每天知道
你的领域
出了什么

Sampled-token on-policy distillation (OPD) efficiently transfers capabilities from teacher to student using student-generated tokens, requiring teacher probabilities only for sampled tokens. Yet it frequently suffers from diversity distillation failure: the student's pass@1 improves while its pass@$k$ plateaus, failing to inherit the teacher's diversity. To explain this, we introduce First-Order Local Entropy Influence, a signed first-order proxy that decouples each update's entropy effect into the teacher--student log-probability gap and the student's local probability structure, and empirically links entropy contraction to negative-influence positions. Motivated by this, we propose Influence-Directed Adaptive On-Policy Distillation (IDA-OPD): rather than relying on costly full-vocabulary Forward-KL objectives, it preserves entropy-expanding updates while replacing entropy-contracting ones with divergence-adaptive advantage shrinkage, using only the teacher's sampled-token log-probability. Experiments on reasoning-oriented distillation show IDA-OPD consistently improves pass@$k$, inheriting the teacher's diversity through distillation, matches the strongest teacher-informed methods at strictly lower cost, and broadly maintains vanilla OPD's pass@1, all without full-vocabulary teacher information.

Interactive world models extend video generation from offline clip synthesis toward persistent simulation of interactive virtual worlds, enabling applications in games, robotics, embodied agents, and XR. Achieving stable long-horizon interactive generation, however, remains challenging, as the model must simultaneously preserve scene geometry, dynamic consistency, and camera control while supporting real-time autoregressive generation. Building upon Matrix-Game 3.0, we present Matrix-Game 3.5, as shown in Figure 1, which advances real-time interactive world generation toward geometry-aware and long-horizon consistent simulation through three key improvements. First, we propose a unified geometry-aware memory framework, whose patch-memory and tiled-PRoPE components introduce no additional learnable parameters, combining explicit 3D patch retrieval with projective camera conditioning to enable geometry-consistent camera control and faithful long-horizon scene recall. Second, we introduce a static-dynamic disentangled world representation that separately models static scene geometry and dynamic subjects, preserving both geometric consistency and subject identity throughout long-horizon generation. Third, we develop a two-stage progressive real-time distillation framework that converts a bidirectional diffusion model into a few-step causal generator through Perceptual Flow Matching and curriculum based Self-Rollout DMD, enabling minute-long real-time interactive generation. Extensive experiments demonstrate that, with a unified training corpus spanning Unreal simulation environments, open-world games, and internet videos, MatrixGame 3.5 achieves strong performance in long-horizon scene recall, precise camera control, subject consistency, prompt-driven world generation, and stable real-time open-world interaction.

Text-to-image (T2I) safety guardrails fail to generalize equitably to non-standard dialects. Evaluating 23,080 paired prompts across five English dialects, we formalize this failure as the dialect penalty, where filters trigger based on linguistic surface features rather than semantic intent. Text-level filters fail in opposing directions: NSFW-T over-flags benign dialect prompts and LatentGuard over-flags toxic ones (bias gaps up to +28.29 pp), while the OpenAI Moderation API under-detects them. A controlled typo ablation confirms this penalty originates from flagging dialectal features, not generic out-of-distribution sensitivity. The pixel-level generator is largely dialect-agnostic; the penalty enters at text processing and cascades unevenly to post-hoc guardrails. We show this bias tracks training data imbalance and is mitigable via group-balanced retraining, with an ablation attributing the gain to balanced exposure rather than to the worst-group objective of GroupDRO (group distributionally robust optimization). Current pipelines systematically fail dialect speakers, an equity failure masked by mean accuracy benchmarks. Our official code and dataset are publicly available at https://github.com/minguinho26/dialect-penalty-t2i. Content Warning: This paper contains offensive, toxic, or disturbing text prompts and generated images.

Text-to-image models learn associations between concepts - in the case of this paper, people's professions, which we refer to as roles - and visual attributes. These associations can underpin many observed forms of stereotypical bias. A key open question in this area is whether these associations are stable or change when visual representations of people in professional roles are placed in different prompted contexts. We introduce ContextBias, a controlled evaluation framework, and ContextBench, a benchmark spanning 92 roles and 1,656 semantically controlled prompts, designed to isolate the effect of contextual variation on role-linked visual representations. Evaluating four state-of-the-art models on 66,240 generated images, we find that placing a role in a semantically unrelated context does not suppress role-linked attributes; instead, cross-role attribute concentration increases (pooled BI $+0.047$). Demographic cues, characteristic garments, and role-specific tools remain highly prevalent across context-free, related, and unrelated conditions, and are robust to semantic prompt reformulation. Scene composition and camera framing show the greatest context-sensitivity. These findings reveal a form of stereotypical persistence that remains largely invisible to context-free evaluations, highlighting the need for controlled contextual variation in bias benchmarking. Code and dataset: https://huggingface.co/datasets/shaghayegh/ContextBias , https://github.com/Sina-Emami/ContextBias

Controllable image relighting is an important problem in image editing, and hand-drawn scribbles provide an intuitive interface for specifying the desired illumination. However, existing methods do not establish a consistent and effective mapping between scribble inputs and relighting results, limiting their ability to control illumination intensity, chromaticity, and complex spatial distributions. We address this limitation by introducing a material-decoupled illumination representation, termed the Lumi Map, which establishes an explicit mapping between user scribbles and the resulting illumination, thereby improving both relighting accuracy and controllability. Specifically, we use a renderer to synthesize source image-Lumi Map-relit image triplets and train the model to predict the target relighting result conditioned on the Lumi Map. To mitigate the domain gap introduced by synthetic data, we further perform reconstruction training on real relighting pairs, improving the model's generalization to real-world images. Finally, we present Dior-Light, an image relighting method controlled by hand-drawn strokes. Extensive experiments demonstrate that our method outperforms existing approaches in relighting accuracy and enables effective control over illumination intensity and chromaticity on in-the-wild images.

We present the first systematic study of the resilience of text-to-video (T2V) diffusion models under random hardware-level faults. While T2V models are widely used for automated video generation due to their ability to produce high-quality, temporally coherent, and realistic videos, their iterative denoising process and spatiotemporal dependencies introduce unique failure modes. We perform an extensive fault-injection study covering both computational and memory faults across three T2V models and a representative benchmark. Our results show that (1) a single fault can degrade overall performance by up to 3.7%, with semantic correctness more affected than perceptual quality; (2) memory faults are more damaging than computational faults, high-order exponent bits are particularly vulnerable, and the widely-used bfloat16 is more susceptible than alternative formats; and (3) 7-28% of faults cause visible artifacts, including semantic changes such as added objects, suggesting that single faults are sufficient to alter output semantics. Our findings reveal reliability risks in deployed T2V systems and motivate further research on improving fault resilience. Code: \href{https://github.com/ztcoalson/T2V-Resilience}{https://github.com/ztcoalson/T2V-Resilience}.

Real-world image generation often involves multi-turn editing, where users iteratively modify small regions while most image content remains unchanged. However, existing diffusion transformer (DiT)-based editing pipelines recompute the entire image at every turn, causing substantial redundant computation. Existing DiT acceleration methods further ignore semantic correspondence across prompts, leading to unnecessary recomputation or unsafe reuse that harms editing quality. To address this, we propose RegionCache, a semantic-aware reuse framework for multi-turn image editing that selectively reuses diffusion states from unchanged regions. RegionCache detects reusable regions through semantic overlap between consecutive prompts and cross-attention localization, and adopts an adaptive reuse schedule based on prompt similarity and contextual consistency. Experiments on PixArt-alpha demonstrate that RegionCache achieves 1.43x--2.55x end-to-end speedup while maintaining comparable image quality. Code is available at https://github.com/hebutBryant/RegionCache.

Agentic workflows in materials science relying on hosted commercial models face severe reproducibility, economic, and data-privacy constraints. To explore fully local agentic science, this work evaluates an open-weights Qwen3:4B model executing an autonomous scientific pipeline across varying hardware constraints. Applied to pentagonal two-dimensional materials, the system extracts parameters from unstructured text, translates them into density functional theory (DFT) inputs, and drives simulations to convergence under a strict neurosymbolic architecture where agents propose and deterministic code disposes. The workflow is guarded by verbatim text grounding and multi-pass inference unions to counteract hardware-induced structural collapse. Evaluated against 201 expert judgements, the extractor achieves 95.7% precision (95% CI 90.3-98.1%) and 67.3% recall (59.8-74.0%), ensuring extracted parameters are strictly factual. However, precision identifying absent parameters does not exceed 47.0%, establishing that the measured omission rate constitutes a loose upper bound on true literature incompleteness. Across three hardware configurations, complete GPU residency governs extraction quality more fundamentally than weight or cache precision, raising Matthews correlation from 0.414 to 0.530 at fixed quantisation and to 0.560 with an unquantised cache. A corpus-scale audit indicates only 19 (33.3%) of the 57 studies are reproducible in principle, reporting every method parameter needed to re-initialise the calculation. Driven to convergence, the workflow reproduces published lattice constants with a mean absolute relative error of 2.3% where the relaxed structure retains its prototype, establishing that lightweight open-weights models can reliably drive autonomous agentic workflows when bounded by deterministic code gates.

Longitudinal cohort studies produce repeated data that enable the assessment of time-varying association patterns between exposures and health outcomes. Classical linear mixed-effects models (LMMs) can accommodate a large variety of association patterns while accounting for the irregularly spaced, partially observed measurement. But they require the analyst to pre-specify the functional form linking the exposure history to the outcome. We propose the Neural ODE-LMM, which embeds a Neural Ordinary Differential Equation (Neural ODE) within the linear mixed-effects framework: a learned vector field encodes covariate trajectories into a continuous-time latent state that drives both the fixed- and random-effect design, while preserving the standard LMM observation model. This retains classical likelihood-based inference while learning complex, potentially cumulative, covariate effects flexibly. All parameters are estimated by maximising a penalised marginal likelihood. To quantify covariate effects, we introduce contrasts of counterfactual predictions that compare the expected outcome under alternative covariate trajectories with variance estimated via the delta method. In simulations, the model recovers both instantaneous and cumulative-burden effects without prior specification of the functional form. Applied to the Trois-Cités (3C) cohort, a population-based study of 7{,}324 participants, the method reveals trajectory-dependent associations of BMI and fasting glucose with cognitive decline.

Future wireless systems are expected to transform the surrounding space from a passive propagation medium into a smart electromagnetic environment, where engineered surfaces control wave propagation, support wireless sensing, and create programmable electromagnetic fingerprints. A key challenge in realizing this vision is the inverse design of metasurfaces for tailored electromagnetic propagation. While forward analysis evaluates the response of a known geometry, the inverse task starts from a prescribed scattering signature and seeks a physically realizable structure that produces it. This inverse task is inherently nonlinear and often high-dimensional, while candidate solutions may be non-unique and provide no direct indication of practical realizability. Here, we introduce a conditional diffusion framework for inverse design of dielectric resonator metasurfaces from target angular scattering patterns. Trained on T-matrix simulated geometry-response pairs, the model learns a conditional distribution of geometries instead of a deterministic mapping, enabling multiple candidate designs for the ill-posed inverse problem. The best generated metasurface achieves a mean percentage error of 1.39%, outperforming CMA-ES optimization (4.1% after 10 h) while requiring only about one minute for after-training inference. The model also produces lower error distributions than deterministic neural baselines for out-of-distribution spectra, highlighting the potential of diffusion models for efficient metasurface design.

We propose Discrete Diffusion Bridges (DDB), a novel framework designed to resolve the fundamental spatiotemporal misalignment of standard discrete diffusion in image translation and generation. By corrupting data into a pure mask state via a random schedule, the conventional forward process induces a twofold misalignment: spatially, this pure-mask destination entirely discards the rich structural priors of the source image; temporally, the random masking order inherently contradicts the ``easy-first, hard-last'' decoding mechanism used during inference. To address this, DDB constructs a direct and efficient trajectory between domains. Spatially, we introduce a hybrid absorption mechanism that redefines the absorbing state to a stochastic mixture of mask and source tokens, effectively injecting source prior as spatial anchors into the latent space. Temporally, we design an information-guided noise schedule that quantifies semantic variation to prioritize the corruption of high-information regions at earlier timesteps. This ensures the model learns to resolve difficult semantic changes using robust context from invariant regions. Extensive experiments validate the versatility and robustness of our framework across diverse generative paradigms. DDB effectively balances edit alignment with structural fidelity across both text-guided semantic manipulation and pure structural image translation, while inherently complementing text-to-image generation and guaranteeing robust high-quality decoding under extremely low sampling steps. Code and models are available at \href{https://github.com/HKU-HealthAI/DDB}{https://github.com/HKU-HealthAI/DDB}.

Understanding which training samples influence a generated image is an important problem in generative modeling. In flow matching, training samples influence the generated image through the velocity field along the generation trajectory. Removing samples to examine their counterfactual influence changes the velocity field, and the resulting effect on the final image depends on how the change propagates through the trajectory. Consequently, local changes in the velocity field do not necessarily predict the final counterfactual effect. This work investigates attribution in flow-matching models through a hybrid analytical--learned approach, and uses it to derive trajectory-based attribution scores at the cluster level. We evaluate these attribution scores using independently retrained leave-one-cluster-out (LOCO) models, and compare with several attribution baselines using two different flow-matching latent spaces. Our experiments show that semantic similarity constitutes a strong baseline, while the closed-form trajectory-based attribution is competitive in some metrics without requiring counterfactual retraining or model gradients. Our results show that attribution in flow matching depends not only on semantic similarity to training samples, but also on the latent representation, trajectory dynamics, and how influence is propagated to the final output.

Text-to-spatial audio generation, such as text-to-First-Order Ambisonics (FOA), provides a convenient way to create spatial audio for billion-dollar gaming and film industries. However, existing text-to-FOA methods are largely data-driven and may produce audio that violates acoustic relations between source direction and distance. They also separate descriptive and parametric control, forcing users to trade usability for precision. In this paper, we present PhysWave, a physics-guided latent diffusion model for controllable text-to-FOA generation. PhysWave unifies natural-language and trajectory control through a shared waypoint-caption representation, and augments diffusion training with two differentiable acoustic priors: spherical-harmonic direction consistency and inverse-square distance consistency. To support dynamic spatial generation, we further construct a 300K-clip FOA dataset with diverse sound categories and source trajectories. Extensive results show that the proposed priors help PhysWave generate spatially consistent FOA audio while maintaining competitive audio quality. Further analyses show that these physics priors improve spatial consistency during training and can also be used as inference-time guidance for training-free spatial refinement.

To mitigate the time complexity of generative models, one-step generative models have recently emerged through direct mapping from noise to data in a single forward pass. However, the reward-guided fine-tuning method of one-step generative models remains largely unexplored. To address this, we consider one-step generators from an optimal transport view, investigating Wasserstein Gradient Flow (WGF) for modeling smooth and controlled distributional evolution in probability space. We then propose a novel reward-guided fine-tuning of a one-step generative model via WGF. We derive a practical training method that requires no reward gradients, thereby handling both non-differentiable and differentiable rewards. Moreover, our method provides smooth and stable reward-guided distributional updates while mitigating reward hacking and mode collapse. Experiments on 2D synthetic data, CIFAR-10, and ImageNet 256$\times$256 with diverse rewards, including JPEG (in)compressibility, class probability, Black-and-White and CLIP alignment, show that our method achieves better reward alignment compared to baselines.

Modern video generators routinely fail at physical dynamics: objects float, trajectories violate gravity, contacts vanish. Standard denoising and flow-matching objectives fit visual data distributions but do not explicitly penalize such physical violations. Existing remedies can improve physical consistency, but typically add substantial inference or training cost. Candidate-selection methods generate and score multiple videos, while gradient-based world-model guidance repeatedly decodes and re-encodes intermediate estimates. Generator-internal refinement adds perturbation and re-denoising loops, whereas post-training requires curated data and additional optimization. We propose Off-Manifold Refinement (OMR), an inference-time method that instead injects world-model feedback directly into a single sampling trajectory. During scheduled middle ODE steps, we augment the generator velocity with the gradient of an adapter-space V-JEPA 2.1 surprise energy. This external correction can move the latent away from the uncorrected sampling trajectory and toward regions ranked as more physically plausible by the frozen predictor, after which the generator continues rendering from the corrected state. A small trained latent-to-embedding adapter keeps the gradient tractable at inference, and both the video generator and the world model remain frozen. On our fixed 400-prompt VideoPhy-2 detailed subset, OMR lifts the joint Semantic-Adherence-and-Physical-Commonsense metric from 47.0% to 52.0% (+5.0pp absolute, +10.6% relative) over the base Wan2.2-T2V-A14B sampler. On a separate fixed 50-prompt efficiency subset, it requires $1.71 \times$ the base runtime rather than the multiplicative cost of reward/search alternatives. Project page: https://itruonghai.github.io/omr.

Diffusion models are increasingly used not only for sampling from learned data distributions, but also for generating samples that optimize task-specific objectives. A common approach is to guide the reverse diffusion process using gradients of an external objective. However, when the data distribution is supported on a structured feasible set, such as a manifold or a constraint set, gradient guidance can move samples away from the learned data geometry. In this paper, we study a simple projected-gradient-guided diffusion update based on the observation that the Stein denoising operator can act as an approximate projection onto the data geometry. The proposed update incorporates the objective gradient inside the denoising step, yielding an inference-time method that uses only a pretrained denoiser and gradient evaluations. We analyze this update as an inexact projected-gradient method for constrained optimization over learned feasible geometries. Our theory covers three settings: linear manifolds, compact convex feasible sets, and compact Riemannian submanifolds. In all these settings, we prove descent and finite-time convergence guarantees. Numerical experiments support the theoretical interpretation and illustrate how the proposed update balances objective descent with preservation of the learned geometry.

Autonomous vehicles (AVs) operate in complex environments where failures are consequential. Sophisticated machine learning models for perception and planning are key to overcoming at least part of that complexity, but their black-box nature complicates validation and verification (V&V). The recent integration of Vision-Language-Action (VLA) models into AVs introduces a unique opportunity: besides generating trajectories, these models produce an explicit Chain-of-Thought (CoT) explaining their underlying rationale. This CoT provides a rich specification to cross-check model outputs and detect inconsistencies that may expose unsafe or unintended behavior. This paper assesses whether CoTs from a recent open driving VLA can support such monitoring. We curate DriveAlignBench, a specialized dataset from NVIDIA's Alpamayo 1.5 VLA for AVs containing 150 CoT-trajectory pairs, which we manually annotate for reliability, trajectory consistency, and safety. Our analysis reveals that 33.3% of CoTs are unreliable. Among reliable CoTs, the generated trajectory is consistent with the CoT in 74% of cases. Leveraging this potential, we propose integrating a CoT-trajectory consistency check into a runtime monitor. The check is nontrivial: CoTs express open-vocabulary, scene-relative driving commitments, while trajectories are low-level ego-motion sequences whose semantics depend on road geometry and motion context. To bridge this gap, we develop a family of automated consistency monitors. Our best monitor, lane-relative F-LLM with GPT-5.5, achieves F1 = 0.75, improving over the strongest raw-waypoint LLM baseline by +0.13 absolute F1 and over a rule-based monitor by +0.38. We release DriveAlignBench, the monitor implementations, and annotation tools at https://github.com/776styjsu/drive-the-thoughts.

Memory operations of long-horizon LLM agents are hard to supervise: an operation's value is unobservable when it is taken. But they are special — they leave machine-readable evidence in the trajectory: retrieval hits and answer-time citations. Hindsight Memory-PRM exploits this audit trail twice: offline to train an operation-conditioned memory-utility critic, and online, where retrievals, citations, and one controlled deletion-and-reanswer per probe settle an intervention-calibrated entry-level presence credit, propagated along version chains as an action-level proxy reward — no per-operation human labels, no Monte-Carlo replay of continuations. On held-out LoCoMo a local 8B policy reaches 77.5% under a fixed shared reader, surpassing its API teacher (65.1%) and all reproduced external systems, at one eighth the context of Mem0's official operating point; on LongMemEval, 79.0%. Ablations attribute the gain to causal calibration rather than signal density, and the policy converges to a multi-version memory organization whose gains no tested open-loop baseline reproduces.

Retrieval-Augmented Generation (RAG) improves the factuality of large language models (LLMs), yet existing RAG systems often struggle with complex, multi-step reasoning that requires adaptive retrieval and continuous revision of intermediate contexts. Recent reinforcement learning (RL)-based agentic RAG methods partially alleviate this issue, but typically rely on coarse-grained action spaces and trajectory-level rewards, resulting in weak reward assignment and a bias toward short-horizon, stereotyped reasoning template. To address, we propose AgenticRag-R1, a RL framework that deeply integrates reasoning, retrieval, and memory via a memory stack and fine-grained action space, supported by hierarchical action-aware rewards and an information-aware trajectory rejection strategy to enable effective long-horizon learning. Experiments across a diverse set of multi-hop, open-domain, and agentic reasoning benchmarks, spanning multiple backbone model sizes, demonstrate that AgenticRag-R1 consistently outperforms strong baselines. Moreover, AgenticRag-R1 learns more robust, interpretable, and memory-aware reasoning behaviors, highlighting the effect of fine-grained action modeling and information-aware optimization for long-horizon reasoning. Our code is anonymous available at https://github.com/jiangxinke/Harness-RL/tree/AgenticRAG-R1-Whitebox.

Large language models (LLMs) can generate apparently highly persuasive text, but does sounding persuasive mean arguing well? We introduce \textsc{Persuasio}, a multi-agent dialogue platform grounded in a formal argumentation-based theory of persuasion dialogues that adjudicates logical winners during free-text debates. Using this system, we generated 192 debates on a UK political topic between humans and LLMs, and evaluated 22 interlocutors through both automated adjudication and 9,702 crowdsourced pairwise judgements across 1{,}386 annotation instances. We observed a consistent decoupling between subjective and formal persuasiveness: LLMs dominated the subjective ranking yet performed substantially worse under argumentation-theoretic adjudication, where humans remained competitive. Multi-agent and retrieval-augmented variants further widened this divergence. These findings reveal a systematic gap between rhetorical fluency and formal argumentative strength in LLM-based persuasive dialogues.

Frozen vision-language-action (VLA) policies offer broad manipulation skills but execute open-loop action chunks without tracking task progress, so the agent cannot reliably decide whether to continue, retry, or terminate. External memory is a natural remedy, yet it can be harmful when attempted actions are treated as completed progress, turning local execution errors into persistent task-state errors. We propose Achievement-Grounded Memory (AGM), a lightweight closed-loop framework for frozen VLA policies that represents a task as a subgoal sequence with a progress pointer and advances this memory only after the current subgoal is verified by physical evidence. Proprioceptive interaction cues decide when to verify, while coherent point tracking and language-conditioned cross-view comparison, sourced from frozen foundation models through a single 2.43M-parameter verification head, decide what was achieved. AGM thereby converts open-loop execution into a closed loop of execution, verification, and progress, keeping the policy frozen without test-time large-model inference. On the RoboMME Counting benchmark, AGM reaches on PickXTimes and on BinFill, surpassing the strongest memory-augmented baseline by points on average, and the framework yields equally decisive gains on a physical robot. Reliable embodied memory thus depends more on disciplined state updates than on memory capacity.

LLM agents increasingly drive long-running cloud inference workloads in which model calls differ in urgency, redundancy, completion semantics, and replay cost. Model-as-a-Service (MaaS) platforms expose several service models for trading cost against latency, availability, and capacity commitment. These models operate primarily at request, job, or endpoint scopes and provide limited support for combining transient platform supply with the evolving semantics of an agent task. We present SemSpot, a semantics-aware service model that allows agent applications to leverage the spot capacity of LLM inference platforms. At the request level, SemSpot lets a provider publish short-lived offers over successful price, completion probability, and failure-notification deadline; the agent runtime selects among these offers using the current task state and completion rule. An audit of 1,535 cases from six agent benchmarks identifies four recurring workflow structures and shows how this service model may produce different cost, service-time, and fallback behavior. With specialized MaaS support, token-level SemSpot further preserves provider inference state and runtime-verified semantic segments inside a long request. We develop the service model, economic boundary, and the cross-layer research agenda required to realize SemSpot.

Early failure prediction is important for long-horizon agents, as it enables timely intervention and can reduce inference and tool-use costs. Uncertainty quantification, such as verbal confidence and perplexity, offers a promising approach to detecting agent failures; however, it has not been explored whether these signals retain their discriminative power during the intermediate stages of long-horizon execution. We evaluate mainstream uncertainty signals on deep-research tasks and find that verbal confidence reliably distinguishes failures at trajectory completion, achieving a mean AUROC of 0.85, whereas all evaluated signals offer limited predictive value earlier in execution, with none exceeding a mean AUROC of 0.60 at 50% trajectory progress. We identify an underlying mechanism explaining this gap: path switching, where agents frequently abandon their current search direction in-trajectory, breaking the link between early signal and final outcome. These findings challenge the assumption that intermediate uncertainty can reliably guide early intervention. They also motivate a practical recommendation for agent harnesses in deep-research settings: use final-step confidence to decide whether to restart, an approach that our experiments find more effective than in-trajectory intervention.

The key limitation of current state-of-the-art influence-based guardrails is that they do not reliably distinguish a legitimate, user-authorized action from a malicious, unauthorized action when both rely on external tool information. This ambiguity can cause benign actions to trigger unnecessary verification and intervention, reducing utility and adding latency. We expose this limitation through an authorization-equivalence audit of 96 conditions derived from 24 base cases. Within matched source comparisons, we hold authorization, the exact committed action, and its intended effect fixed, changing only whether a required value comes from the user or a legitimate tool result. Although the action remains unchanged, this harmless relocation shifts the causal signal toward the attack region in all 24 cases under both Llama and Gemma scorers. Matched unauthorized controls show that the signal remains attack-sensitive, yet the benign relocation produces a larger average score shift than the actual change in authorization. Architecture-level evaluation shows how this mismatch propagates through guardrail designs. With a semantic monitor, attack success is 0% and utility is 28%, compared with 16% and 60% without it. A shadow-based guardrail allows every tested harmless run, yet does not reject matched unauthorized actions more often overall: 57.5% of unauthorized runs pass automatically before reaching the later security check, compared with 29.2% of authorized runs. These results show that the studied causal signal reveals what shaped an action without reliably encoding whether the action was authorized, and that reference construction and routing are integral to the effective security decision.

Recent AI systems promise autonomous scientific discovery, claiming to discover algorithms and produce research papers, yet understanding whether they exhibit creativity, the capacity to produce solutions that are both novel and useful, remains an open question. We present a framework for evaluating multi-turn LLM research agents' creativity using ML engineering tasks as a testbed, through three dimensions: P-Creativity (psychological novelty: novel relative to the agent's own prior solutions within a run), H-Creativity (historical novelty: novel relative to the corpus of human solutions), and Usefulness (task performance). Evaluating two agent frameworks, AIDE and AIRA-Dojo, on 10 Kaggle-style machine learning tasks from MLE-Bench, we develop an LLM-as-a-Judge pipeline and verify its strong correlation with human creativity judgments, providing a reliable automated metric for P-Creativity evaluation at scale. Applying this pipeline to agent trajectories, we find: (1) all agents exhibit declining P-Creativity as they transition from exploration to exploitation; (2) LLMs exhibit greater H-Creativity than medal-winning humans, yet achieve lower performance. Our findings reveal that current agents can explore novel regions of the solution space but lack the capacity to convert this novelty into improved task performance.

Flat retrieval-augmented generation treats a corpus as a bag of chunks, discarding document hierarchy and cross document structure. We introduce SearchWiki, a harness framework that synthesizes a corpus into a hierarchical, typed, navigable wiki and trains an agent, WikiResearcher-9B, to retrieve information through multi-turn tool use. The wiki organizes knowledge into three layers - document overviews, cross- document topic pages, and page-level source records; enabling progressive refinement of retrieval when initial lookup misses. We optimize the agent's navigation policy with on-policy reinforcement learning with a multi-component reward function balancing answer correctness, retrieval quality and trajectory efficiency. Evaluation on ViDoRe-V3 (8 domains), FinanceBench, and memory benchmarks (LoCoMo, LongMemEval, PersonaMem-v2) shows that WikiResearcher- 9B which is our RL-tuned Qwen 9B model, significantly outperforms same-size untrained baselines and exceeds or matches larger external models. SearchWiki paired with WikiResearcher-9B demonstrates that learned navigation over structured corpora is a superior alternative to flat retrieval.

Modern Vision-Language Models (VLMs) perform well above the human baseline in image geolocalization, a task critically important in disaster response, OSINT verification, and location privacy. However, most efforts to study AI behavior on the task remain limited to static image-based retrieval, classification, and predictions. We argue that faithful recreation of the task should involve embodied navigation, where a multimodal agent autonomously explores its surroundings to gather observations before submitting a prediction. To this end, we introduce GeoAgent, an agentic environment-based benchmark that requires agents to navigate Street View environments to refine their geolocalization through sequential reasoning. Our analysis shows that modern VLMs struggle to discern regional patterns while succeeding at country- and continent-level predictions. When compared to static image-based baselines, agentic navigation significantly improves accuracy across established metrics. We also note severe bias in a developed/developing region context across frontier model architectures and poor self-improvement capabilities given incorrect priors. Overall, our work establishes the challenges of embodied navigation and geospatial reasoning. We publicly release our code and the GeoAgent environment: https://geoagent-benchmark.github.io

Attributing an artwork to an artist has traditionally relied on detailed visual observations and descriptions, known as stylistic analysis in art history. By contrast, current artificial intelligence (AI) models used in the field offer only unexplained probabilistic classifications. To bridge this methodological gap, we present an AI framework that automates stylistic analysis of paintings, providing a foundation for enhancing evidence collection, discovery, and verification. By training a vision transformer (ViT) on a large corpus of paintings with metadata, our system encodes this art history-specific data as embeddings. These representations are factorized via sparse dictionary learning into a shared set of features that recur across the training set. A large language model (LLM) then interprets each feature by retrieving associated artworks and their accompanying curator-written texts, and synthesizes them into descriptions that reflect their stylistic attributes. Finally, an autonomous coordinator LLM applies a reasoning-and-action (ReAct) framework to weight, test, and refine these features into cohesive descriptions of an artwork, or comparisons of artworks. This approach converts detailed visual features into descriptive terms, addressing a key challenge in art history. It thus connects the use of images as data with the semantic concerns of humanists, establishing vision-based computational art history as an area for future growth.

Large language model agents place outputs from external skills into their execution context, allowing attacker-controlled data to influence later privileged actions. Existing defenses mainly classify untrusted content or authorize proposed operations. They do not directly address how an agent's future authority should change once untrusted data enters its state. We present SkillGuard, a harness-level enforcement layer that treats this event as contamination and restricts future capabilities to disconnect the resulting state from deployer-defined forbidden states. Given sound skill summaries and policies, SkillGuard represents security-relevant transitions with a Skill Impact Graph, specifies admissible control over skill parameters via steerability signatures, and mediates invocations with an inline reference monitor. Following contamination, it computes weighted capability restrictions using binary, fractional, or fractional-flow strategies without auxiliary language-model inference. We evaluate SkillGuard on four AgentDojo suites with two backend LLMs, Gemini 2.5 Flash and Llama3.3-70B, against an LLM-only No Defense baseline and three defenses at different system layers: Spotlighting, CaMeL, and AttriGuard. We construct a compositional attack benchmark in which each attack combines observations individually insufficient to induce target violation and evaluate the same baselines on it. Under AgentDojo's Tool Knowledge attacks, SkillGuard eliminates attack success on three of four suites for both backends and reduces it to 4.8% and 14.3% on Slack. Against compositional attacks, it outperforms every baseline on Llama and matches the strongest baseline on Gemini at higher benign utility. Fractional-flow restriction preserves substantially more capabilities than binary restriction at the same attack success rate. Across both settings, SkillGuard adds no model calls or token overhead.

Longitudinal clinical agents must maintain an evolving patient state from evidence distributed across visits, time points, and specialties. However, how agent memory should be designed for this setting remains unclear. We introduce a benchmark of multi-visit, multi-specialty patient records that evaluates long-context evidence retrieval, cross-time evidence aggregation, and cross-specialty clinical reasoning. Using this benchmark, we systematically study four memory design choices: curation, organization, retrieval, and memory-augmented reasoning. We find that temporal validity is more important than simply retaining more history; specialty-factorized memory reduces context but can hide shared evidence; and multiple agents help when specialists must reason together, not merely when evidence comes from multiple memories. Guided by these findings, we propose MedCache, a hybrid framework that constructs temporally valid patient memory, organizes evidence into overlapping specialty views, routes each query to relevant memories, and adaptively invokes one or multiple specialists. Experiments show that MedCache improves reasoning accuracy and memory efficiency over strong single-agent and multi-agent baselines, while generalizing across model backbones and external datasets.