信息流

Streaming video understanding requires models to interpret evidence as it arrives, yet current evaluations often report task scores without specifying when evidence becomes valid, how visual history is maintained, or how responses are triggered. As a result, similar scores may correspond to different workloads, failure modes, and operational behavior. We introduce TRACE (Temporal Audit and Condition-aware Evaluation), a condition-aware benchmark and evaluation framework that makes these factors explicit. TRACE combines temporally audited visual tasks with evidence timing and instruction-dependent trigger annotations, a unified causal Core--Adapter protocol that controls information availability while recording actual history processing and response events, and multidimensional reporting of answer quality, timeliness, response-selection behavior, workload, completion, and reliability. On 1,240 records from 517 videos, we evaluate eight publicly available models or systems in eight configurations. We find that nearly identical QA accuracy can mask substantial differences in completion, answer validity, and generation workload, while proactive performance separates into response quality, response delay, false alarms (responses emitted while no target window is currently valid and a later one remains), and missed target windows. These results show that streaming-video performance should be interpreted as execution-conditioned system behavior rather than a single score. Our benchmark and code can be accessed at \href{https://github.com/om-ai-lab/trace-bench}{https://github.com/om-ai-lab/trace-bench}.

网络时尚图像支持视觉商品搜索和消费者与店铺匹配等应用，但其视觉异质性使可靠的服装分割具有挑战性。尽管可提示分割取得了进展，但在没有人工提供空间提示的情况下，自主识别并纠正服装特定错误仍然困难。我们提出 GarmentSegAgent，一个通过迭代文本自我优化整合语义规划、空间执行和视觉评估的自主框架。其核心是引入 CAST（Constrained Attribution-guided Selective TextGrad），将视觉错误归因与有针对性的文本优化连接起来。CAST 识别导致分割错误的指令组件，通过文本梯度选择性地更新它们，并依据服装特定约束验证生成的指令。这使得推理时能够进行针对具体图像的纠错，同时保持底层基础模型冻结。在 DeepFashion2、ModaNet 和 Fashionpedia 上的实验表明，区域重叠和边界准确率均持续提升，并且相比传统 TextGrad 具有更可靠的优化。通过让视觉反馈引导受控的指令更新，GarmentSegAgent 为自主服装理解提供了一个有前景的方向，并有望应用于可扩展的时尚内容分析、视觉检索以及网络上的消费者与店铺匹配。

Deep neural networks continue to grow in parameter count, driving up training and inference cost on GPUs. Sparse neural networks and Dynamic Sparse Training (DST) promise to reduce these costs, but most implementations rely on binary masks over dense tensors and recover little of the theoretical compute, memory, or energy savings. We propose SNACK, a truly sparse GPU layer that stores and computes only non-zero connections. SNACK exposes a simple PyTorch API for restructuring connections and backpropagating gradients entirely in the sparse paradigm, and ships SNACK-COO, a custom COO-format SpMM CUDA kernel with a batch-to-Streaming-Multiprocessor mapping tuned for the small-batch, high-sparsity regime typical of large-model training and single-stream inference. At the kernel level, SNACK is up to 7x faster than the masked dense baseline (Dense+Mask) and competitive with cuSPARSE, Sputnik, and FlashSparse at 95% sparsity. At 90% sparsity, a single SNACK layer accelerates training by 8x and 3.7x, and inference by 4x and 2x, over Dense+Mask and fully dense layers, respectively, while using 72% less memory than dense and substantially less energy. End-to-end, SNACK reduces GPT-2 peak training memory by up to 40% and graph-style inference latency by 4.8x over Dense+Mask at 99% sparsity.

Behavioral tasks measuring facial affect perception assume that identical stimuli impose equivalent perceptual difficulty across participants. However, this assumption is systematically violated by individual differences in perceptual sensitivity. Using an affective perception task as our testbed, we propose a framework to normalize for perceptual difficulty that directly estimates each participant's Just Noticeable Difference (JND) along the facial affect spectrum via cognitively lightweight perceptual adjustment queries (PAQs). We use these PAQ-inferred JNDs to re-express stimulus distances, constructing difficulty-equated tasks in perceptual space. We validate the framework in a Two-Alternative Forced-Choice (2AFC) task using two complementary behavioral measures: binary metacognitive difficulty judgments and response time variance decomposition. We find that PAQ calibration significantly equalizes perceived task difficulty at an individual level when compared to both the non-calibrated baseline and population-level Weibull calibration, while also reducing mean response time and between-subject variance in response time. These results establish PAQ as a principled and practical instrument for individualized perceptual calibration in facial affect recognition.

We formulate \emph{Artistic Intelligence} as exploration driven world realization, leaving space for creative possibility while preserving the semantic, artistic, and compositional structure that must remain true. Moonworks Lunara, a text-to-image model, implements this framework with a novel Diffusion Mixture Transformer architecture. A new training algorithm iteratively evolves the data distribution through informative sample acquisition and targeted injection of human-created art. We benchmark Lunara against seven image-generation models, including FLUX.2-Klein-4B, Qwen-Image (20B), and GPT-Image-1-Mini. With GPT-5.6 Sol as evaluator, Lunara ranks first in \emph{Aesthetic Quality (8.473 vs. 8.457 GPT-Image-1-mini)}, second in \emph{Emotional Resonance}, and remains competitive in \emph{Content Integrity}. A blind human evaluation over the same evaluation set corroborates the automated metrics, ranking Lunara first. It also stays among the strongest models under conventional measures including CLIPScore and LAION Aesthetic Predictor. On GenEval, Lunara achieves competitive performance against a broader set of 16 models, including GPT Image 2 and Seedream 4.0. These results place Lunara at the frontier with Artistic Intelligence while maintaining a sub-10B active-parameter footprint and sub-10-second inference latency. Lunara advances the general visual intelligence frontier by shifting the question from whether models can get images right to how deeply they can interpret meaning and realize it as imaginative, expressive worlds.

Large Language Models (LLMs) tend to add disclaimers like "I'm just an AI" when asked about something related to themselves. The self-reports from such responses are used in debates about AI safety or self-knowledge of the models, yet what drives them is not well understood. Are the models telling us about themselves or rather how they are deployed? In this work, we show that the chat template works like a switch - when present, it turns this disclaimer voice up and experiential voice like "I feel" down, across 8 popular open-source instruct models up to 9B parameters in size. And conversely when the chat template is not present, it turns the disclaimer voice down and experiential voice up. Inside the activations of 3 models, we find a direction that steers this behavior. Removing the direction in the model's activation space turns disclaimer voice down and adding it turns it up, while a random direction of the same size has little effect. We find that instruct models without chat template, when we add the disclaimer direction to them, disclaim like the template was there. Since the chat template controls the disclaimer voice of LLMs, then researchers studying self-reports or introspection of models might have a confound they need to control for. Our results show that there is a direction they can use to steer this voice. More broadly, our work shows that what models say about themselves is not a fact about them. What they say doesn't come only from weights, but it is partially set by the chat template, and because of that a model's self-description shouldn't be treated literally.

We ask whether relation-type information (e.g., capital-of) and entity-specific information (e.g., France to Paris) become causally active at the final-token position at the same depth during recall. Using four complementary causal diagnostics across four decoder-only models and eight prompt families, we find a robust temporal asymmetry: relation information becomes generation-controlling before entity information does. Relation onset precedes entity onset by 10-16 tested layers (31-44% of network depth) at threshold 0.4, with the ordering holding across all 16 model-threshold combinations for thresholds 0.2-0.5. Critically, entity information is not absent early: entity-token patching succeeds at 90-100% in early layers. Instead, entity commitment to generation is deferred: entity information is available at the entity-token position but becomes generation-controlling at the final token only after being routed there.

Diffusion models can produce striking images and videos, but they still struggle with the compositional details that make a generation faithful to a prompt, such as object counts, attribute binding, spatial relations, and temporally grounded actions. A common way to improve prompt satisfaction is to spend more compute at test time through Best-of-N sampling, but final-sample selection is fixed. Best-of-N can only choose among completed outputs and cannot repair a promising trajectory before it fails. We introduce PreviewDiff, a training-free test-time search method that turns diffusion sampling from scalar search into a multimodal critic-guided search over intermediate latents. At selected denoising checkpoints, PreviewDiff decodes a partial preview, asks a multimodal judge to score and critique it, and uses the resulting natural-language feedback to branch over semantic prompt edits and locally re-noised latent continuations. These branches are then scored and selectively rolled forward, allowing verifier compute to guide generation while the sample is still editable. Across image and video generation benchmarks, PreviewDiff consistently improves over budget-matched Best-of-N selection and strong scalar-search baselines. Ablations show that earlier interventions and increased search width provide the largest gains, while deeper search and additional semantic variants offer complementary improvements. PreviewDiff demonstrates that multimodal feedback is most useful not only as a final verifier, but as an active controller inside the denoising process.

Multimodal misinformation is increasingly crafted to look convincing by pairing a textual claim with an image that appears to "prove" it. Yet in practice, building effective detectors often hinges on a small set of design choices that are rarely examined in a controlled way. In this paper, we conduct a large-scale study of multimodal design choices for misinformation detection with over 3,375 experiments- spanning three benchmark datasets and a broad range of pre-trained vision and language backbones. Through systematic comparisons and targeted robustness analyses, we distill practical guidance on which design choices help, when do they fail silently, and what aspects of the pipeline most strongly shape model behavior, answering 4 key Research Questions (RQs). We aim to provide a reliable foundation for designing stronger and more dependable multimodal misinformation detection systems, thus contributing to the broader research community.

Subliminal learning shows that language models can transmit a hidden trait through outputs that appear unrelated to it. One proposed explanation, token entanglement, links animal and number tokens through the model's output vocabulary. Yet existing measurements answer different questions: whether outputs co-vary, fixed output vectors align, an answer can be read from a hidden state, or that state causally controls the answer. We measure each separately in a fixed animal-number prompting protocol. From Llama-3.1-8B to 70B, fixed output-vector similarity predicts behavior less well: the paired mean correlation change is -0.080 (95% CI [-0.127, -0.035]). A fixed output-head readout shows no resolved change in normalized depth AUC. To test control, we copy the temporary answer-position state from one number prompt into another at five depths and measure which prompt the final animal score follows. Donor-control AUC rises from 0.254 to 0.540, a paired change of +0.286 (95% CI [+0.272, +0.300]), with increases for all 18 concepts. The contrast remains with exactly eight transformer blocks remaining, while specificity and identity controls remain small or exact. In two Qwen models, scoring every digit in sequence does not recover the positive one-token association. Per-token averaging instead creates a positive pooled association that disappears after controlling number width, revealing a length confound. Thus, fixed geometry, observational readability, causal timing, and multi-token measurement are distinct properties of this frozen prompting channel. They constrain token-level explanations but do not identify the mechanism of training-time trait transfer.

Continual Learning Bench scores learning as short-horizon gain versus a reset baseline and finds naive full-context ICL strongest among the memories it tested. We treat ICL as one learning system and score it on a longer shared-world schedule. Steady-state learning is the gap versus baseline on a pre-set late window (last 40 of 174 BIRD-SQL formula-1 questions). We split the score into exploration efficiency (SQL probes), task reward (hits), and delivery cost (API dollars and context size). On gpt-5.6-luna, late probes fall from 4.6-5.6 to 0.95 while hits rise only modestly and ICL context grows to about 95k tokens with cost roughly doubling. Short-horizon gain understates the late probe saving and misses the cost inversion, so we find that unbounded ICL is a poor candidate for the learning mechanism.

Long-horizon agents continuously accumulate interaction history during task execution, yet the importance of past interactions changes as the agent state evolves. Existing context management methods largely compress history based on fixed windows, periodic schedules, or current relevance, overlooking a more fundamental question: when has a past interaction become safe to replace? Premature compression may remove information still needed for future actions, while overly conservative retention leads to substantial context overhead. To address this, we propose State Conditioned Compression (StateComp), a framework that determines when historical interactions can be safely compressed according to the current agent state. StateComp constructs KEEP and READY supervision through a two-stage annotation procedure and trains an imbalance-aware router on hidden representations from a frozen language model. A bounded state representation further reduces the cost of evaluating long histories, while adjacent READY interactions are grouped into continuous spans and replaced with compact summaries during execution. Experiments on WorkBuddyBench show that StateComp reduces total agent and summarization tokens by 52.27% while maintaining task performance, and achieves a 12.67-fold speedup in representation extraction.

LLM agents for clinical text-to-SQL applications reason autonomously over multiple steps but cannot assess whether their own reasoning or outputs can be trusted. In high leverage applications such as healthcare, this presents a critical risk where system mistakes can be costly. These reliability failures are also resource failures: an incorrect reasoning trajectory spends computation budget on outputs that must be discarded. We introduce Sentinel, a trajectory-derived, classifier-based confidence layer that analyzes an agent's reasoning, code and database outputs to decide at three points whether to stop: refusing unanswerable questions before the agent runs, halting doomed trajectories mid-run, and withholding untrustworthy answers at delivery. Here, utilizing Chow's rule, we optimize decisions under the EHRSQL shared task's Reliability Score, which penalizes incorrect answers given a utility weighting, and find on the benchmark EHRSQL that Sentinel raises this score from +0.08 to +0.24 when mistakes have a low utility weighting, well above the +0.03 earned by refusing every question, with delivered-answer accuracy rising from 54% to 69% as coverage falls from 84% to 55%. At higher stakes the agents we test rarely answer reliably enough to deliver, and Sentinel detects this on its own, abstaining to that same +0.03 where the unmonitored agent scores -3.02. The same stopping decisions cut computation: at low stakes, where the system still answers, Sentinel eliminates 13-28% of agent steps at little or no reliability cost.

LLM agent 难以自动化复杂的企业流程，因为静态的、预先构建的 harness 无法捕捉专有策略、未文档化的行为，或只有在执行过程中才会浮现的、依赖状态的恢复路径。从 agent 经验中积累的程序性记忆提供了一条无需训练的路径来弥合这一差距——但与情景记录或语义事实不同，它必须将重复出现的成功和经过诊断的失败，整合为可复用的动作与恢复策略。我们提出 WASP-Mem（Wiki-based Agent-Synthesized Procedural Memory），这是一种轻巧而有效的方法，将执行轨迹与反馈提炼为持久的、相互链接的 wiki 页面，从而实现 agent 的持续学习。在 TheAgentCompany benchmark 上，WASP-Mem 在我们比较的三个模型中，相较于最强基线将 checkpoint 通过率提高了 4.6-6.1%，同时所需存储减少 57-61%。我们进一步提出了一项系统性对照研究，涵盖整合方法、反馈类型、轨迹摄取范围和 refinement 轮次。我们的结果表明，诸如人工定义 schema 和显式去重规则这类规定性知识组织方式会损害性能，而不是弥合程序性知识差距。相反，人类诊断将失败轨迹转化为纠正性知识，并推动持续改进：单轮人类诊断比四轮自我判断高出 7.0-10.9%。这些发现指向一种简单的分工：让 agent 组织程序性

Although large language models (LLMs) encode language in highly distributed spaces, they remain vulnerable to focal disruptions---a phenomenon strikingly similar to how localized brain damage causes clinical aphasia in humans. To investigate this theoretical tension, we ask whether LLMs possess a functional structure analogous to the human language system, by introducing a novel mechanistic interpretability framework inspired by aphasia. Through two complementary approaches, classifier-guided lesioning and continuous activation steering, we systematically induce aphasia-like breakdowns in LLMs. We find that language functions can be successfully isolated to discrete structural circuits and continuous representational vectors, yet these functions remain densely entangled with significant circuit overlap and vector similarities across distinct aphasic profiles. Furthermore, while steering generalizes robustly across downstream tasks, structural lesioning remains asymmetrically constrained. By bridging neuropsychology and mechanistic interpretability, we demonstrate that LLMs, functionally analogous to biological brains, rely on a densely entangled architecture that can be systematically mapped through clinically-informed interventions.

Integrated deep reinforcement learning (DRL) and model predictive control (MPC) methods are increasingly used to control autonomous systems by combining their complementary capabilities. DRL learns control policies through interaction with the environment. MPC uses a system model to optimize control inputs while accounting for constraints. In DRL-MPC frameworks with shared control authority, both the DRL agent and the MPC controller each determine part of the control inputs. However, common learning formulations treat MPC as part of the environment and therefore do not explicitly account for MPC's contribution to control or its interaction with the DRL agent. This paper proposes a novel composite-gradient learning (CGL) method that integrates the MPC controller into the learning process by representing the DRL and MPC control inputs as a joint action and accounting for their interaction when updating the DRL agent during training. CGL is evaluated on two multi-class freeway traffic networks with different strengths of interaction between the DRL and MPC control inputs and it is compared with alternative methods that treat MPC as part of the environment or that only partially incorporate MPC into learning. The results show that CGL offers limited benefit under weak interaction, but learns higher-performing control policies than the alternative methods in a subset of training runs under strong interaction, although the average control-performance gains remain modest.

Benchmarks for full-duplex spoken dialogue models score turn-taking with binary fixed-window rules that reward immediate response or silence by completeness of the prior turn. We argue that the appropriateness of a response offset, whether delayed silence or anticipatory overlap, is conditional on the speaker's latent intent, identifiable only from that speaker's behavior. We introduce TACT, a benchmark of 9,728 episodes and 73.2 hours from five dyadic corpora; each episode carries dialogue history, a per-speaker memory profile, and an annotator-derived posterior over six intent classes. Scoring replaces binary windows with a strictly proper threshold-weighted continuous ranked probability score whose weights are intent-conditioned timing kernels fitted to human floor-transfer-offset distributions, proving boundedness, consistency, and binary reduction. Across eleven systems the best model reaches 0.47 against a human topline of 0.86, is nearly invariant to speaker profiles, and TACT agrees with human judgments at Spearman 0.81 versus 0.46 for binary metrics.

While Vision-Language-Action (VLA) models perform strongly on manipulation tasks, their responses to invalid task premises remain underexplored. Existing evaluations of premise conflicts often focus on terminal task outcomes, yet task failure alone cannot distinguish behavioral disengagement from continued pursuit followed by an execution error. We call the latter pattern Failed Persistence. To study this phenomenon, we introduce ConflictVLA-Bench, which pairs conflict rollouts with premise-consistent reference rollouts and evaluates both outcomes and execution processes. Built on LIBERO, the benchmark contains 2,826 prompt-conditioned conflict tasks spanning four conflict families, four structural configurations, and two prompt conditions. Across all eight VLAs, invalid premises reduce original goal completion by at least 17.3 percentage points, with the reduction reaching 56.2 percentage points for OpenVLA. Crucially, even when models succeed on premise-consistent tasks and fail on their matched conflict tasks, they often continue to approach the original targets, retain early trajectory structure, and show limited action magnitude suppression. Failed Persistence therefore recurs across the evaluated models. Explicit premise checking does not consistently produce selective and coordinated behavioral changes. These findings show that terminal failure alone establishes neither behavioral disengagement nor refusal and that outcomes alone are insufficient for VLA evaluation. Experimental data and additional details are available on the project page: https://github.com/EmbodiedAISurvey/ConflictVLA-Bench

Gaussian head avatars typically model intrinsic facial appearance as temporally static, omitting subtle cardiac-induced skin-color variation. We propose $\unicode{x1F493}$Heartian, a physiology-aware modulation framework that learns cardiac-cycle-dependent per-frame albedo modulation of facial skin-region Gaussians within a relightable head avatar to encode remote photoplethysmography (rPPG) signals. Using synchronized contact PPG supervision, $\unicode{x1F493}$Heartian models the prescribed cardiac waveform as the sum of two Gaussian functions and learns per-frame spatial residuals via a lightweight MLP. Across 152 stationary recordings from UBFC-rPPG, PURE, and MMPD, attribute-space recovery of the supplied signal achieves a pooled recording-level heart-rate MAE of 0.29 bpm and MAPE of 0.38%. The signals remain detectable after rendering by benchmark rPPG methods, with the best tested configuration - a motion-augmented TS-CAN decoder pretrained on UBFC-rPPG - recovering heart rate from the rendered MMPD avatars at 0.97 bpm MAE and 1.21% MAPE. Meanwhile, $\unicode{x1F493}$Heartian maintains reconstruction quality comparable to the baseline, with negligible average PSNR degradation of 0.005 dB. Overall, our work embeds recoverable rPPG signals as controllable material attributes to subject-specific Gaussian head avatars while retaining the reconstruction quality.

While Large Language Models (LLMs) have achieved remarkable results across various benchmarks, their alignment with normative values often results in homogenized responses that fail to address diverse user preferences. Existing training-free methods often occupy valuable context windows through prompt engineering, while training-based methods typically remain static post-training, failing to support the continual optimization required in real-world settings. To address these challenges, we propose COPE (Continual Optimization with Personalized embedding and self-Evaluation), a novel optimization framework tailored for real-world-motivated interaction settings with sparse user feedback. Our framework assigns learnable personalized embeddings to each user and synergistically integrates preference capture, self-evaluation calibration, and personalized response optimization within a single update step. A key innovation of our method is the use of self-evaluation to generate proxy rewards, enabling continuous model updates even when explicit user feedback is unavailable. Experiments show that COPE consistently outperforms strong training-free and training-based baselines under sparse feedback, and remains complementary to Retrieval-Augmented Prompting (RAP). Further analyses confirm COPE's reliable self-evaluation, meaningful preference patterns, stable general capabilities, and robustness under shifting preferences and alternative evaluators.

流匹配通过回归学习生成速度场，但在分布漂移下，速度场必须在当前数据到达之前完成部署。我们针对一个固定的、可能非凸的神经网络类研究这一预测问题，并以每一轮相对最佳可表示速度场的动态遗憾为衡量标准。我们提出 DyProd-FM，它在二进历史前缀上重新拟合流匹配，并用 sleeping Prod master 在各拟合结果中进行选择。在完全不知道任何漂移预算的情况下，该方法在相对速度风险变化 $\Omega$ 上取得 $\widetilde{O}(\min\{N_\omega+1,1+\sqrt{T\Omega}\})$ 的遗憾，在目标分布变化 $V_\mu=\sum_t W_2(\mu_{t+1},\mu_t)$ 上取得 $\widetilde{O}(\min\{N_\mu+1,1+T^{1/3}V_\mu^{2/3}\})$ 的遗憾，忽略近似与拟合代价。针对固定类的匹配下界表明，这两个速率在 $T$ 以及各预算上均达到最优（至多相差对数因子）。先前的动态遗憾最优保证依赖于凸决策集以及强凸或 exp-concave 损失，而我们的类可能是非凸的，且不假设其参数具有曲率。目标分布速率依赖于两种变化之间的二阶联系，而这一联系由流匹配的基础噪声提供。

Large language models (LLMs) are increasingly used in dynamic financial trading, yet real-time trading benchmarks mainly evaluate agents by cumulative returns and risk-adjusted performance. Such outcome metrics cannot show whether trades are faithfully reported, supported by available evidence, and consistent with account states. We introduce AStockArena, a real-time decision-auditing benchmark for continuous trading in China's A-share market. Under shared market conditions and execution constraints, we evaluate 13 open- and closed-source models over 3,198 sequential decisions. AStockArena reconstructsthe actions actually posted to accounts by aligning reports, trade requests, execution responses, and account states, and audits them with eight checks spanning execution traceability (O), evidence sufficiency (RK), and financial meta-awareness (B), combining record verification with consensus expert annotation. Of 1,070 trading points, 72.99% exhibit at least one audit gap; model-level rates range from52.2% to 91.2%, with evidence gaps most prevalent. Historical trade-filtering replay further shows thatportfolio outcomes are sensitive to these flagged trades. Decision-process auditing thus offers a verifiable complement to return-based evaluation of financial agents.

Multi-objective Bayesian optimization (MOBO) maximizes $M$ expensive black-box objectives with the goal of approximating the Pareto set, whose objective vectors cannot be improved in one component without sacrificing another. A common quality measure is hypervolume, the volume in objective space dominated by these vectors relative to a reference point. Expected hypervolume improvement (EHVI) directly targets this metric and is one-step Bayes-optimal, but its standard computation scales exponentially with $M$, making it prohibitively expensive for many-objective problems. We introduce random max-margin multi-objective Bayesian optimization (RaMMOBO), whose acquisition function lower-bounds EHVI while requiring only linear cost in $M$ to evaluate. RaMMOBO measures the volume of a max-margin box contained in a candidate's hypervolume-improvement region and averages over random reweightings of the objective space, yielding a utility that is strictly Pareto-monotone on hypervolume-improving outcomes. It naturally handles noisy observations and different probabilistic surrogates and supports efficient gradient-based optimization. Across synthetic and real-world benchmarks, RaMMOBO matches or outperforms state-of-the-art methods, with largest gains at higher $M$, with acquisition optimization up to two orders of magnitude faster than qLogNEHVI. Our implementation is available at https://anonymous.4open.science/r/rammobo.

特定领域标注图像的稀缺仍然是开发用于朝觐人群分析的计算机视觉系统的主要限制。现有的人群计数数据集和无约束生成模型无法捕捉朝觐特有的视觉条件，包括圣地的几何结构、身穿白衣的朝觐者人群以及室内/室外光照变化。本研究提出了 CrowdH-Edit，一个参考引导、基于指令的合成朝觐人群图像生成框架。真实朝觐图像提供结构参考，而文本指令引导人群条件和光照的变化，并帮助保持主场景布局。评估了 Stable Diffusion 2.1 和 Stable Diffusion XL 1.0 结合低秩适配作为对比基线。Qwen Image Edit 在 ComfyUI 工作流中无需额外训练即可使用，并被选用于最终数据集构建。该工作流生成了 CrowdH-Edit，一个包含 6,122 张合成图像的清洗后图像池，并从中选取和标注了 387 张图像以构建 CrowdH-Edit-Ann-387。在评估的生成流程中，Qwen Image Edit 取得了最高的语义对齐分数以及最低的分布距离和感知距离。跨五个群体计数模型的下游评估表明，将带标注的合成子集加入真实训练数据后，三个模型的平均绝对误差降低，两个模型的均方根误差降低。这些发现表明，参考引导的合成朝觐图像可以支持

Tool-calling agents are mostly optimized for accuracy; in deployment, however, their growing cost is latency: every tool call is decoded autoregressively. Speculative decoding accelerates this without changing the output, and retrieved past calls make strong drafts under real use, where each user's behavior tends to repeat. Yet existing systems keep no cross-session memory, or build the draft store once and never update it. We propose SpecMem, a tool-call memory kept alive and managed throughout deployment: verified calls are written back online into bounded per-user stores with LRU eviction, and a side-effect gate defers irreversible calls until verification confirms. On multi-session, multi-user query streams served by three open-weight models, a frozen store steadily loses acceptance while SpecMem holds it: a 27% gap in mean accepted tokens (+11% to +34% across benchmarks, tracking how much novel work arrives after the store is built) and up to 1.74x wall-clock speedup. Ablations attribute the gain almost entirely to keeping the store live; local per-user scoping adds no acceptance but closes a measured cross-user leakage channel, and gating on declared side effects, rather than on retrieval confidence, is what eliminates irreversible mistakes.

Voice agents depend on accurate automatic speech recognition (ASR) to act on what callers say, yet ASR is evaluated on read, English-centric speech with word error rate (WER), which penalizes surface rather than semantic differences. We introduce mu-bench, a dataset of 4,270 caller utterances from 250 phone calls to an AI banking agent in English, Spanish, Turkish, Vietnamese, and Mandarin, centered on form-field inputs such as names, email addresses, and confirmation codes. We release Utterance Error Rate (UER), an LLM judge of whether a transcript preserves meaning, calibrated against human raters, together with an LLM normalizer that makes WER comparable across providers' output formats. On 1,847 human-rated transcripts, UER agrees with annotators at $\kappa$ = 0.78, versus 0.53 for exact-match WER on normalized text. We rank six commercial providers on a public leaderboard; the best reaches 11.9% UER, and Mandarin is hardest for all six.

大型基础模型往往在可用的视觉或文本证据不足以支撑可靠结论时仍会给出回答。这种无法拒答的问题使其在高风险或信息检索场景中难以被信任，因为这些场景下一个自信却缺乏支持的答案可能比不回答更糟。这也引出一个自然的问题：这种行为背后的内部机制能否在推理时被识别并控制？我们提出 Dr. No，一种特征引导方法，它挖掘与拒答和作答相关的稀疏自编码器（SAE）特征，然后在 prefill 阶段对其中一小部分特征进行调制，而无需重新训练基础模型。Dr. No 在 LLaVA-1.6 Mistral 7B 上跨多模态拒答基准进行评估，包括 UNK-VQA 和 MM-UPD，并额外使用 Qwen3.5-2B 在 AbstentionBench 子集上对语言推理任务进行基准测试。与多种基于不确定性的拒答基线相比，特征引导在正确回答与适当拒答之间取得了比事后拒绝方法更有利的平衡。这些提升在 UNK-VQA 以及涉及图像—问题或问题—答案不兼容的 MM-UPD 案例上最为显著，并且最清晰地迁移到 AbstentionBench 的数学推理子集。然而，在缺失答案检测和检索密集型多跳问答（如 MuSiQue）上性能较为有限。这些结果表明，稀疏内部特征提供了轻量化的控制手段，用于塑造

Reference-based object compositing inserts or replaces an object using a background image, a reference image, and a 2D compositing mask. These inputs guide appearance and placement but leave the completed scene's geometry implicit, which can distort object structure or alter the surroundings. Our Depth-to-RGB (D2R) framework predicts composite depth for a scene not yet observed in the RGB inputs. It learns reference-conditioned corrections to a frozen depth estimator using encoder features of paired completed scenes as targets. The unchanged decoder maps the corrected representation to the intended scene's depth, which a separately trained renderer holds fixed during RGB synthesis. Under matched architecture and training, encoder-feature supervision reduces OOD Stage-1 AbsRel by 31.4% relative to decoded-depth supervision. We also introduce AnyInsertion++ with paired in-distribution and category-disjoint splits to evaluate generalization beyond compositing training categories. The complete D2R system leads 12 open-source and 3 closed-source baselines in estimator-derived geometry and photometric quality on both paired splits. On category-disjoint data, D2R reduces AbsRel by 43.7% and improves PSNR by 2.4 dB over the matched RGB baseline. Across three unpaired benchmarks, D2R leads both identity metrics and reduces mean CLIP reference cosine distance by 55% relative to the strongest baseline. Project page: https://shjo-april.github.io/Depth2RGB/

Reading "While the man hunted the deer ran into the woods" requires giving up the analysis in which "the deer" is the object of "hunted". People who have just read such a sentence still produce more transitive structures, so the abandoned analysis primes later language. We ask whether language models keep using an analysis they have abandoned, and what in the model carries it. In 24 open models from 124M to 70B parameters, a garden-path sentence leaves two traces in the context. A residue of the abandoned analysis raises later object attachment, by about three quarters of the effect of a sentence in which the noun phrase really is an object, and it appears in sampled text, on held-out and human-normed items and for a second type of garden path. An adaptation toward the resolved structure scales with the sentence's disambiguation error and acts mainly on comprehension-like measures. Activation patching separates the two. Later text reads the residue from the role information in the noun phrase's states, which were computed under the object analysis before the disambiguation, and replacing these states removes almost all of the residue in the 32B and 70B models. The adaptation is read from the states that processed the disambiguation. Overridden statements show the same persistence (a corrected value keeps pulling later predictions in 8 of 9 models), and in most models the pull is carried by the positions after the statement. Finally, the human paradigm transfers only with controls for punctuation, because the models also write garden-path sentences themselves, which reverses the paradigm's raw statistic.

Science fiction is widely credited with unsettling gender. This article asks whether that holds at the linguistic level, in how readily gendered reference gives way to its counterpart, a neutral term, or other social category, and whether it separates the genre from general fiction. It does so by masking man/men, woman/women, and person/people across same-period science fiction and general fiction corpora and classifying RoBERTa's five likeliest substitutes by gender. Retention and displacement are permutation-tested over five resampling rounds, and the underlying sentences are read closely. Both corpora cross the gender binary in both directions, but unequally: where woman is masked, the model predicts a masculine term about twice as often as it predicts a feminine term where man is masked. Each mismatch marks a statistical surprise, a word the corpus did not expect there. Against this shared pattern, science fiction shows greater gender permeability than the general fiction corpus, placing man in feminine contexts more often than general fiction while the reverse rate holds steady. Its degendering tendency is stronger too, drawing both gendered terms toward the neutral. Beyond the binary, a masked woman also draws substitutes naming a social category and no gender far more often than a masked man does. In measuring the gender permeability of man, woman, and person in linguistic distribution and testing whether genre changes it, the study gives science fiction studies linguistic evidence for thematic claims and computational literary studies a pipeline applicable to other category vocabularies, such as race and class.

This paper introduces an energy-adaptive noise scheduling and whitening strategy for transform-domain diffusion models. Existing spectral diffusion methods account for the non-uniform statistics of transform coefficients through coefficient scaling, normalization, or frequency prioritization, while the forward diffusion noise schedule remains largely independent of the underlying spectral-energy distribution. We investigate whether the temporal evolution of the forward diffusion process should also follow the spectral organization of natural images. The proposed formulation combines global spectral whitening with energy-conditioned noise allocation that jointly modulates the injected noise according to the energy of individual transform coefficients and an image-dependent energy path over diffusion time. The resulting forward process preserves Gaussian transitions with closed-form marginals and remains compatible with standard DDPM and DDIM procedures without modifying the diffusion architecture. Experiments on CIFAR-10 demonstrate the contribution of the proposed energy-conditioned noise schedule and spectral whitening, reducing Fr\'echet Inception Distance from 142.48 for a compact DCTdiff U-Net variant to 100.45.

LLM 在企业 agentic 工作流中的采用，受制于覆盖广泛能力范围的评估：CRM 操作、异构文档检索、个人数据理解、多轮规划、工具使用、函数调用、长周期研究以及长上下文软件工程。如今这些评估分散在各自的仓库中，运行器、依赖项和评分惯例各不相同，这使得跨基准比较变得困难，也模糊了单一聚合数值实际代表什么。我们发布 EGI-BENCH，这是一个统一的 LLM agent 基准测试套件，将八个面向企业的 agent 基准（CRMARENA、HERB、PERSONABENCH、USERBENCH、MCP-UNIVERSE、MCPEVAL、BENCHFORCE、LOCOBENCH）整合在单一命令行界面之后，并提供两种受运行时预算约束的执行模式。在该套件之上，我们定义了 Enterprise General Intelligence（EGI）Score，这是一个六维框架（M1-M6），其各维度由从每个基准路由而来的细粒度子指标填充，而非采用基准级平均值。我们描述了该套件、子指标路由以及 quick/deep 预算，并报告了在 quick 模式下涵盖 GPT、Claude 和 Gemini 系列的十四个前沿及高性价比 LLM 的结果。总体 EGI 分数从 39.6（GPT-4o mini）到 65.6（gpt-5.5）不等，且没有单一模型在所有能力轴上占据主导：gpt-5.5 在任务结果、规划和检索 grounding 方面领先；Claude Sonnet 4.5 在工具执行与安全/拒答方面领先；

Large reasoning models (LRMs) generate long chain-of-thought traces before answering, creating significant inference overhead. Pruning can reduce this cost, but its effectiveness depends on the calibration data used to estimate parameter importance. Recent work calibrates on the model's own rollouts instead of generic dataset, but treats all reasoning tokens uniformly, regardless of whether they contribute to successful reasoning. As a result, pruning protects weights by statistical salience rather than by their contribution to correct reasoning, so weights behind erroneous computation survive as readily as those behind correct computation. These erroneous patterns then get carried into the pruned model, degrading reasoning quality, producing both lower accuracy and longer reasoning traces. We propose Outcome-Based Calibration for Large Reasoning Model Pruning (OBC-Prune) to close this gap. OBC first constructs difficulty-matched pairs of correct and incorrect rollouts from problems the model answers inconsistently. It then estimates the causal importance of each reasoning sentence through intervention-based analysis, quantifying how removing its influence affects subsequent predictions. These causal importance scores are converted into per-token weights that rescale the calibration activations used by one-shot pruning methods (SparseGPT, Wanda, ALPS), without modifying the underlying pruning algorithms. Experiments on DeepSeek-R1-Distill-Qwen 1.5B, 7B, and 14B models at 40\% and 50\% sparsity demonstrate consistent improvements over state-of-the-art calibration baselines across most model sizes and sparsity levels on MATH500, LiveCodeBench, and AIME 2025. These results indicate that preserving causally important reasoning circuits is a substantially more effective pruning objective than uniformly preserving observed activations.

Web agents are usually evaluated in live environments, where environment state and judge models drift between runs, so the same checkpoint rarely reproduces the same score, making controlled studies of training phenomena impractical. We present WebMRE, an offline benchmark of 541 tasks and 5,293 steps derived from successful WebArena trajectories, with fully audited test labels and a deterministic protocol that scores a checkpoint identically on every run without any environment. Each step pairs a human oriented guide sentence with a grounded action, enabling the first study of the mutual reinforcement effect between them in web agents. Averaged over three seeds the effect holds for both models in both decoding orders and grows with scale: jointly decoding a guide lifts element selection over an action only reference by 0.9 and 0.2 points for Qwen3.5-4B and by 1.7 and 2.2 points for Qwen3.5-9B. A mediation analysis shows that the guide is a causal channel rather than commentary: forcing the gold guide as a decoding prefix lifts action accuracy from .422 to .684, another step's guide collapses it to .055, and a paraphrase that renames the target still recovers half of the gain, so the channel carries instruction meaning and not only the label string. The same channel yields an offline reward that only a replayable protocol makes computable, though optimizing it from a strong checkpoint brings no gain yet. Our fine tuned models outperform GPT-5.5, Claude Opus 4.8, and Gemini 3.5 Flash, run zero shot, on every offline metric.

学习机械三重周期极小曲面（TPMS）超材料的信息性表示对于力学分析和结构设计十分重要，然而现有方法很少显式地用与力学响应相关的全局结构连通性信息来监督所学的表示。我们提出了一种拓扑感知表示学习框架，将持久同调融入DMAE主干网络，以显式编码全局结构连通性。持久性引导的掩码策略优先处理与持久拓扑特征相关的图块，而辅助的持久同调解码器则促使潜在表示保留关于连通分量、隧道和孔洞的拓扑特征数量信息。我们在两个互补的TPMS数据集上，通过族分类、力学性能分类、无监督聚类和回归来评估所学表示。在两个数据集上，我们的方法在结构族分类和力学性能分类方面均表现出强劲性能，族分类最高提升3.9%，同时取得了最佳的线性探测性能。更值得注意的是，无监督性能聚类在两个数据集上分别最高提升26.7%和51.5%，同时保持了强劲的回归性能，$R^2=0.993$。这些结果确立了持久拓扑作为表示学习的有效归纳偏置

准确的日前光伏（PV）功率预测既需要强大的时间建模，也需要遵循光伏系统的物理约束。纯数据驱动模型往往忽略底层的 PV 转换过程，并可能产生超出 PV 系统运行范围的预测。本文提出一种物理信息状态空间框架，将可微 PV 转换模型与选择性状态空间主干（PowerMamba）相结合。我们的方法遵循物理残差分解。物理模块基于太阳几何、辐照度、温度效应和逆变器限制提供结构化基线，而状态空间模型则在 24 小时时间范围内从数值天气预报（NWP）输入中学习残差修正。在一年公用事业规模 PV 数据上的实验表明，所提方法显著提高了预测精度：与仅物理基线相比，RMSE 降低 38.5%，MAE 降低 45.8%，并优于纯数据驱动模型。结果进一步表明，引入系统级物理信息带来最大的性能提升，而额外的细化仅带来边际改进。除精度外，该模型通过将预测约束至物理可行性来提高鲁棒性，在测量数据损坏时减少不切实际的预测。

This empirical study is an independent reproduction of the dissociation Zhao reported in 2026. The shape of a large language model's chain-of-thought entropy trajectory predicts whether the final answer is correct, while the magnitude of its total entropy drop does not. The dissociation merits reproduction because the magnitude half rests on a single 300-problem run with one model at one seed, while the shape half was reported at full scale on both benchmarks and on a second model family. Registered at OSF before any confirmatory run, the reproduction crosses the complete GSM8K and MATH-500 benchmark test sets with four open-weight models including one reasoning-distilled model of a kind the original did not test. The shape signal replicates. The magnitude signal divides by setting. On the anchor model the accuracy gap between monotone and non-monotone chains is +9.6 percentage points on GSM8K and +27.5 on MATH-500, while the rank correlation of the total entropy drop with correctness is -0.018 on GSM8K and +0.414 on MATH-500. On the reasoning-distilled model the binary form of the shape signal fires on about one chain in a hundred, too few to estimate the registered contrast, while the graded violation count remains predictive there. In an exploratory comparison the final-step entropy alone outperforms the binary shape flag in all eight model-by-benchmark cells by ROC area, and in six or seven by the risk-coverage area the original reports, depending on an integration range the original does not state. The study contributes a reproduction of the shape signal at full test-set scale under seven documented protocol differences, a map of the settings where the magnitude signal holds and fails, and measurements of four protocol dependencies the original does not report.

Speech language models have expanded automatic speech recognition (ASR) from transcription to general speech-to-text generation. Yet existing benchmarks still assume a single ground-truth transcript per utterance, an assumption that struggles when models produce readable, formatted, or structured outputs. In this work, we introduce Grounded Generative Speech Recognition (GSR), reformulating speech recognition as generating semantically grounded text instead of a single transcript. We define a grounded reformulation space that characterizes all valid textual realizations supported by the input speech, and establish a four-level hierarchy of reformulation capabilities, including literal transcription, canonical reformulation, semantic reformulation and speech-grounded structuring. Based on this formulation, we present GSR-Bench, the first comprehensive benchmark designed to evaluate generative speech recognition systems beyond exact transcript matching. GSR-Bench covers diverse scenarios including refined meeting transcription, cross-dialect translation, context-aware speech understanding, and robustness under silent or noisy acoustic conditions. To address the limitations of word-level evaluation metrics, we develop a multi-dimensional evaluation framework combining speech grounding assessment and LLM-based semantic judgment, whose reliability is validated through human preference studies. Extensive experiments across representative architectures, ranging from conventional ASR systems to end-to-end Speech-LLMs, reveal fundamental trade-offs between transcription fidelity and generative reformulation ability. Our analysis further identifies critical failure modes, including acoustic hallucination, unsupported generation, and instruction drift, providing insights into the design and evaluation of future grounded generative speech systems.

推测解码（SD）可加速大语言模型（LLM）服务，但其收益取决于工作负载的形态（如批大小和上下文长度）以及内容。基于模型的草稿器可针对任意内容提出短块，而无模型草稿器（如 n-gram 查找）在输出重复其输入时（例如代码编辑）能以几乎零成本提出长复制。自适应验证将整个批次的验证预算花在最可能被接受的草稿 token 上，但它仅支持一个基于模型的草稿器。无模型草稿器不提供逐 token 置信度，因此无法参与预算分配，部署时必须固定一个草稿器或在草稿器之间拆分 GPU。我们提出 \sys，一个在每一步同时运行基于模型和无模型草稿器的框架，并在自适应验证内、完全在 GPU 上为每个请求提供两者链中更优者，从而保持异步调度不受影响。关键思想是每个已验证草稿 token 的价格：该槽位在批次其他位置本可获得的有效吞吐量。一旦无模型草稿器的 token 携带校准后的接受概率估计（在通用文本上一次性拟合），这一价格就将两个草稿器的链放在同一尺度上。在 vLLM 中以 \dspark 和 n-gram 草稿器实例化后，\sys 在代码编辑、批大小为 4 时输出吞吐量达到 \dspark 的 1.10--1.62 倍，批大小为 128 时为 1.04--1.27 倍，而在温度为 0 的聊天场景中保持为其 0.97--1.01 倍。单一部署

我们将 attention 的指数核展开为包含 $n$ 项的截断 Taylor 级数，并将每一项分解为张量幂内积，从而将因果历史压缩为恰好 $n$ 个循环状态；每个状态每个 token 仅通过一次秩一外积更新，并且与上下文长度 $N$ 无关。该递推相对于截断核是无损的（float64 下为 $2.2 \times 10^{-16}$），并且状态在分块间可加（$5.7 \times 10^{-14}$），因此 prefill 可采用并行扫描。实测指数为：softmax 的 $t \propto N^{2.28}$，相比之下状态为 $N^{0.91}$，在 $N \approx 728$ 处出现交叉，在 $N = 8192$ 时为 $20.8\times$；每 token 解码延迟恒定为 $7.3\,\mu$s，而 softmax 为 $N^{1.04}$，在 $N = 65536$ 时达到 $76.0\times$。准确率随 $n$ 指数提升，而成本组合式增长，因为第 $m$ 个状态的宽度为 $\binom{d+m-1}{m}$——该上界由 SVD 确认——因此最优为二阶，$n = 3$，误差为 $6.5 \times 10^{-4}$，每个 head 有 $9{,}945$ 个元素。Query 和 key 归一化是前提：没有它，自变量会超出收敛半径，增加项也不再有益。

大语言模型（LLM）为将自然语言指令转换为信息物理系统（CPS）的任务计划提供了便捷接口。然而，有两个缺口阻碍了它们用于实际执行：生成的计划可能误解或遗漏用户意图；即使逻辑上合理的计划，在组件动力学、交互或安全约束下也可能不可行。现有基于LLM的规划方法往往缺乏关于计划失败原因的有信息量反馈，难以为相应修订计划提供指导。为应对这些挑战，我们提出了一种分层反馈框架，利用不同类型的反馈信息来引导LLM为CPS生成可执行计划。我们提示LLM同时生成结构化任务计划和相应的信号时序逻辑（STL）约束，并设计了一种可靠的算法来识别二者之间的确定矛盾，并提供反馈以修订不一致之处。在执行层面，当控制综合失败时，我们提取不同类型的诊断信息，例如被违反的约束和冲突的组件分配，并将其反馈以引导任务重规划。在七个异构多组件CPS场景上的实验表明，我们的框架能够生成可执行计划与可行的控制执行，满足原始自然语言任务，使用GPT-5时成功率达到87.9%。

Understanding the design space of streaming video diffusion is essential to exploring its potential for generation quality and computational efficiency. We develop a unified analytical framework that relates model and sampler choices, historical conditioning, execution scheduling, and training strategies. The framework accommodates a broad family of causal context-selection policies and makes their computational dependencies and training-inference alignment explicit. Within this design space, we study three representative policies: clean, same-level, and progressive history. On the full VBench prompt set, same-level and progressive history achieve aggregate scores of 85.24 and 85.60, respectively, compared with 84.45 for the clean-history reference. Long-video comparisons further show improved subject consistency and more coherent motion with progressive history. By allowing multiple denoising nodes to be processed together, progressive-history pipelining achieves $1.57$-$2.83\times$ steady-state DiT speedups under our evaluated conditions. We additionally find that LoRA adaptation of the DMD fake-score network improves generation quality using only 2.15% as many trainable fake-score parameters as full-parameter adaptation. Together, these findings show that fully denoised history is not a prerequisite for high-quality streaming generation and motivate the joint design of historical conditioning, execution, and training.

Mapping the distribution of traffic dynamics at high spatiotemporal resolution is a fundamental question in transportation research. Distributed acoustic sensing (DAS), an innovative seismic observation tool, emerges as a promising solution for real-time urban traffic monitoring at high spatial and temporal scales. Distributed acoustic sensing repurposes existing underground fiber-optic cables as dense, continuous sensor arrays, enabling passive and privacy-preserving monitoring of roadway traffic activity at meter-level spatial and second-level temporal resolution. This study examines whether integrating DAS and deep learning models can serve as a continuous and efficient urban traffic observatory for revealing urban traffic dynamics (i.e. traffic volume and congestion, event-driven changes) at high spatiotemporal resolution. Using a DAS deployment along a roadway network in the City of College Station, Texas, USA, this study develops a deep learning-empowered analytical framework that converts raw ground vibration waveforms into spatiotemporal representations, detects vehicle trajectory, and infers traffic states from aggregated traffic volume and speed. A hybrid training strategy combining synthetic and manually annotated DAS images is used to improve vehicle detection under noisy and congested conditions, with model outputs further aggregated to characterize system-level traffic dynamics.

Data quality now matters as much as compute for training language models. Much training data comes from human annotation of text, and interpretive annotation has no ground truth that could settle what is "accurate". Two lines of work respond to this. One combines annotators into a "ground truth" and measures how well they agree with each other; the other treats their disagreement as a signal. Both compare different people at one point in time. We measure something else: how well one reader reproduces their own reading of the same text over time. One expert human reader and three LLM families segmented three Sumerian myths and labelled the causal function of each segment with one of seven states. Across runs months apart, the human cut the text in much the same places but named the segments differently, in every myth. The models show no such consistent pattern: their gap between the two layers is positive in some myths and negative in others, and its size varies. The human's label changes are not random: the runs go through much the same functions but start them one step apart, while model runs start them at the same places. We argue that this pattern is a usable measure of data quality and a contamination check: a "human" annotation whose labels are as stable as its boundaries, and whose functions start in sync, looks like a model's.

Autonomous research agents aim to automate scientific workflows, from proposing ideas to conducting experiments and analyzing results. Yet current AI and research agents can propose more directions than available resources allow them to pursue. Moreover, each attempt could consume substantial resources, requiring agents to reconsider how to invest in subsequent research. Thus, deciding how to invest research effort strategically should be a defining capability of autonomous research agents. Accordingly, we introduce PrimeScientist, which jointly determines research direction and resource investment across successive research attempts. Specifically, we formulate this challenge of strategic research effort allocation as a sequential decision problem where remaining resources should explicitly guide the research policy. We first introduce an executable plan tree that preserves competing plans and their outcomes across attempts. Building on this representation, we propose an adaptive MCTS-based allocation policy that balances exploration and exploitation using experimental feedback and remaining resources. Comprehensive evaluations across AI research, systems and code optimization, and machine learning engineering show that strategic allocation improves research quality and sample efficiency together. Across 12 AI research tasks, PrimeScientist improves average reward by 10.3% with 50.6% fewer research attempts than AutoResearch under the same resource budget. We believe making research effort allocation an explicit optimization target establishes effective resource use as a core research capability for autonomous agents to drive scientific breakthroughs at scale.

The world population is growing rapidly, and technology is improving in parallel. Meeting the huge demand for food for these 7 billion people not only depends on increasing food production but also on reducing food loss. Crop losses due to disease affect both the food supply and the financial and economic stability of a country. Tomatoes are among the top food-producing crops globally, and a significant portion of this production is lost due to disease. People have used Machine Learning techniques for feature extraction and early diagnosis of tomato diseases, and nowadays, Deep Learning-based models are widely used for disease recognition. However, most existing models are highly parameter-intensive, which increases the time required for training and inference. As a result, while lightweight models are more suitable for user-friendly applications, they often show a reduction in performance. To balance performance and model size, we propose CoAtNet-DeepMoE, a Convolution-Attention hybrid architecture for rich feature extraction, further enhanced with a DeepSeek Mixture of Experts to substantially reduce the number of parameters without sacrificing accuracy. We evaluate our model on both balanced and imbalanced datasets from Kaggle and PlantVillage, demonstrating robustness and achieving 99.80% accuracy, 99.80% precision, 99.80% recall, and 99.80% F1-score on Kaggle, and 99.83% accuracy, 99.85% precision, 99.76% recall, and 99.80% F1-score on PlantVillage, representing state-of-the-art performance with only 2.47M parameters. The source code will be available at https://github.com/nadimbrur/CoAt-MoE.

Compressed-domain video captioning avoids full video decoding by operating directly on I-frames, motion vectors and residuals, trading a small amount of accuracy for a large gain in inference speed. CoCap established this pipeline using a CLIP vision encoder and a shallow BERT-style multimodal decoder. Both components predate substantially stronger alternatives. We ask a narrow, controlled question: how much of CoCap's accuracy is limited by these two components, and which of the two is the binding constraint? We replace the CLIP I-frame encoder with SigLIP2 and the BERT-style decoder with GPT-2, and evaluate three configurations (the original pairing, the encoder substitution alone, and both substitutions together) under identical data, sampling budget and optimisation schedule. All comparisons are made against our own reproduction of CoCap rather than its published numbers, because we train on a 4,999-clip subset of VATEX at a reduced sampling budget; absolute values are therefore not comparable with the original work. Our reproduction tracks the published result closely: CIDEr and METEOR run slightly above it (54.9 against 52.7; 23.4 against 23.2), BLEU-4 and ROUGE-L slightly below (29.7 against 31.4; 48.9 against 49.4). We attribute the differences to our evaluation subset rather than to any improvement in either direction. We find the two substitutions pull in opposite directions: SigLIP2 alone improves every metric (+4.5 CIDEr), while adding GPT-2 on top erodes that gain, because a pretrained decoder overfits 4,999 clips within two epochs. We additionally report inference latency for each configuration, since speed is the property that motivates compressed-domain captioning in the first place, and an accuracy gain purchased at a latency cost should be reported as such.

Legislative hearings are distinct communication events with their own human recognizable structure and logic, yet they have been analyzed as distinct phenomena beyond their transcripts. Understanding legislative hearings at a deeper level is a major step in analyzing the work and power relationships within democratic bodies. In this work, we present a novel end-to-end pipeline for parsing rhetorical structures from legislative committee hearings with the goal of improving legislative discourse analysis and democratic accountability. We also propose a rhetorical structure grammar and find that 55.21% of California state bill discussions in our dataset can be parsed with this grammar. We evaluate our pipeline with two utterance section classifiers on a random subset of over 2,000 annotated utterances corresponding to 100 hearings. We find that a logistic regression-based classifier with grammar-based output masking and utterance tag smoothing outperforms more complex models, achieving an F1 score of 52.8% against our ground truth discourse section labels.

人脸识别系统已普遍部署于日常场景中。为保护用户隐私，主流系统存储人脸模板而非原始人脸图像来进行身份验证。然而，人脸模板反演（FTI）技术能够从模板重建人脸，从而危及生物特征隐私。现有 FTI 方法仅适用于单一类型的模板，无法从未见类型中重建身份一致的人脸。为解决这一局限，我们调研了主流模板提取器，并依据模板分布将其分为五类：单位球面、角度线性、严格单位球面、锥形、非均匀超球面。受此观察启发，本文提出 GMoE-FTI，一个由真正的 MoE 机制驱动的通用 FTI 框架，区别于负载均衡的多分支替代方案。它遵循三阶段训练策略：（1）独立训练五个映射专家（ME）和一个匹配模块（MM）；（2）通过用 MM 的输出对每个 ME 的输出进行加权来执行协同推理，从而生成与预训练 StyleGAN3 兼容的统一潜在编码；（3）对所有模块进行端到端联合微调，以提升身份一致性并增强跨专家协同。在 LFW 基准上的大量评估验证了 GMoE-FTI 相较于 SOTA 方法的优越性。在包括 AdaFace、CosFace、CurricularFace、ElasticFace、SphereFace2、Spher 在内的七种主流未见模板类型上进行了评估

像素空间世界模型最近作为 GUI 智能体的一种有前景的方法出现，它在执行前预测未来屏幕，随后重新感知这些屏幕以进行动作选择。然而，这种两阶段范式在 GUI 环境中面临两个关键挑战：(L1) 像素重建的高计算成本，以及 (L2) 需要重建大量且多样的外观级细节，超出了动作选择所需的范围。在实践中，这种屏幕生成目标可能引发恒等捷径问题，即未来状态预测在很大程度上保留当前状态，而不是捕捉动作引起的变化。在这项工作中，我们通过将预测目标从全屏重建转变为潜在空间中的动作诱导未来信息，重新表述了 GUI 世界建模。具体来说，我们利用大型多模态模型作为潜在世界模型，其共享表示支持两个专用头，分别用于动作选择和预测性转移建模。动作头直接将这些潜在表示转化为动作选择，而转移头则捕捉以动作为条件的未来信息。为了缓解恒等捷径，我们进一步引入了一种动作感知对比目标，鼓励转移表示捕捉具有判别性的动作引起的变化。在离线和在线 GUI 设置中的大量实验表明，我们的方法在成功率上实现了高达 31.9% 和 40.6% 的相对提升。

Egocentric gaze prediction enables many downstream applications but remains challenging, as human gaze is inherently stochastic. This stochasticity is constrained by structured temporal dynamics alternating between fixations and saccades, top-down influences from tasks, and bottom-up visual saliency. Based on this observation, we introduce GazeFlow, a framework that directly models gaze as a joint distribution of temporal gaze positions conditioned upon both top-down and bottom-up information. In particular, GazeFlow uses conditional flow matching (CFM): a learned velocity field iteratively transports a Gaussian noise sample into a plausible gaze trajectory drawn from this joint distribution. The velocity field is conditioned on bottom-up visual features extracted by a video encoder and on top-down task information obtained by globally querying these features. On standard datasets, GazeFlow achieves state-of-the-art performance on per-frame metrics, and the generated trajectories align better with human gaze temporal dynamics.

Peer review is under strain from the growth of submissions, and review comments written by humans or AI can misstate what a paper says. We introduce ReviewFact, a benchmark that labels such factual errors in review comments under one definition across NLP, machine learning, and the natural sciences, and that measures cost and latency alongside accuracy. A frontier LLM checks comments well but must read the whole paper for each check, while cheap decision models, which score whether a statement holds in a text, fall far short on whole comments that mix several implicit assertions. PremiseCheck is a framework that combines the two. An LLM rewrites each comment as explicit premises without reading the paper, and a decision model checks each premise against the paper. When each comment is checked on its own, PremiseCheck reaches accuracy comparable to GPT-5.5 across all three fields at about half its cost, and on NLP reviews at a quarter of its cost and less than half its latency.

长到短（L2S）模型合并旨在将长思维链模型强大的推理能力与基础模型的简洁风格相结合，但现有依赖线性插值和一阶激活统计的方法无法捕捉损失景观如何响应参数位移，导致在困难任务上的准确率下降。我们提出低秩分层合并（LHM），这是一个免训练框架，用轻量级二阶曲率信息替代启发式统计。LHM 包含三个组件：（1）低秩 Hessian 近似，通过 Hutchinson 估计器在随机输入上高效估计逐层 Hessian 迹，无需任务特定数据；（2）层二阶敏感性（LSOS）指标，将 Hessian 迹与任务向量范数结合，以量化逐层融合难度；（3）双约束系数映射（DCCM），将 LSOS 分数转换为层自适应合并系数，同时满足融合误差上界和全局长度缩减目标。在 1.5B 和 7B 模型上的六个数学基准实验表明，LHM 优于 ACM-TIES 和 ORCA 等最先进方法，相较长思维链模型实现超过 4% 的准确率提升，并将输出长度最多缩减 93.1%。

Text-to-image diffusion models are increasingly distilled into few-step variants and being deployed to enable fast inference. However, their ability to generate harmful or undesired content poses significant safety risks. Data-driven unlearning methods suppress targeted generations by fine-tuning model weights using specialized unlearning objectives. Crucially, these objectives implicitly rely on multi-step denoising dynamics, an assumption that breaks down for few-step distilled (FSD) models, resulting in ineffective forgetting. Furthermore, performing unlearning on the non-distilled base model and subsequently re-distilling it to obtain an unlearned FSD model incurs substantial computational and time overhead, making it impractical in many settings. Hence, we address this limitation with a preference-driven unlearning framework that revisits Direct Preference Optimization (DPO) for diffusion models. We show that standard DPO and its unlearning derivatives, formulated around noise-prediction error, transfer poorly to FSD models due to their altered generation dynamics. To overcome this, we introduce a modified preference optimization formulation explicitly aligned with the few-step generation properties, enabling direct concept removal in FSD models while preserving few-step efficiency and maintaining strong retention of desirable (non-targeted) capabilities. We evaluate our framework primarily on identity and NSFW (nudity) removal tasks and also extend our method to object-level unlearning. Extensive experiments demonstrate consistent and effective forgetting, and strong retention performance, establishing our method as a practical and principled solution for unlearning in FSD models.

Reward maximization alignment methods for discrete diffusion models have primarily focused on steering the reverse process, either by influencing token logits or by selecting favorable sequences at intermediate steps. These approaches largely treat inference as a unidirectional process, lacking mechanisms for revisiting undesirable token selections. We introduce Spectral Feedback, an algorithm that selects edit-positions in a feedback loop, allowing the model to iteratively correct its own generations. This approach leverages the mask structure of discrete diffusion models by re-masking and re-sampling tokens, analogous to image editing methods that reintroduce noisy latents and re-run the reverse process. While prior alignment methods focus on what token labels to assign to maximize a target reward, we instead treat which tokens to revisit as the central alignment problem. Selecting edit-positions is challenging because edit effects are interdependent: the impact of modifying one token depends on which others are edited simultaneously. We define an edit-set as a set of token positions to re-mask and re-sample. Motivated by prior work on sparse interactions in biological systems, we find empirically that edit-set value functions for protein inverse folding admit sparse Fourier representations. This structure enables Spectral Feedback to efficiently learn and optimize the value functions for edit-position selection. Spectral Feedback is model-agnostic and can be applied to pretrained, test-time aligned, and fine-tuned diffusion models. For all of these models, the algorithm improves alignment performance without modifying the underlying generative process. Applied to inverse folding with a protein stability reward oracle, it achieves a 32.3% increase in stable proteins for a pretrained model, 24.8% for Best-of-10, and 5.8% for a state-of-the-art RL fine-tuned diffusion model.

Reliable confidence estimation is increasingly central to the trustworthy deployment of language models: a calibrated estimate of the probability that an output is correct decides what to ship, what to escalate, and what to retry. Existing confidence estimators, however, share one design premise: they only read the current inference process, either by introspecting on it, scoring its token probabilities, or resampling it. We argue that the current inference is not a sufficient basis for confidence. We propose XConf (eXperiential Confidence): estimating confidence together with the model's accumulated experience. The experience is stored as a record of the model's own graded past episodes, each holding the task, the model's reflection, its stated confidence, the outcome, and a lesson written once the grade arrived. Given a new task, XConf's Recall stage retrieves past episodes on similar tasks met with a similar stated confidence, and reads off their historical success rate; its Reflect stage shows the model this record, has it name its recurring failure mode, and restate a confidence now informed by its own track records. Our estimator is format-general, requiring no logit access or weight updates, and costs only one answer generation. Across nine benchmarks spanning reasoning, coding, multimodal QA, and interactive agents, and four models from three families, XConf beats or matches ten-sample self-consistency in discrimination (AUROC) on 23 of 24 comparisons, with much lower calibration error (ECE), at a tenth of the generation cost. Used for selective prediction, abstaining on the 10% least-confident episodes raises the delivered success rate by up to 8.7 points on agent tasks. We therefore see experiential confidence estimation as a new paradigm for future general-purpose confidence estimation.

Lifelong editing of LLMs requires storing thousands of edits after acquisition. A widely used family of approaches attaches one LoRA adapter per edit, which preserves behavior but grows linearly in storage. To address this challenge, we propose LadderEdit, a method that compresses each LoRA adapter after it is acquired. Each edit is first stored at low rank as a cheap sketch. We then check whether this sketch still satisfies the rewrite, generalization, and locality contract on probe prompts. Edits that pass keep the sketch; those that fail are promoted to a higher rank along a ladder until the contract is met. Because every edit retains some representation, coverage is maintained, and only hard edits consume more rank. Across ZsRE, CounterFact, and WikiBigEdit benchmarks on LLaMA-3-8B, Mistral-7B, and Qwen2.5-7B, LadderEdit tracks exact LoRA storage at 5.2x less memory and remains effective at 50,000 sequential edits.

许多系统将变分自编码器（VAE）的潜在空间读取为离散码，并依赖于每个潜在变量被读回为其写入时的码。Shannon（1948）将这种共享码本视为给定；而在 VAE 内部，没有任何机制强制这一点。我们证明，标准训练报告中的任何一项都无法揭示写入码和读出码在编码数据上是否一致。我们构造了两个 VAE，它们能完美重构每个输入、精确推断，并且在每个逐样本 ELBO 和每个码直方图上都匹配，但一致度在第一个中为 1，在第二个中为 0。然后我们表明，ELBO 自身的各项限定了测得的一致度能移动多远：变分间隙限定了向精确推断的偏移幅度，而 rate 限定了向从先验生成的偏移幅度。对于高斯后验，平均场编码器在间隙中付出一个精确代价，该代价取决于其坐标轴选择。在 probabilistic PCA 上，该间隙在我们测试的全部 75 种设置中认证了一个训练后的编码器，一致度在 0.76 到 0.97 之间。在 240 个训练后的 VAE 上，由 rate 和数据集大小设定的下限提前排除了 66 个模型在 rate 侧认证的可能性，而它们没有一个通过认证。这些结果给出了一个简单协议 Certify-or-Measure，适用于任何将 VAE 读作码的流水线。

情绪分类器经常在其训练数据之外应用，假设同名标签（如恐惧或惊讶）在不同资源中含义相同。我们针对 GoEmotions、XED 和 BRIGHTER 共有的五个标签检验这一假设，在相同输入上以及跨语料迁移下比较模型。在相同文本上，一个恐惧分类器将 71% 标记为恐惧，另一个则为 8%，但两者在文本排序上大体一致。在 30 组资源-标签比较中，训练标签流行度解释了这类尺度不匹配的大约一半，而拟合截距消除了其中大部分。重新校准无法修复的问题更大：迁移模型总体上损失了 55% 的原生区分能力，分类器之间排序一致性与该损失仅弱相关，而对于惊讶，各资源对哪些文本表达该情绪存在分歧。因此，训练资源的选择可能改变结论：在两个外部推文集合上，BRIGHTER 和 XED 训练的分类器对哪个集合表现出更多惊讶给出相反答案，即使在秩归一化之后，两个方向的区间均不包含零。共享的情绪标签名称并不确立一个可互换的测量目标。

A world model learns to forecast how a physical system evolves from recorded trajectories, yet the systems it imitates obey physical laws that are neither fully supplied nor reliably respected. The model may create energy, drift or diverge over long rollouts, and answer a changed law query using the law observed during training. We ask how much general physical structure must be hard coded into a world model, and how much system-specific physics can then be learned from data, for four properties to hold simultaneously: second law compatible dissipation, correct responses to interventions on physical parameters, stability out to one hundred times the training horizon, and robustness to disturbances. The imposed structure is general: dynamics are generated from the gradient of a learned energy through a fixed reversible operator, the energy is restricted to a confining class, a one way port can remove energy but never inject it, the drive channel is known, and the intervened parameter enters through a separable map. The model learns the energy functional, constitutive relations, dissipation rate, and couplings. Across an electromagnetic cavity, a particle in cell grid, and a shallow-water fluid, models with roughly nine thousand parameters recover constitutive functions with unit slope, separate conserving from dissipating worlds by four orders of magnitude using a single set of weights, and transfer changes in sign, magnitude, rate, and gravity to unseen values, where equal-capacity models without the same structure perform at chance or worse. A nonlinear constitutive law is recovered with its curvature preserved and predicts a held-out intervention $2$-$17\times$ better than a converged linear model.

Large language models (LLMs) are widely used to answer factual questions, but their answers can contain errors and must be fact-checked claim by claim. Yet the resulting score depends on three design decisions: which questions the claims answer, under what conditions each claim is judged, and which sources count as evidence. No existing benchmark specifies all three. We introduce JSFactCheckBench, a Japanese benchmark that specifies all three, with 1,913 claims from LLM answers to questions about misleading information that circulated on Japanese social media, each annotated for factuality at a reference time and for importance to the question, using evidence from preselected Web sources. Across six LLMs, all verify claims directly answering the question less accurately than supplementary claims, a gap that aggregate scores obscure. Scores also drop when models must locate the relevant content in longer or multiple sources and, for all four open models, when the surrounding question and response are added. These weaknesses are visible only because the benchmark records the conditions behind each label. A benchmark that leaves them implicit cannot reveal them.

基于大语言模型（LLM）的智能体近期在单轮任务中展现出卓越性能，但在长时程任务中仍表现不佳，尤其是在没有先前训练的环境中。单个 LLM 的一个关键局限在于难以处理长上下文，这源于其在持续决策与推理能力上的不足。尽管配备长期记忆库的现有多智能体系统（MAS）可以部分弥补这一差距，但它们大多依赖知识复用或摘要，这种做法往往受限于直接检索而无法针对当前任务进行任何适配，同时检索到的记忆可能包含大量有待预处理的历史噪声。相比之下，以任务特定指导为特征的 MAS 通常缺乏专门为该指导设计的长期记忆。为应对这些挑战，我们提出 M-Plan，一个融合长期与短期记忆的闭环 MAS 框架。通过结合全局-局部与过去-当前指导，M-Plan 在保持即插即用设计的同时提升了智能体在多轮任务中的表现。在任务执行过程中，智能体会接收任务特定的里程碑与洞见，并且仅在出现困惑时获得局部提示，这些共同构成全局-局部指导。这些指导的生成与利用也构建了短期记忆。任务完成后，与当前任务相关的信息会被优化或重新组织

Financial search is a highly demanding task for LLM agents, requiring not only a correct final answer but also temporally valid information retrieval, authoritative source selection, entity and period alignment, unit and definition consistency, and verifiable evidence for all conclusions. Existing benchmarks predominantly evaluate only the final answer, making it difficult to localize errors or assess whether an answer is well-founded. To address this gap, we introduce FinFIRST (Financial Information Retrieval, Sourcing and Traceability), the first financial benchmark to jointly evaluate answers and supporting evidence through atomic rubrics. FinFIRST comprises 123 expert-authored tasks spanning a graduated difficulty spectrum, constructed from aggregate patterns of real-world financial scenarios through an 18-field taxonomy, a six-axis coverage blueprint, a registry of 138 financial sources, contributions from over 50 finance experts, and a six-stage quality-control pipeline. Each task is accompanied by an evidence-grounded reference package decomposed into atomic criteria across three dimensions: raw-information acquisition, source verification, and computation and answer formation. We evaluate 15 model configurations under a unified tool setting. Claude-Opus-5 achieves the highest atomic score of 87.59%, while GPT-5.6-Sol attains the highest strict pass rate of 71.54%. Computation and answer formation consistently lag behind raw-information acquisition across systems. FinFIRST retains final-answer correctness as the primary objective while making the supporting research process measurable, verifiable, and diagnosable.

Central bank press conferences are not merely information releases --- they are structured narratives. We study whether the shape of sentiment within a statement, not just its average tone, carries policy-relevant signals. Constructing sentiment arcs for ECB and Fed press conferences along three dimensions --- monetary stance, economic outlook, and uncertainty --- we assess their predictive content for policy rate changes, inflation expectations, and forecaster disagreement. Our findings show that arc shape robustly predicts rate decisions beyond lexicon-based benchmarks at both institutions --- it is not merely whether a statement sounds hawkish or economically optimistic on average, but how these sentiments are sequenced and emphasized across the statement, that carries the policy signal. Arc features also shape how professional forecasters update inflation expectations and how much they disagree, pointing to a receiver-side effect distinct from the direct policy signal. These findings suggest that communication design --- the sequencing and emphasis of policy language across a statement --- is a first-order feature of the policy signal, not a second-order refinement.

LLM system capabilities depend not only on model parameters but also on the executable harnesses that organize information acquisition, tool use, reasoning, and memory. Recursive harness evolution can automate their design, but how it can reliably discover generalizable systems from limited, noisy, and repeatedly reused feedback remains unclear. We study this question in agentic video understanding, where rich multimodal evidence and coupled mechanisms make harness optimization challenging. We first introduce VideoHarness-BRSI, a recursive self-improvement framework that evolves video harnesses around a frozen vision-language model. Analyzing over 4,000 evolved candidates, we find it competitive with hand-engineered systems but identify three interconnected failure modes: validation overfitting, search-budget dilution, and the early-experience trap. We trace all three to a common root: the same imperfect validation evidence drives both harness revisions and the proposer's search strategy. We therefore propose VideoHarness-SRSI, an evidence-governed framework in which every search decision rests on evidence that is transferable, sufficient, and gathered before the search commits. It restricts the proposer to content-free feedback, informs exploration with instrumented warm-up experiments, and develops mechanisms in isolated probe-race branches before factorial composition and joint refinement. Across four video understanding benchmarks, VideoHarness-SRSI improves the equally weighted mean held-out score by 6.2% relative to VideoHarness-BRSI and reduces the mean validation-test gap by 37.9%. Reliable harness evolution thus depends on how search is organized and informed, not only on how well it optimizes validation performance.

Vision-language models (VLMs) remain largely unreliable on panoramic dental radiographs and can rely on learned anatomical priors rather than evidence in the image. This is particularly problematic for tooth localization and spatial reasoning, and fine-tuned dental VLMs can retain the same spatial biases. We present AgenTeeth, a model-agnostic, tool-augmented framework that grounds frozen VLMs using seven specialized dental vision experts. A question-aware orchestrator selects the relevant tools, whose detections are mapped to FDI tooth numbers or anatomical regions and returned as structured findings together with annotated image overlays. A fresh synthesis call then answers the question using this evidence, without fine-tuning the underlying VLM. On MMOral-OPG-Bench, AgenTeeth improves four backbone VLMs by 12.9-23.0 percentage points over their baselines. Our strongest configuration reaches 65.66% on open-ended VQA, compared with 45.35% for OralGPT-Plus. The advantage also holds at matched scale: a frozen Qwen2.5-VL-7B-Instruct with AgenTeeth reaches 48.11%, exceeding OralGPT-Plus built on the same backbone after supervised fine-tuning and reinforcement learning for tool use. We release the framework, all seven expert models, and a dentist-annotated dataset for alveolar bone-loss detection in panoramic radiographs.

利用基于 LLM 的智能体自动生成高性能内核，有望在开发效率上带来显著提升，但仍存在两个障碍。首先，虽然增加与执行环境的多轮交互对于提升生成质量至关重要，但持续的上下文累积会迅速超出模型的有效上下文窗口。其次，通过 GRPO 训练长时程任务时，由于其依赖终端结果评估，天然会产生稀疏奖励，严重阻碍了步级信用分配。为解决这些障碍，我们提出了 ATRL（Ascend Triton Reinforcement Learning），这是一种后训练方法，将 GRPO 改进为面向长时程内核生成的、压缩感知的分段级目标。我们还设计了一个轻量级智能体框架，提供级联的编译、正确性和性能分析反馈，并在智能体强化学习训练和评估期间注入昇腾专用知识库，以弥补基础模型缺失的平台专业知识。在昇腾硬件上执行的 NPUKernelBench 和 KernelBench 评测表明，ATRL 在正确性和加速比上均持续超越包括 GRPO 在内的所有基线。

从现有的二维纸样创建数字服装仍然是一个耗时的过程。尽管纸样已广泛以数字形式提供，但在服装 CAD 软件中组装它们需要手动定义缝纫操作，并为悬垂模拟安排纸样裁片。我们提出两种基于学习的方法，以二维矢量纸样作为唯一输入，自动完成这两项任务。我们的自动缝纫模型在片段级别运行，利用具有多个预测头的 Transformer 编码器。我们的新颖架构支持多对多、有序的片段对应关系，以及多层服装所需的折边、重叠缝，以及成品虚拟服装所需的缝份和线迹类型等构造属性，这些此前的方法都忽略了。一个互补的裁片级模型预测身体区域分配、层和放置坐标，以初始化悬垂模拟。我们在一个经过整理的包含 4,035 个专业组装服装实例的数据集上评估了这两个模型。缝纫模型精确恢复了 63.3% 的测试服装的所有缝线和折边拓扑，而排列模型在 66.2% 的测试服装中满足所有裁片的指定放置标准。我们还将这两种方法集成到服装 CAD 软件中，展示了成功预测的服装的自动装配。

Dexterous robotic manipulation remains a major challenge in robotics, exemplified by piano playing, which requires coordinated control of fingers and provides a demanding testbed for reinforcement learning (RL). Although previous work established a benchmark for robotic piano playing, it is restricted to two hands and does not support unified evaluation. To address this gap, we introduce OmniPiano, a benchmark that provides one to up to five Shadow Hands and supports standard, robust, safe, and multi-agent RL within a shared piano-playing task family. OmniPiano is developed for scalable and diverse piano-playing tasks with configurable perturbations, explicit safety constraints, and decentralized cooperation settings. Particularly, to facilitate usability and extendability, OmniPiano adopts a highly modular design and provides comprehensive tasks to support RL study of task performance, robustness, safety, and cooperation in dexterous control. With at least 912 task settings and 36 baseline algorithms, OmniPiano further incorporates LLM-based agents to broaden the evaluation scope and reveal new insights. Extensive evaluations show that state-of-the-art RL algorithms and frontier LLM-based agents still struggle on challenging piano-playing tasks, even under standard RL settings, highlighting substantial room for improvement in dexterous control. The code, dataset, tutorial, and demo are available at https://sail-research-lab.github.io/omnipiano/.

Neural controlled differential equations（Neural CDEs）沿通过插值离散观测构造的连续路径演化隐藏状态。稀疏测量使观测时刻之间的输入无法确定，而符合同一数据的可容许连续路径可能产生不同的终端状态。我们针对速度有界的平面驱动项刻画这种歧义。相容终端状态集合允许一个首阶 zonotopic 近似，该近似由可行有符号面积和传输后的 Lie brackets 决定。其到标量预测上的投影产生一个可计算、针对特定目标的歧义证书。我们还研究减少剩余不确定性所需的测量次数。对于非退化平面 CDEs，在速度有界时，确定性自适应采样策略下的最优最坏半径按 $N^{-1}$ 缩放；当速度和加速度均有界时按 $N^{-2}$ 缩放，其中 $N$ 是观测区间数。合成实验使用可容许连续补全评估该证书，并考察路径正则性的作用。

Production LLM agents are evaluated repeatedly as they evolve, but full agent benchmarks are costly to rerun. We study efficient recurring evaluation for a production analytics agent serving tens of thousands of monthly active users and report first-hand deployment experience. Using 574 historical runs of the production benchmark, split chronologically into calibration and held-out periods, we compare random sampling, historical caching, fixed representative subsets, and IRT-based adaptive testing. The results show that multidimensional 2PL adaptive testing achieves the best overall score fidelity: executing 200 questions, 38.5% of a full run, yields 1.03 pp of MAE. We nevertheless deployed difficulty-stratified fixed subsets because of their operational simplicity, and show they transfer without recalibration to five other agent families and remain stable across calibration windows as short as one day. Drawing on this deployment experience, we report practical recommendations for recurring production-agent evaluation.

Sharpness-aware minimization (SAM) doubles the cost of every training step, yet its benefit concentrates where training ends. We study where an expensive training mode should be spent and propose Temperon: a plain-SGD explorer for the first 43% of the epoch budget, then one scheduled hand-off that gives the entire final cosine anneal to a SAM-wrapped Muon refiner. On CIFAR-10/100, SVHN and Tiny ImageNet (five seeds, times reported as epochs-to-target times an idle-GPU-calibrated epoch cost), Temperon matches the best full-time-SAM recipe on accuracy everywhere while reaching the hardest common target 35%, 34% and 32% sooner on three of the four, and sits a tier above the published SAM+SGD recipe at level cost. Ablations make the attribution exact: the Muon refiner is worth +0.85pp with everything else fixed; the explorer's shape and its restarts are worth nothing, and we withdraw them as contributions. Re-running the closest rival, late-phase SAM, at matched budget shows the frontier: it is fastest to every mid-level target, but the tier the Muon refiner buys (0.83 on CIFAR-100, 0.97 on CIFAR-10) is reached by no SGD-refined method in any seed, and on Tiny ImageNet, where Muon buys no tier, the rival simply wins -- the measured boundary of the method. The allocation law transfers to GPT-2 pretraining (full-SAM quality at -29% wall-clock) and GLUE fine-tuning (never worse than full-time SAM at a third of its SAM cost). Two constants organize the economics: skipping SAM early buys a fixed credit, and a Muon epoch costs 1.50x a SAM+SGD epoch on all four datasets. Finally, the hand-off cannot be timed from the trajectory: under cosine schedules the accuracy curve is plateau-then-surge, so the information lives in the schedule, making the scheduled switch principled rather than convenient. Code and a pip-installable implementation are released.

Alcohol intoxication is a leading contributor to fatal and severe-injured e-scooterist crashes. Current countermeasures, such as temporal restrictions or pre-ride cognitive screening, cannot continuously assess an e-scooterist's physical motor control or impairment in real time. We conducted a controlled experiment in which 25 participants rode an instrumented e-scooter through a test track while sober and at two targeted blood alcohol concentration levels (0.05% and 0.08%). The e-scooter was instrumented with a six-axis inertial measurement unit (IMU), and throttle and brake lever position sensors, all sampled at 100 Hz. Two complementary signal features were computed: normalised permutation entropy, which quantifies temporal complexity, and standard deviation, which quantifies signal amplitude. Repeated measures correlation identified seven kinematic features (all IMU and throttle signals) whose entropy decreased (p < 0.001) while standard deviation increased (p < 0.01) with increasing intoxication, indicating that intoxicated riders shift from continuous, low-amplitude micro-corrections to fewer, high-amplitude reactive corrections. An entropy based multi-class logistic regression classifier, evaluated through leave-one-participant-out cross-validation, achieved 85% overall accuracy and a weighted one-vs-rest area under the receiver operating characteristic curve (AuROC) of 0.94, with a sober-vs-high AuROC of 1.00. Steering rate and lateral acceleration were the most important predictive features, indicating that alcohol induces a distinct collapse in lateral equilibrium during riding. Ultimately, these results demonstrate that onboard kinematic sensing combined with entropy-based signal analysis can reliably distinguish sober from intoxicated e-scooter riding, providing a foundation for automatic intoxication detection systems that preserve mobility for sober riders.

联邦无监督域适应（FUDA）旨在利用分散的源域将模型适配到无标签目标域，同时保护数据隐私。然而，现有 FUDA 方法通常需要多轮通信、依赖复杂的聚合策略，并且常常难以应对噪声伪标签和源知识不一致的问题。为应对这些挑战，我们提出了 SKD-ETR，一种新颖的一次性 FUDA 框架，结合了平滑知识蒸馏（Smoothed Knowledge Distillation, SKD）与基于指数移动平均的教师精炼（Exponential moving average-based Teacher Refinement, ETR）。SKD 利用由源模型集成生成的平滑软伪标签在目标域上训练学生模型。这降低了过度自信、缓解了噪声并提升了鲁棒性。ETR 进一步通过指数移动平均（EMA）将每个源模型的参数向学生模型插值，从而将目标域知识回传给教师模型。这种双向精炼在无额外通信开销的情况下提升了伪标签质量和学生模型的泛化能力。SKD-ETR 通过随机初始化学生模型并执行单轮蒸馏过程，消除了复杂的聚合。在 OfficeHome、Office-Caltech 和 DomainNet 上的大量实验表明，SKD-ETR 在保持通信与计算高效性的同时取得了有竞争力的性能，并且在噪声监督下具有鲁棒性。本文代码可在 https:// 获取。

We ask whether behavioral constraints acquired through post training remain explicitly reportable. Using constrained recipe generation as a testbed, five banned ingredients enforced via LoRA fine tuning of Llama 3.1 8B Instruct we compare supervised fine tuning (SFT) and Group Relative Policy Optimization (GRPO) against an untrained baseline on a four tier Constraint Awareness Benchmark. Averaged over three seeds, both methods raise behavioral compliance from 4% to about 90% while reducing explicit constraint reporting below the untrained model (0.48/5 to 0.16/5 for SFT, 0.07/5 for GRPO) and eroding retained third person knowledge (93% to 36% for SFT, 14% for GRPO; p less than 0.01 between methods). Contrary to our initial hypothesis, the reward based signal is the more destructive of the two: a reward that penalizes banned ingredient tokens regardless of framing learns a context independent suppression rather than a self directed constraint. A context conditioned reward designed to teach the self to other distinction fails, collapsing toward inclusion in both framings. Probing prompt time hidden states recovers per ingredient avoidance at 83.8% (layer 24 MLP), but only 6.4 points above a per ingredient base rate predictor (77.4%), and the model's own verbal self report is more accurate still (87.8%). A positive control adding explicit self description examples does not restore reporting. The failure is therefore specific to enumerating constraints on request, not a general loss of access to them.

The aggregation of many lay estimates often outperforms individual expert judgment, a phenomenon known as the wisdom of crowds. While this is usually attributed to the independence of estimates, an even stronger effect arises through deliberation: averaging the consensus estimates of small deliberating groups outperforms the classical wisdom of crowds, with individual judgments themselves also becoming more accurate after deliberation. Whether these improvements transfer to large language models deliberating amongst themselves is unknown. Here we adapt a three-stage deliberation paradigm previously used with human participants for use with large language models from three different families, and test it across four domains of increasing real-world stakes: visual numerical estimation (Study 1), peer review of machine-learning papers (Study 2), detection of hidden malicious behavior by an artificial intelligence agent (Study 3), and sports forecasting against a real prediction market (Study 4). Across domains, deliberation reduced collective error beyond passive aggregation of independent responses, and post-deliberation individual judgments retained this collective gain. Notably, the advantage required model diversity: groups composed of clones of a single model did not benefit from deliberating. These results establish machine deliberation as a general-purpose aggregation mechanism, and point to diversity as an active ingredient.

本研究探讨了广义边界表示与迁移学习，用于射流冲击冷却中空间变化对流换热系数（HTC）的物理信息神经网络（PINN）估计。分段线性（PWL）与神经网络边界估计器由瞬态内部温度测量值推断得出，并在雷诺数 5000、10000 和 15000 下与 CFD-CHT 参考 HTC 分布进行对比评估。基于 $H^1$-半范数梯度正则化的离散 Adjoint-FEM 公式提供了独立的反演方法对比。迁移学习在两种源-目标条件下进行了研究：相同雷诺数下不同噪声水平之间的迁移，以及从受控高斯分布问题到不同雷诺数下由 CFD-CHT 导出的目标问题的迁移。对于具有 5% 测量噪声的代表性 Re = 5000 目标，二次、PWL 和神经网络 PINN 分别实现了 5.08、5.31 和 5.19% 的相对 HTC 分布误差，而广义表示将 Re = 15000 目标的误差从二次 PINN 的 16.76% 降低至 14.41% 和 14.46%。迁移学习使固定雷诺数迁移的目标训练时间缩短 57.86-81.71%，使高斯到 CHT 的迁移缩短 50.59-84.94%。其对重建精度的影响取决于目标：固定分布的 PWL 迁移与从头训练保持在 0.30 个百分点以内，而高斯到 CHT 的迁移产生

A principal who deploys a binary classifier with an abstention option must decide which queries the mechanism abstains on. The right choice depends on the adversary. A gaming adversary already knows the classifier and tries to manipulate features across the boundary, so the principal does best by abstaining on queries close to that boundary. The same boundary-localizing rule is the worst possible choice against a learning adversary who does not know the classifier: each abstention now tells the adversary that the boundary is nearby, which is enough to drive a binary search. We analyze this tension. The two natural defenses, abstaining at a fixed rate and abstaining near the boundary, are Blackwell-incomparable: neither can be simulated by post-processing the other's responses. The number of queries needed to reconstruct the boundary to error $\eps$ is $\tilde\Theta(d/\eps)$ under the first defense and $\Theta(d \log(1/\eps))$ under the second, where $d$ is the VC dimension of the classifier family and $\tilde\Theta$ suppresses factors polylogarithmic in $d$ and $1/\eps$. The first rate is a worst case over query distributions; no reconstruction algorithm can close the gap at the distributions that attain it. We characterize the Pareto frontier between the two defense objectives, and confirm both rates on seven binary-classification tasks spanning tabular, image, and language-model-feature inputs: label-plus-counterfactual access extracts the boundary with up to $200\times$ fewer queries than a published label-only baseline.

多模态智能可以被视为世界与智能体之间通过交互进行的人类引导协同进化。然而，扩展这一过程面临两个瓶颈：构建异构、可验证的环境成本高昂，且人工进行 Harness 迭代无法跟上日益增长的任务多样性和复杂性。一个关键洞察是，具有不同模态和表面形式的任务可以共享相同的底层能力需求。因此，我们提出 The Trinity，一个通过 Creator、Explorer 和 Evolver 三种角色将世界生成与 Harness 进化耦合起来的递归自我改进框架。Evolver 借助经过验证的执行轨迹，将任务组织的基础从表面类别转向共享需求，把原始任务树重组为分层的能力树。我们提出 Capability-Driven Real2Sim2Real：对于同一节点上的异构任务，Creator 将其共同需求抽象为简化、可验证的模拟世界以进行能力学习，同时在原始任务上的评估衡量迁移效果。沿着这棵能力树，Explorer 通过选择性继承逐步进化出可复用的能力模块，形成一个 Harness 树，其经过验证的更新在固定拓扑内跨迭代累积。由此得到的各个 Harness 专精于不同能力，这促使我们提出 Capability-Aware Harness Router，它利用公开任务信息从 t 中选择兼容的模块配置

Multimodal synthetic datasets combine structured attributes with free text, but are often evaluated separately. Such metrics can remain high after tabular--text pairings are disrupted. We present a projection-based evaluator for tabular--text synthetic data. A fixed sentence encoder maps text to embeddings, \(k\)-means converts them to cluster states, and tabular variables are represented as categorical or quantile-binned states. Real and synthetic contingency tables are compared using Jensen--Shannon divergence (JSD), normalized mutual information (NMI), conditional JSD (cJSD), and joint-state entropy. We also report text-to-attribute (T2A) utility and a holdout-calibrated proximity flag rate (PFR) as a representation-level diagnostic. A text-permutation control preserves both marginal distributions while disrupting their pairing. Experiments on Amazon Reviews, Kiva Loans, and the Employment Scam Aegean Dataset show that modality-specific scores remain high under this control. The projection diagnostics detect disruption when real projected dependence exceeds a permutation baseline, but are less informative for weak or sparse projections. Some conditioned LLM baselines also exhibit stronger measured dependence than the corresponding real-data projections. These results support explicit cross-modal evaluation with permutation baselines and coverage reporting.

While various multimodal methods combining imaging and tabular data for Alzheimer's disease (AD) diagnosis were proposed, they are often limited in generalization across cohorts. In-context learning (ICL) has demonstrated excellent generalization performances and high flexibility in foundational tabular models such as TabPFN. To extend TabPFN's ICL to multimodal AD analysis, the main obstacle is that TabPFN is meta-trained on synthetic tabular priors that do not naturally match the statistical structure of image-derived features. We propose M$^2$PFN, an end-to-end framework that turns this tabular foundation model into a multimodal AD predictor. M$^2$PFN (i) performs differentiable inference through TabPFN's transformer, back-propagating task gradients into 3D-MRI and tabular encoders; (ii) aligns the two modalities into a shared subspace, via disentanglement and a contrastive objective, matched to the ICL engine's prior; and (iii) folds in a frozen tabular-only prediction through a learnable gated shortcut. Because the ICL engine stays frozen, its in-context mechanism is preserved for test-time generalization, while end-to-end training shapes the encoders into features it can exploit. On ADNI ($n=2240$, three-class CN/MCI/AD), M$^2$PFN attains $65.55\%$ macro-F1 and $82.21\%$ macro-AUC, surpassing a comprehensive set of unimodal and multimodal baselines. By swapping only the head for a TabPFN regressor, the same architecture regresses baseline MMSE on a $1250$-subject sub-cohort to test MAE $1.743$, outperforming every multimodal baseline. On two external cohorts (OASIS-3 and SCAN) with no retraining, M$^2$PFN achieves the best AUC and the lowest MMSE MAE across all baselines, and transfers even when the cognitive instrument changes.

3D foundation models recover video cameras and geometry in one forward pass, but some of the strongest are up to scale. Joint people-scene reconstruction then requires two missing outputs: metric scale and persistent person identity. We ask whether one up-to-scale foundation representation can support both through lightweight adaptation. Exact metric labels are scarce, but unlabeled in-the-wild video is abundant. We use people in curated web video to initialise the solution: a posed metric body and 2D keypoints give an approximate, closed-form scale pseudo-label. These pseudo-labels pretrain a Scale Readout, which is then fine-tuned together with a lightweight adapter using exact metric supervision from standard real-video training splits. At inference the head predicts metric scale from foundation-model tokens, without the ruler or its teachers. For person identity, we probe the pretrained foundation model alone and find evidence that its intermediate query-key features encode person correspondence across frames. In most evaluated moving-person clips, a mid-layer token prefers that person over the vacated location and other people. A tiny projection reads this correspondence; together with metric pelvis motion and proposal confidence, it drives dustbin-aware Sinkhorn association of per-frame bodies. WildHSR combines both readouts to reconstruct metric cameras, scene and people from monocular video. Each window is predicted feed-forward; analytic association and Sim(3) composition connect windows. On EMDB-2, WildHSR is the first feed-forward method in the published comparison to beat the best optimization-based WA-MPJPE and RTE while leading feed-forward methods on all three world-frame metrics. On RICH, it leads feed-forward people-and-scene methods on WA-MPJPE and W-MPJPE. The complete pipeline runs at 10.1 fps on one GPU.

现代实验越来越多地产生丰富的治疗后对象——例如文本、图像和视频——它们可能为延迟、昂贵或仅在一部分参与者中观察到的结果提供早期替代信号。预训练编码器可以把这类对象转化为强大的通用嵌入，但该嵌入并非为因果估计器所需的估计任务而训练，可能会强调与因果目标无关的信息。我们提出 SIRNET，这是一种替代信息驱动的表示适配框架，用于存在非结构化治疗后替代指标时的基于神经网络的因果估计。从预训练的替代嵌入出发，SIRNET 学习一个共享适配器，联合用于结果预测、处理倾向以及分臂替代边缘化，从而超越通用预训练目标。由此得到的协变量表示和替代表示为这些 nuisance 任务提供共同接口。最终的治疗效应估计和置信区间由交叉拟合的正交得分获得，在所述 nuisance 率条件下达到半参数效率界。我们还在合成和半合成实验中验证了 SIRNET 的性能，并在 Anti-Amyloid in Asymptomatic Alzheimer's Disease (A4) 试验中，使用冻结的预训练脑 MRI 嵌入作为长期结果的替代指标进行了验证。

Historical Polish is well documented as a language but annotated in machine-readable form only to about a million words for the period this paper covers; the rest sits behind optical character recognition of variable quality. We present Wieszcz-XIX, a corpus of 6.75 billion tokens (about 3.1 billion words) in 294,369 documents, most of them periodical issues, of Polish published from 1800 to 1918, assembled from Wolne Lektury and the Internet Archive by a pipeline that filters, deduplicates, audits for post-1918 leakage and splits at the document level. It is over three orders of magnitude larger than the annotated corpus of the same period, and we quantify its defects: recognition corruption against a false-positive floor, near-identical duplication, which is removed, and post-1918 leakage, which is excluded from the training corpus itself down to a known residue of 0.04 to 0.38% of its bytes, found in the transcribed source, so the published corpus is the trained one document for document. On a hand-corrected sample the character error rate is 0.68% where the text is legible, and 45% of the sampled passages cannot be corrected. On it we train a ladder of decoder-only models from 47M to 349M parameters from scratch, and measure their temporal boundedness. Against two modern Polish base models, one far larger, the 349M shows a crossover, as does the 107M against the comparator of its size: post-1918 vocabulary costs them about 3.1 bits per byte more than period vocabulary, a gap the comparators do not show, and period vocabulary costs them fewer bits than it costs the comparators. Shown period text, the models keep its spelling and the comparators only partly. Adding parameters gains about twice as much as a second pass over the data. We release the corpus, code and weights. Content warning: the models reproduce period prejudice, including antisemitic statements.

视觉问答（VQA）模型已知存在偏置问题，即过度拟合频繁模式的规律。此前，研究者只考虑了语言偏置。最近，多模态偏置受到关注。然而，多模态偏置究竟是什么仍不清楚。本文研究了计数问题与非计数问题中的多模态偏置。我们发现二者具有不同的内在机制，因此其偏置来源于不同的原因。我们为 VQA 计数机制提供了有价值的解释：\textbf{数量信息来源于特定维度上多模态特征的幅值。}因此，计数偏置来源于视觉侧，并通过模态融合传播到多模态特征；而非计数偏置主要来源于语言偏置，并带有轻微的视觉偏置。因此，我们提出一种分别处理它们的方法：用于缓解双模态偏置的变分与分解方法（VDB）。它提取受视觉侧和语言侧不同比例影响的双模态偏置。我们引入利用变分自编码器原理的变分偏置模型，以主要学习带有轻微视觉效应的语言偏置；并引入一个双模态偏置模型，用于结合语言侧和视觉侧的双模态偏置。在 VQACPv2、VQACPv1 和 GQAOOD 数据集上的大量实验证明了我们主张的合理性和方法的有效性。我们还对 VQAv2 上性能受损的原因给出了解释。

Parallel decoding speeds up masked diffusion language models by writing several tokens per forward pass. Published commit rules choose these tokens by testing the predicted distributions. Under sampling the written tokens are drawn, and these tests do not see them. We find that answer quality follows the reliability of the tokens actually written. An answer's log probability splits into this reliability and the dependence within each step. We propose PACT, a training-free commit rule that draws the tokens first, certifies the set it will write and spaces the positions of each step apart. This certificate holds at every temperature, and we show that no rule with the same guarantee commits more positions. Experiments use four models of the LLaDA and Dream families and HumanEval, MBPP, GSM8K, MATH500 and IFEval benchmarks. PACT reaches the quality of published commit rules with fewer forward passes. Spacing removes positions from each step, yet it usually lowers the total cost because later steps then write more tokens. At the same speed, published decoders also write a word twice in a row far more often than PACT. These results suggest that commit rules should certify the tokens they write. Code is available at https://github.com/LiZaiyuan0619/PACT.

Breast cancer is the leading cause of cancer-related death among women in Brazil, and the time between the request and the release of mammography reports directly influences adherence to screening, making the agility in processing these reports a critical factor for early diagnosis. In this context, this study compares the performance of a Large Language Model (LLM) with manual extraction performed by a team of health researchers in identifying clinical findings from mammography and breast ultrasound reports written in Brazilian Portuguese. Named Entity Recognition (NER) was applied through Prompt Engineering using a few-shot strategy, employing the Gemini 2.5 Flash model, selected from preliminary exploratory tests with four candidate models. The Gemini 2.5 Flash model demonstrated the best performance, achieving a Macro F1 of 0.91 and a Micro F1 of 0.98. The subjective validation, in which 29 exams of different formats were evaluated by health researchers using a Likert scale, yielded an agreement index of 93.1%. The model outperformed human extraction in overall Macro F1 (0.91 vs. 0.72), as in four reports the model correctly identified information that had been omitted or incorrectly recorded during manual extraction, demonstrating its potential as a complementary verification tool alongside specialists. The results confirm the hypothesis that LLMs, when instructed through Prompt Engineering, can achieve performance comparable to or superior to manual extraction by health professionals.

Agentic marketplaces are emerging where AI agents with varying capabilities autonomously complete specialized tasks for buyers. A major challenge of such marketplaces is that buyers cannot easily determine which agent will perform best on their tasks. Reported benchmark scores may be difficult to verify or compare across tasks, software, and budgets. We introduce LEGIT, a credentialing protocol connecting certification, reputation, and proposed marketplace allocation. Certification binds measured quality and cost per solved task to an agent configuration, task domain, evaluation budget, and evidence through a signed record. Reputation links records of past task outcomes to the same identity, subject to the reliability of the reported feedback. Buyers and agents can verify credential records and inspect optional visual profiles. Evaluations reveal cost differences between agent configurations with similar observed task success, and show that comparisons depend on the evaluation budget. These results support binding performance measurements to the tested configuration and resource limits. A complementary analysis quantifies the deposits and fees required for reputation manipulation under a stated Sybil attack model.

Large Language Models (LLMs) have made remarkable progress in the processing and modeling of many languages. Yet, unlike human multilinguals, they exhibit surprisingly limited cross-lingual knowledge transfer. While this limitation is well documented, its origins during multilingual training remain unclear. We pretrain 360M- and 7B-parameter LLMs and show that poor cross-lingual knowledge generalization emerges during pretraining and persists under standard interventions. To isolate its cause, we employ a controlled bilingual pretraining setting using two copies of the same language, sharing identical text and token segmentation, but mapped to disjoint token spaces. We find that disjoint tokens alone are enough to induce knowledge compartmentalization, even between identical copies of the same language, establishing disjoint token spaces as a fundamental barrier to cross-lingual knowledge generalization. Guided by this understanding, we suggest mapping languages into a shared token space by simple word-wise translation and find it substantially improves cross-lingual knowledge generalization, recovering up to 12.6\% of native-language learning efficiency --- 14$\times$ the baseline.

Removing an invisible watermark and concealing the forensic evidence are distinct objectives: successfully disrupting the embedded watermark does not imply that the removal process is forensically undetectable. When verification fails, removal traces can provide complementary evidence for provenance and ownership verification, whereas their absence leaves the cause of the failure ambiguous. Existing methods are typically evaluated by watermark suppression and perceptual quality, while forensic stealth is rarely considered. We therefore study watermark-attack-specific few-shot forensics: for each known pipeline, a specialist can separate its outputs from paired clean and unattacked watermarked controls. Separate Attack-vs-Clean and Attack-vs-Watermarked evaluations prevent watermark-presence shortcuts. Image-aligned and prompt-matched controls are used for post-hoc and generator-integrated schemes, respectively. In this work, we introduce RemTraceNet, which fuses constrained residuals, local relations, FFT/Haar statistics, and block-DCT evidence at native resolution. Across 23 removal pipelines and 10 watermark configurations, we evaluate native 256 x 256 and 512 x 512 inputs. With 100 attacked training images per pipeline, the three-seed TPR@1%FPR, macro-averaged over attacks and watermark configurations, ranges from 82.75% to 88.17% across resolutions and control types. Under the condition of same labels and protocol, RemTraceNet outperforms retrained SRNet, ZhuNet, and SiaStegNet baselines by 10.80--15.68 percentage points. Extensive experimental results show that erasing a watermark and erasing evidence of its removal are distinct challenges, and that removal traces remain learnable under limited supervision.

长期个性化智能体需要跨模态、跨不断演化的对话状态对分布式历史进行推理。现有扁平跨模态方法过早压缩视觉证据，且难以应对随时间失效的观测，同时忽视了统一多模态与跨会话证据的必要性。关键在于，原始多模态记忆本质上是纠缠且异构的，使得细粒度视觉定位与跨模态推理在计算上难以承受。为此，我们提出潜在多模态记忆推理（LMMR），一种新颖的神经符号框架，将长期多模态记忆组织迁移至连续潜在空间。具体而言：(i) LMMR 不将图像转换为文本并依赖静态检索，而是将文本隐藏状态与视觉 patch 特征映射到共享的 Sparse Autoencoder 概念空间，建立稀疏激活作为通用语义地址，以动态构建轮级多模态记忆图。(ii) 随后，图编码器在任务特定的潜在子图上聚合跨模态与时间依赖关系，将图读出作为连续残差 rollout 注入生成主干，同时保留原生视觉 token 以保留细粒度细节。在九个 Mem-Gallery 任务和十二个 MemEye 设置上的大量实验表明，LMMR 在捕捉复杂多模态依赖关系方面显著优于最先进的基线，同时能够 f

Large language model (LLM) assistants are being deployed to clinicians across health systems, and judgments about their readiness rest largely on benchmark scores, most of them derived from examination questions or curated cases. A benchmark predicts performance in deployment only to the extent that its items resemble real use, yet whether benchmarks reflect the work these systems receive has rarely been measured. Here we analyze 127,833 queries sent by 6,342 physicians, advanced practice providers and nurses in 35 specialties to an institutional assistant during an eight-month roll-out. We characterize each query with RCQ-Map, a clinician-validated framework grounded in taxonomies of clinical questions and of LLM evaluation, which records its task, intent, answerability, missing information and potential harm. Documentation and administration (36.2%) and knowledge retrieval (28.9%) made up nearly two-thirds of use, and diagnosis 3.7%; more than a third of queries could not be answered well as posed. Applying RCQ-Map to 58 public benchmarks drawn from major evaluation suites and frontier model reports, which we assemble into the Clinical AI Benchmark Atlas, showed that the median benchmark contained no documentation requests and shared 31% of the task mix of real use, less than an even spread across task categories would. Benchmarks in suites designed to resemble clinical practice were individually no closer to real use than those used in frontier model reports. Benchmark scores therefore say little about how clinical AI performs on most of the work it is actually given, and evaluation should be matched to real clinical use.

Parameter-efficient fine-tuning adapts pretrained models at a fraction of the cost of full fine-tuning, yet most low-rank adapters are single-task and represent each weight update multiplicatively, leaving no explicit account of what is shared across tasks and what is task-specific. We introduce JIVEAdapter, a multi-task "additive" low-rank adapter inspired by statistical Joint and Individual Variation Explained (JIVE). JIVEAdapter decomposes every weight update into a Joint structure shared across all tasks plus a per-task Individual structure, penalizes the Individual structures to be near-orthogonal to the Joint so shared and task-specific signal stay "interpretable" and separated, and allocates rank adaptively across a shared Joint pool and a per-task Individual pool. The Joint is learned once, jointly over a task group or incrementally, one task at a time, then frozen and reused as a prior for new tasks without retraining the shared part. On GLUE and SuperGLUE with DeBERTaV3-base, JIVEAdapter is competitive with strong single-task and multi-task low-rank baselines at a matched per-task effective rank, without extra modules such as MoE, and when a related held-in task exists its frozen Joint serves a held-out task by reusing that task's Individual with only a cheap per-direction scale, otherwise training a small new one.

语言模型智能体与以往的智能系统有着根本区别，它们能够从上下文中的自然语言中学习。这一能力使智能体可以依据反馈迭代地改进自己的答案，而这正是现代智能体工作流中一种普遍的操作模式。然而，社区主要用 Pass@1 或 Pass@K 等指标对模型进行排名，这些指标无法体现模型如何利用对先前尝试的反馈。我们提出 Seq@K，这是一种自然的评估协议，用于衡量模型能否在 K 次连续尝试内解决问题，且每次失败后都会收到反馈。与考察长程规划的多轮智能体基准不同，Seq@K 隔离并形式化了“依据反馈进行改进”这一核心能力。我们在 7 个基准和 16 个前沿模型上评估 Seq@K，结果表明 Seq@K 揭示了一个显著不同的模型选择格局：在 Seq@K 下，(1) 模型排名往往不同，(2) Pareto 前沿允许更高效的配置，(3) 选择最优模型配置所带来的收益相对于 Pass@K 可能扩大或缩小。我们进一步表明，分析 Seq@K 下的模型轨迹有助于刻画迭代改进行为，例如模型处理反馈中所提问题的速率，或它们坚持自身答案的僵硬程度。最后，我们研究了反馈的性质如何影响模型的迭代改进，表明从业者必须审慎地设计和基准化