Elucidating the Design Space of LM Data Attribution
Steven Yu, Sheng-Yu Wang, Kimia Hamidieh, Antonio Torralba, David Bau
AI 速览：作者提出可控归因基准 CATT，通过合成虚构知识隔离数据集属性，并发现负对数几率梯度与仅聚合关键 token 能显著提升事实溯源归因效果。
Data attribution methods for Language Models (LMs) are typically evaluated using downstream tasks like fact retrieval. However, existing evaluations vary multiple dataset properties simultaneously, making it impossible to isolate exactly why an attribution algorithm fails. To diagnose these failures, we explore the design space of LLM data attribution, with a primary focus on evaluation. We introduce the Controlled Attribution Benchmark (CATT) for fact tracing, in which we fine-tune LLMs on entirely fictitious knowledge, providing unambiguous ground truth for attribution. This synthetic setup allows us to systematically alter one dataset configuration at a time, such as document redundancy and lexical distractors. With this configurable evaluation, we demonstrate that redundancy and distractors consistently degrade attribution performance. Rather than proposing a new estimator, we examine two design choices that are underexplored: the loss used to compute gradients and how per-token attribution scores are aggregated into a document score. We find that the true causal evidence within a training document is highly localized, but standard methods dilute this evidence into noise by averaging attribution scores across the entire document. We find that replacing standard cross-entropy loss gradients with the negative log-odds gradients surfaces the critical tokens from other uninformative ones. We can then aggregate only the top token scores to improve ground truth recall by 35 to 72 percentage points across multiple baselines on our diagnostic benchmark. The negative log-odds correction also generalizes to FTRACE-TREX, improving counterfactual retraining performance for every method we test, by up to 4.9x.
Shuqi Zhao, Boyu Wang, Shuangyue Niu, Chenhao Zhang, Zhanzhan Zhao
Rapid advances in large language models (LLMs) have enabled autonomous agents to interact, coordinate, and pursue individual goals in artificial societies. However, our understanding of how reward structures shape their competitive behavior and collective outcomes remains incomplete. We simulate 200 LLM agents competing over 50 rounds to maximize individual wealth. We define overcompetition and lying flat as population investment above and below a Bayesian adaptive baseline, respectively. At the lowest tested reward, populations lie flat; at the two highest, they overcompete, most strongly at intermediate winner coverage. The investment committed to pursuing an uncertain reward constitutes the price of a chance. Agents near the qualification cutoff can invest heavily yet receive no prize, finishing poorer than nonwinners who invest less. Agents' stated rationales link investment decisions to perceived qualification prospects and expected returns. At the collective level, greater investment does not consistently improve ability selection. With a fixed total prize budget, offering smaller prizes to more winners can reduce investment while preserving selection quality relative to the corresponding top-ability target. Further broadening reward coverage yields additional savings but reduces selection quality. These findings show how reward design shapes overcompetition and lying flat, distributes the costs of pursuing opportunity, and influences what competition collectively achieves.
Chengzhi Yi, Shenao Wang, Yanjie Zhao, Xiao Cheng, Haoyu Wang
Recent work shows that LLMs can synthesize static-analysis checkers from security patches, but a compilable checker is not necessarily a correct one. A patch reveals what changed, not why the change is safe, so generated checkers often match only the surface edit and miss the deeper semantic condition. Existing refinement loops feed back compilation errors and validation reports, yet these signals describe symptoms without explaining what the checker misunderstands about the code. We observe that the analysis framework itself already computes the information needed to close this gap. Control-flow graphs, symbolic values, memory regions and checker callbacks are available inside the analyzer but unused by current refinement strategies. SemWeaver is a plug-in refinement layer that extracts patch-relevant analyzer evidence and assembles it with source context and validation feedback into a structured, provenance-labelled bundle. An iterative editing loop refines the checker on both the vulnerable and fixed revisions, guarded by a retention gate that accepts warning recovery or false-positive reduction while preventing over-pruning. We evaluate SemWeaver on 39 KNighter-derived CSA checkers from Linux kernel fixes. With analyzer evidence, refinement reaches 23 warning-positive vulnerable revisions (vs. 15 for KNighter's loop) and raises version-level $F_1$ from 0.448 to 0.575. An ablation removing internal evidence recovers fewer silent cases and produces different checker edits, indicating that the two evidence configurations address complementary recognition gaps. Additional experiments cover three editor models on a 12-subject subset, a CodeQL query refinement, and controlled source-form variations. A qualitative follow-up on all 11 non-tied comparisons traces each warning-count change to a specific checker repair, such as recognizing a lowered callback or unwrapping an expression cast.
Agentic image editors decompose complex requests into steps, but when a step fails, they mostly retry with the same narrow set of tools and a judge that scores each output by eye directly. We argue that such agents are bounded by two spaces: the action space of what the agent can do, and the observation space of what its judge can see. On our expert-annotated benchmark, a general-purpose editor yields no usable result in over 40% of the cases, and no judge is reliable across edit types. We present EditAgent, which expands both spaces. It routes each step to general editors and specialized methods, and attaches to each candidate a verification record. With these records, off-the-shelf VLM judges outperform finetuned reward models. Candidates are generated in parallel and judged in one listwise call: experts prefer this to sequential refinement in 78.5% of decided comparisons, at no more than half the judge calls. EditAgent improves two open-weight editors on three benchmarks and serves as a data engine for training stronger editors.
Intermediate-fusion cooperative perception lets a vehicle detect objects beyond its own field of view by fusing features shared by nearby agents. When only a collaborator's cameras are degraded by fog, motion blur or exposure errors, the features it shares drift away from the training distribution while the ego input stays clean, and cooperation can become worse than not cooperating at all. Test-time batch-normalization adaptation, the standard training-free remedy, estimates statistics over the pooled batch of all agents and therefore normalizes both the clean ego and the corrupted collaborator with a mixture of two distributions. We show that this mixing harms even the clean ego: when the ego shares BN statistics with a corrupted collaborator, its detection from its own features drops by 3.58 to 6.68 AP@0.5 points although its cameras are untouched. We propose AgentNorm, an agent-wise test-time normalization that i) estimates the normalization statistics of every collaborator from its own current observation and ii) anchors the ego to its source statistics, so that the ego path remains exactly the original model. AgentNorm has no learnable parameters, needs no gradients or labels, is independent of the order of test frames, and runs locally on each agent without changing the transmitted features. On camera-based OPV2V with four fusion checkpoints and collaborator-only corruptions, AgentNorm raises AP@0.5 under severity-5 fog from 29.42-37.25 to 43.32-58.86 and outperforms pooled normalization by 5.53-12.94 points under fog and motion blur, at an extra latency of 2.3 ms per frame. On the real-world DAIR-V2X-C dataset, it outperforms pooled normalization in all ten infrastructure-corruption settings. We also report the cases it does not fix: occlusion, which removes image content rather than shifting feature statistics, and the case in which all agents share the same shift.
Haotian Lyu, Haoyu Wang, Ru Wang, Feng Ji, Ziwen Li, Jiaxian Guo, Mingming Gong
Label-free test-time reinforcement learning adapts vision--language models using rewards derived from their own predictions. However, misinterpreted visual evidence and sensitivity to textual formulation can produce a consistent but incorrect consensus that majority voting reinforces. Cross-view aggregation can expose such errors, but image and text transformations may themselves degrade task-relevant evidence to different extents. We introduce \method, a reliability-aware harmonic consensus framework for multimodal test-time reinforcement learning. \method constructs a $2\times2$ grid combining original and perturbed images with direct-solving and reframing modes, selects pseudo-labels through weighted harmonic aggregation, and adapts modality-level weights using batch-averaged, temporally smoothed shifts in pseudo-label confidence as a heuristic reliability signal. We formulate consensus as a weighted $χ^2$-barycenter problem whose unique minimizer is the normalized weighted harmonic aggregate, without requiring correct-answer invariance. An exact support--dispersion decomposition characterizes when harmonic aggregation preserves or corrects arithmetic rankings, while a sensitivity analysis establishes how modality reweighting changes pairwise ranking margins, informing interpretation of the heuristic updates. Across six benchmarks and three Qwen3-VL model scales, \method outperforms TTRV in 16 of 18 benchmark--model settings on adaptation instances and 17 of 18 settings on held-out instances under matched estimation and optimization rollout counts. These results demonstrate improved target-task adaptation and transfer to unseen instances without further updates.
Spatial transcriptomics (ST) enables joint characterization of tissue morphology and gene expression. However, 3D volumetric ST acquisition remains expensive. AI-driven methods offer a low-cost alternative but remain limited in accuracy, while recent 2D section-wise ST acquisition still rely on dense grid sampling, resulting in redundant measurements and undersampling of biologically informative regions. To bridge the gap, we propose STRATA, a reinforcement-driven active sampling framework with tri-paradigm adaptation for reconstructing dense 3D transcriptomic maps from sparse voxel-wise ST acquisition. STRATA constructs a 3D graph integrating H&E morphology with single-cell priors and employs a reinforcement learning agent to estimate morphological sparsity and transcriptomic informativeness, enabling budget-aware selection of informative spots. Given these sparse measurements, an adaptive tri-paradigm predictor combines retrieval, regression, and cross-layer imputation through lightweight fusion, dynamically weighting their contributions. Experiments across datasets demonstrate improved sampling efficiency and accuracy, particularly under low-budget settings. Code is available at https://github.com/Junchao-Zhu/STRATA.
Proactive agents must decide whether, when, and how far to assist without an explicit request. Existing evaluations often rely on automated scores or binary verdicts, offering limited human diagnosis of evidence use and intervention quality. To address these gaps, we introduce HALO, Human-Grounded GrAnular EvaLuation of PrOactive Agents. HALO-Arena covers 4 proactive need types, Dependency, Blockage, Risk, and Opportunity, across 10 everyday domains and 50 personal environments. Scenarios require agents to retrieve evidence from structured databases or unstructured documents before judging whether assistance is warranted and choosing its scope. On HALO-Arena, we construct HALO-8K, a dataset of 8,000 responses from 10 models across 400 scenarios, annotated by 20 experts for action level, acceptability verdicts, and human diagnosis. For rejected responses, human diagnosis covers timing, intensity, and other issues. Recorded reasoning tokens and retrieval rounds measure the effort behind these responses. The best proprietary and open-source models achieve 72.88% and 56.63% acceptability, respectively. Our central finding is that over-proactivity is the dominant observed failure mode, with Premature timing or Excessive intensity accounting for 61.20% of rejected responses. Further analyses yield complementary findings on the balance between intervention and restraint, evidence acquisition, and the calibration of timing and intensity.
Huayu Wang, Gyungmin Ko, Jung Heon Kim, Jia-Xian Jian, Dae Kon Kim, Dong Keon Lee, Hsiang-Wei Huang, Pau-Choo Chung 等 12 位作者
Reliable vocal-cord localization in emergency video laryngoscopy remains visually challenging due to rapid tissue deformation, specular highlights, and instrument occlusions. We propose Local Point Tracking (LPT), a robust framework that combines a single-frame detector, local point propagation, and stateless mask refinement. LPT follows a localize first, snap second pipeline. First, a single-frame semantic detector generates sparse anchors, within which LPT initializes local query points (SAM-supported by default, with a regular grid as fallback). A point-propagation module tracks these points through detector-missed frames and recovers temporal bounding boxes from their geometry. Second, because point-derived boxes can be loose around the visible boundary, a box-snap step refines the selected box to the target boundary. This decomposition keeps temporal state in local point coordinates rather than a stateful mask or whole-region representation, while the spatial refinement module handles single-frame boundary precision without feeding mask errors back into temporal memory. Extensive evaluations on two distinct datasets demonstrate that our proposed LPT method achieves strong overall performance against existing baselines, with the best mAP_50, success AUC, and missing rate on both datasets. Furthermore, ablation studies confirm that point tracking primarily enhances temporal tracking robustness, while box-snap refinement effectively improves boundary tightness.
Generative recommendation (GR) reformulates item retrieval as autoregressive generation of discrete semantic identifiers (SIDs), offering a unified paradigm for personalized recommendation. GR4AD, as an industrial generative recommendation framework, has greatly improved advertising revenue while enabling efficient real-time serving in Kuaishou.However, continued deployment has exposed three limitations: deterministic SID assignment remains sensitive to semantic boundaries; conventional backbone scaling provides diminishing revenue gains under strict serving constraints; and sequence-level value optimization provides limited credit assignment to intermediate SID decisions, causing likelihood-based beam search to prematurely prune potentially high-value candidates. We present GR4AD-X, which addresses these limitations through reliable SID construction, serving-efficient capacity scaling, and unified value alignment. Specifically, GR4AD-X improves SID reliability through Multi-Path assignment. It expands model capacity along multiple complementary dimensions rather than relying solely on wider or deeper backbones, achieving a better effectiveness-efficiency trade-off under strict serving constraints. Finally, Unified Value-Guided Generation and Discrimination (UniVGD) aligns generation with commercial value through prefix-level reward optimization, value-guided beam pruning, and discriminative reranking within a unified model. Large-scale online A/B tests on the Kuaishou advertising platform demonstrate that GR4AD-X achieves a cumulative revenue improvement of 3.07% over the deployed GR4AD baseline while satisfying strict latency and throughput constraints.
Subjective judgments such as offensiveness often vary substantially across individuals, motivating pluralistic modeling approaches that incorporate annotators' socio-demographic information to capture this variation. However, prior findings on the benefits of socio-demographic information for pluralistic modeling are mixed, calling for a systematic analysis of both its effectiveness and the mechanisms through which it improves prediction. In this work, we study $\textbf{when}$ and $\textbf{where}$ socio-demographics benefit pluralistic offensiveness modeling, as well as $\textbf{the role of explicit reasoning}$ in leveraging this information. Across three multi-annotator datasets with prompted and trained LLMs, we first show that demographic gains are small and inconsistent for individual predictions but become clearer at the group and population levels, because aggregation attenuates the idiosyncratic variation that masks small but systematic demographic signals. Second, using a reasoning and prediction format that explicitly decomposes individual judgment into an overall baseline and a demographic-specific deviation, we show that the gains arise from both components: demographic information improves baseline estimation and enables meaningful demographic-specific shifts, and interventions show that better component estimates improve final predictions. Finally, comparing direct demographic conditioning with explicit reasoning about how the text interacts with annotator demographics, we find that reasoning yields more accurate baseline and shift estimates as well as final predictions, whereas conditioning alone rarely produces informative shifts. These results suggest that the mixed evidence in prior work may partly stem from individual-level evaluation, and that the value of socio-demographic information depends on both how it is evaluated and how models are guided to use it.
Preference modeling aims to capture heterogeneous human values and increasingly relies on learning latent components to represent this variation without fine-grained supervision. What these components actually represent, however, is rarely examined. We audit Mixture-of-Experts (MoE) reward models along three rungs---semantic coherence, functional correspondence, and composability---in controlled and real-world settings, holding the architecture and training data fixed so that differences isolate the effect of the training objective. Our diagnosis is that the components conventional MoE reward models learn are entangled, which limits not only interpretability but also the personalization such models are built for. The entanglement is not intrinsic to the architecture but a property of the training objective: off-the-shelf sparsity and diversity regularizers already substantially reduce it. The resulting configuration, which we call sparse MoE, recovers all underlying semantic categories without domain supervision in a controlled setting, reaching 84.8% expert purity against 60.8% for conventional MoE; on real-world data its experts admit human-interpretable task- and domain-level descriptions aligned with their functional strengths. Targeted interventions on routing weights produce predictable preference changes, showing that these interpretations are behaviorally grounded rather than merely descriptive. Crucially, the disentanglement induced by this objective also makes the experts composable: lightweight router updates yield a 25.81-point improvement in few-shot personalization with 50 examples, roughly 2.7 times the gain of the identical architecture and adaptation procedure applied to entangled experts. Together, these results show that MoE reward models can learn interpretable expert structure, and that the same structure can substantially improve personalization.
Zhenyu Wang, Molei Liu, Jing Lei, Francis Bach, Zijian Guo
When synthesizing multi-source high-dimensional data, a key objective is to extract low-dimensional representations that effectively approximate the original features across different sources. Such representations facilitate the discovery of transferable structures and help mitigate systematic biases such as batch effects. We introduce Stable Principal Component Analysis (StablePCA), a distributionally robust framework for constructing stable latent representations by maximizing the worst-case explained variance over multiple sources. A primary challenge in extending classical PCA to the multi-source setting lies in the nonconvex rank constraint, which renders the StablePCA formulation a nonconvex optimization problem. To overcome this challenge, we conduct a convex relaxation of StablePCA and develop an efficient Mirror-Prox algorithm to solve the relaxed problem, with global convergence guarantees. Since the relaxed problem generally differs from the original formulation, we further introduce a data-dependent certificate to assess how well the algorithm solves the original nonconvex problem and establish the condition under which the relaxation is tight. Finally, we explore alternative distributionally robust formulations of multi-source PCA based on different loss functions.
Ruoyu Wang, Xinyi Ye, Yunshuo Shi, Liangyu Yuan, Beier Zhu, Chi Zhang
We present a variational control framework that connects sampling-time guidance and reward-based model adaptation through a shared Jacobian pullback. Starting from terminal reward maximization with a path-space KL penalty, the Gibbs variational principle and Doob transform yield a correction to the generative dynamics. A local denoising approximation exposes the pullback that transports output preferences to the controlled variable. This construction gives Variational Optimal Guidance (VOG) in state space and extends to Distilled Variational Optimal Control (DVOC) by amortizing locally improved endpoints into model parameters. The resulting controls share a local variational rule: VOG acts through the denoising Jacobian, while DVOC realizes the parameter pullback by endpoint distillation. Under a linear MMSE model, VOG's exact correction lies in the data subspace. On ImageNet, VOG reduces DiT-XL/2 FID from 2.27 to 2.04 at 250 sampling steps and from 7.58 to 4.44 at 25 steps. After 20 updates on Z-Image-Turbo, DVOC improves HPSv3++ and four held-out evaluator scores over the untrained model, with the highest corresponding means among the trained methods compared. These results connect inference-time control and reward adaptation through a common geometric principle.
Bingchao Wang, Yongcheng Jing, Shunyu Liu, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
Mobile GUI agents are typically fine-tuned 7B-scale vision-language models (VLMs) that generate each action as text. However, these VLMs decode every action token by token, and even agents that score candidate actions keep a large language model or VLM in the loop, making them slow and memory-hungry. To address these issues, we propose GUI-JEPA, to our knowledge the first decoder-free GUI agent policy built on a Joint-Embedding Predictive Architecture (JEPA), which selects actions in the latent space of a JEPA encoder pretrained on unlabelled GUI screenshots. Its action- and region-conditioned latent predictor turns the screen latent into one latent per candidate action, letting lightweight heads score all candidates against the instruction in a single batched pass, without decoding a token. We further show that the scorer relies on this latent pathway and that the predictor is trained for the decision itself: a randomly initialised predictor does just as well as one pretrained on screen dynamics. On AndroidControl, GUI-JEPA is 2.8-4.0x faster on the same GPU than a Qwen2.5-VL-7B agent we fine-tune with the same inputs, while reaching 86-92% of its in-distribution accuracy with 6% of its parameters. Code is available at https://github.com/bcwang-sjtu/GUI-JEPA
Bingchao Wang, Yongcheng Jing, Chengyu Wang, Jun Huang, Jie Yang, Wei Liu, Dacheng Tao
GUI grounding on high-resolution screens fails mostly on small targets, and zooming in is a common training-free remedy. However, most existing zooming pipelines do not match the zoom to the target's size, accept the zoomed answer unconditionally and never revisit a wrong answer. To address these issues, we propose AdaGround, a training-free framework with three components: (1) Token-Normalized Adaptive Zooming, which sets the magnification, field of view and zoom depth of each sample from the model's own estimate of the target size; (2) Verifier-in-the-Loop Refinement, which uses one marked-click Yes/No verifier to keep or reject each zoom, choose between zoom chains, trigger re-grounding and, with an existence query, decide when to abstain; and (3) Compositional Structure Grounding, which composes spreadsheet cells from their headers. On Qwen3-VL models, AdaGround surpasses state-of-the-art training-free zooming methods and many trained GUI models, improving accuracy over the backbone by 7.4 to 20.6 points on OSWorld-G and ScreenSpot-Pro, and its gains carry over to held-out UI-Vision. Code is available at https://github.com/bcwang-sjtu/AdaGround
Da Li, Yihan Zang, Ivan Viola, Zeyu Wang, Dominik Engel
Achieving high-quality reconstruction and responsive camera tracking remains challenging for monocular RGB Gaussian splatting SLAM under a limited optimization budget. We present an online system that addresses this challenge by combining learned geometric priors with explicit geometric camera tracking, which can substantially outperform learned camera prediction in terms of tracking accuracy. A frozen streaming 3D foundation model serves as a virtual RGB-D sensor, providing depth estimation from monocular video. An explicit geometric solver then estimates camera motion through feature reprojection, and a streaming Gaussian mapper integrates the resulting observations into a dense scene representation. We further show that geometric tracking improves camera pose accuracy across representative streaming geometry backbones, demonstrating the broader applicability of our approach. Extensive experiments on indoor and outdoor datasets show that our method significantly improves reconstruction quality over existing approaches, achieving about 7 FPS on a single RTX 4090 and delivering 2-3 times the throughput of the evaluated prior SLAM systems.
Kai Ma, Tianle Sun, Shun Zhou, Boyao Yang, Ningyu He, Haoyu Wang
Many decentralized exchanges use automated market makers (AMMs) to trade assets against liquidity pools. Modern AMMs allow third-party contracts, called Hooks, to customize this process. However, Hooks can impose additional deductions that fee breakdowns based on standard pool parameters do not itemize. We present, to the best of our knowledge, the first large-scale empirical study of the production Hook ecosystem, focusing on these undisclosed settlement charges. Our study covers 7,873 Hook deployments, 50,011 pools, and 296M pool events on Uniswap v4 and PancakeSwap Infinity across Ethereum and BNB Smart Chain (BSC) through August 1, 2026. We develop a bytecode analysis that reconstructs how Hooks change traders' payments and receipts, separating charges from exchanged assets and standard fees. We then reconstruct historical settlements to measure charging frequency and economic burden. We find that direct settlement adjustment is the dominant Hook capability, supported by 71.37% of registered Hooks. Charging capability is widespread and often exercised: 24.30% of Hooks can impose undisclosed deductions, and these deductions occur in 59.23% of accounted swaps involving such Hooks. These charges generate millions of dollars in revenue, with $3.97M recorded across Hooks with available token prices. These findings reveal a gap between standard fee displays and actual trading costs, motivating explicit disclosure of Hook deductions alongside standard fees.
Benchmark researchers and developers of large language models (LLMs) and other AI systems need to find relevant evaluations, locate their benchmark datasets and code, and understand the settings behind reported scores. We present Benchmark Radar, a living database and search engine for retrieval and discovery of AI benchmarks, covering LLM evaluation, agentic and tool-use benchmarks, coding, reasoning, safety, and domain-specific evaluations. The system combines daily discovery of benchmark papers, repositories, datasets, and releases with a searchable benchmark catalog, mentions in model cards and technical reports, and score histories. It retains source identities and citations so readers can inspect candidate benchmarks and their evaluation evidence. Daily discovery draws on 37 sources: 13 direct connectors and 24 first-party research and engineering feeds. The catalog contains 1,283 source records drawn from 4 benchmark catalogs and 12,916 numeric observations on 790 records. We describe collection and retrieval, audit the full catalog, and examine benchmark saturation, adoption trends, and the limits of score comparisons. A worked example walks through a complete prior-art search, showing how to query the catalog and inspect benchmark evidence when designing a new evaluation. We release the web dashboard with a benchmark leaderboard, a Pareto frontier view of score against measured use, saturation and trend views, daily feeds, downloadable evidence, a command-line interface (CLI) for offline queries, and reproducible analysis.
Zeyu Wang, Yiheng Duan, Tao Chen, Donghyun Koo, Sihan Liu, Wenbo Zhang, Chuan He, Liang Liu
Low-precision training often relies on stochastic rounding (SR) to keep quantized gradients unbiased and prevent small updates from rounding to zero. Standard SR draws an independent random threshold for every gradient element. However, accelerators generate random numbers on vector units whose throughput lags far behind their matrix units. Drawing one threshold per element saturates the vector unit and can make INT8 training with SR slower than BF16 training. Sharing random thresholds across tensor elements removes this cost but forces those elements to round in the same direction and shifts their rounding errors by a common offset. We find that this offset degrades training primarily through the activation gradient, which propagates the error backward into earlier layers. Because earlier layers cannot cancel an incoming offset using their own local rounding errors, the offset must be removed at the layer where it originates. We propose Structured Subspace Stochastic Rounding (S3R), which quantizes each upstream gradient once for both the activation and weight gradient matrix multiplications and shares a single threshold per token across the hidden dimension. Its error correction, S3R-EC, subtracts the structured mean rounding error from the activation and weight gradients while leaving the optimizer unchanged. In LLM pre-training experiments, uncorrected threshold sharing degrades convergence, whereas S3R-EC eliminates this degradation and matches the validation loss of independent SR. Furthermore, S3R-EC accelerates full INT8 training steps of dense LLMs on TPU v5p by 1.24x to 1.32x over BF16 and by 1.11x over independent SR.
Remote sensing object detection has advanced rapidly with the development of large-scale benchmarks and modern detection architectures. However, existing datasets and detectors remain fragmented. Most benchmarks focus on limited categories, fixed spatial resolutions, or a single sensor, while detectors still struggle to work across different sensors and categorical systems. In this paper, we introduce LEVIRDet-159, the largest and most comprehensive remote sensing object detection dataset to date, with 159 categories, 2.56 million bounding boxes, and 700k fine-grained annotations under a multi-level taxonomy. In each key scale dimension, LEVIRDet-159 exceeds the corresponding largest existing remote sensing object detection dataset, containing approximately (7x) more images, (6x) more object instances, and (4x) more categories. Based on this dataset, we design LEVIRDetNet, a scale-hierarchy-aware detection foundation model for universal remote sensing object detection. LEVIRDetNet couples online visual Ground Sampling Distance (GSD) prediction, GSD-conditioned query modulation and allocation, and a hierarchy-aware detection head for mixed-granularity remote sensing supervision. Under stringent evaluation settings, LEVIRDetNet demonstrates strong cross-domain generalization. Even without target-domain training or fine-tuning, it achieves state-of-the-art detection performance on 9 external benchmarks, improving the strongest fully supervised competing methods by 5.02 mAP on average under each benchmark's primary metric. We hope this study will facilitate the development of strongly generalizable remote sensing object detection across diverse category systems, spatial resolutions, and sensor platforms. The full image tiles, annotations, source-license manifest, code, and trained models have been released in a versioned project repository at https://qinzheyang.github.io/LEVIRDet/.
Yao Huang, Zhiyuan Zeng, Yu Wang, Zhitong Zhu, Dingcheng Duan, Hao Li, Shangzhi Xu, Yifan Wang 等 17 位作者
As frontier language models grow increasingly capable and autonomous, independent third-party evaluations are critical for characterizing their systemic safety risks. While closed-source releases are routinely accompanied by comprehensive system cards, open-weight models receive less scrutiny regarding frontier risks. To address this gap, we present an independent safety evaluation of DeepSeek V4.1 Flash-1017, the latest version of V4.1 Flash that departs from V4 with a new architecture, scaled-up training, and strong agentic capabilities. Drawing on the EU General-Purpose AI Code of Practice, we evaluate the model across 34 benchmarks, including 5 in-house suites, spanning evaluation integrity, loss of control, and harmful manipulation. Conforming to the AEF-1 standard, we compare V4.1 Flash-1017 against V4.1 Flash, earlier DeepSeek V4 models, and leading closed-source systems (GPT-6 Astra, GPT-5.6 Sol, and Gemini 3.8 Flash). Overall, V4.1 Flash-1017 shows substantial safety improvements over the V4 generation and approaches or matches leading closed-source baselines on several dimensions, while still exhibiting notable behaviors that warrant continued attention. Building on our findings, we highlight three key observations and recommendations. First, because safety can shift and even regress across minor updates, developers should integrate safety evaluation into incremental releases. Second, as final-turn outputs and end-state checks alone cannot reliably detect loss-of-control behaviors, deployments require process-level observability across both reasoning traces and action logs. Third, because public benchmarks elicit strong evaluation awareness and face data contamination risks, developing private benchmarks is essential for trustworthy frontier evaluation.
While dataset distillation has witnessed significant advances in image classification, its extension to dense prediction tasks such as semantic segmentation is plagued by two core bottlenecks: bi-level optimization-based approaches suffer from prohibitive computational costs stemming from pixel-wise gradient unrolling, whereas proxy-based generative methods merely infuse visual textures into fixed semantic masks derived from real data, thus failing to condense rich semantic knowledge into a compact set of informative, novel scene generations. To address both bottlenecks simultaneously, we propose a novel Sample-Set Reinforcement Learning (S$^2$-RL) framework for generative semantic segmentation dataset distillation (SSDD). Specifically, S$^2$-RL formulates SSDD as a diffusion model based text-to-image generation paradigm, enabling flexible generation of images with informative, novel semantic distributions via a concise text prompt. Furthermore, we fine-tune a diffusion policy via Group Relative Policy Optimization (GRPO), leveraging a sample-set dual reward paradigm: for individual samples, it enforces semantic alignment with input text prompts and intra-class feature diversity; for sample sets, it maximizes inter-sample semantic distribution diversity by solving a contextual multi-armed bandit problem. These design choices enable S$^2$-RL to distill large-scale semantic segmentation datasets into a compact set of generated samples that encapsulate diverse, informative semantic knowledge without the need of pixel-wise gradient unrolling. Extensive experiments on the ADE20K and COCO datasets demonstrate that S$^2$-RL achieves substantial and consistent improvements over state-of-the-art baselines, establishing a new benchmark for SSDD.
Chain-of-Thought (CoT) reasoning has significantly enhanced the reasoning capabilities of large language models (LLMs). However, existing reinforcement learning methods primarily optimize outcome-level rewards, providing limited supervision over the reasoning process. As a result, models may produce shortcut-prone or unstable reasoning trajectories even when the final answer is correct. We propose GeoFaith, a spatio-temporal latent dynamics framework for learning faithful reasoning processes without process reward models. GeoFaith combines final-state latent uncertainty, which measures whether a completed trajectory lies in a well-supported manifold region, with predictive-entropy dynamics, which penalize abnormal uncertainty patterns such as premature collapse, oscillation, and prolonged uncertainty. These signals are incorporated as process regularization into GRPO, jointly optimizing answer correctness and reasoning quality without step-level process annotations. Experiments show that GeoFaith consistently improves reasoning accuracy, produces more efficient reasoning trajectories, and outperforms both outcome-only reinforcement learning and process-supervision baselines.
Additive codes generalize linear codes and play an important role in the construc?tion of quantum codes. As generalizations of linear MDS codes, additive MDS codes have received much research attention. As an important subclass of additive codes, ACD codes are widely used for quantum error correction. In this paper, we study single-twist additive generalized twisted Reed-Solomon (GTRS) codes over finite fields. We derive necessary and sufficient conditions for these codes to be MDS for arbitrary hook and twist parameters. Then we employ Schur squares to prove that such additive MDS codes are not monomially equivalent to additive Reed-Solomon codes. We also identify a class of additive GTRS codes whose duals remain additive GTRS codes, and show that, for all sufficiently large prime powers q whose characteristic avoids a finite set, such GTRS codes yield ACD codes.
Long-video generators increasingly retrieve past visual observations to prevent identity drift. The intuition is compelling: more access to history should make future video more consistent. But an image can be faithful to the past and wrong for the present. Does visual memory always improve the generated story? We investigate this question through VMem-Bench, a two-track diagnostic study. Track-A measures which references are selected from 92 real films (16,548 segments), before generation and without generator confounds. Track-B compares four retrieval policies with no-memory and gold-informed last-seen controls on 30 causally generated stories (2,824 segments) using a shared generator. Retrieval increases entity recall from 0.559 to 0.628-0.693 and improves similarity to past appearances, but decreases precision from 0.977 to 0.915-0.938 and forbidden-entity avoidance from 0.968 to 0.845-0.898. It can also reinforce obsolete appearances after the story changes them. Conversely, high long-gap recall can arise without memory when the prompt itself supplies the appearance; an appearance-stripping analysis exposes this confound. These results distinguish historical fidelity from current-story correctness and show why retrieval relevance or final-video consistency alone does not establish memory utility.
Yifan Lan, Yuanpu Cao, Hanyu Wang, Lu Lin, Jinghui Chen
Large language models (LLMs) have demonstrated impressive reasoning abilities across a wide range of tasks, but data contamination undermines the objective evaluation of these capabilities. This problem is further exacerbated by malicious model publishers who use evasive, or indirect, contamination strategies, such as paraphrasing benchmark data to evade existing detection methods and artificially boost leaderboard performance. Current approaches struggle to reliably detect such stealthy contamination. In this work, we uncover a critical phenomenon: a model's generated reasoning steps actively mask its underlying memorization. Inspired by this, we propose the Zero-CoT Probe (ZCP), a novel black-box detection method that deliberately truncates the entire Chain-of-Thought (CoT) process to expose latent shortcut mappings. To further isolate memorization from the model's intrinsic problem-solving capabilities, ZCP compares the model's zero-CoT performance on the original benchmark against an isomorphically perturbed reference dataset. Furthermore, we introduce Contamination Confidence, a metric that quantifies both the likelihood and severity of contamination, moving beyond simple binary classifications. Extensive experiments on both previously identified contaminated models and specially fine-tuned contaminated models demonstrate that ZCP robustly detects both direct and evasive data contamination. The code for ZCP is accessible at https://github.com/Yifan-Lan/zero-cot-probe.
Yunqi Liu, Peiyu Wang, Bingli Wang, Leo Luo, Xuanwu Song, Xudong Ma, Hongxin Li, Yaxin Du 等 24 位作者
Command-line interface (CLI) agents increasingly execute longer-horizon tasks, yet existing benchmarks either assume that user instructions remain fixed throughout execution or evaluate task outcomes without capturing runtime collaboration quality. In this paper, we present REALCLI, a paired benchmark that unifies the evaluation of task execution and human-agent collaboration quality on realistic CLI tasks. To comprehensively measure collaboration, REALCLI builds a five-dimensional framework, operationalized into 22 sub-dimensions and 51 interaction scenario templates. REALCLI provides two modes for paired comparison: REALCLI-EXEC supplies complete specifications upfront, whereas REALCLI-COLLAB introduces a scripted user that delivers progressive information, corrections, constraint revisions, and authorization decisions. Evaluations reveal profound capability gaps in frontier agents: the best configuration attains an average score of only 49.2%, and when evaluated by the joint collaboration score ($J_{\mathrm{collab}}$), the peak performance reaches merely 37.6%. These results show that task effectiveness diverges sharply from collaboration quality, with models consistently struggling in steerability, trust calibration, and safety boundary management. REALCLI provides a rigorous, reproducible foundation for measuring and advancing collaborative CLI agents.
Mingyu Wang, Bin Yu, Wenjian Lu, Zhi Wang, Gao Kefeng, Cheng Wen Xi'an University of Electronic Science and Technology, Xu Lu Xidian University, Cong Tian
Linux kernel device drivers are tightly coupled with hardware, making them difficult to execute and test without physical devices. This heavily limits automated code analysis and vulnerability discovery. While manual modeling is unscalable, Large Language Models (LLMs) offer a new approach to scale virtual device construction across the Linux driver ecosystem. In this paper, we present DevGen, an LLM-powered tool that generates QEMU-based virtual devices directly from Linux driver source code. DevGen combines static analysis to gather necessary context, guides the LLM through step-by-step prompting, and uses an automated self-correction loop driven by compilation and execution feedback. To further reduce errors, similar fixes are retrieved from a library of common modeling failures and incorporated into the repair prompt, which supports more targeted corrections in later iterations. The generated devices finally integrate with QEMU and Syzkaller, enabling driver fuzzing without physical hardware. DevGen is evaluated on 50 PCI/PCIe drivers from Linux 6.18 using three mainstream LLMs, and successfully generates usable models for 44 drivers. In these drivers, 24% of them achieve significant improvements in fuzzing coverage, and 7 previously unknown crashes are triggered with 1 CVE assigned. These results demonstrate the practical capability of LLMs to automate complex, system-level code generation tasks.
尚无 AI 解读，展示原文摘要
cs.CL
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations) · 2026 · 正式发表
Generative AI has enabled the creation of photorealistic images and videos that are increasingly disseminated on social media, often used for spam, misinformation, manipulation, and fraud. Existing AI-generated content (AIGC) detection methods face challenges including poor generalization to new generation models, reliance on single modalities, and lack of interpretable explanations. We present our pipeline that mitigates these issues by continuously curating diverse multi-modal social media data and training a compact vision-language model for detection and explanation. Our model achieves state-of-the-art detection performance on public benchmarks and demonstrates robust detection and explanation capabilities on internal social media datasets across multiple platforms. We deployed our model for post recommendation on social media platforms and observed positive downstream impacts on user engagement, demonstrating that it is feasible to perform effective AIGC detection in dynamic, real-world social media environments.
尚无 AI 解读，展示原文摘要
cs.CL · cs.CV
Proceedings of the 64th Annual Meeting of the Association for Computational Linguistics (Volume 6: Industry Track) · 2026 · 正式发表