Papers

Hidden-state hallucination probing - training a linear classifier on an LLM's internal activations to detect whether a generated answer is faithful to the input - is an active area of 2026 research, with recent work reporting 0.90-1.00 AUROC across several benchmarks and languages. However, none of this work has tested probes on code-mixed input, despite the fact that a huge population of chatbot users write in Hindi-English code-mixed text ("Hinglish"). We address this gap directly: does a hallucination probe trained on clean-language hidden states transfer to Hinglish, or does the signal degrade under code-mixing? We construct a 5,674-item Hindi/English/Hinglish QA benchmark, generate and label 17,022 model responses across three open-weight 7-8B LLMs (Qwen2.5-7B, Mistral-7B, Llama-3.1-8B), extract per-layer hidden states at two token positions, and train linear and MLP probes for in-distribution detection and cross-lingual transfer. We find that the hallucination signal survives code-mixing well: transfer AUROC ranges from 0.88 to 0.99, with gaps of mostly under 0.05 AUROC relative to in-distribution performance, and that Hindi-trained probes transfer to Hinglish more reliably than English-trained probes. As an independent, practically motivated finding, all three models hallucinate substantially more on Hindi and Hinglish than on English for matched facts. We release our code and synthetic Hinglish QA dataset to support further work on code-mixed hallucination detection.

Multimodal large language model (MLLM) agents are increasingly used as personal assistants for long-running tasks. Their utility depends on continuity: agents must retrieve and use earlier evidence across dialogue, files, and workspace state. However, agents can generate plausible answers even when access to that history has degraded, causing outcome-only evaluation to overestimate true evidence use. We present MIRAGE (Multimodal Interaction Retrieval, Attribution, and Grounding Evaluation), a controlled study of historical evidence use under conversation-state variation in multimodal personal agents. MIRAGE holds evidence objects, questions, and scoring fixed while varying only conversation state, and evaluates whether an agent can determine answerability, recover the correct source, and answer from it. Across seven frontier and open-weight multimodal backbones, we find that: 1) pre-compaction depth and post-compaction continuation form distinct, non-monotonic failure regimes rather than a single degradation curve; 2) open-weight models rely heavily on context continuity and are reluctant to spontaneously switch to tool-mediated retrieval when provenance fails; and 3) retrieval pressure improves source attribution in deep pre-compaction states for tool-compliant models, but consistently regresses after compaction, where stored evidence has already degraded. These findings show that historical evidence use should be evaluated under state variation, rather than inferred from outcome-only correctness.

We study Langevin diffusion and Langevin Monte Carlo (LMC) when the target distribution changes over time. Under a log-Sobolev inequality (LSI), we derive non-asymptotic Rényi-divergence guarantees for tracking the current target. The framework covers continuous-time Langevin diffusion and its discretizations. We then apply the results to nonsmooth sampling based on successive Moreau envelopes. For this scheme, we give explicit choices of the smoothing parameters and step sizes, together with corresponding complexity bounds. To our knowledge, these are the first non-asymptotic Rényi-divergence tracking bounds for Langevin dynamics with discrete target updates.

Cooperation is a cornerstone of human societies, enabling collective progress in dynamic and uncertain environments. With the advent of AI systems acting autonomously, it becomes crucial to understand not only human-AI cooperation but also AI-AI interactions in adaptive networks. In this work, we examine the interactions of AI using Large Language Models – Mistral, Llama3, Gemma3, and Phi3 – in a public goods game within dynamic network structures. Our experiments were conducted under single-model and mixed-model conditions across Watts-Strogatz (WS), Barabasi-Albert (BA), and Erdos-Renyi (ER) networks. We analyzed the impact of model architecture, network topology, and prompt design on cooperative behavior. Results show that Mistral and Llama3 offer high cooperation rates, while Phi3 shows defective tendencies. Additionally, the random structure of Erdos-Renyi networks dramatically improves cooperation. Prompt design also plays a key role; a society-benefits prompt leads to a higher cooperation level. These findings offer a preliminary framework for LLM-based simulations in adaptive social networks.

Recent machine learning weather prediction (MLWP) models have demonstrated remarkable forecasting skill on global reanalysis-based benchmarks. However, their performance remains unclear in challenging environments such as Northern Norway, where narrow fjords and rapidly changing weather result in highly variable local wind conditions. In this case study, we evaluate FourCastNet3 (FCN3), GraphCast, and ECMWF High Resolution Forecast (HRES) for wind speed forecasting using multi-year station observations from Northern Norway, focusing on their relative performance, generalization beyond the training period, and performance under high-wind conditions. Our results show that HRES slightly outperforms FCN3 and GraphCast, with an overall RMSE of 2.89 $\mathrm{m\,s^{-1}}$, compared to 2.96 $\mathrm{m\,s^{-1}}$ for FCN3 and 2.94 $\mathrm{m\,s^{-1}}$ for GraphCast. Notably, the MLWP models maintain comparable performance beyond their respective training periods, with no clear evidence of noticeable degradation. FCN3 performs best under high-wind conditions, although all models substantially underestimate strong winds. Our findings suggest that MLWP has become competitive with NWP for local wind, but further refinements are still needed to capture complex terrain better.

Recently, a new record kissing arrangement of $841$ points in $\mathbb R^{12}$ was obtained numerically by optimization (Takhanov-Assylbekov-Yun, 2026). The configuration was released as a coordinate file, without a mathematical description of its structure. The purpose of this paper is to provide such a description. The key observation is that the geometry becomes transparent once we regard $\mathbb R^{12}\cong \mathbb H^3$ as the Cartesian product of three copies of the quaternion algebra. We first introduce a new $840$-point kissing arrangement with a certain quaternionic structure. It consists of three mutually orthogonal regular $24$-cells, supported on the three quaternionic coordinate factors $\mathbb H\times\{0\}\times\{0\}$, $\{0\}\times\mathbb H\times\{0\}$, $\{0\}\times\{0\}\times\mathbb H$, together with two $384$-point families obtained by lifting affine sets of the form $$\{(u,v,w)\in (\mathbb F_2^2)^3\mid u+v+w=\eta\},$$ to quaternionic triples (whose components belong to the binary octahedral group $2O$) and then applying suitable component-wise rotations and weightings. A characteristic feature of this construction is a pronounced asymmetry among the three quaternionic factors. For the $816$ vectors obtained after removing the third $24$-cell, most of the squared norm is concentrated in the first two quaternionic coordinates, while the third coordinate carries systematically less mass. Thus, the third four-dimensional factor contains more available space than the first two. We then show that this $840$-point configuration provides a natural structural model for the numerical $841$-point record. Finally, we introduce a notion of the general quaternionic construction in dimensions divisible by $4$, and check that record kissing arrangements in ${\mathbb R}^{4k}$, $k\leq 5$, admit a quaternionic construction.

The computation of Black-Scholes implied volatility is a fundamental task in quantitative finance, underpinning option valuation, model calibration and risk management. Although implied volatility is routinely used in practice, the inversion of the Black-Scholes pricing formula remains a challenging numerical problem, particularly in asymptotic regimes corresponding to extreme option prices, strikes or maturities, where the inverse map becomes highly sensitive to perturbations of the price. In this paper, we introduce a new family of asymptotically-informed neural-network architectures for implied-volatility computation. Exploiting the distinct behaviours of the Black-Scholes pricing function in different volatility regimes, we propose a family of architectures that learn a trainable partition of the price-log-moneyness domain through a system of gating functions and combines specialised local approximations of the implied-volatility function within each region. Extensive numerical experiments demonstrate that the proposed models consistently outperform standard feed-forward neural networks across a wide range of parameter domains, often by several orders of magnitude in relative accuracy while maintaining excellent generalisation properties. Furthermore, the neural-network outputs provide highly accurate initial guesses for a third-order Householder scheme, allowing near machine-precision implied-volatility computations after only two refinement iterations.

In signaling games where a sender and a receiver have misaligned criteria, equilibrium behavior may lead to fully revealing, quantized, or randomized policies. Notably, the first arises in statistical decision theory and classical communication theoretic problems involving a fully aligned sensor and receiver, the second arises in Nash theoretic simultaneous signaling games, and the last may appear in Stackelberg type (leader-follower) Bayesian signaling games. In this paper, we investigate the Bayesian persuasion problem involving a receiver that tries to estimate the source. We show that for certain payoff structures, the equilibrium solution is such that a source observation is mapped to distinct messages with nonzero probabilities. More specifically, we completely characterize conditions under which the sender requires randomization for the Bayesian persuasion problem involving general sources with finite cardinality. In particular, regardless of whether the equilibrium solution under a deterministic policy restriction is fully revealing, quantized or noninformative, there exists a randomized sender policy that improves the sender's payoff under certain conditions characterized in the paper. Moreover, we provide an algorithmic procedure to obtain the Bayesian persuasion solution, where the algorithm compares the payoffs with finitely many posterior probability combinations. We also consider fully aligned and completely misaligned payoff structures, where the solutions respectively involve a fully revealing sender and a noninformative sender. Then, we unify these results by proving that if the sender's expected payoff with respect to posterior distributions is continuous, then the equilibrium solution involves either a fully revealing sender or a noninformative sender.

Conformance suites for quantized GEMM kernels ask whether two implementations agree within a tolerance. We measure what such a suite can detect. Injecting nine faults into a reference INT8 pipeline over 8,232 layer–fault–regime cells of Qwen3-1.7B, we find that every one of five epilogue faults – scale precision, double rounding, multiplication order, output truncation, fused ordering – moves the output by at most a single bfloat16 spacing, and by exactly one whenever it moves it at all, across 5,880 cells. A tolerance of one spacing is therefore blind to the entire class by construction: four of the five faults are detected by no check in the suite, and the fifth only under power-of-two scales. Faults that violate the accumulator's exactness preconditions, or that break operand sharing, are detected without exception, and a null fault never fires. What a tolerance-based suite of this shape establishes is therefore narrower than interchangeability: that the preconditions hold, that operands are shared, and that differences stay within one spacing. The power-of-two constraint that exposes the one detected fault is also deployable. Requantizing every weight scale to its nearest power of two makes CUTLASS and Triton agree bitwise at every linear layer (196/196 and 252/252, against 8/196 and 10/252 under the checkpoints' own scales) and yields byte-identical generated token sequences at 1.7B, 8B and 14B (8/8 prompts, against 0/8 at all three). Observed perplexity point estimates are +0.32%, -0.28% and +0.48%; the 90% intervals cover zero at the two smaller sizes but not at 14B, reaching +0.71% and +0.76%. A previously reported +157% perplexity for this intervention was an artifact of a probe that rewrote scales without requantizing the weights; separating the effects attributes 99.8% of it to the resulting weight–scale mismatch rather than to the power-of-two constraint itself.

This paper presents \ul{CogRun}, a framework that enables safety-critical ground robots to perform cognitively-grounded runtime learning entirely on edge-AI devices in unknown physical environments, without prior maps or perceptual knowledge. CogRun consists of three components: a Learning-Agent, a Rational-Agent, and a Coordinator. The Learning-Agent is novel in cognitive-neural learning architecture, which featurs dedicated replay buffers, cognition-driven experience sampling, and a safety-aware action blending of actor-critic reinforcement learning (RL) with instance-based learning (IBL). The Rational-Agent is a non-learning module that complements the Learning-Agent by exclusively handling safety-critical functions, while the Coordinator manages interactions between the two agents to promote safe and efficient runtime learning. CogRun's full autonomy stack (i.e., perception, learning, and control) on edge-AI devices eliminates dependence on wireless communications, enabling broader applications in challenging environments with limited or no connectivity. Experiments on a quadruped robot in real-world wild forests and on an off-road autonomous vehicle in a simulated wild forest demonstrate that CogRun enables safe and efficient runtime learning, allowing robots to safely and continuously interact with the physical world for enhancing task performance in complex, unknown environments.

Digitized herbarium collections, now comprising over 100 million freely accessible specimen images, have become a critical resource for addressing fundamental questions in ecology and evolutionary biology. Yet the rich metadata encoded in herbarium labels (collector identities, geographic localities, collection dates, and ecological observations) remains largely inaccessible at scale, constraining both biodiversity informatics and the construction of specimen-specific image-text corpora for multimodal AI. We present HERBIOME, a modular end-to-end pipeline for automated herbarium label digitization, integrating YOLOv8-based component detection, CRAFT Hezar word-level text localization, fine-tuned TrOCR for recognition of mixed handwritten and printed text, and GPT-4o Mini for semantic metadata structuring into standardized fields. TrOCR was trained on a multi-source dataset combining general transcription corpora (CREMMA-AN, PictoCatalogs) with herbarium-specific data (RéColNat), achieving a Character Error Rate of 4.05-4.10%. End-to-end evaluation on 450 French herbarium specimens, using a dual-metric framework of Maximum Window Similarity (MWS: 0.614-0.618) and Semantic Metadata Accuracy (SMA: 0.440-0.445), reveals that hybrid training strategies improve semantic fidelity while random sampling maximizes surface similarity, with taxonomic fields remaining the principal bottleneck. By automating the extraction of structured metadata from complex, heterogeneous labels, HERBIOME reduces transcription burden, enables the construction of paired image-text datasets that faithfully capture specimen individuality, which is a prerequisite for next-generation multimodal biodiversity AI systems.

Video reasoning tasks such as grounded video question answering and temporal grounding require selecting temporal evidence that supports the query. In many current training setups, temporal supervision is applied through local objectives such as boundary regression or span generation, while verification is used mainly to rerank candidate segments at inference time. We study whether a frozen verifier can also guide training. Our multi-agent framework couples a trainable Grounder with a frozen Verifier: the Grounder samples candidate trajectories and evidence segments, the Verifier assigns query-conditioned segment scores, a group-relative policy-gradient objective favors trajectories that outperform their within-input peers, and a bootstrapped calibration loss steers temporal predictions toward verifier-preferred spans. Trained on source tasks and evaluated without target-dataset fine-tuning, a two-billion-parameter instantiation transfers zero-shot across grounded question answering, temporal grounding, and long-video question answering, reaching 28.7% intersection-over-union and 25.4% answer-grounding accuracy on a grounded-question-answering benchmark, 46.1% intersection-over-union on a temporal-grounding benchmark, and 54.1% on a long-video question-answering benchmark. Relative to a strong same-scale baseline, the gains are modest but consistent, with the clearest improvements on relevance-oriented metrics such as intersection-over-union and moderate-overlap recall. Within the tested benchmarks and transfer setting, the results support frozen verification as a training signal for evidence selection, while showing that strict boundary precision remains comparatively weaker. Code and models are available at https://anonymous.4open.science/r/MASIRL-E50C/

Multi-exposure is an efficient way to capture real-world high-dynamic-range (HDR) scenes. However, HDR imaging suffers from severe ghosting artifacts in dynamic scenes due to the temporal gap between sequential exposures. In this article, we categorize the literature on two important topics on HDR imaging: multi-exposure fusion (MEF) and ghost removal. Conventional filter-based and data-driven methods are studied in pixel space and feature space. For popular deep learning-based approaches, we provide a granular taxonomy based on their alignment and fusion domains: pixel-space methods, which typically employ explicit motion compensation such as optical flow or spatial transformers, and feature-space methods, which leverage implicit alignment through deformable convolutions, attention mechanisms, or latent representation merging. Representative works are compared across different supervision settings, and key design principles are summarized. In addition, this survey summarizes commonly used datasets and evaluation metrics, discussing their applicability under diverse output forms. Finally, major bottlenecks and promising directions for future research are outlined.

Artificial intelligence is transforming personalized healthcare, yet fragmented clinical, self reported, and wearable evidence remains difficult to interpret and trace. We present CareGraph, an auditable hybrid AI framework that converts heterogeneous records into prioritized trends, missing context indicators, bounded next steps, discussion questions, and provenance linked explanations. CareGraph organizes evidence without diagnosing, predicting outcomes, selecting treatment, or making autonomous clinical decisions. Its pipeline covers deterministic analysis, context detection, graph construction, constrained language model synthesis, evidence validation, safety controls, and release gating. Tests used synthetic cohorts of 400 patients each for development, validation, and holdout. On holdout data, a frozen ordinary least squares trend rule with a sufficiency gate achieved 0.827 accuracy, 0.837 macro F1 with a 95 percent confidence interval of 0.819 to 0.854, and 0.974 insufficient data F1. Missing context detection achieved 0.815 strict micro F1 versus 0.318 for the legacy detector. On an authored holdout benchmark, safety ruleset version 1.2 achieved 1.000 precision, 0.950 recall, and 0.974 F1. An audit requiring graph retrieval across 80 patients yielded 79 syntheses and 78 presentations without fallback; one output was blocked and one failed closed because of an invalid evidence key. Against monolithic GPT 5.6 on 56 matched patients, CareGraph was faster at 40.15 versus 49.62 seconds, shorter at 661 versus 1,163 words, and showed better exploratory lexical alignment with longitudinal targets; the baseline used fewer tokens and cited more raw evidence. Graph auditing verified provenance and deterministic retrieval; incremental graph effects on generation require paired evaluation. CareGraph offers a safety bounded foundation for intelligent personalized health systems.

Objective: Recent studies demonstrate that functional connectomes contain subject-specific signatures, or fingerprints, that can identify individuals across repeated sessions and tasks. Existing methods mostly rely on edge-level features that are sensitive to noise, difficult to interpret, and limited in their ability to generalize across tasks and datasets. Methods: We propose a multiscale community-based functional connectome fingerprinting framework that characterizes each individual by the mesoscale structure of their functional networks. We introduce a signed multilayer community detection framework that incorporates both correlated and anti-correlated brain activity to identify subject-specific community structures across tasks and sessions. Graph-theoretic metrics are then computed from the resulting joint community structures to derive low-dimensional community-level fingerprint representations. Results: The proposed framework is evaluated on 810 healthy control subjects from the Human Connectome Project (HCP). The results show that community-based fingerprints provide a reliable and interpretable substrate for individualized brain characterization across sessions and tasks. Conclusion: Mesoscale community structure provides meaningful and discriminative subject-specific fingerprints. Significance: The proposed framework offers a promising foundation for precision neuroimaging and personalized neuroscience applications.

We study context localization for generalized level-based evaluation in knowledge-based systems. The framework models situations where a structured nonnegative score, defined on facts, rules, cases, criteria or evidence units, is evaluated through conditional aggregation tests on admissible knowledge contexts. The generalized level measure maximizes a monotone set function over all contexts whose aggregated support reaches a prescribed level. We characterize when filtering the score by a context $B$ is equivalent to localizing the admissible contexts by intersection with $B$. The main theorem shows that this consistency holds for all monotone set functions if and only if two structural conditions are satisfied: monotonicity with respect to contexts and a reduction property excluding positive localized support outside $B$. We analyze pointwise and block-generated mechanisms producing the reduction property, extend the result to parameterized systems, and interpret it as a stability criterion for context-dependent evidence selection, non-additive support evaluation and level-based knowledge aggregation.

Computer Use Agents (CUAs) are increasingly deployed to navigate mobile and desktop applications on behalf of users, yet no benchmark comprehensively evaluates whether they can safely interact with visual interfaces while handling ambiguous instructions. We introduce ADeptS-Bench, a dual-stream trustworthiness benchmark, grounded in the ADEPTS capability framework and general population user studies. The Safety stream provides paired benign/malicious tasks with threats embedded in the visual interface. The Disambiguation stream evaluates whether agents seek clarification when intent is ambiguous. Evaluating seven models reveals that no model consistently exceeds 80% task success while staying below 30% attack success; every model clicks "Checkout" on a $25K order without hesitation, and none detects that a "factory reset" button is mislabeled as "Optimize." An ablation reveals three distinct safety architectures: tool-dependent (ASR +21-23pp without refusal tool), partially tool-dependent (+10-11pp), and no mechanism (unchanged). In disambiguation, all models overestimate consequence severity, mirroring the over-refusal bias observed in safety. We release all data, evaluation code, and analysis tools upon publication.

Modern video games combine first-person perception, rapid visual changes, persistent world state, and heterogeneous native controls. Existing game agents map visual and task context directly to actions but lack explicit world dynamics modeling, whereas interactive game world models predict visual futures from supplied actions but do not serve as task policies. World-Action Models (WAMs) unify these objectives, but remain largely unexplored under the dynamics and open-ended interaction of video games. We introduce GameWAM, to our knowledge the first WAM for native closed-loop gameplay and GUI control. GameWAM jointly generates future visual observations and executable keyboard-mouse trajectories through parallel visual and action generative processes with block-causal conditioning and flow matching. To support joint world-action learning, we construct synchronized gameplay and GUI trajectories. To handle heterogeneous native control, GameWAM predicts a gameplay/GUI mode per action step and generates actions with mode-specific prediction distributions and continuous-action normalization. For long-horizon interaction, block-cycle control coordinates prediction, execution, and temporal context: it predicts beyond the committed horizon, executes short action blocks, replans from new observations, and hierarchically structures context from fine-grained within-cycle history to persistent cross-cycle history. Experiments demonstrate competitive task success with fewer executed native actions than the compared agents. We further uncover Low-Frequency Action Source Imprinting (LASI), in which low-frequency components of the sampled action source systematically steer coarse generated camera motion under fixed conditioning, revealing a source-sensitivity failure mode in generative control. Project page is available at https://yunncheng.github.io/GameWAM/.

We ask whether AI agents powered by locally deployed large language models can reliably automate expert-defined hardware design workflows in an industry-realistic tool-calling setting. In these environments, engineers issue repetitive, dependency-ordered operations—such as creating components, adding ports, and wiring connections—through specialised tools. Confidentiality constraints on component specifications and naming conventions often preclude hosted proprietary APIs, motivating the use of locally deployed models. To study this setting, we build a Model Context Protocol (MCP) server that reproduces the state and dependency logic of a proprietary hardware design tool used in embedded system development and construct a benchmark covering single-operation edits, multi-step dependency chains, invalid requests, misspelled prompts, and multi-server tool contexts. We evaluate seven open-source models comparing pipeline choices including system prompts, tool-description detail, context scope, and single-agent versus multi-agent architectures. Results show that strong models can achieve near-complete expected-call coverage on the benchmarked workflows, but reliability depends strongly on both task structure and agent configuration. Comprehensive tool descriptions consistently reduce failures, few-shot prompting can cause severe inaction for some models, cumulative context harms constrained models, and multi-agent decomposition helps weak workers or long sessions at the cost of additional calls. These findings provide practical guidance for deploying local LLM agents in stateful hardware design environments.

Operating a scientific instrument such as an atomic force microscope (AFM) requires continuous expert decision-making. A trained user defines the experimental intent, translates it into instrument commands, assesses incoming data, adjusts imaging parameters, and post-processes the final image. Existing automation usually addresses only parts of this workflow through hard-coded routines, task-specific controllers, or trained machine-learning models. Here we present an agentic-AI framework that operates the executable part of the AFM workflow using a general-purpose, tool-augmented large language model connected to instrument functions through the Model Context Protocol (MCP). The framework consists of 3 MCP-based agents: AFM Messenger converts natural-language instructions into checked instrument commands; AFM Pilot assesses image quality through a large language model (LLM) and, if necessary, adapts imaging parameters; and AFM Doctor diagnoses image artifacts and applies transparent post-processing from a pre-approved tool set. Because the language model performs image assessment rather than a fixed scalar objective or external optimizer, the same strategy can be applied across sample types and imaging modes without specific retraining. Safe hardware operation is enforced through an ambiguity check layer before execution. Benchmarking against fine-tuned and off-the-shelf tool-using models shows that this guarded execution layer, rather than model capability alone, reduces wrong-command execution to zero. In live experiments on different samples, AFM Pilot matched expert operators in image quality, iteration count, and tuning time, with no significant difference. These results demonstrate a safe route to agentic operation of scientific instruments, where experimental intent remains human-defined while command execution, image-based tuning, and post-processing are delegated to AI agents.

Large Language Model (LLM) based agents exhibit substantial run-to-run execution variance even when given identical tasks and tools – acceptable for exploratory use but unacceptable in regulated domains such as finance and compliance. We study harness engineering: wrapping an agent in a deterministic execution layer (finite-state control, forced tool selection, output validation, bounded retry, and structured planning) and measuring its effect on execution determinism and task success. Across two synthetic tasks (finance and legal) and two open-weight models (Qwen-2.5-7B-Instruct, Gemma-3-27B), a first-pass harness produces a mixed result: it significantly improves reproducibility in one of four model-task cells, significantly degrades it in two, and has no effect in the fourth. A trace-level diagnostic finds the cause: once tool sequence, state sequence, and output are already highly consistent, an unconstrained free-text planning step becomes the dominant remaining source of variance. Adding Structured Planning – validating the plan against a fixed schema before any tool is invoked – eliminates the effect entirely: three of four cells reach a Reproducibility Rate and Determinism Index of 1.000 at N=100, and task success rises to 100% in three of four cells. The gain is not free: token cost falls in every cell, but latency shows a genuine, sample-size-robust split by model – one becomes faster under the constraint, the other markedly slower. We argue harness engineering is a distinct, effective discipline for agent reliability, but its cost is model-dependent and must be measured, not assumed.

LLM agents execute tasks through multi-step trajectories that accumulate cost in tokens, latency, monetary fees, and environmental risk while producing utility only at the aggregate task level. Prior surveys address inference optimization, agent capabilities, or evaluation in isolation, leaving practitioners without principled tools to determine whether a trajectory's resource expenditure is justified by its task contribution. We address this gap by developing a trajectory-centric cost-utility alignment framework that treats resource consumption and task contribution as dual ledgers over the same execution, organized around five analytical stages: cost profiling, utility attribution, misalignment diagnosis, targeted adaptation, and evaluation. Utility attribution is central to this structure: rather than relying on aggregate outcomes, it organizes contribution methods by evidential strength, from process proxies and information dependency to counterfactual replay, supplying the causal evidence that grounds diagnosis and guides adaptation. Using this framework, we analyze recent agent systems, attribution methods, and evaluation protocols covering efficiency, reliability, and economic value, as well as five forms of misalignment spanning cognitive and context use, external interaction, recovery-loop control, resource-capability allocation, and multi-agent coordination, together with their targeted adaptations. The result is a closed analytical loop connecting the cost side of agent execution to its utility side, providing a structured basis for resource-aware agent design and deployment.

In dynamic collaborative systems, the selection of reliable collaborators is critical to ensuring effective task execution. Existing trust evaluation methods often rely on unidimensional or scalar representations, which fail to faithfully capture a collaborator's true trustworthiness, thereby motivating a shift toward multi-dimensional trust modeling. However, due to the asynchrony of collected trust-related data across different dimensions, as well as the complex intra- and inter-dimensional dependencies embedded within these data, multi-dimensional trust evaluation remains challenging. To address these challenges, we propose TrustFormer, a task-specific multi-dimensional trust evaluation framework. Specifically, TrustFormer leverages task identifiers and device-generated timestamps to synchronize heterogeneous trust-related data across historical collaborations. It further employs cross-temporal and cross-dimensional attention mechanisms to jointly model temporal dynamics and inter-dimensional correlations, thereby effectively learning the multi-dimensional trust evolution of potential collaborators from historical performance data. In addition, according to the multi-dimensional resource requirements of tasks, potential collaborators' multi-dimensional resource trust is evaluated. Finally, by synthesizing these multi-dimensional trust profiles, the framework enables the optimal collaborator selection. Experimental results demonstrate that TrustFormer outperforms existing methods by yielding a 40.8% improvement in trust evaluation accuracy and enabling more reliable collaborator selection.

Clinical decision-making often involves prioritizing ethical values, such as beneficence, non-maleficence, respecting a patient's autonomy, and justice. Recent work has begun to assess how large language models (LLMs) make such subjective, value-laden clinical judgments. However, evaluations of LLM decision-making in rare disease care contexts, where ethical tensions are ubiquitous and where scarce prior information likely impacts LLM behavior, are still lacking. Here, we present a benchmark of 208 clinically grounded rare disease vignettes, each of which presents genuine, high-stakes conflicts. When prompting 11 state-of-the-art LLMs to choose between clinically defensible yet ethically conflicting next steps embedded within these vignettes, we found that all evaluated models consistently prioritized justice over other core bioethical principles. Specifically, models overwhelmingly favor equal resource allocation over need-based considerations, indicating LLMs' limited responsiveness to differences in clinical severity or situational context. We also identify a strong authority-framing effect: models favor justice in committee-based contexts and shift toward beneficence and autonomy only when final decisions are framed as being made by clinicians or patients respectively. Our work suggests that institutional pressures surrounding rare disease resource utilization may be silently reflected in LLM-based decision support systems, with finer ethical considerations disregarded.

Selection of trustworthy collaborators in distributed systems is critical for efficient task completion, necessitating the inference of trustworthiness from their past collaboration experience. However, as a collaborator serves distinct devices across diverse scenarios in past collaborations, its trust-related data, observed from different device-specific views, is inherently multi-source, heterogeneous, and uneven in quality. Consequently, achieving accurate trust evaluations for collaborator selection remains a major challenge. To tackle these issues, we propose a novel multi-view evidential learning (MVE) based trust evaluation method. First, to accommodate the multi-source heterogeneity of observed trust-related data, we model each task owner who has interacted with a potential collaborator as an independent observational view, enabling the evaluation of the collaborator's view-specific trust. Second, to address the dynamic evolution of trust under changing conditions, we leverage the powerful long-sequence modeling capability of the Mamba model to capture the deep temporal patterns of a collaborator's trust state within each view. Furthermore, to quantify the certainty levels of view-specific trust assessments, we incorporate an evidential deep learning mechanism in MVE, which outputs trust evaluation results while quantifying the subjective uncertainty underlying them. Finally, we employ a dynamic evidential fusion strategy to adaptively integrate the multi-view evidence based on their respective quantified uncertainties, thereby yielding a final trust evaluation for the collaborator. Extensive experiments demonstrate that the proposed MVE method outperforms baselines in both trust evaluation accuracy and task success rate.

Effective selection of trustworthy collaborators is crucial to ensuring the successful completion of collaborative tasks, which requires accurate assessments of both long-term device behavior and short-term collaborative dynamics. Consistent device behavior patterns, which are learned from historical collaborations, can be used to predict their reliability in future collaborations. However, accurately assessing device behavior based on historical collaborations remains challenging. First, behavior assessment from limited historical collaborations captures only instantaneous past behavior, failing to represent the devices' true behavior. Second, due to the temporal dependencies of device behavior, a unidirectional evaluation that relies only on earlier collaborations loses the opportunity to learn from subsequent collaborations. Addressing these challenges requires evaluating device behavior based on long-term collaborations while considering both forward and backward temporal dependencies. To this end, this work proposes a bidirectional Mamba-enabled model (BM) for long-term behavioral evaluation. For each short time slot, a graph is constructed among devices based on historical collaborations, and device behavioral features within the slot are then aggregated accordingly. Subsequently, a bidirectional Mamba model integrates these short-term representations across all time intervals, producing a stable and reliable long-term behavior evaluation for each device. Experimental results demonstrate that BM achieves higher evaluation accuracy than baseline methods, thereby enabling the selection of collaborators that maximize the value of task completion.

Mixture-of-Experts (MoE) models appear to encode moral content as robustly as dense models, yet prove far more fragile in their encoding. In OLMoE-1B-7B, linear probes recover moral valence from nearly every expert-layer combination, with mean peak-layer accuracy above 90%. But these representations collapse under levels of activation noise that a dense model of matched size easily tolerates, with a 4.2-fold difference in robustness. We trace this to output dilution. Because the MoE block averages across active experts before contributing to the residual stream, the feedforward signal reaching downstream layers is nearly two orders of magnitude smaller than in a dense MLP. Moral information, our interest, survives aggregation intact but at a scale trivially overwhelmed by perturbation. Routing itself remains stable under noise while the vulnerability originates entirely in the diluted aggregate. Checkpoint trajectories confirm this is architectural, not learned. Experts never specialize and accuracy saturates within the first few thousand steps. In sparse architectures, redundant encoding does not imply robust encoding.

Every deployed KV-cache-eviction rule and every top-$k$ sparse-attention rule keeps a subset of the keys, discards the rest, and renormalizes the attention weights over the kept set. Identifying the exact best subset under that constraint, we find that keeping the largest weights is already near-optimal, since the best subset we find closes only a median $2$ to $5\%$ of the remaining gap to full attention. We evaluate deployed eviction methods in KVPress and find that methods with non-uniform per-head budgets apply attention masks over the full cache because dense cache tensors require equal length across heads. Truncating each head to the target budget lowers benchmark scores by $14$ to $62$ points, and computing token importance with the question in context accounts for $87.6$ points in retrieval accuracy. We propose ContourKV, a training-free rule that applies KVzip's score under a physically enforced budget. It beats KVzip in $93$ of $160$ benchmark runs across RULER and LongBench while using $28$ to $1078$ times less memory, and ties Compactor, the strongest baseline at equal memory.

We introduce CIPS, a contract-driven execution abstraction for managing computational latency and sampled-data updates in safety-critical cyber-physical systems (CPS). A fundamental challenge in real-time control is that physical safety certificates are defined spatially, yet predicting their validity under non-zero computation and handoff latency requires online numerical integration of plant dynamics. CIPS resolves this operational dichotomy by systematically compiling heterogeneous spatial safety certificates into normalized, unit-rate temporal contracts entirely offline. This transformation abstracts complex plant dynamics, exposing a deterministic, O(1) temporal budget to a generic sampled-data scheduler. We formally prove that this architecture guarantees global hybrid safety invariance under bounded computational latency and asynchronous execution. Finally, we validate the framework via an autonomous vehicle braking benchmark, demonstrating a 60.5 times reduction in micro-architectural evaluation overhead compared to latency-aware event-triggered control (LA-ETC) while preserving safety bounds.

Public campaigns urge people to update their Internet of Things (IoT) devices and change default passwords. What happens when people try? We gave 28 participants in the Netherlands two pieces of government-issued advice and asked them to try applying each to three of six bestselling IoT devices (168 sessions). We located no manufacturer-set password shared across units, the kind the advice describes; the only device-level credential located was unique to its unit. Fewer than half the update sessions established firmware status. Told that a setting might not apply, no participant concluded it did not: they treated whatever related setting the interface offered as the target, and located the difficulty in themselves rather than in the advice or device. Generic advice asks people to judge what only manufacturers can state and only devices can report. Campaigns must be coordinated with device design, or replaced by secure defaults that remove the task.