuration produced a large, significant gain in Cognitive Support over a minimal baseline (r = 0.97, p_FDR = 0.003; LMM p_FDR < 0.001). Thematic analysis traces the same mechanism, and a validator gate structurally blocks unratified content. An exploratory blind review by three professional concept artists provides complementary descriptive evidence that the rewritten briefs were more complete and executable.
Humans possess a remarkable ability to mentally explore and replay 3D environments they have previously experienced. Inspired by this mental process, we present EvoWorld: a world model that bridges panoramic video generation with evolving 3D memory to enable spatially consistent long-horizon exploration. Given a single panoramic image as input, EvoWorld first generates future video frames by leveraging a video generator with fine-grained view control, then evolves the scene's 3D reconstruction using a feedforward plug-and-play transformer, and finally synthesizes futures by conditioning on geometric reprojections from this evolving explicit 3D memory. Unlike prior state-of-the-arts that synthesize videos only, our key insight lies in exploiting this evolving 3D reconstruction as explicit spatial guidance for the video generation process, projecting the reconstructed geometry onto target viewpoints to provide rich spatial cues that significantly enhance both visual realism and geometric consistency. To evaluate long-range exploration capabilities, we introduce the first comprehensive benchmark spanning synthetic outdoor environments, Habitat indoor scenes, and challenging real-world scenarios, with particular emphasis on loop-closure detection and spatial coherence over extended trajectories. Extensive experiments demonstrate that our evolving 3D memory substantially improves visual fidelity and maintains spatial scene coherence compared to existing approaches, representing a significant advance toward long-horizon spatially consistent world modeling.
Leyao Wang, Xutao Mao, Qinwen Ge, Xuhui Zhan, Yuying Zhao, Bo Ni, Ryan A. Rossi, Nesreen K. Ahmed 等 9 位作者
Textual reviews enrich recommender systems with fine-grained preference signals and enhanced explainability. However, in real-world scenarios, users rarely leave reviews, resulting in severe sparsity that undermines the effectiveness of review-aware recommenders under partial observability. A natural remedy is to synthesize the missing reviews to restore the textual signal these recommenders rely on. Yet, conventional imputation techniques, such as matrix completion and LLM-based augmentation, either lose contextualized semantics by embedding texts into vectors or overlook structural dependencies on the user-item graph. To address these shortcomings, we propose TWISTER (ToWards Imputation on Sparsity with Textual Edge Graph Representation), a unified framework for sparse-text recommendation that synthesizes missing reviews by jointly modeling semantic and structural signals. Specifically, we represent user-item interactions as a Textual-Edge Graph (TEG), treating reviews as edge attributes. To capture relational context, we construct line-graph views and employ an LLM as a graph-aware aggregator. For each interaction lacking a textual review, our model aggregates natural-language representations from neighboring interactions to generate a coherent and personalized review. Experiments on multiple datasets show that TWISTER consistently outperforms traditional numeric, graph-based, and LLM baselines, delivering higher-quality imputed reviews and, more importantly, improved downstream recommendation. In summary, TWISTER is a drop-in, structure-aware imputer that recovers the textual signal that review-aware recommenders need when reviews are scarce.
Real-world vision is dynamic and scene-rich, while EEG and fMRI resolve it at complementary temporal and spatial scales. Most neural decoding systems, however, align one modality in isolation and mostly against static images or short clips. We study the synchronized case: EEG and fMRI are recorded during the same long-form naturalistic viewing, giving a shared stimulus but also recurring narrative context and many semantically similar candidates. We present BrainBind, a framework that maps synchronized EEG and fMRI into frozen long-form video embedding spaces and treats those spaces as common visual-semantic references, not literal neural codes. On CineBrain, a five-seed Recall@K 2 x 4 matrix across two video anchors and four system configurations shows distinct retrieval profiles. With Omni, fMRI organizes the full candidate list most strongly (MRR = 23.82, nDCG@10 = 25.31), while Joint-Dual gives the highest mean label-aware R@1 (13.52%) with a distinct rank-wide profile. The complementary episode-disjoint evaluation quantifies the contribution of repeated stimulus identity and narrative context to the random split. Across these experiments, proxy choice and system configuration produce distinct top-rank and rank-wide retrieval profiles within the present synchronized dataset. The code is available at https://github.com/Forainest789/Brainbind.
Di Wu, Jingyu Liu, Xuechao Wang, Jian Liu, Yingjie Xue, Kui Ren, Chun Chen
Cross-chain transactions today remain slow, costly, and fragmented. Existing custodial exchanges expose users to counterparty and centralization risks, while non-custodial liquidity bridges suffer from capital inefficiency and slow settlement; critically, neither approach guarantees users a unilateral path to recover assets if the infrastructure fails. We introduce the Delegated Ownership Transfer (DOT) paradigm, which decouples key ownership from value ownership to enable secure, high-performance cross-chain payments. In DOT, a user deposits funds into a sandboxed on-chain Temporary Account (TA) (value ownership) while delegating its private key (key ownership) to an abstract Trusted Entity (TE). Payments and swaps are thus reframed as near-instant, off-chain ownership handoffs. Security follows from dual guarantees: the TE's exclusive control prevents double-spending, while a pre-signed, unilateral recovery transaction ensures users retain ultimate authority over their assets. Building on this foundation, we design a novel off-chain atomic swap that executes optimistically in near real time and remains fair even if the TE fails. We formalize the security of DOT in the Universal Composability framework and present two concrete instantiations: a high-performance design based on Trusted Execution Environments (TEEs) and a cryptographically robust variant leveraging threshold cryptography. Our geo-distributed prototype shows that cross-chain payments complete in under 16.70 ms and atomic swaps in under 33.09 ms, with costs fully decoupled from Layer-1 gas fees. These results provide a practical blueprint for building secure, efficient, and interoperable cross-chain payment systems.
AI 速览：作者提出首个面向多模态有状态交互的大规模多轮对话数据集 YANAMI-400K 与评测基准 YANAMI-Bench，并报告微调 BAGEL 后在图像编辑、生成与有状态交互上均有提升。
Unified Multimodal Models (UMMs) integrate visual understanding, generation, and editing within a single unified framework. However, existing unified multimodal datasets and benchmarks remain dominated by single-turn tasks. Even in recent multi-turn datasets, many turns can be solved independently without tracking information across turns. Consequently, they provide limited supervision and evaluation for multimodal stateful interaction, which requires models to understand, maintain, update, and faithfully realize evolving dialogue states across turns. To address this gap, we introduce YANAMI, the first large-scale multimodal dialogue dataset and benchmark for stateful interaction, comprising YANAMI-400K and YANAMI-Bench. YANAMI-400K contains 400K multi-turn multimodal dialogues, with more than 2M images and 2.6M dialogue turns. The dialogues span four representative scenarios: Multi-Image Fusion, Image Editing, World Modeling, and Visual Reasoning, and require models to continuously track, preserve, and revise evolving multimodal dialogue states across turns. YANAMI-Bench comprises 400 multi-turn dialogues for systematically evaluating multimodal stateful interaction. It jointly evaluates Dialogue State Understanding and Visual Realization, characterizing the gap between models' ability to understand evolving dialogue states and their ability to faithfully realize them in visual outputs. Experiments show that fine-tuning BAGEL on YANAMI-400K improves image editing, generation, and multimodal stateful interaction by 10.3%, 2.0%, and 13.23%, respectively, while preserving its general visual understanding capabilities. Evaluations on YANAMI-Bench further reveal that existing UMMs still struggle to maintain, update, and visually realize evolving multimodal dialogue states.
Large language models are deployed as long-horizon agents that compose diverse tools across MCP services into executable workflows. Existing MCP environment-scaling methods are data-centric: they first synthesize a database and scenario, then derive tools and tasks from them. As a result, tool interfaces may not match real ones, tool compositions are not grounded in execution, and tasks rarely span more than one environment. We propose ToolSpace, a tool-centric framework that reverses this order. ToolSpace mines data, state, and retry dependencies from adjacent calls in execution traces of real MCP services and open agent corpora, using deterministic rules rather than LLM judgment, and clusters the resulting weighted tool-composition graph into environments. Databases are synthesized only for state-mutating environments, while query and stateless tools replay recorded returns. ToolSpace then performs state-aware random walks with rollback and decompiles each chain into a natural-language task with assertion-based checks, keeping only samples that replay successfully from the initial state. Applied to AutomationBench and tau3-Bench Banking, the pipeline produces harder Pro versions on which 7 frontier models drop by up to 38.2 points. Across seven long-horizon tool-use benchmarks, ToolSpace raises Qwen3.6-35B-A3B from 23.0 to 41.0 average score (43.2 with RL) and GLM-5.2 from 44.6 to 56.6, on par with Claude Opus 5. Ablations show that execution-witnessed edges clearly outperform LLM-judged ones, and performance keeps improving with data and environment scale.
Zhengdong Hu, Chao Wang, Kaituo Feng, Huanjin Yao, Mingfei Lu, Hehe Fan, Yi Yang
Multimodal large language models (MLLMs) can now process hour-long videos, making extended multi-turn dialogue over a single video a natural mode of use. Sustaining such a dialogue requires more than answering each question in isolation: the model must retrieve and reuse information established earlier in the conversation, both what it has seen and what it has said, an ability we term in-context video memory. Existing video QA benchmarks leave this ability unmeasured: their questions are mutually independent, never reference one another, and each is answerable from the video alone, so nothing ever requires carrying information across turns. We introduce VideoMemBench, in which each sample pairs a single video with a long multi-turn dialogue, whose turns are connected by typed cross-turn dependencies: an Episodic Dependency refers back to a visual episode described in an earlier turn, probing memory of what the model has seen, while an Answer Dependency builds on the model's own earlier answer, probing memory of what the model has said. Every dependency is verified considering two properties: 1) necessity, meaning the question cannot be answered without its earlier turns, and 2) non-leakage, meaning the question does not restate the information from its earlier turns. To enforce these properties, we design a construction pipeline with three LLM roles: a planner, a generator, and a judger, with each dialogue reviewed by human experts. VideoMemBench comprises 326 videos with 2,476 turns, of which 1,387 depend on earlier turns, including multi-parent dependencies and long-range call-back. Among over 20 evaluated open- and closed-source MLLMs, Claude Opus 5 performs the best with an overall score of only 45.7%, highlighting the challenge of in-context video memory. Our dataset and evaluation are released.
Video generation has progressed from isolated clips to coherent short-form audiovisual sequences, and the bottleneck is shifting from rendering to semantic control. This report summarizes the observations and insights on multimodal understanding gained from successive generations of Kling. We argue that generation quality is the projection of multimodal understanding through the entire data--conditioning--alignment--evaluation chain, and we examine this chain module by module. Kling-MLLM, a shared multimodal foundation model that unifies image, video, audio, synchronized audio--video, and interleaved inputs, is specialized for each downstream module. Because a DiT is trained on far fewer training samples than a language model, we treat captioning as a semantic contract that maximizes description density, and extend it to reference-grounded captioning that makes reference bindings explicit. Prompt enhancement translates user intent into this caption distribution: basic PE is derived from captions of compelling videos and refined with text-level and generator-in-the-loop reinforcement learning, while agentic PE refines the story and retrieves compelling examples. We show that a reliable text encoder comes from perceptual training rather than from stronger language ability, and that candidates can be screened from text alone before end-to-end training. Reward modeling pairs a general reward model for explainable, localized critiques with a decision reward model for fast multidimensional scoring, treating perceptual bandwidth as a first-class design parameter. Finally, we review task-specific automatic evaluation, focusing on protocols that expose scoring shortcuts and help trace failures to their upstream source. We hope these lessons offer practical guidance for building the understanding stack behind controllable video generation.
User-generated videos, including opinion vlogs and video reviews, provide a rich source of people's attitudes toward products, services, and social topics. Understanding the sentiment expressed by speakers is important for analyzing online opinions, as transcripts alone may not capture the attitude and intensity conveyed through vocal tone and facial expressions. Multimodal sentiment analysis (MSA) addresses this need by integrating textual, acoustic, and visual cues. However, these cues are often subtle, noisy, and inconsistent across modalities. Existing local affective cue modeling methods typically operate within individual modalities and overlook global contexts from other modalities. This may cause cross-modal affective cue inconsistency. Hierarchical global-local fusion methods integrate global semantics with fine-grained local affective cues. However, they commonly determine the relative contributions of the two representations implicitly, preventing the global-local contributions from adapting to the sample-dependent reliability of local cues and resulting in either insufficient use of informative local cues or excessive interference from unreliable ones. To address these issues, we propose DGC-Net, a Dynamic Gating and Global-Guided Compensation Network for MSA. DGC-Net mitigates cross-modal affective cue inconsistency through cross-modal contribution-aware dynamic expert routing. This mechanism uses auxiliary global contexts to guide local expert selection. Moreover, global-guided dynamic compensation fusion alleviates sample-dependent global-local contribution imbalance while preserving global semantic as the primary anchor and adaptively incorporating expert-refined local cues. Experiments on five multimodal benchmark datasets demonstrate the effectiveness and general applicability of DGC-Net, and ablation studies verify the contribution of each proposed component.
Post-training updates change what large language models (LLMs) can answer, but may leave their answerability judgments misaligned. We study this knowledge-judgment desynchronization through functional feeling of knowing (FoK), which supports selective generation, retrieval triggering, and model routing without implying subjective experience. We propose two trainable channels: a knowledge channel that changes answerability and a FoK channel that judges the current knowledge state. Three stages train them: on-policy supervised fine-tuning (OP-SFT) updates knowledge from corrected on-policy answers; knowledge-delta FoK training (K$Δ$-FT) adapts judgments across intermediate knowledge states; and FoK-guided policy optimization (FGPO) trains the subsequent answer or explanation. Factual-QA experiments, cross-dataset evaluation, and three sequential updates show that explicit FoK alignment improves judgment quality on updated knowledge. MMLU and ARC-Easy separately assess capability preservation.
Wentao Wang, Shuo Shao, Ruibin Li, Jingjing Qian, Jun Chen, Bin Xia
The distortion-rate-common randomness-perception (DRCP) theory characterizes how limited common randomness (CR) improves reconstruction fidelity under perceptual constraints, thereby providing an information-theoretic foundation for generative compression. In this paper, we characterize the DRCP tradeoff for vector Gaussian sources with limited CR under the squared-error distortion measure and the squared Wasserstein-2 perception measure. Specifically, we first establish a generalized reverse waterfilling characterization of globally optimal allocations. Second, we derive an analytical upper bound on the DRCP function through sequential greedy allocation of rate and CR. For fixed rate and CR budgets, the resulting representation is independent of the perception threshold. Finally, we establish its approximate universality through uniform performance guarantees: the achievable distortion exceeds the optimal distortion by at most $2ρ^*\approx0.2897$ times the total source variance, whereas the optimal-to-greedy distortion ratio is at least $η^*\approx0.7517$.
Indirect Prompt Injection (IPI) poses a critical threat to LLM agents. Existing defenses largely focus on whether an injection induces a new tool invocation or unintended action. In this paper, we study a different attack surface, parameter poisoning, where the attacker preserves the intended action but manipulates one or more tool arguments, such as a recipient or other task-critical value. Such attacks are difficult to detect because no new action is introduced, while parameter changes themselves are often legitimate parts of the user's request. Our key insight is that the legitimacy of a parameter change should be determined by the trust of the information flow that produces it under the user query. The query induces a task-specific trust hierarchy over available information fields: some are trusted to determine particular parameters, while others should have little or no influence. Based on this insight, we propose AuthorityFlow, a three-stage parameter-level defense. First, it interprets the user query to construct a task-specific trust hierarchy over the information fields available to the agent. Second, for each tool argument, it traces the information flow that produces its current value and attributes the value to its supporting field. Finally, it accepts the value only when that field is trusted to determine the parameter under the current task, allowing high-trust changes while flagging low-trust influence as potential parameter poisoning. We further introduce FactAuthBench for evaluating parameter poisoning in LLM agents. Without defense, the attack success rate reaches 93.06%. AuthorityFlow reduces it to 9.03%, while retaining 82.64% benign task utility compared with 90.28% without defense. Our results show that effective IPI defense requires reasoning not only about what action an agent takes, but also about which information is trusted to determine its parameters.
Ziyang Zeng, Tingyang Chen, Jia Deng, Qiyao Wang, Shuo Tang, Yike Yang, Tao Ren, Wayne Xin Zhao 等 13 位作者
Scaling supervised fine-tuning for working agents requires not only broad coverage of professional tasks, but also expert demonstrations that faithfully carry out professional workflows to completion. However, missing or inconsistent task information can disrupt expert demonstrations, weakening the resulting supervision. We introduce WorkFactory, a scalable task-construction framework that turns professional intents into executable task environments through hybrid intent sourcing, semantic workflow planning, and execution-based repair. Starting from these intents, WorkFactory constructs a Semantic Workflow Graph that explicitly represents information requirements, dependencies, and locations within the task environment. The graph guides the instantiation of task environments, while an adversarial execution-refinement loop exposes missing or inconsistent information and steers targeted repairs that preserve the original intent. Using WorkFactory, we construct 112,686 tasks spanning 612 occupations and train the WorkPilot family at 4B, 27B, and 35B-A3B scales on the resulting teacher trajectories. All three models improve over their respective base models across four professional-work benchmarks, with WorkPilot-35B-A3B rising from 879 to 1565 Elo on GDPval-AA v2.1. Scaling experiments with up to 110k tasks further show that larger models benefit more from additional task data but exhibit diminishing returns at smaller data scales, while WorkPilot's long-horizon execution evolves from verbose trajectories toward more efficient task completion. Controlled component studies quantify the negative effects of task-environment defects on downstream performance and show how semantic workflow planning and execution-based repair yield downstream gains through improved task-environment quality. The subset of data are available at https://github.com/IQuestLab/WorkFactory.
Video multimodal large language models achieve strong results on existing benchmarks, but answer accuracy alone does not establish whether they can locate the evidence needed to answer a question. We introduce VideoZeroBench, a challenging long-video benchmark with manually annotated question-answer pairs spanning 13 video domains. Questions target fine-grained cues, fleeting events, and evidence distributed across multiple segments. Temporal intervals and key-frame boxes are annotated where applicable. All questions undergo two rounds of cross-verification for answer validity and evidence quality. Our five-level diagnostic protocol compares answering with and without evidence hints, then combines answer correctness with independently evaluated temporal and spatial grounding. Across 19 evaluated models, the best standard QA accuracy is 24.8% (Level-3), achieved by Gemini-3.7-Flash. No model exceeds 1.8% when correct answers and accurate spatio-temporal localization are jointly required (Level-5). Analyses of atomic abilities, evidence spans, input modalities, and thinking-with-videos inference further characterize where the evaluated systems struggle. These findings motivate more precise evidence search and localization for long-video question answering. Our code and data are publicly released.
Jie Liu, Xusheng Tang, Chao Wang, Zhengxiao Yan, Yuchen Dai, Qi Zhang, Conghe Song
High-resolution land-cover maps support ecological monitoring, but consistent historical time series remain scarce because aerial imagery varies across years and sensors. We developed a dual-track workflow to produce temporally coherent 1-m land-cover maps from National Agricultural Imagery Program (NAIP) imagery without historical training samples. For higher-quality imagery from 2009-2017, we used a pretrained U-Net/ResNet-34 model and refined boundaries with the Segment Anything Model (SAM). For lower-quality imagery from 2004-2008, we reconstructed land cover through an NLCD-based spatiotemporal bridging and label back-casting pipeline. Evaluation across six counties in North Carolina and Pennsylvania yielded overall accuracies of 0.887 (2014), 0.886 (2017), and 0.733 (2004). Structure, Water, Wetland, and Cropland consistently achieved high F1-scores. The workflow integrates GeoAI and multiple data sources to address temporal heterogeneity in historical aerial imagery, supporting high-resolution land-change analysis and landscape planning across the conterminous United States.
Planning has been identified as an important factor for improved code generation with Large Language Models (LLMs). However, the interaction between planning capability and coding capability remains largely underexplored, especially when the two capabilities are from different LLMs. In this paper, we conduct the first comprehensive empirical study of planning-coding interaction in LLM-based code generation, where planning and coding are assigned to different LLMs (i.e., planning-coding separation). Specifically, we explore different planner-coder combinations to investigate the effectiveness for coders to utilize planning information from various planners and the impact of further training of coders on planner-coder combinations. Our results demonstrate that almost all planner-coder combinations are as competitive as using the same LLM for planning and coding, and the combination of a stronger planner and a stronger coder typically achieves improved effectiveness. Furthermore, using a large model as the planner paired with a much smaller model as the coder can sometimes be as competitive as using the large model as both the planner and the coder, and further training of the coder can typically boost various planner-coder combinations. These findings provide empirical evidence that planning and coding can be separated to a considerable extent in LLM-based code generation, which may offer a more flexible way to construct code-generation systems.
Kangning Zhang, Wenxiang Jiao, Bowen Fang, Shuai Shao, Jiachen Zhu, Rong Shan, Shijian Wang, Hao Wang 等 11 位作者
Harness evolution from a common base exhibits Locality, Complementarity, and Nonstationarity: variants repair different failures, and their utility changes as the policy learns. Rather than seeking a universally superior harness, we introduce Harness-S to selectively rescue failures of a base execution system through sibling harnesses and transfer the resulting capabilities into a single policy that remains effective under the original harness. Directional Harness Evolution develops specialized sibling harnesses along distinct repair directions, forming a reusable Sibling Harness Pool that remains fixed during policy training. Complementary Harness Rescue Training invokes this pool only when all initial attempts under the base harness fail. The same policy generates fresh rollouts under selected auxiliary harnesses, and these trajectories replace the failed group for reinforcement learning. An Adaptive Harness Router selects the Top-k variants using Harness Behavior Profiles and voting across multiple LLMs. Policy-aware profile re-anchoring refreshes this evidence as the policy learns. Across AppWorld, SWE-Gym, and DBBench, the trained policy substantially outperforms policy-only training when returned to the original base harness. Behavioral analyses and held-out cases support the transfer of harness-elicited behaviors into the policy itself. Reusing the router for Rank-1 inference provides additional gains at the same executor rollout count.
Gia Phong Tran, Thai Duong Huynh, Randolph Loh, Balachandar Gowrisankar, Zihao Wang, Moe Thu Kyaw, Jun Sun, Adams Wai-Kin Kong 等 9 位作者
Artificial Intelligence (AI) has seen significant developments and increased usage in recent years. AI benefits various tasks, automates decision-making, and enhances efficiency. However, AI also introduces additional attack surfaces for threat actors to perform attacks like prompt injection and jailbreaking, making security a great concern. AI security testing aims to imitate adversarial attacks or malicious inputs via manual testing or with the assistance of tools. Thus, it is critical that organizations are aware of the capabilities and limitations of the tools before using them. Existing studies provide insights on the overall architecture of AI security testing tools and raise concerns on their reliability, but their analysis is limited to attack performance. Selection of a tool for AI risk assessments should also depend on their alignment with the recommendations of industry standards, but to our knowledge, previous studies have not utilized such guidelines. To bridge this gap, we analyze several industry standards and guidelines to identify the requirements of AI testing tools. We discuss and highlight desirable features across various AI security testing tools for a variety of different AI systems. We evaluated six tools across six Large Language Models, reported on our insights, and identified gaps that should be considered by future tools.
Low-rank adaptation (LoRA) has become a widely used parameter-efficient fine-tuning method for large pretrained models. However, standard LoRA typically assigns a fixed rank profile despite layer- and stage-dependent adaptation needs, while recent adaptive-rank variants often rely on heuristic or first-order signals that provide only an indirect link between rank reallocation and local loss behavior. This leaves the rationale behind each rank reallocation insufficiently interpretable. Motivated by this gap, we propose LIRA (Loss-Informed Rank Allocation), a dynamic-rank LoRA method designed to make online rank reallocation more interpretable. LIRA scores each layer by the estimated magnitude of the local loss perturbation associated with its recent LoRA-factor movements under a second-order approximation, making the basis for each rank change interpretable term by term. By periodically applying DELTA (Displacement-Evaluated Loss-Trajectory Allocation), a displacement-based algorithm for LoRA rank reallocation, LIRA produces discrete effective ranks under matched budgets while remaining compatible with standard (A,B) adapters. On RoBERTa-base GLUE under the La-LoRA/LoRA protocol, LIRA outperforms the uniform-LoRA and adaptive-rank baselines reported under this protocol; Trajectory EMA attains performance competitive with Fixed-point Fisher at lower estimation cost.