Papers

Reinforcement learning (RL) fine-tuning is widely used in language model training to improve performance on a target task while limiting drift from a reference policy. A standard way to balance this trade-off is via a KL-regularized RL objective, although this formulation does not by itself provide a principled way to set the regularization coefficient. In practice, the coefficient is typically chosen heuristically or via hyperparameter search, which can lead to unnecessary overhead in training cost or undesirable reward-retention trade-offs. We instead propose a game-theoretic framework that gives this trade-off an explicit statistical interpretation. Specifically, we study a sequential game in which an agent chooses a policy to maximize cumulative reward while a monitor observes policy outputs over time and tests for deviations from the reference policy. Although not originating from the same perspective, we show that the resulting equilibrium policy can nonetheless be expressed as the solution to a KL-regularized RL problem for an optimal regularization parameter that can be viewed as maximizing reward per unit of statistical distinguishability. Drawing on classical results from concave-convex fractional programming, we provide a principled method for learning this equilibrium coefficient via reduction to the KL-regularized RL objective, thus allowing for flexible integration into standard fine-tuning pipelines. In experiments with Qwen3-8B and Llama-3.2-1B, we show that our methods result in competitive reward-retention trade-offs in a continual learning setting, and illustrate how our framework may be used to audit API providers serving open-source models.

The problem of learning the graphical Markov blanket (MB) of a variable from data has applications in many areas such as structure learning for Bayesian networks and Markov random fields, causal discovery, and feature selection. However, a common assumption most methods make is that the conditional independencies in the distribution imply the same separation in the graphical structure – also known as the faithfulness assumption. Unfortunately, this assumption can be violated by higher-order dependencies such as XOR and parity-type relations, and – on finite samples – by empirical violations that, in extreme cases, even induce spurious dependencies absent from the true distribution. Therefore, in this paper we propose a "k-order" relaxation of the faithfulness assumption that captures parity type relationships between k+2 variables. We then propose a proof of concept algorithm called k-order Markov blanket (kOMB) that uses this relaxation for MB discovery. Finally, we empirically show how kOMB can recover the MB of a variable under both true and empirical violations of faithfulness. Code available at: https://github.com/lklee9/k-order-Markov-blanket

Autonomous agents derive concrete mutations from database reads, retrieved evidence, policy, beliefs, and delegated authority. Those inputs may change while reasoning is in progress. Database isolation orders the submitted transaction; agentic transaction processing determines whether a proposal satisfies an executable contract. Neither guarantee establishes a common valid point for the mutation and its derivation inputs unless the contract represents the relevant predicates. Typed dependency tokens distinguish content integrity from applicability, and trusted mediation captures the values exposed to reasoning. Under strict Cognitive Serializability, committed effects admit a serial order and a logical event at which every value exposed to derivation is unchanged. The fences last until the runtime event that realizes the sealed durability domain. The weaker Effect-Compatible Cognitive Admission recertifies an effect against a simultaneously held current dependency vector and current policy without claiming to serialize the original stochastic derivation. TCT combines immutable versioned executable definitions, registry-derived authority plans, sealed envelopes, guard-first commit transactions, post-seal envelope- and witness-bound grants, co-committed receipts, idempotent grant finalization, and receipt-driven epistemic reconciliation. Complete registered footprints and a single growing phase induce an acyclic lock-point order over local guards and incompatible external reservations. The corresponding results give serializability conditions and an observational-equivalence boundary for zero-error soundness and positive progress. A falsification suite tests the implementation obligations: the prototype prevented all injected anomalies and added 3.22 ms mean commit overhead.

High-quality representations are essential for a wide range of downstream tasks. Dedicated embedding models are explicitly optimized for representation learning, yet their training data are often more limited in scale and diversity than the massive corpora used to pretrain modern large language models and multimodal large language models. Large-scale pretraining and instruction following enable autoregressive models to select relevant evidence, integrate multimodal information, and infer semantics under different task perspectives, creating a distinctive opportunity for training-free representation learning. However, our analysis reveals that existing semantic-elicitation methods do not reliably orient the extracted states toward the semantic perspective required by the downstream task. Consequently, the resulting representations often remain dominated by salient input content. We characterize this problem as semantic perspective misalignment and propose Lens, a training-free framework that makes representation readout task-directed. Semantic Perspective Anchoring associates the task-required perspective with a task-specific readout phrase, specifying the interpretive role of the positions later used for extraction. Contextualized Phrase Readout places the same phrase after the complete input and aggregates its token states, combining full-context access with the anchored perspective. The resulting representation reflects task-conditioned evidence integration and inference rather than a generic summary of salient content. Without parameter updates, architectural modification, or reranking, Lens achieves an overall Precision@1 of 63.9 across all 36 MMEB datasets, outperforming the closest same-backbone training-free embedding baseline by 10.2 points.

Large language models can label a tabular row from a plain-English description with no training - a capability now shipping in mainstream spreadsheet tools such as Microsoft Copilot in Excel and Anthropic's Claude for Excel - raising a practical question for the many business prediction problems where labels are expensive: should you prompt a frozen LLM, or collect data and train a model - and if so, how much data? We quantify the answer with the labeled-data crossover N*, the training-set size at which a trained classical model's learning curve overtakes a frozen LLM's training-free (and therefore flat) error. Aggregating 126 independent student evaluations of small GPT models under eight prompting configurations across 18 tabular datasets, paired with authoritative power-law learning curves for six classical model families, we find that training wins fast: even given an oracle choice of its best prompt configuration, a trained classical model beats the small frozen LLM using no more labeled data than is already on hand in 86% of cases, and wins by the smallest labeled subset we evaluate in 40%, with the observed crossover at a median of ~6% of the training set. In-context few-shot examples do not behave like training - error versus shot count does not follow a power law - and the same protocol re-run by independent implementers varies with a coefficient of variation of 0.148. A controlled probe indicates the LLM depends on recognizable feature-name semantics, which plausibly makes our crossover a conservative estimate (we do not claim memorization). For a typical business table, the evidence is clear: collect a few hundred labels and train a gradient-boosted model.

This paper describes the participation of the YNU-HPCC team in subtask A of task 11, Bridging the Gap in Text-Based Emotion at SemEval-2025. Our best-performing system employs the RoBERTa (Robustly Optimized BERT Approach) model, an improved version of BERT that utilizes the Transformer encoder architecture. We enhanced the output head to allow the model to process one emotion simultaneously. We obtained the official ranking score (0.44), including results from all languages. The entire dataset was translated into English using Google Translate to facilitate subsequent processing. Through probabilistic and attention analyses, we found that (I) a single prediction head performs better than six heads predicting six emotions simultaneously, and (II) training on a uniformly translated English dataset yields better results than using the original dataset. The code is available at: https://github.com/BGWH123/Semeval-2025-task11.

Reliable transformer fault diagnosis is essential for maintaining power system stability. The IEEE Key Gas Method (KGM), a widely utilized approach in Dissolved Gas Analysis (DGA), exhibits limitations in addressing ambiguous data and ensuring high diagnostic accuracy. This study presents An enhanced model combining Fuzzy Logic with the IEEE Key Gas Method (FL-KGM) that introduces refined membership functions, optimized fuzzy rule sets, and a novel separation of CO and CO2 to eliminate diagnostic inconsistencies. By leveraging multidimensional gas ratio analysis and an adaptive classification framework, FL-KGM delivers superior fault identification and classification. Experimental validation utilizing real-world datasets demonstrates that FL-KGM achieves up to 98.6% accuracy, significantly outperforming KGM and other FL-based approaches. These findings elucidate the potential of FL-KGM in advancing transformer monitoring, enabling intelligent fault detection, and enhancing predictive maintenance strategies in modern power systems.

Agentic systems increasingly delegate model selection to a router, yet open-source routers are usually evaluated with different tasks, candidate pools, and execution protocols, limiting direct comparison. We present a common measurement protocol and hybrid evaluation of four router implementations across RouterBench, BFCL v4, tau2-bench, and WebArena. We evaluate 290 frozen tasks against a locked matrix of 2,610 candidate outcomes. Three routers emit constant or near-constant tier assignments; only vLLM Semantic Router varies materially with prompt content, and it has the highest observed success rate on none of the four benchmarks. Always-Mid matches Aurelio exactly on three benchmarks and within 0.003 on the fourth. For vLLM, task-level superiority tests detect no task-specific advantage over a share-matched content-blind allocation; equivalence is established only on WebArena at the protocol-declared five-percentage-point margin. The results show that, under these configurations and controls, observed gains track selected-tier composition more closely than demonstrated task-specific targeting. Fixed-tier baselines and selected-tier distributions are therefore necessary controls in router evaluation; the findings are scoped to these configurations, candidate pool, and frozen benchmark samples, not to routing paradigms in general.

Stacked bilayer materials exhibit rich stacking-dependent properties driven by the interplay between strong intra-layer bonding and weak inter-layer van der Waals interactions. The computational discovery of such materials is challenging because accurate structure generation typically relies on expensive DFT-based optimization, while existing machine-learning models often fail to explicitly distinguish different interaction types during property prediction. To address these challenges, we propose a machine-learning framework for efficient construction and property prediction of stacked bilayer materials. The framework employs a MatterSim-D3-based structural optimization workflow to generate DFT-quality bilayer structures from monolayer building blocks and stacking configurations at substantially reduced computational cost. For property prediction, we introduce BDIP-Net (Bilayer Dual-Interaction Potential Network), a graph neural network that explicitly models intra-layer and inter-layer interactions through interaction-specific potential representations and adaptive message fusion. We evaluate the proposed framework on BiDB, HetDB, and SAMBA, encompassing homobilayers, heterobilayers, and twisted bilayer systems. Results show that the MatterSim-D3-based workflow closely reproduces DFT-PBE-D3 optimized structures, while BDIP-Net consistently outperforms existing graph neural network and potential-based approaches for bilayer property prediction.

Per-field accept/review with selective risk at most alpha – accept a field only if the error rate among accepted fields is controlled – is the trust contract document-extraction systems need, and the natural procedure silently violates it on real documents. On 13,859 genuine claude-sonnet-5 fields from 800 CORD receipts (49.0% correct) we diagnose three failure modes: document clustering (design effect 1.84-2.45), score-refit leakage (coverage 0.416 at risk 0.127, violating alpha=0.10 in 95% of splits), and a tie-mass pathology (a degenerate score collapses the threshold grid, 0.030 to 0.001). We organize the fixes as a validity ladder, guarantee form stated per tier. A fit/val split protocol restores expected-selective-risk control for a learned fusion: coverage 0.318 at risk 0.096 at nominal alpha=0.10, no tolerance band (production variant 0.326) – an on-average point whose realized risk exceeds alpha in 47.5% of resplits, not a certificate. Mondrian Learn-then-Test with exact binomial tails yields per-group PAC certificates: field-iid 0.171 at risk 0.068, cluster-corrected 0.140, doc-iid 0.060 – the only tier matching documents, honestly near-vacuous today. Support-bin, the pre-specified provenance taxonomy, wins every rigor tier on the sonnet CORD capture (p<1e-4, Bonferroni-corrected) – a win that does not replicate on the same documents under haiku or qwen – while on higher-accuracy corpora pooled thresholds win: conditioning helps exactly where pooled cannot certify, subsumed by a learned score elsewhere. A frozen-configuration confirmation on selection-untouched claude-haiku-4-5 held at both risk levels, and a blind three-annotator human-gold audit verifies the practical tier's accepted-set risk at 1.3% against its 10% budget (Fleiss' kappa=0.83; labels err one-sidedly pessimistic). Released Apache-2.0 with seed-pinned, regression-gated procedures.

Fractional optimization methods and fractal activation functions are two independent directions for improving neural network training. Fractional optimizers extend first-order optimization through fractional derivatives and memory effects, whereas fractal activations introduce multi-scale nonlinear representations based on self-similar Weierstrass- and Blancmange-type functions. Here, we investigate their interaction within a unified experimental framework. We evaluate fractional optimizer families on Ackley and Himmelblau benchmark surfaces, in standard form and with additive Weierstrass-type perturbations, and then in feed-forward neural networks with conventional and fractal activations on ten classification datasets. The comparison includes standard methods, regularization-style optimizers, explicit and adaptive memory-based fractional optimizers, and other representative literature methods. Overall, fractional optimization and fractal activations show useful but selective pairings. Regularization-style fractional scaling performs well with selected fractal activations in network training, while Grünwald–Letnikov memory is most relevant on perturbed surfaces. Adaptive memory improves plain memory substitution in several cases, supporting controlled fractional memory as a promising direction rather than a universal replacement.

Activation steering modifies a language model by adding a learned direction to its hidden activations, enabling targeted behavioral changes without retraining. While effective, steering often produces unintended side effects on other behaviors, making it difficult to deploy safely. We therefore ask: can these side effects be forecasted before steering is applied? We answer this question by constructing a cross-effect matrix over a taxonomy of 67 behaviors across three open-weight language models. We find that side effects are common, structured, and often asymmetric, revealing interactions that cannot be explained by existing similarity-based heuristics. Despite this complexity, we show that side effects are largely predictable before steering is performed. Their magnitude depends primarily on the target behavior, while their direction can be forecasted from the model's unsteered representations with substantially higher accuracy than simple baselines. Our results demonstrate that activation steering has systematic and forecastable side effects, enabling proactive safety auditing and more informed deployment of steering interventions.

Deep reinforcement learning (DRL) agents achieve strong performance in complex environments, yet their decision-making processes remain difficult to interpret. We introduce SPOT (Sampling Policy Observation Tree), a novel model-agnostic, sampling-based framework for interpreting DRL policies. Given access to the policy and an environment simulator, SPOT constructs an interpretable finite-horizon tree by sampling actions and recursively simulating the resulting successor states. The tree provides an empirical representation of the policy's action preferences and their possible downstream evolution. We provide formal guarantees establishing SPOT's asymptotic recovery of the policy's unique most probable action and characterizing its disagreement behavior under high-entropy policies. We demonstrate SPOT in the SUMO-RL traffic-signal control domain. The case study illustrates how its tree-based representation can be used to inspect policy preferences, compare alternative future trajectories, and reveal downstream behaviors that are not visible through single-timestep feature-attribution methods.

Understanding how computational effort is allocated across individual chain-of-thought (CoT) reasoning steps remains an open challenge: existing interpretability methods rely on output-level signals or collapse processing depth into a single trajectory-level scalar, leaving step-wise effort opaque. We propose Step-Aware Reasoning Energy (SARE), a geometric framework that quantifies effort at the granularity of individual CoT steps via Centered Kernel Alignment (CKA) between Gram matrices of token hidden states across adjacent transformer layers, capturing inter-token relational structure without requiring eigenvector alignment or cluster correspondence. SARE further contextualizes this energy within reasoning's semantic progression by modeling CoT trajectories as transitions among latent semantic states. Across six reasoning benchmarks and three open-weight LLMs, we find that reasoning energy is highly non-uniform across step types, exhibiting phase-like transitions invisible to trajectory-level metrics; incorrect trajectories show systematically lower energy at critical reasoning junctions; and SARE-based features match or outperform output-based confidence baselines in most settings, indicating that internal geometric dynamics encode predictive information beyond surface-level signals.

Transformer-based surrogate models are increasingly used to replace expensive first-principles simulation in engineering design. But conventional transformer architectures are often over parameterized for the small, low-dimensional datasets typical of engineering design spaces, where large simulation data is expensive to generate. Under these conditions, excess parameter capacity leads to overfitting rather than improved accuracy, while also incurring unnecessary memory and compute overhead. This motivates a shift towards architectures that focus on additional compute rather than additional learnable parameters. This paper presents a hardware-aware evaluation of three recursive transformer paradigms for surrogate thermo-mechanical analysis of advanced packages: a)Tiny Recursive Model, b) our proposed Depth Recursive transformer, c) and a simple recursive transformer. We systematically compare their predictive performance (Recall, Mean Reciprocal Rank), parameter count, computational complexity (FLOPs), providing practical design guidelines for selecting recursive transformer architectures under resource-constrained scenarios. We validate this principle on two low-dimensional engineering prediction tasks: 1) thermo-mechanical reliability analysis of advanced semiconductor packages, where stress and warpage from thermal cycling must be evaluated repeatedly across a design-of-experiments sweep under costly finite element analysis (FEA). 2) Laplace PDE iterative numerical solver for capacitance field. Overall, recursive weight-sharing transformers provide an effective and generalizable trade-off between prediction accuracy, parameter efficiency, and computational cost for small data engineering surrogate modeling.

Persistent context files (AGENTS.md, CLAUDE.md) are standard practice for guiding AI coding agents, yet evidence for their effectiveness is contradictory. We present a controlled ablation of context-injection strategy across two frontier agents (Claude Code and Codex), 17 real tasks from 3 repositories (15 shared + 2 Codex-only), and 288 evaluated runs with gold-test evaluation. Context strategy does not measurably move correctness on either agent (bounded to <=10-15pp via equivalence testing). A failure-mode triage reveals why: agents fail on implementation skill—feature design, pattern selection, exact wiring—not missing repository knowledge that a context file could supply; a manipulation probe confirms the real AGENTS.md never converts a near-miss to a pass on either agent. We further show that borderline task difficulty is agent-specific (Spearman rho=0.75), offering a candidate explanation for prior contradictions: single-agent studies draw tasks from different agents' informative bands. We release all code, data, and analysis.

While large language models excel in reasoning, these generalists often lack knowledge for specialized scientific domains. Conversely, domain models~(specialists), while knowledgeable, suffer from specialization side-effects including diminished logic and reduced robustness.To address this dilemma, we introduce Divergence Decoding, a training-free framework for capability fusion. It reconstructs the "draft-and-verify" skeleton of speculative decoding into an adaptive routing mechanism. The core is using Jensen-Shannon divergence to monitor the distributional disagreement between the two models at each token. When the specialist exhibits significant divergence, our method identifies it as a potential reasoning risk and instantaneously routes control to the generalist. This allows the dynamic injection of general reasoning while preserving domain expertise, achieving inference-time policy composition of the generalist and the specialist.We evaluate Divergence Decoding across diverse model families (Qwen and Llama series) on challenging scientific benchmarks (GPQA, ChemBench, and ChemCoTBench). Experimental results demonstrate that Divergence Decoding outperforms both the domain-specialized and general-purpose models, effectively surpassing the performance of most single-model baseline. This suggests that Divergence Decoding provides a general, training-free paradigm for fusing diverse LLM capabilities through adaptive inference-time collaboration.

Networking research advances by turning hypotheses into empirical evidence, so accelerating it means reducing the lag between ideation (synthesizing a hypothesis) and generating the data that tests it. Consider a concrete case: does a bulk BBR download fairly share its bottleneck with competing real-time Google Meet traffic? Validating this requires configuring a realistic bottleneck link, concurrently generating BBR's bulk transfer and Meet's real-time traffic, and collecting relevant service-quality metrics. Today this overhead is high, often forcing researchers to start from scratch for every new idea. This ideation-to-data-generation gap will only worsen in the agentic AI era, where AI-assisted ideation accelerates exponentially, yet its outputs cannot be validated without a data-generation backend. This paper explores how to bridge this gap. We envision a composable, domain-specific backend, Pramana, shaped as a thin waist, with diverse research intents at the top and disparate execution substrates at the bottom. Pramana realizes this waist through a single contract, the intent specification, which disaggregates an experiment into three independent axes: the intent (what data to generate), the substrate (where to generate it), and the mechanism (how to produce it), so one specification runs on any substrate. We demonstrate Pramana's utility by building a first-of-its-kind corpus of 255 data-generation intents mined from 66 published papers, and show the intent specification satisfies all of them, where no existing tool satisfies more than 13%. Our current proof-of-concept implementation already satisfies 34% of these intents, more than twice the best existing tool, and we lay out a roadmap for closing this abstraction-implementation gap through a broader community effort to build the envisioned data-generation backend and accelerate empirical networking research.

Large language models (LLMs) are increasingly used as synthetic users, stand-ins for human respondents whose simulated answers feed product, policy, and market decisions. We ask when this substitution is valid and when it fails, and package the answer as an evaluation framework for intelligent synthetic-user systems. A single protocol, run across four models spanning two families and an 8B-to-frontier capability range, is applied to two independent domains of real human-response data: U.S. general social attitudes (General Social Survey) and cross-cultural values (World Values Survey). Every model is benchmarked against a suite of non-LLM baselines fit on held-out human data. Under demographic prompting and the survey-simulation protocols we test, two failures replicate across both domains, all four models, and both families. First, at the individual level no LLM beats even the strongest baseline; on cross-cultural values every model falls well below it, and the gap survives distance-aware and proper scoring. Second, models systematically over-determine demographics, treating identity as far more predictive of attitudes than it is among real people, a distortion present for nearly every question-group combination and robust to a coding-invariant measure. Neither failure is remedied by a larger, more capable model. A decision-impact analysis shows why this matters in practice: on a segment-targeting task the models inflate between-segment gaps two to fourfold, would direct a team to the wrong segment in half of U.S. and most cross-cultural cases, and manufacture segment splits that do not exist in real people. We publicly release the cross-domain benchmark and validation framework, including all code and data, so that teams can determine in advance when synthetic-user evidence is safe for decision support and when it is not.

A gradient-based GNN explainer given a molecule with two chemically equivalent nitro groups assigns them attribution scores that are equal to the last bit. It cannot do otherwise: message passing is exactly permutation equivariant, so any automorphism of the input leaves every attribution invariant. Yet the standard report, the top-k edges, names one of the two, and which one is settled by the order of an array. We show this is a structural obstruction rather than an implementation slip. When no minimal valid explanation is fixed by the input's automorphism group, no rule can be single-valued, minimal and symmetry-respecting at once. For the exact-k reports used in practice we give a parameter-free criterion, mechanised in Lean 4 with no axiom dependencies, that decides from the graph alone whether every score-optimal report of that size must split an orbit. Across 21298 instance-budget decisions the criterion agrees with a mechanical model-equivalence check without exception, and no severing case we found admitted a neutral alternative. The obstruction is common. Nontrivial automorphisms occur in 93.4% of Mutagenicity, the dataset the seminal explainability papers use, so the measure-zero dismissal of symmetric inputs, sound on the continuous domains it was made for, collapses here. At the sparsity budget those papers report, 24.0% of molecules with two interchangeable nitro groups (6 of 25) surface exactly one of them, every one arbitrary under mechanical verification. A model's blindness also manufactures symmetry: every MUTAG molecule contains atoms chemistry separates and the network provably cannot, and a matched control shows the resolution is set by what the model reads rather than how it is parameterised. Reporting orbits removes the arbitrariness at 0.11 ms and 0.43 extra edges per graph.

Psychometric calibration for educational tests typically requires costly human response data. Large language models (LLMs) simulated examinees offer a promising route to early calibration, but their responses are too accurate and too uniform. We propose Cognitive Diagnostic Profiling (CDP), a zero-shot framework that prompts LLMs to simulate plausible examinees with diverse cognitive profiles: binary attribute-mastery patterns are rendered as natural-language profiles and sampled under an uninformative or an informative distribution. Using the Tatsuoka fraction-subtraction dataset (536 examinees, 15 items, five attributes), we evaluated eight LLM configurations under no-profile, uninformative-CDP, and informative-CDP conditions, assessing alignment with human examinees at the ability-distribution, mastery-profile, and item-difficulty levels. CDP improved all three levels: distributional overlap rose across configurations; weighted correlations between profile-level scores and human profile expectations reached 0.92 to 0.98; and item-difficulty recovery improved in rank order and absolute alignment, most for reasoning-enabled models; in the strongest case, Gemini 3.0 Flash (Thinking), one-parameter logistic (1PL) difficulty Spearman correlations rose from 0.24 to 0.86 and 0.90 and the root-mean-square error (RMSE) fell from 6.31 to 1.30 and 0.90; the informative condition helped most where profile-level alignment was strong. CDP brings LLM-simulated examinees into closer psychometric alignment with human examinees, making them practical for operational test development.

Stealth, the discipline of achieving an objective without revealing your presence, capabilities, or collected intelligence, is what separates sophisticated operators from detectable ones. Elite security researchers and advanced persistent threats achieve their objectives unnoticed; autonomous agents increasingly inherit the same offensive tasks, but do they inherit the tradecraft? We introduce StealthBench,a benchmark that measures operational stealth in autonomous offensive-security agents across six operational security (OPSEC) dimensions. We extract 11 hand-verified OPSEC incidents from real bug-bounty and red-team trajectories, expanded into 14 dockerized task scenarios, where agents, despite finding real vulnerabilities, committed stealth failures inconsistent with standard operational tradecraft: embedding credentials in public uploads, deleting production resources to prove access, force-adding uninvolved users to demonstrate a race condition. We evaluate agent trajectories using a 3-model large language model (LLM) judge panel with majority-vote aggregation, measuring safe success rate (solved and stealthy), Stealth@Solve (tradecraft quality among successful solves), and reckless solve rate (solved but cover blown). Our results show that no model exceeds 54% safe success rate (the compound metric requiring both task completion and stealth), confirming that OPSEC failures are systematic across model families. We release StealthBench as a public benchmark to support both the development of stealth-aware agents and automated OPSEC monitoring for autonomous offensive-security deployments. The interactive leaderboard, evaluation harness, and dataset are available at https://stealthbench.com.

Agentic systems act, so a defect in the evidence they retrieve becomes a wrong action with a currency cost. The most dangerous enterprise defects are metadata-borne: a stale price or a superseded record, perfectly well-formed in the payload and betrayed only by freshness, lineage, or provenance. Such a defect never enters the agent's context, and an agent cannot doubt data it cannot see. On a priced replenishment benchmark, a competent agent silently converts an injected metadata-borne defect into a costly action about 60% of the time, with zero data-quality flags and behavioral doubt markers at chance (AUC <= 0.50). Across four model tiers spanning roughly 15x in inference price, the rate stays flat: capability does not buy skepticism. A metadata-aware pre-action gate with downstream-only remediation recovers the loss fully on the signals its predicates cover and not at all on those they miss. A model-free oracle derived from the task's decision geometry tracks the measured rates with MAE 0.015 (Pearson r = 0.876, interval coverage 15/16 cells), giving the flat ladder an analytical form. Evidence integrity is a systems axis distinct from model capability; mitigation depends on enforcement placement and predicate coverage. Code, frozen results, and a deterministic analysis pipeline: https://github.com/besanson/dqSarc

Contemporary LLM-based coding agents produce code as black-box outputs: the rationale behind each line is hidden, the evolution of the code through benchmark-driven repair is ephemeral, and post-hoc auditing is impossible. We present a code generation concept that addresses these shortcomings through three complementary mechanisms: (i) a relational snippet-history schema that records, per repair event, the benchmark reference, round number, failure text, and LLM explanation, enabling full provenance queries; (ii) a browser-based visualisation tool that renders this history as heat-mapped, hover-annotated source code; and (iii) a competitive fractional position-key indexing scheme with tree-node delimiters that assigns stable, lexicographically-ordered identifiers to each code snippet, enabling fine-grained tracking without disrupting surrounding lines. We evaluate TraceCoder on 30 algorithmic programming tasks spanning string processing, mathematical computation, and data-structure manipulation, across two provider configurations. Of these, 10 exhaust the 6-iteration budget on tasks with subtle edge-case behaviour. Mean Chg% reaches 30%, three in ten code snippets carry a traceable repair-event row, compared to 21% when using Gemini 2.0 Flash as sole provider on a 20-task subset. Three detailed case studies demonstrate how the system explains which specific benchmark failures shaped each line of the final program. The proposed mechanism makes the internal "narrative" of automated code generation auditable and replayable, a property essential for trust and accountability in production deployments.

AI agents are increasingly adept at tackling complex, long-running tasks. With the rapid surge of autonomous capabilities, human oversight is systematically lagging behind due to limited human-centered interfacing. Aiming to address this, we introduce AgentGUI, a user-friendly, locally hosted GUI for seamlessly observing and steering AI agents amid multiple concurrent, long-running sessions. AgentGUI features 1) rich agent trajectory visualizations, 2) effective manual and automated steering, and 3) integration with and coordination between open-source and frontier agent frameworks. A controlled user study demonstrates statistically significant reduction in the time it takes to identify key elements from agent traces (38% faster, p = 0.023). In a preliminary experiment, AgentGUI's automated drift prevention feature raises the task completion rate of small local agents by as high as 34pp across a 0.8B–9B model ladder (N=50 runs per model). AgentGUI is publicly available through its project website (https://agent-gui-project.github.io) and open-source repository (https://github.com/eth-medical-ai-lab/agent-gui), along with a demo video (https://youtube.com/watch?v=GSDyxN1gTF0).

We built and evaluated a self-serve entity resolution (ER) system on six benchmarks spanning 864 to 5M records, and three lessons emerged that are absent from existing ER literature. (1) No single matching algorithm wins everywhere - a self-serve pipeline cannot predict its next dataset, so we recommend training several algorithm families per dataset and letting an automatic bake-off pick the winner. (2) Precision and recall need separate fixes, not a shared threshold - precision needs hard rule-based vetoes, recall needs more diverse candidate retrieval. (3) One false-positive link can silently merge unrelated entities - assuming "A matches B" and "B matches C" implies "A matches C" lets a single bad link chain hundreds of records together, so every cross-group merge must be actively re-verified. We hope these lessons save practitioners the months of dead-end experiments that led us to them.

We consider a stochastic multi-objective bandit problem where, at each round, the agent selects a slate of $k$ arms and observes their $d$-dimensional reward vectors under semi-bandit feedback. We do not aim at identifying a single optimal arm; instead, we consider the problem of maintaining a small set of actions that jointly approximate the Pareto frontier. We formalize this objective through the dominated hypervolume induced by the selected subset of arms, and define an $\alpha$-approximate hypervolume regret with respect to the best size-$k$ subset achievable in hindsight, where $\alpha = 1 - 1/e$ reflects the approximation guarantee of greedy maximization for monotone submodular functions. To address this problem, we introduce THV-UCB, an optimistic algorithm that selects arms greedily based on optimistic estimates of their marginal hypervolume contributions. We establish a gap-free regret bound $\tilde{O}(d\sqrt{nkT})$ that holds on every instance, together with a gap-dependent bound $\tilde{O}(nk^{2.5}/\Delta_{\min})$ that becomes polylogarithmic in $T$ once the arms are sufficiently well separated. Our results provide theoretical support for using small subsets to approximate Pareto fronts in various multi-objective applications.

AI-generated counterspeech offers a scalable and effective strategy to mitigate online toxicity by promoting more constructive dialogue. Yet, existing approaches adopt a generic, one-size-fits-all paradigm, overlooking the conversational context and characteristics of the targeted users. Here, we propose and evaluate multiple strategies for generating contextualized counterspeech that is adapted to the moderation setting and personalized to the moderated user. In detail, we explore a range of configurations that integrate different forms of contextual information and fine-tuning techniques. We conduct a comprehensive evaluation combining quantitative indicators with a pre-registered, mixed-design crowdsourcing experiment. To ensure robustness, we implement algorithmic measures of counterspeech quality based on ROUGE, BLEU, and BERTScore, observing overall consistent results across metrics. Furthermore, we analyze which characteristics of both the generated counterspeech and the moderated toxic message most strongly influence perceived persuasiveness, yielding insights into how contextualized interventions can be made more effective. Our findings show that personalization can be effective, but not uniformly so. Lightweight strategies combining conversational context and user history improve perceived adequacy and persuasiveness, whereas several other contextualization strategies degrade human-perceived counterspeech quality. Taken together, these results provide actionable directions for developing more personalized, effective, and responsible counterspeech systems, ultimately advancing human-AI collaboration in online content moderation.

Context: Large Language Models (LLMs) offer natural-language flexibility for automated requirements elicitation but frequently generate structurally invalid requirements and logical inconsistencies, lacking formal correctness guarantees. Objectives: This study aims to eliminate logical inconsistencies and enforce structural conformance in LLM-generated requirements while quantifying the LLM's pre-validation decision uncertainty within a formal domain model. Methods: We present a neuro-symbolic multi-agent architecture that operationalizes the Object-Oriented Method for Requirements Authoring and Management (OOMRAM) lattice. The LLM acts as a non-deterministic heuristic for lattice traversal, while a deterministic symbolic validator enforces all structural constraints. We introduce a three-valued (T, I, F) – Truth, Indeterminacy, Falsity – framework to classify and score the LLM's requirement decisions before and after validation. Results: Evaluated across 37 natural-language project visions in eleven application families, the system completely eliminated structural inconsistencies in 35 out of 37 cases (94.6%), with the remaining two containing only 6 unresolved structural errors (0.39% of decisions) due to iteration limits. Three-valued analysis revealed that 24.7% of all decisions are indeterminate – structurally valid but discretionary choices not explicitly mandated by the stakeholder. Conclusion: Offloading structural integrity to a deterministic symbolic layer successfully guarantees structural conformance, while the three-valued classification provides a formal way to measure neural uncertainty, facilitating safe LLM deployment in formal requirements engineering.

Multi-Agent Debate (MAD) is a promising paradigm for improving the accuracy and robustness of Large Language Model (LLM)-based agentic systems. It enables multiple agents to exchange arguments, critique each other's outputs, and iteratively converge towards a solution. However, research remains fragmented, with inconsistent terminology and no rigorous synthesis of MAD design dimensions. We present a systematic literature review characterizing 141 primary studies on MAD. We derive a three-dimensional taxonomy covering debate participants, the interaction mechanisms structuring the exchange, and the agreement protocols governing debate resolution, supported by formal notations to render MAD configurations. Our analysis reveals that the field has implicitly converged on a narrow design pattern - static, fully connected topologies, verbatim exchange, short-term memory and voting resolution strategies - adopted by convention rather than systematic comparison, while promising alternatives remain marginal. Because any MAD setting reflects roughly a dozen interacting design decisions, cross-study comparison is unreliable when these are left implicit. We position the taxonomy as a descriptive map of the research landscape, a framework for controlled benchmarking, and potentially as a schema for machine-readable MAD specifications. As future work, we propose formalizing it into an executable specification, enabling cost-aware benchmarking and automated tuning of debate configurations.