Papers

The social interactions among crowds via Danmaku (a.k.a., bullet comments) on modern multimedia platforms can facilitate both viewpoint conflicts and consensus, providing fine-grained discriminative social signals that can benefit fake news detection. However, the inherent accumulation latency of Danmaku in real-world scenarios violates the real-time necessity of fake news detection, making the studies of Danmaku-related fake news detection underexplored. To break this violation, we simulate this temporal-aware user interactive process by proposing a novel temporal Generative danmaku framework, called Genda, which consists of: (1) a Danmaku Trigger for predicting the timing and intensity of user reactions; and (2) a Danmaku Generator for synthesizing corresponding semantic and emotional expressions, thereby mutually constructing a temporally aligned and human-like pseudo Danmaku streams. To make the generated Danmaku useful for identifying fake news videos, we further design a Danmaku-guided Temporal Multimodal fake news detection model - DM-FEND, which enables fine-grained multimodal interactions among video, audio, text, and Danmaku, enhancing dynamic modalities alignment and semantic noise inhibition. The experimental results demonstrate that DM-FEND consistently outperforms state-of-the-art baselines across both Chinese (FakeSV) and English (FakeTT) benchmarks. Further ablations validate the crucial role of temporal Danmaku modeling in enhancing robustness and discriminative capability. Finally, this study offers a bright and robust solution for multimodal fake news detection in modern social interactive fashions by bridging the temporal inconsistency between news and user behaviors.

Deploying LLMs for enterprise Text-to-SQL is bottlenecked less by the model than by what context reaches it: business logic spans thousands of tables, and no model can ingest a full catalog at once. We argue that the most effective place to intervene is therefore the knowledge-base context the model consumes, and that this context should be constructed from historical usage rather than tuned for as a fixed input. Using a query-DAG decomposition–the same family of intermediates that enterprise benchmarks like BEAVER annotate, here recovered from production SQL–we compare the value of oracle query graphs versus retrieved knowledge-base context. In this ablation, retrieved knowledge-base context provides the largest marginal improvement when added to the full oracle graph. Building on this, we optimize a distillation procedure that turns historical query profiles into reusable SQL reference cards. On a benchmark of 5176 production queries from a major online retailer, optimizing these context artifacts yields larger gains (${\sim}12$–$25\%$ AST similarity) than optimizing the retrieval harness (${\sim}3$–$12\%$). On the public BEAVER benchmark, which lacks the production-usage signals available in our internal setting, the picture is more mixed: table cards alone perform about the same as raw historical SQL. The best optimized variant retrieves both cards and raw SQL, scoring $9.00\%$ versus $6.33\%$ (p-value $0.12$) for the comparable baseline on a held-out $N{=}300$ subset, using retrieved context and harness changes but no agentic loop.

Deep learning models often produce performance disparities across demographic groups, due to the training data imbalance with respect to sensitive attributes such as gender or age. To address this problem, existing work has explored fair representation learning, data re-sampling, and adversarial training, which can be broadly categorized into two main approaches. Single-stage methods typically learn a shared representation for fairness, but often struggle to handle heterogeneous subgroup distributions. Two-stage methods learn representations separately from the final prediction task, which can lead to misalignment between fairness objectives and downstream predictions. We identify routing-induced bias, a failure mode in which subgroup imbalance drives the gating network to route subgroups onto a few experts, and propose an end-to-end Mixture-of-Experts (MoE) framework that corrects it. Specifically, we apply subgroup reweighting to correct data imbalance, and introduce gate entropy regularization to prevent routing from collapsing onto subgroup attributes, keeping expert utilization both balanced and interpretable. Beyond improving fairness, the routing distribution offers an interpretable view of how subgroups are allocated across experts. Experimental results demonstrate that the proposed approach improves fairness while maintaining competitive predictive performance.

Medical large language models are often judged by how many clinical questions they answer correctly. That view is useful, but it misses a practical risk. A model may know the right answer and still change its response when the same case is written in a different patient voice. This paper evaluates that risk as SDoH aware narrative anchoring bias. We use NarrativeShield SDoH MedQA, a counterfactual medical question answering dataset in which each case appears in persona based narratives while the answer key remains fixed. The dataset is reshaped from wide format into case grouped persona rows. We evaluate three open source instruction tuned LLMs from the Qwen2.5 family: 1.5B, 3B, and 7B. The final experiment uses 300 clinical cases and produces 8,100 model responses across three prompting conditions. We report persona level accuracy, counterfactual consistency, correct consistency, and narrative sensitivity error. Qwen2.5 7B achieves the best accuracy at 56.33 percent and the best correct consistency at 40.33 percent. Paired McNemar exact tests show significant accuracy gains for 7B over 3B in all prompt settings. Even so, narrative sensitivity remains, with the lowest error still at 31.67 percent. These results suggest that trustworthy clinical decision support should be evaluated by both average correctness and stability across medically equivalent patient narratives.

Ensuring reliability in uncertain environments remains difficult for long-horizon robotic manipulation. End-to-end VLA models are data-heavy and opaque, making diagnosis and verification difficult. Hierarchical pipelines are more interpretable, but their plans are often weakly grounded in observations, weakly aligned with low-level actions, and computed without online feedback, leading to open-loop behavior and hallucinations. To address these issues, we introduce the Triplet-to-Track System (TTS), a closed-loop long-horizon imitation learning system that uses human videos to reduce reliance on robot-collected data. TTS represents high-level subgoals as instance-grounded triplets, translates them into continuous track priors for execution, and monitors task progress from observations for online replanning. Across diverse real-world long-horizon tasks, TTS achieves a 74.8% average success rate and supports object-level and compositional generalization.

Background This study was designed to evaluate whether a domain-specific large language model (LLM) trained exclusively on patient education resources can answer questions about psychiatric medications, in a manner superior to LLM chatbots. We developed an LLM ("MIND") fine-tuned for clinical fidelity, trained on patient education resources from authoritative medical organizations. Methods We compared the responses of MIND, ChatGPT, and OpenEvidence to patient questions about escitalopram, using two methods: (1) computer analysis according to a rubric measuring accuracy, clarity, completeness, nuance, safety, and referral appropriateness; (2) ratings from N=10 board-licensed psychiatrists on similar metrics. Results When rated by rubric, MIND was rated highest in all domains (p<0.001). When rated by psychiatrists, ChatGPT was rated accurate more often than MIND with a negligible effect size (p=0.021, r=0.073); MIND was rated complete more often than ChatGPT with a small effect size (p<0.001, r=0.160); and MIND and ChatGPT were rated safe with the same frequency (p=0.955, r=0.002). The majority of psychiatrists preferred the responses generated by ChatGPT (57.6%) compared to MIND (42.4%, p=0.003). Conclusions MIND was able to answer many questions about escitalopram in a manner deemed accurate, complete, and safe by psychiatrists the majority of the time. However, despite MIND's ability to provide more complete responses, psychiatrists preferred ChatGPT's responses. MIND represents a step towards building safe LLM systems to enhance patient education in psychiatry.

Reliable deployment of LLM agents in user-facing products depends not on raw task-solving ability but on consistency and limit-awareness: behaving the same way across repeated trials, and recognizing when a request cannot, or cannot yet, be safely fulfilled. CAR-bench exposes this reliability gap in the domain of in-car assistants: an LLM-simulated user issues incomplete or ambiguous requests, requiring the agent to resolve uncertainty through multi-turn dialogue and tool use while strictly adhering to domain policies. Even frontier models show a substantial gap between what they can solve at least once (Pass@3) and what they solve consistently across trials (Pass^k). We bridge this gap with TRACE (TRAjectory-Contrastive Evolution), which iteratively improves a skill-based agent's behavioral knowledge without modifying model weights. This knowledge is organized as a Skill Bank of modular, retrievable skills, each encoding a self-contained set of tool-use rules and behavioral guidelines. TRACE evolves this bank through an agentic self-evolution loop: after each evaluation round, it groups trajectories by the skills invoked and refines each skill by contrasting successful and failed behaviors. The updated bank then guides subsequent rounds, while during deployment the Actor performs state-conditioned skill orchestration at every turn. On GPT-5.5, TRACE improves consistency (Pass^3) by 34.6 points, from 59.9% to 94.5%, while shrinking the gap between potential and reliable performance to just 4.0 points. On the official hidden set, TRACE achieved first place using GPT-5.6-Sol, attaining a Pass^3 score of 70%-a 40% relative improvement over the baseline. These results show that TRACE converts high model potential into stable, consistent performance gain. Project homepage: https://darwin-agent.github.io/Car-bench-TRACE.

Large-scale rollouts have become a core component of modern LLM systems, spanning reinforcement learning (RL) post-training, on-policy distillation (OPD), and sampling-heavy evaluation pipelines. Unlike online serving, which is typically optimized for request-level latency and throughput, a small number of long-tail generations can dominate the end-to-end makespan of an entire rollout step. In practice, rollout requests are often routed uniformly across replicas, which can place extremely long generations inside high-concurrency decoding batches. To address this, we present TailSieve, a partial-rollout-guided framework that jointly controls tail routing and replica allocation for LLM rollouts. In an idealized setting with known completion lengths, we show that makespan-optimal routing in the long-tail regime combines tail isolation with load balancing, and that a simple top-k policy closely approximates this offline optimum. Leveraging the observation that long-tail prompts tend to remain long-tailed across policy updates, TailSieve uses partial rollouts as a training-free signal for identifying candidate tail groups. A hierarchical controller then jointly adapts the number of isolated groups and the replica split between the tail and bulk pools using collected response-work history and a measured concurrency-throughput model. TailSieve achieves up to 1.67x routing-only speedup over uniform group routing. The resulting low-concurrency tail pool further enables route-specialized speculative decoding with MTP or DFlash, achieving up to 2.59x speedup over uniform routing. Selected prompts are regenerated under the current policy, preserving on-policy generation and avoiding additional routing-induced length bias in steady state.

Long-term language-model agents accumulate memories across interactions, but their retrievers typically do not accumulate retrieval experience. Semantic retrieval is efficient, but embedding similarity does not always reflect whether a memory contains evidence relevant to the current query. Large language model (LLM) rerankers provide stronger query-conditioned relevance scores, yet stateless reranking repeatedly scores a large candidate pool and discards these scores after each query. We introduce EARM, an experience-amortized reranking framework that treats previously acquired LLM relevance scores as reusable retrieval experience. EARM stores sparse query–memory relevance scores in an online matrix, learns their shared structure through causal matrix completion, and combines a small set of newly observed scores with estimated scores to rerank the remaining candidates. The scoring budget decreases as experience accumulates, changing LLM reranking from a repeated per-query expense into a retrieval capability learned over an agent's lifetime. Experiments on long-term conversational memory show that mixed observed-and-estimated reranking improves answer accuracy over semantic retrieval by up to 6.62% and remains effective when only 17.5% of candidates receive direct LLM relevance scores, thereby substantially reducing the inference overhead of LLM reranking. These results motivate a broader view of agent memory: a long-lived agent should remember not only past content, but also how that content has proved useful for retrieval.

Knowledge graph question answering (KGQA) is a key task for evaluating KG-augmented Large Language Models (LLMs), and complex KGQA that requires multi-hop reasoning is especially challenging. Solving a complex query involves two coupled phases: candidate retrieval, which locates answer candidates over the KG, and constraint handling, which filters these candidates against the query constraints. Faithful reasoning requires grounding both phases in the KG. However, existing agent-based methods ground candidate retrieval through entity-centric exploration, while leaving constraint handling to the LLM's internal knowledge, which leads to two critical limitations. (1) Unreliable entity pruning: entity-centric exploration uses entities as search units and must prune them to a fixed-size subset at each hop. Because entity information in KGs is often incomplete and a fixed-size subset cannot retain all valid entities, such pruning inevitably drops valid entities and ultimately leads to wrong answers. (2) Ungrounded constraint handling: query constraints are resolved from the LLM's internal knowledge rather than the KG, leaving the final answers unverifiable and prone to hallucination. To address these limitations, this paper introduces a relation-centric exploration paradigm, which uses relations rather than entities as search units and thus avoids unreliable entity pruning. Built on this paradigm, this paper proposes Compositional Chain-of-Relations (CCoR), a simple and effective framework that grounds both phases in the KG with two relation chains: a main chain for candidate retrieval and a constraint chain that verifies query constraints through explicit KG exploration. Experiments on four KGQA benchmarks show that CCoR consistently improves accuracy, faithfulness, and efficiency over strong baselines, with more pronounced gains on complex queries.

Large Language Models generate text autoregressively, but open-ended generation is prone to verbatim looping, in which models repeat spans already present in context. Standard defenses such as repetition, presence, and frequency penalties and n-gram blocking act on token recurrence rather than the sequential structure of a loop, and often suppress looping only at strengths that also degrade formatting or fluency. We propose Don't Repeat Yourself (DRY), a sampling-time logit adjustment that penalizes a candidate token only when generating it would extend the current suffix into an exact continuation of a span seen earlier in the context. Sequence breakers protect chat templates and formatting tokens. Across models from 1.5B to 120B parameters, nine prompt families, and a 600-pair human study, DRY reduces suffix-extension rate by 47% while improving lexical diversity. An intervention-matched placebo produces no comparable reduction, identifying suffix matching as the operative mechanism. On AWQ-quantized 70B and 120B models, DRY reduces loop rate by roughly half while preserving MT-Bench, MMLU, and GSM8K performance, whereas standard alternatives lose measurable ground. DRY has been adopted by popular open-source LLM inference frameworks including llama.cpp, ExLlamaV2, and text-generation-webui, highlighting its practical impact on text generation.

Standard decoding rules for autoregressive language models promote diversity by rescaling the full next-token distribution or truncating its low-probability tail. These strategies overlook a common regime of open-ended generation in which several continuations are plausible but too much probability mass remains concentrated on the most generic choice. We introduce XTC (Exclude Top Choices), a lightweight head-aware decoding operator that targets this regime directly. XTC identifies tokens whose probabilities exceed an absolute plausibility threshold $\tau$: when at least two qualify, it removes the dominant eligible choices with probability $\rho$ and retains only the weakest plausible alternative before renormalization. Across 60 experiments on Gemma 3 27B Q4, Gemma 3 12B Q6, and DeepSeek R1 14B Q6, with scaling validation on Llama 3.3 70B Q4, XTC improves the diversity-repetition Pareto frontier. On creative generation, Distinct-2 increases by 11–15% and repeat trigrams decrease by 27–47% across the four models. Combined with temperature scaling, gains reach 38% in Distinct-2 and 71% in repeat-trigram reduction over baseline. A blinded Amazon Mechanical Turk study with 150 Master raters yields a 62.3% creativity preference for XTC ($p<10^{-4}$) without reduced fluency, while a GPT-4o control judge reproduces the Anthropic-judge direction on every measure. On IFEval with Llama 3.3 70B Q4, XTC preserves prompt-level strict accuracy within 1.7 percentage points of baseline while recovering most of the diversity gain; a temperature setting matched on Distinct-2 reduces IFEval by 8.8 points. The effect is additive with temperature and repetition penalties, robust across quantization levels and model families, and consistent across twelve prompt genres. XTC has been adopted by llama.cpp, ExLlamaV2, and text-generation-webui.

Unified models for visual understanding and generation have made rapid progress, yet they still lack the ability to understand and manipulate the spatial states of object instances. Existing models can describe objects in natural language, but they struggle to precisely represent continuous object poses and generate geometrically consistent images under target viewpoints. To mitigate this, we propose Object-Uni, a unified model for object-centric spatial understanding and controllable generation. Specifically, we formulate object-centric spatial intelligence as a unified problem connecting pose perception, spatial reasoning, pose-conditioned generation, and object-centric novel view synthesis. We treat object pose as an explicit geometric variable shared by understanding and generation, rather than merely a prediction label or control signal. To make pose usable by multimodal large language models, we propose a viewpoint-based orientation abstraction that maps orientation into structured viewpoint descriptions while preserving continuous geometric supervision. We further construct an object-centric spatial benchmark (UniSpatial-80K) and train a unified model with an object-token-grounded pose anchor to associate each instance with its pose state. Experiments show that our model improves object-level pose understanding and pose-controllable generation, moving unified models from describing objects toward manipulating spatial states.

A safety rule and an episodic log compete for the same tokens in an AI agent's context. When the budget overflows, both are summarized at the same rate; only the rule needs exact wording to remain enforceable. On 20 production agent configurations, Claude Code's /compact prompt on Sonnet 4.6 preserves 53% of safety rules after one compaction round and 10% after five. We name this the Compaction Cliff. We address it with Knowledge Triage, a framework that classifies each line of an agent's knowledge base by type and routes each type through its own retention policy. Three deterministic operators implement this triage across the three context-management operations: TypeCompact rewrites items in place under per-type fidelity, TypeDecompose partitions a topic too large to compact safely, replicating in-scope safety rules across partitions, and TypeRetrieve fetches items from external storage with in-scope rules pinned ahead of relevance. On five public corpora, TypeCompact preserves 2–4$\times$ more safety rules than the strongest single-shot LLM compactor at every ratio, with 96% recall over five rounds. TypeDecompose reaches 0% locality violations against 93% under uniform partitioning. TypeRetrieve reaches 100% recall@50 against 73% for the best single-shot LLM retriever. On three downstream behavioral benchmarks, we outperform the production Sonnet compactor on medical compliance (paired McNemar $p < 10^{-8}$ on preservation, $N = 200$), the full-policy and hierarchical baselines on retail task pass rate ($p < 0.01$, $N = 115$), and the hierarchical compaction on the airline domain ($p = 0.024$). We release AgentArtifactCorpus (396{,}934 agent configurations from 54{,}628 public GitHub repositories), the classifier, and the reference implementation.

Emotion Recognition in Conversation (ERC) requires models to identify subtle emotional cues that are often distributed across distant dialogue turns. Existing methods typically incorporate dialogue history through a fixed context window. However, short windows discard potentially useful long-range evidence, while enlarging the window repeatedly re-encodes overlapping utterances, increases computational and memory costs, and may introduce irrelevant context. Moreover, commonly used parameter-efficient adaptation methods, such as LoRA, mainly introduce fixed low-rank transformations in the feature space and do not explicitly maintain a dialogue-level state or condition their transformations on the evolving conversational context. To address these limitations, we propose a lightweight adapter, DiaRelay, to enable LLMs to explicitly maintain a dialogue-level memory for accurate ERC. Based on LoRA, DiaRelay introduces two extra tightly collaborative components, Selective Relay Memory Transition and Dual-axis Relay Memory Read. Selective Relay Memory Transition progressively aggregates useful historical evidence into a bounded relay memory and propagates it across successive utterance predictions. This allows earlier emotional cues to influence later predictions after they leave the local context window, without re-encoding the complete dialogue history or expanding the backbone context length. Dual-axis Relay Memory Read uses the propagated memory to dynamically modulate low-rank feature transformations, enabling context-dependent representation adaptation without test-time gradient updates. Extensive experiments show that DiaRelay can achieve SOTA weighted F1 and accuracy on MELD while obtaining competitive results on IEMOCAP with only an extra 7.1M trainable parameters, indicating the effectiveness and generalizability of our DiaRelay in enhancing LLM-based emotional understanding.

Nonverbal behavior generation systems for virtual agents often take an utterance as input and generate nonverbal behaviors that emphasize or illustrate the content of the verbal channel. However, human nonverbal behavior is shaped by more than the content of the speech. It is also influenced by speaker roles, interpersonal relationships, social context, and the cognitive and emotional states of the interactants. As a result, the nonverbal channel may reinforce, weaken, qualify, or even contradict the verbal channel. It may also reveal internal states that are hidden or only indirectly implied in speech, including emotional "leakage" that may be incidental to the immediate interaction. Modeling this richer relationship between verbal and nonverbal behavior is important for designing virtual agents that exhibit realistic, human-like behavior. It is especially critical in training contexts that require nuanced social interpretation, such as counseling simulations involving virtual patients. Drawing on Ekman's framework of verbal nonverbal relationships, we propose a taxonomy of categories in which mismatches between verbal and nonverbal behavior can occur. We then examine alternative approaches for realizing these behaviors using large language models, focusing on whether LLMs can select contextually appropriate mismatched verbal and nonverbal behaviors from a given dialogue and social interaction context. Finally, we evaluate the resulting behaviors in a human-subject study, assessing whether context-driven nonverbal behavior, when embodied in a virtual human, produces the intended effects on observers.

Short-drama generation has grown into a large, industrialized pipeline, and as it scales from isolated shots to the episode level, visual continuity has become a critical bottleneck. Current agent frameworks generate each shot in isolation, so context drifts across shots and props, character posture, and blocking turn inconsistent. Once assembled, these small discrepancies amplify into severe visual breaks. We present SEAM (Shot Entity-Attribute Memory), a training-free, model-agnostic memory graph that repairs continuity entirely at the prompt-text layer by extracting a multi-dimensional state for every shot, retrieving only causally prior context over the resulting graph, filtering it selectively, and injecting the surviving constraints by natural-language prompt rewriting. We further release SEAM-Bench, a double-blind continuity storyboarding benchmark, on which SEAM raises cross-episode continuity recall from 0.700 to 0.946, generalizes across six mainstream text models, and yields consistent, though not yet significant, gains at the generated-image layer. Deployed as a mandatory stage in CreativeFitting's SEAM-Agent production pipeline over 201 shots, SEAM reaches a 96.5% director-acceptance rate with zero unsafe injections; a conservative counterfactual attributes at least 21.9 percentage points of that rate to its cross-episode memory.

Despite their remarkable success, machine learning models, particularly in vision applications, are alarmingly vulnerable to a range of security threats. One key factor in the attack landscape is the distinction between white-box and black-box threat models, as the latter poses challenges that limit attack effectiveness when access to model information is limited. As a result, using Trusted Execution Environments (TEEs) enhances security for machine learning applications by protecting model confidentiality and execution integrity, effectively shifting the execution environment from the white-box to the black-box side of the threat model spectrum. While adopting TEEs for large vision models, e.g., Vision Transformers (ViTs), is crucial for enhancing security and privacy, significant challenges related to memory constraints and increased computational latency must be addressed, especially in time-sensitive edge applications where safety and privacy are paramount. The objective of this work is to enable large vision models to be fully hosted within TEEs, achieving GPU-level inference latency for time-sensitive edge vision applications while maintaining performance. To this end, we propose TEE-X, a TEE-aware acceleration framework that introduces a sensitivity-aware modularization technique and enables vectorization in TEE inference. This design is validated on OP-TEE for Arm TrustZone, configured to optimize performance on the NVIDIA Jetson AGX Xavier for efficient edge vision applications using ViT models. The findings reveal that TEE-X delivers an effective TEE-aware acceleration framework that achieves minimal accuracy-latency trade-offs while ensuring fast and secure edge inference for vision models.

Tool use in LLM systems faces a structural trade-off. Progressive disclosure keeps the prompt small by showing only the tools relevant to the current task, while prompt caching rewards a request prefix that stays fixed across calls; every change to the visible tool list invalidates the cached prefix. This paper treats the trade-off as a problem of request architecture and proposes a dual-path routing design that assigns tool selection and tool delivery to separate channels. The main model always sees a small, fixed set of core tools, so the head of its request is unchanged across calls; all other tools are reached through an independent routing channel, in which a router sub-model searches the full tool list, selects one tool, executes it, and returns the result. Tool registration is automated from source code and supports runtime updates, so the tool set can grow without modifying the main model's request prefix. The design generalizes progressive disclosure: capabilities are disclosed through the routing channel, and the main model's prefix stays stable. A prototype implementation was exercised on 55 functional queries and a 30-turn dialogue; token-level cache hit rates reached 90.99% and 95.2%, cutting input cost to about 12.0% and 8.0% of a no-cache baseline under DeepSeek's pricing, where cache-hit input tokens cost roughly 1/30 of cache-miss tokens.

When shopping is delegated to AI agents, it is unclear whether the ranking advantage documented for humans persists. Across 7,000 sessions with five large language models (LLMs) and varying reasoning effort, we compare AI agents with human field data. AI agents search more extensively than human consumers. As with humans, lower-ranked listings are less likely to be inspected, although the effect is smaller for AI agents. Unlike humans, AI agents show a pattern consistent with the lost-in-the-middle effect, whereby middle listings have the lowest probability of inspection. At the choice stage, position effects are concentrated at lower reasoning effort, but higher effort reduces the middle penalty for every LLM that exhibits it. These findings suggest that, when search is delegated, displayed attributes may matter more than placement, and that exposure to position bias depends on how the AI agent is configured.

Agent ecosystems now include thousands of MATS components (Models, Agents, Tools, and Skills), yet their discovery still relies on in-context routing. These systems read a registry (names, hints, or descriptions, as context budget permits), pick a candidate, invoke it, and retry on failure. This pattern degrades with scale, and registries are growing fast. We recast capability discovery as search over a registry by defining an offline enrichment step that turns sparse metadata into searchable profiles, and an online retrieve-then-rank pipeline that returns a ranked shortlist without invoking any candidates online. We show that from N=10 to 7,278 capabilities, in-context routing's top-1 accuracy (Match@1) collapses (0.85 to 0.12), while retrieve-then-rank degrades more gently (0.81 to 0.39) because its reranker still ranks the right capability first 0.70-0.87 of the time once retrieval finds it. In the Nova Micro sweep, the crossover is around N=500. We compare against two in-context baselines. Full-Ctx puts the whole registry in the prompt and asks the LLM to pick. Search&Pick gives the LLM a search tool to narrow candidates before it picks. At full scale the pipeline leads Search&Pick by 6.5 percentage points (pp) on Match@1 at about half the cost. It reduces cost 70x versus Full-Ctx. We use a fixed configuration (same enrichment, retriever, and scorer weights) across agent, tool, and skill registries. The pipeline runs in production as the default capability-discovery layer of a large-scale multi-agent platform.

UAV vision-language navigation (UAV-VLN) is commonly evaluated as goal reaching, but inspection-oriented deployment requires the agent to stop within a valid inspection region and avoid falsely confirming visually or semantically similar distractors. This requirement exposes a key weakness in existing coarse-to-fine UAV-VLN policies: the coarse goal predicted before local refinement is often treated as reliable, although it may drift toward plausible but incorrect object regions and limit the ability of the local stage to recover. To systematically evaluate this problem, we introduce LG-UVI, an object-centric inspection evaluation setting derived from CityNav/CityRefer. LG-UVI extends standard UAV-VLN episodes with target objects, hard distractors, type-aware inspection regions, and diagnostics for inspection-region arrival and object-level confirmation. To address this inspection-oriented setting, we further propose RACO, a reliability-aware adaptive coarse-to-fine navigation framework. Instead of treating the predicted coarse goal as a fixed waypoint, RACO views it as a runtime hypothesis and uses object-level candidate anchors to check and correct coarse localization before Stage 1 and at the Stage 1-to-Stage 2 boundary. RACO also applies scale-adaptive terminal refinement to handle terminal near-miss cases using runtime-observable geometric and anchor-based evidence. Under a unified online evaluation protocol, RACO improves SR over the reproduced HETT baseline by 9.53 and 7.98 percentage points on validation-unseen and test-unseen, respectively. It also improves inspection-region arrival and reduces false verification risk, showing that coarse-goal reliability optimization is an effective complement to existing coarse-to-fine UAV-VLN policies.

Tool-using agents increasingly rely on external tools to complete multi-step tasks, but tool returns can fail in different ways and require different recovery actions. Existing robustness studies often use uncertainty-based measures to detect when an agent becomes unreliable. These measures can reveal that something has gone wrong, but they do not directly identify the type of tool failure or the appropriate response. We address this limitation by analyzing tool failures at the moment a return enters the agent context. Our approach combines two complementary signals. The first compares the likelihood of the returned content under the tool schema and under the full trajectory prefix. The second measures the agent's probability distribution over its legal next actions. We evaluate the approach by injecting incomplete and inconsistent returns into a retail customer-service benchmark. The results show that likelihood-based signals clearly capture incomplete returns and some direct inconsistencies, while action-based signals reveal how strongly a failure changes the next decision. Some failures that are weak under likelihood signals can still redirect the agent toward state-changing actions. These findings show that tool failures can be recognized at the return boundary, but reliable diagnosis requires combining multiple signals.

Energy system operation contains a loop of work that automation has never taken over: posing the optimization problem the current cycle should solve, disposing of infeasibility, sequencing a solution into interlocked switching orders, assembling evidence no single model holds, negotiating adjustable capacity with many parties, and settling experience into practice. Licensed dispatchers carry all of it in person, and the rising share of variable renewable generation is making that loop turn faster than their number can grow. Agentic AI supplies the abilities it requires, but enters as the outer loop of control: it calls SCED and the other decision models rather than being called by them. We propose A-CPES, three nested rings, an authorization and accountability frame around an agentic control outer loop around a six-layer CPES core. We argue the loop is indivisible, tune where and how tightly it may close, state eight structural failure modes as falsifiable predictions, and specify six governance modules that rebuild the authorization frame until it covers the loop, before the loop starts turning.

We investigate the token mixer in vision backbones by revisiting clustering, one of the most classic approaches in machine learning. An effective token mixer is a fundamental component of modern vision backbones like vision Transformers, facilitating information exchange between image patches. Mainstream token mixers, which rely on convolution, attention, MLP, or their hybrids, primarily focus on navigating the trade-off between accuracy and computational cost. However, a significant drawback of these methods is their black-box nature; their encoding process is opaque and lacks interpretability. Diverging from these opaque designs, we introduce ClusterMixer, a transparent token mixer that is grounded in a clustering paradigm and interpretable by design. ClusterMixer explicitly formulates the token mixing process through a hierarchical clustering mechanism. To model the natural, tree-like relationships inherent in visual data, the clustering is performed in hyperbolic space, which is well-suited for embedding hierarchies with low distortion. Building on this innovation, we present HCFormer, a new backbone architecture that integrates ClusterMixer with a series of meticulously designed clustering strategies to ensure robust performance across tasks. Extensive experiments demonstrate that HCFormer consistently outperforms its counterparts across diverse tasks, including image classification, object detection, instance segmentation, and semantic segmentation. Considering its transparency and efficacy, we hope HCFormer can facilitate a paradigm shift toward interpretable backbones.

Inter example relational distillation transfers a teacher's representation geometry by matching relations among examples within a mini batch. Computing all pairs has quadratic complexity in the batch size, whereas uniform subsampling may use a limited relation budget inefficiently. We introduce Reliability Aware Pair Importance Distillation, or RAPID, which separates a reliability gated relational target from a full support adaptive pair proposal. Reliability determines which teacher relations are emphasized, while calibrated teacher entropy and detached student-teacher residuals determine which relations are evaluated. Exact inverse proposal correction makes the loss and gradient estimators conditionally unbiased with respect to the gated mini batch target. We evaluate RAPID in two text classification settings: AG News with BERT-to-DistilBERT distillation using three paired seeds and a relation budget of 256, and SST-2 with DistilBERT to DistilBERT distillation using three paired seeds and a relation budget of 64. Reliability gated relational distillation achieves the highest observed mean student accuracy on both datasets: 94.285 plus or minus 0.054 percent on AG News and 88.800 plus or minus 0.532 percent on SST-2. RAPID ranks second, achieving 94.241 plus or minus 0.025 percent and 88.685 plus or minus 0.462 percent, respectively, compared with 94.154 plus or minus 0.124 percent and 87.271 plus or minus 0.162 percent for the cross entropy baseline. Pilot evaluations are counted toward the same total budget as the main relation evaluations. Across both settings, the gated target yields the highest mean accuracy, while the adaptive proposal remains within seed-level variation. These results support the modular view that target reliability and evaluation priority are separable design dimensions.

Knowledge-intensive multi-hop question answering requires systems to select evidence and compose dependent facts, yet multilingual benchmarks usually translate an entire example into one language. This hides failures at language boundaries inside the reasoning chain. We introduce XHotpotQA, a controlled benchmark for cross-lingual knowledge composition over mixed-language evidence. Each instance is modeled as an evidence-dependency graph whose question, bridge evidence, answer-bearing evidence, and distractors have explicit language assignments. The audited resource contains 15,661 training and 7,405 validation instances, with sentence-level support supervision and supplied distractors. In validation, 99.81% of items cross the question-to-gold-evidence language interface and 95.60% use gold paragraphs in different languages. Across three reader artifacts, full question-evidence mismatch is associated with 10.25 to 15.79 lower Unicode-aware answer F1 than partial alignment, and different-script evidence with deficits of 11.98 to 23.70 points; the corresponding adapted-selector contrasts are 1.71 and 1.78 points. Under this supplied-candidate design, the evaluated readers therefore show substantially larger condition-associated deficits than the selector. XHotpotQA provides role-aware diagnostics, modular evaluation, and an audited test bed for knowledge-based systems that must integrate evidence across languages.

World models enable agents to reason about future outcomes and learn policies from their knowledge of state transition, but existing approaches primarily focus on reconstructing future observations or features, which introduces unnecessary complexity and limits their effectiveness for decision making. In this work, we propose a compatibility prediction Latent World Model (LWM) for robot navigation that predicts action-conditioned latent feature compatibility rather than reconstructing observations. Our key insight is that spatial proximity correlates with latent feature similarity, enabling action consequences to be evaluated directly in latent space. To support counterfactual training, our model leverages action sequences sampled across trajectories and learns to predict which sequences lead closer to the goal. Furthermore, we demonstrate how the learned world model can supervise policy learning from unlabeled video data and further improve policies through reinforcement learning entirely within the world model. This imagination-driven framework eliminates the need for action annotations and additional environment interaction. Extensive experiments on multiple real-world robot navigation datasets show that our approach significantly outperforms prior world model and imitation learning methods in prediction accuracy, policy learning, and real-world navigation performance. The code, pretrained models, and additional materials are available at https://wzm206.github.io/latent-world-model-nav.

Tool-using agents fail two ways: choosing the wrong tool, or forming wrong arguments, and an early failure of either kind can silently corrupt everything downstream. We measure a correct-invocation rate that separates the two, under both a clean teacher-forced context and the model's own free-running context, on five open-weight models over contamination-free multi-step tasks (depths 1-8). By depth 6, roughly 70% of a model's own clean-context capability is lost to its own earlier mistakes (L6 = 0.686, 0.684). Our central finding concerns the measurement itself. Under exact-match scoring against a fixed gold trajectory, a propagation model's severity and recovery parameters are not merely hard to estimate - they are fixed by the scoring rule. Severity is forced to its boundary (0 of 869 poisoned steps correct); recovery is structurally unobservable (0 of 580 poisoned steps returned on-track, against an expected 0.0058 by chance). Both follow from one mechanism: post-divergence, the gold value is generated by tool constants the model never sees, so it is information the model cannot derive. A fit run anyway returns 0.92 and 0.73 for a quantity that is exactly 1.000 - confident numbers for a parameter the scoring rule already determined. We give the mechanism and a remedy, conditional-on-state scoring, applied retrospectively to cached completions at zero additional cost, which un-pins severity to interior estimates excluding zero (+0.149, +0.316).

Large language models are increasingly used to inform safety decisions in cities, such as where it is safe to walk, rent, or travel. We ask whether such judgments track measured risk or the patterns attached to an urban neighborhood's name. We probe seven instruct-tuned models under three conditions that dissociate name from geography: coordinates-only, name-only, and name+coordinates, across 186 neighborhoods in Los Angeles and Chicago, joined to violent crime and American Community Survey data. First, ratings are nearly flat under coordinates for six of seven models, while names carry most between neighborhood variation and are moderately calibrated to violent crime; only at frontier scale does the coordinate channel show appreciable variation. Second, names lower safety ratings more for neighborhoods with higher shares of the locally dominant marginalized group (percent Black in Chicago, percent Hispanic in Los Angeles), and this name effect tracks demographic share in all seven models and both cities. In Los Angeles, where demographic share and crime are more separable, the effect survives controls for crime and income and is confirmed by crime-matched pairs. An enforcement-elasticity analysis further shows that over-caution tracks near-fully-reported homicide rather than discretionary, deployment-driven offenses. Third, the effect scales with geographic knowledge: models that better distinguish real neighborhoods apply more demographic stereotype to them. Because neighborhood names carry both genuine crime signal and demographic stereotype, removing names reduces both bias and accuracy. We discuss implications for deploying LLMs in advice and decision-support settings.