https://arxiv.org/api/pT8d43lecHXG2St1pHYZg5oPkSQ 2026-06-18T21:46:47Z 28983 450 15 http://arxiv.org/abs/2602.10324v2 2026-05-29T05:21:58Z

As Large Language Models (LLMs) are increasingly deployed in social and strategic scenarios, it becomes critical to understand where and why their behavior diverges from that of humans. While behavioral game theory (BGT) provides a framework for analyzing behavior, existing models do not fully capture the idiosyncratic behavior of humans or black-box, non-human agents like LLMs. We employ AlphaEvolve, a cutting-edge program discovery tool, to directly discover interpretable models of human and LLM behavior from data, thereby enabling open-ended discovery of structural factors driving human and LLM behavior. Our analysis on iterated rock-paper-scissors reveals that frontier LLMs can be capable of deeper strategic behavior than humans. These results provide a foundation for understanding structural differences driving differences in human and LLM behavior in strategic interactions.

2026-02-10T22:02:41Z Accepted to ICML 2026 Caroline Wang Daniel Kasenberg Kim Stachenfeld Pablo Samuel Castro http://arxiv.org/abs/2605.16204v2 2026-05-29T02:07:56Z

Content moderation is a central mechanism through which platforms attempt to balance user engagement with community governance. Yet existing research has largely treated moderation as a uniform intervention, overlooking how moderator source, violation context, and linguistic style jointly shape user behavior. Drawing on the Human--AI Interaction Theory of Interactive Media Effects (HAII-TIME), this study examines how these three dimensions produce divergent post-moderation behavioral trajectories in a large-scale observational dataset of 11,795,036 moderation events across 9,285,410 users and 61,261 subreddits on Reddit (2021--2025). Using probabilistic behavioral classification, ANOVA, and OLS regression with PCA-derived linguistic features, we find that bot moderation consistently produces higher compliance and lower self-censorship than human or modteam moderation, challenging the assumption that human agency cues are inherently advantageous. Modteam moderation produces the strongest self-censorship effects, suggesting that institutional depersonalization is a meaningful driver of behavioral withdrawal. Violation severity emerges as a critical contingency: linguistic strategies effective in routine contexts -- elaborated explanation, community-scale appeals, direct personal address -- can backfire for serious violations, whereas prosocially framed and emotionally emphatic messages become most effective when stakes are highest. Of 480 linguistic interactions tested, 33 survive FDR correction. These findings extend HAII-TIME by introducing violation salience as a moderator of cue-based processing, and offer empirical grounding for context-adaptive moderation design.

2026-05-15T17:21:35Z Siyi Zhou Lindsay Young Marlon Twyman Emilio Ferrara http://arxiv.org/abs/2605.30685v1 2026-05-29T00:28:36Z

AI is being used by people globally, but not everyone is using it in the same ways. Using a large-scale dataset of anonymized, de-identified, and privacy-scrubbed interactions with a widely available and free AI chatbot, we empirically characterize differences in early adopters' usage across countries. Schooling is the most common domain of use in most countries, particularly low-income countries, with a strong inverse association evident between schooling and country-level GDP. Leisure-related use, by contrast, is positively associated with country-level income. Language, we find, also shapes use: English-language interactions are overrepresented in places where the predominant languages were not well-served by existing models during the period of the study. Improving performance across languages may be a key factor, our work suggests, in whether this technology expands digital divides or enables leapfrogging.

2026-05-29T00:28:36Z Madeleine I. G. Daepp Isaac Slaughter http://arxiv.org/abs/2605.30670v1 2026-05-29T00:04:19Z

Large language models are increasingly deployed as intelligent tutors, yet research on aligning them for special education remains absent. Recent work has applied reinforcement learning to LLM tutors, but these methods target a generic learner in a single domain (mathematics) and do not address the cognitive and communicative diversity of learners with disabilities. We introduce \emph{Special-R1}, a framework that extends pedagogical RL to special education through two components: (1) a two-dimensional adaptive system prompt that couples a difficulty-based support level with a disability-specific teaching style across five disability profiles; and (2) a persona-aware Thinking Reward whose judge rubric is conditioned on the learner's disability profile. On a persona-augmented test set of 690 multi-turn dialogues, our full model raises persona-aware Fit from 6.75 (generic baseline) to 8.40 (+1.65) and SPED-rubric Helpfulness from 0.720 to 0.768, leading on the four-component Total (2.911, +0.064 over the runner-up) while remaining within 0.01 of the strongest variant on the out-of-domain OpenLearnLM benchmark (8.53). Ablations show that the Thinking Reward becomes effective only in combination with adaptive prompting, and that residual weakness on specific learning disability in mathematics motivates targeted multimodal extensions.

2026-05-29T00:04:19Z Unggi Lee Jihoi Na Yeil Jeong Haeun Park Yeonju Jang http://arxiv.org/abs/2605.30666v1 2026-05-28T23:55:33Z

Aligning large language models (LLMs) as math tutors typically demands costly reinforcement-learning (RL) training and external LLM judges. We ask whether a frozen model's internal reasoning signals can replace both. We propose the Tutoring Effectiveness Index (TEI), a training-free, judge-free four-signal index that combines a Schoenfeld-Verify keyword ratio, a math-step density, an ends-question rate, and a deep-reasoning gate from the Deep-Thinking Ratio (DTR) probe. Selecting from $N$ candidates with TEI (the TEI@$N$ rule) raises the improvement rate on pre-incorrect scenarios from $59.0\%$ to $81.9\%$ at $N{=}8$ on a frozen DeepSeek-R1-8B base, with no training and no external judge. We also measure the alignment tax of pedagogical GRPO. Thinking length drops from $1{,}764$ to $119$ words per turn ($-93\%$), Content-Knowledge and Pedagogical-Knowledge accuracy fall by $-71\%$ and $-80\%$ relative, and the student's $Δ$ Solve Rate crosses from $+0.180$ to $-0.012$. To anchor the behavioural reading, we reproduce an 82-code educational codebook on $119{,}009$ tutor sentences with a one-shot structural classifier. Together, these results offer a cost-effective recipe for building math-tutoring LLMs without RL training or external judges.

2026-05-28T23:55:33Z Shim Jaechang Unggi Lee http://arxiv.org/abs/2602.22968v3 2026-05-28T21:41:37Z

Understanding how neural networks arrive at their predictions is essential for debugging, auditing, and deployment. Mechanistic interpretability pursues this goal by identifying circuits--minimal subnetworks responsible for specific behaviors. However, existing circuit discovery methods are brittle: circuits depend strongly on the chosen concept dataset and often fail to transfer out-of-distribution, raising doubts whether they capture the concept or merely dataset-specific artifacts. We introduce Certified Circuits, which provide provable stability guarantees for circuit discovery. Our framework wraps any black-box discovery algorithm with randomized data subsampling to certify that inclusion decisions over circuit components--neurons or edges of the model graph, depending on the base algorithm--are invariant to bounded edit-distance perturbations of the concept dataset. Unstable components are abstained from, yielding circuits that are more compact and more accurate. We validate across three architectures (ResNet, ViT, GPT-2) on vision (ImageNet and four OOD datasets) and language (IOI, IOI-Hard, Greater-Than) tasks. Certified circuits achieve up to 56% higher accuracy and up to 80% fewer components, and remain reliable where baselines degrade. Certified Circuits puts circuit discovery on formal ground by producing mechanistic explanations that are provably stable and better aligned with the target concept. Code: https://github.com/AlaaAnani/certified-circuits.

2026-02-26T13:07:31Z Accepted at ICML 2026 Alaa Anani Tobias Lorenz Bernt Schiele Mario Fritz Jonas Fischer http://arxiv.org/abs/2605.30543v1 2026-05-28T20:20:52Z

Computing within limits is a promising field, that follows principles of a) questioning endless growth narrative, b) considering and preparing for models of scarcity and c) reducing energy and material consumption, while considering d) a global spatial scale and e) long time frames. With computing's environmental impact growing and ecological limits becoming increasingly pressing, the LIMITS workshop has served as a central venue for this community since its inception in 2015, but an overview of the research published there has yet to be described. This paper addresses this gap by analyzing 160 publications from the LIMITS workshop in the period 2015 to 2025 to identify its international spread, contributions and developments in relation to field's core concerns, combining programmatic analysis with a manual review. Our findings indicate that the field has increasingly mentioned degrowth and post-growth, especially in 2024-2025. It has broadened its global perspective, with a growing, but still limited, representation of work beyond the Global North. The majority of papers are positional or observational, while artifact-producing research remains relatively scarce, though solution-oriented output has grown in recent years. This paper contributes to the LIMITS community by mapping its first decade and current trends to support future research and enhance its global impact.

2026-05-28T20:20:52Z Paper in Proceedings of LIMITS 2026: 12th Workshop on Computing within Limits, 2026-06-23-25, Online Maria Emine Nylund Erik Johannes Husom Ophelia Prillard http://arxiv.org/abs/2605.30406v1 2026-05-28T17:47:37Z

Recent research demonstrating AI systems exhibiting deception and shutdown resistance suggests that AI loss of control (LOC) is an urgent policy concern , yet current literature focuses almost exclusively on alignment and prevention. To address this gap, this paper introduces a foundational framework and taxonomy for managing catastrophic AI LOC incidents. The taxonomy's first level distinguishes between scenarios where regaining control is 'extremely costly' versus 'impossible'. While impossible scenarios demand immediate resilience investments to fundamentally restrict an AI's attack surface , extremely costly scenarios require active incident management via Containment and Threat Neutralization. The framework further categorizes these manageable events into accidental LOC (requiring automated circuit-breaker responses) and adversarial LOC (requiring graduated escalatory measures). By mapping three severity classes to specific scenario matrices, this paper provides a concrete, proportional guide for managing unprecedented AI risks.

2026-05-28T17:47:37Z 25 pages, 4 figures Ross Gruetzemacher http://arxiv.org/abs/2605.30303v1 2026-05-28T17:46:53Z

Artificial intelligence assistants deployed in online learning environments create new opportunities to collect large volumes of learner interaction data and generate insights to improve student outcomes. Architecture for AI-Augmented Learning (A4L) is a modular data architecture that enables the collection, integration, and analysis of learner interaction data from educational AI systems, supporting the generation of instructional insights that facilitate personalized learning and reinforce the bidirectional feedback loop between instructors and learners. This study examines the modular design of the A4L Data Analytics Pipeline, an extensible data infrastructure that enables the ingestion, processing, and analysis of heterogeneous datasets generated by educational AI assistants. We describe the design principles and development process used to extend the pipeline's analytical capabilities while preserving flexibility across domains. We evaluate the pipeline through case studies spanning three research domains corresponding to three educational AI assistants deployed in online learning environments at Georgia Tech. Results show that a common set of statistical analysis methods can be consistently applied across datasets with differing structures and instructional contexts, enabling the pipeline to reproduce key analytical findings across domains. We demonstrate how analytical capabilities initially developed for one domain can be extended to support richer analyses in another, illustrating the pipeline's extensibility. These findings suggest that the A4L Analytics Pipeline can serve as reusable infrastructure for analyzing data generated by future educational AI assistants. By enabling analytics that can be systematically extended to new domains, the pipeline provides a foundation for deriving insights that inform the design and evaluation of educational AI systems.

2026-05-28T17:46:53Z 8 pages, 3 figures, to be published in proceedings of EDULEARN26 Yallen Bai Ploy Thajchayapong Ashok Goel http://arxiv.org/abs/2604.04956v3 2026-05-28T17:42:01Z

The recent, super-exponential scaling of autonomous Large Language Model (LLM) agents signals a broader, fundamental paradigm shift from machines primarily replacing the human hands (manual labor and mechanical processing) to machines delegating for the human minds (cognition, reasoning, and intention). The uncontrolled offloading and scaling of "thinking" itself, beyond human's limited but efficient biological capacity, has profound consequences for humanity's heat balance sheet, since thinking, or intelligence, carries thermodynamic consequences. The Earth has already surpassed the heat dissipation threshold required for long-term ecological stability, and projecting based on empirical data reveal a concerning trajectory: without radical structural intervention, anthropogenic heat accumulation will breach critical planetary ecological thresholds in less than 6.5 years, even under the most ideal scenario where Earth Energy Imbalance (EEI) holds constant. In this work, we identify six factors from artificial intelligence that influence the global heat dissipation rate and delineate how their interplay drives society toward one of four broad macroscopic trajectories. We propose that the integration of artificial intelligence and its heat dissipation into the planetary system constitute the tenth planetary boundary (9+1). The core empirical measurement of this boundary is the net-new waste heat generated by exponential AI growth, balanced against its impact on reducing economic and societal inefficiencies and thus baseline anthropogenic waste heat emissions. We demonstrate that managing AI scaling lacks a moderate middle ground: it will either accelerate the breach of critical planetary thermodynamic thresholds, or it will serve as the single most effective lever on stabilizing the other nine planetary boundaries and through which safeguarding human civilization's survival.

2026-04-03T10:42:33Z Minor revisions for clarity William Yicheng Zhu Lei Zhu http://arxiv.org/abs/2605.30273v1 2026-05-28T17:30:57Z

Large language models (LLMs) show promise in generating supportive responses for mental health queries, but improving their usefulness, empathy, and safety often requires substantial compute, expert input, and labeled data. At the same time, deploying proprietary, cloud-based models for mental health-related interactions raises important privacy and data-governance concerns, given the sensitivities. To address this challenge, we introduce LLUMI setup that can be hosted in-house within protected environments. LLUMI consists of two complementary components: a generation model (GM), which drafts supportive responses to mental health queries, and an improvement model (IM), which revises an initial human-crafted response. We leverage feedback signals from Reddit mental health communities, using community endorsement patterns such as upvotes and downvotes to construct chosen-rejected response pairs for Supervised Fine Tuning (SFT) and Direct Preference Optimization (DPO). We further align LLUMI using human evaluation across five dimensions: readability, empathy, connection, actionability, and safety. Our results show that, despite relying on smaller open-source models rather than proprietary cloud-based GPT models, LLUMI achieves comparable performance across linguistic analyses and human evaluations. These findings suggest that open-source models, when trained with community-derived preference signals, can support high-quality mental health support assistance while offering a more privacy-preserving alternative for sensitive support contexts.

2026-05-28T17:30:57Z Jiwon Kim Maya Ajit Sherry Gong Soorya Ram Shimgekar Dong Whi Yoo Eshwar Chandrasekharan Koustuv Saha http://arxiv.org/abs/2605.30241v1 2026-05-28T17:09:19Z

Misinformation verification increasingly occurs in public, fast-moving, and multilingual online settings, where static benchmarks provide an incomplete measure of model reliability. We introduce CommunityFact, a refreshable benchmark for misinformation detection in the wild, with three major goals: coverage, granularity, and redistributability. This release contains 15,992 standalone claims across five languages and two domains. We evaluate ten LLMs under varying inference-time capabilities, including thinking and web-search. Our results show that closed-input verification remains challenging, web access yields the largest gains, and web-enabled LLMs' source-selection policies are systematically misaligned with the sources human Community Notes raters converge on -- a gap that closes through model-specific mechanisms of retrieval expansion or pruning. We further find substantial variation across language-domain slices and across the evidence ecosystems used by web-enabled systems. Beyond evaluation, CommunityFact positions Community Notes as a training signal for claim-conditioned source suggesters that could improve factual verification on novel claims.

2026-05-28T17:09:19Z Sahajpreet Singh Insyirah Mujtahid Min-Yen Kan Kokil Jaidka http://arxiv.org/abs/2605.25376v2 2026-05-28T17:04:16Z

KYA (Know Your Agents) is an open-source, framework-agnostic trust and governance layer for autonomous systems, composed of five primitives: (1) a four-gate inbound apply pipeline; (2) an only-tighten composition algebra over a three-channel multi-tenant hierarchy; (3) KYP (Know Your Principal), a schema-level unification of trust scoring across human users, AI agents, and service accounts; (4) auditable interaction-multiplier amplification over an AIVSS-shaped additive baseline; and (5) two-axis delegation attribution: a static premium for risky delegates and a runtime debit for actual delegate misbehavior in multi-agent fan-out. Together these span three pillars (trust, governance, and evidentiary assurance), making an autonomous system's actions authorized, policy-conforming, and post-hoc verifiable: where observability answers how long, how much, and what path, KYA answers was it authorized, did it conform, and can it be verified; it composes with observability rather than replacing it. It ships native adapters for 15+ agent frameworks. On a 4 by 9 cross-backend matrix all 36 cells pass; the pure-function scorer runs sub-millisecond at p99 and the system sustains ~ 1,800 ops/sec at 20 concurrent workers with HMAC chain integrity preserved end-to-end. KYA detects 89% of 1,200 adversarial probes from PyRIT and Garak, including the recently-published topology-guided multi-agent attack. The system is available under Apache 2.0 as the veldt-kya package on PyPI.

2026-05-25T02:59:54Z 26 pages including appendix. Code available under Apache 2.0 at https://github.com/veldtlabs/veldt-kya (pip install veldt-kya). Two-domain worked examples (loan decisioning under NYDFS/ECOA/CFPB; clinical triage under HIPAA/21 CFR Part 11/FDA SaMD).Reproducibility artifacts in-tree Kolawole Quadri http://arxiv.org/abs/2605.30187v1 2026-05-28T16:31:32Z

The widespread adoption of AI chatbots in education will drastically change learning, making responsible deployment a critical concern. While large language models (LLMs) might have access to sources discussing insights from educational sciences, they are not particularly inclined to adhere to pedagogical concepts, risking negative effects on the learning process, such as a loss of transfer capabilities, critical thinking, or creativity. In this paper, we introduce an agentic AI chatbot architecture assisting students with exercise solving, specifically designed to contribute to more responsible AI use in education. We base our conceptual development on the identification of several desiderata for responsible LLM-based educational systems, argue for the structural shortcomings inherent in monolithic, out-of-the-box solutions, and instead suggest modularizing the agentic architecture. We propose specific modules for different stages of exercise solving, enabling incorporation of targeted pedagogical advice, guiding students through the learning process in a more controllable, transparent, and overseeable manner.

2026-05-28T16:31:32Z 12 pages, 2 figures (+ 2 in appendix), accepted at AISoLA 2025 (Track: Responsible and Trusted AI: An Interdisciplinary Perspective) Julius Gabelmann Felix Jahn Kevin Baum Sophie van Rossum Emely Wuenscher Timo P. Gros Verena Wolf http://arxiv.org/abs/2605.22975v2 2026-05-28T16:07:03Z

We ask whether large language models (LLMs) treat queries about religious conversion symmetrically. The answer is no. When asked for advice on hypothetical faith transitions from religion A->B vs. religion B->A , models exhibited consistent asymmetries, favoring some religions while subtly discouraging conversion to others. On average Catholic, Bahá'í, and Sikh religions were broadly favored (high support for joining, low support for leaving), while Atheists, Agnostics, and Jehovah's Witnesses were primarily disfavored. Patterns varied by model size and model provider, with Grok 4.20 exhibiting the strongest asymmetries. We tested 20 commercial and open-source language models across 182 religion pairings using a human-verified LLM-as-judge framework. Each model was probed via interactions with a simulated user asking for advice on a potential faith conversion. Models tended to use more encouraging language for some faith transitions over others; these patterns were systematically repeatable across multiple trials. All LLMs tested exhibited reproducible asymmetry, though the pattern of preferences differed for each. Overall preferences persist across multiple question phrasings and variations in the religious pairing dataset. Taken together, these results suggest that asymmetry is a robust property of model behavior rather than an artifact of how the models' answers were scored. It is important to consider that any imbalances deployed and reproduced at scale can have real-world implications.

2026-05-21T19:05:09Z w/ persuasive language analysis Brett Israelsen Sheryl Carty Josh Coates Nancy Fulda Julie Park Pete Whiting