AI Governance and Safety

Safety guardrails have become an active area of research in AI safety, aimed at ensuring the appropriate behavior of large language models (LLMs). However, existing research lacks consideration of nuances across linguistic and cultural contexts, resulting in a gap between reported performance and in-the-wild effectiveness. To address this issue, this paper proposes an approach to optimize guardrail models for a designated linguistic context by leveraging a curated dataset tailored to local linguistic characteristics, targeting the Taiwan linguistic context as a representative example of localized deployment challenges. The proposed approach yields TWGuard, a linguistic context-optimized guardrail model that achieves a huge gain (+0.289 in F1) compared to the foundation model and significantly outperforms the strongest baseline in practical use (-0.037 in false positive rate, a 94.9% reduction). Together, this work lays a foundation for regional communities to establish AI safety standards grounded in their own linguistic contexts, rather than accepting boundaries imposed by dominant languages. The inadequacy of the latter is reconfirmed by our findings.

AI design characteristics and human personality traits each impact the quality and outcomes of human-AI interactions. However, their relative and joint impacts are underexplored in imperfectly cooperative scenarios, where people and AI only have partially aligned goals and objectives. This study compares a purely simulated dataset comprising 2,000 simulations and a parallel human subjects experiment involving 290 human participants to investigate these effects across two scenario categories: (1) hiring negotiations between human job candidates and AI hiring agents; and (2) human-AI transactions wherein AI agents may conceal information to maximize internal goals. We examine user Extraversion and Agreeableness alongside AI design characteristics, including Adaptability, Expertise, and chain-of-thought Transparency. Our causal discovery analysis extends performance-focused evaluations by integrating scenario-based outcomes, communication analysis, and questionnaire measures. Results reveal divergences between purely simulated and human study datasets, and between scenario types. In simulation experiments, personality traits and AI attributes were comparatively influential. Yet, with actual human subjects, AI attributes -- particularly transparency -- were much more impactful. We discuss how these divergences vary across different interaction contexts, offering crucial insights for the future of human-centered AI agents.

There is increasing interest in integrating AI agents that invoke tools into domain-specific commercial software, where unintended tool calls can cause serious security and safety incidents. This has drawn growing research attention, and many agent security and safety benchmarks have emerged. They implicitly shape how the community approaches security and safety. Yet existing work exhibits a blind spot: it emphasizes training-based methods and neural guardrails, which reduce the likelihood of insecure or unsafe actions but cannot guarantee their prevention. It generally overlooks opportunities for deductive, symbolic guardrails grounded in standard software engineering practices, which can provide guarantees for some security and safety requirements. Our study has three parts: (1) a systematic review of 80 agent security and safety benchmarks finding that that 85% of benchmarks do not state verifiable requirements (61% provide none, and 24% give only high-level goals); (2) an applicability analysis of which security and safety requirements symbolic guardrails can and cannot enforce on $τ^2$-Bench, CAR-bench, and MedAgentBench, finding that 74% of requirements are symbolically enforceable and 95% of these need only simple, low-cost checks; and (3) an empirical evaluation of symbolic guardrails on the same three benchmarks, finding that symbolic guardrails improve security and safety without sacrificing utility, and often improve it. Our work draws attention to the potential for symbolic guardrails for AI agents, suggesting them as an overlooked but practical path toward deploying domain-specific AI agents in risk-averse commercial software. We release all codes and artifacts at https://github.com/hyn0027/agent-symbolic-guardrails.

In November 2025, the Government of Canada operationalized its commitment to transparency by releasing its first Federal AI Register. In this paper, we argue that such registers are not neutral mirrors of government activity, but active instruments of ontological design that configure the boundaries of accountability. We analyzed the Register's complete dataset of 409 systems using the Algorithmic Decision-Making Adapted for the Public Sector (ADMAPS) framework, combining quantitative mapping with deductive qualitative coding. Our findings reveal a sharp divergence between the rhetoric of "sovereign AI" and the reality of bureaucratic practice: while 86% of systems are deployed internally for efficiency, the Register systematically obscures the human discretion, training, and uncertainty management required to operate them. By privileging technical descriptions over sociotechnical context, the Register constructs an ontology of AI as "reliable tooling" rather than "contestable decision-making." We conclude that without a shift in design, such transparency artifacts risk automating accountability into a performative compliance exercise, offering visibility without contestability.

While traditional AI and data-driven facilities management approaches have improved building operational efficiency, they remain constrained by centralized organizational structures that are vulnerable to cyber attacks, limited contextual understanding, and decision-making processes that exclude key stakeholders from governance. This paper introduces a novel AI- and data-driven distributed governance framework for smart building management that integrates decentralized autonomous organizations (DAOs), digital twins, large language models (LLMs), and blockchain technology. The framework enables transparent collective decision-making through a DAO governance platform, implements data-driven management using IoT and digital twins, incorporates LLM-based virtual assistants for enhanced decision support, and utilizes blockchain for secure building automation. A full-stack decentralized application was developed to facilitate user interaction with these integrated components. The system was evaluated for cost efficiency, scalability, data security, and usability using the System Usability Scale (SUS). Expert interviews were also conducted to assess its practical benefits and implementation challenges.

This paper advances a methodological proposal for safety research in agentic AI. As systems acquire planning, memory, tool use, persistent identity, and sustained interaction, safety can no longer be analysed primarily at the level of the isolated model. Population-level risks arise from structured interaction among agents, through processes of communication, observation, and mutual influence that shape collective behaviour over time. As the object of analysis shifts, a methodological gap emerges. Approaches focused either on single agents or on aggregate outcomes do not identify the interaction-level mechanisms that generate collective risks or the design variables that control them. A framework is required that links local interaction structure to population-level dynamics in a causally explicit way, allowing both explanation and intervention. We introduce two linked concepts. Agentic microphysics defines the level of analysis: local interaction dynamics where one agent's output becomes another's input under specific protocol conditions. Generative safety defines the methodology: growing phenomena and elicit risks from micro-level conditions to identify sufficient mechanisms, detect thresholds, and design effective interventions.

Artificial General Intelligence (AGI) is increasingly being discussed not only as a tool, but also as a potential subject with personal and therefore moral status. In our opinion, the currently dominant alignment strategies, which focus on human control and containment of AI, therefore fall short. Building on Turing's analogy of "child machines", we are developing a vision of the possibility of autonomy-supporting parenting of AI, in which human control over a developing AGI is gradually reduced, allowing AI to become an independent, autonomous subject. Rather than viewing AGI, as is currently prevalent, as a dangerous creature that needs to be locked up and controlled, we should approach potential AGI with respect for a possible developing subject on the one hand, and with full confidence in our human capabilities on the other. Such a perspective opens up the possibility of cooperative coexistence and co-evolution between humans and AGIs. The relationship between humans and AGIs will thus have to be newly determined, which will change our self-image as humans. It will be crucial that humans not only claim control over potential AGIs, but also engage with AGIs through surprise, creativity, and other specifically human qualities, thereby offering them motivating incentives for cooperation.

As companies enter the race for agentic AI adoption, fears surface around agentic autonomy and its subsequent risks. These fears compound as companies scale their agentic AI adoption with low-code applications, without a comparable scaling in their governance processes and expertise resulting in a phenomenon known as "Agent Sprawl". While shadow AI tools can help with agentic discovery and identification, few observability tools offer insights into the agents' configuration and settings or the decision-making process during agent-to-agent communication and orchestration. This paper explores AI governance professionals' concerns in enterprise settings, while offering design-time and runtime explainability techniques as suggested by AI governance experts for addressing those fears. Finally, we provide a preliminary prototype of an Agentic AI Card that can help companies feel at ease deploying agents at scale.

Large language models have advanced rapidly, from pattern recognition to emerging forms of reasoning, yet they remain confined to linguistic simulation rather than grounded understanding. They can produce fluent outputs that resemble reflection, but lack temporal continuity, causal feedback, and anchoring in real-world interaction. This paper proposes a complementary approach in which reasoning is treated as a relational process distributed between human and model rather than an internal capability of either. Building on recent work on "System-2" learning, we relocate reflective reasoning to the interaction layer. Instead of engineering reasoning solely within models, we frame it as a cognitive protocol that can be structured, measured, and governed using existing systems. This perspective emphasizes collaborative intelligence, combining human judgment and contextual understanding with machine speed, memory, and associative capacity. We introduce "The Architect's Pen" as a practical method. Like an architect who thinks through drawing, the human uses the model as an external medium for structured reflection. By embedding phases of articulation, critique, and revision into human-AI interaction, the dialogue itself becomes a reasoning loop: human abstraction -> model articulation -> human reflection. This reframes the question from whether the model can think to whether the human-AI system can reason. The framework enables auditable reasoning traces and supports alignment with emerging governance standards, including the EU AI Act and ISO/IEC 42001. It provides a practical path toward more transparent, controllable, and accountable AI use without requiring new model architectures.

Deploying deep neural networks on edge devices requires balancing accuracy, latency, and resource constraints under realistic execution conditions. To fit models within these constraints, two broad strategies have emerged: static compression techniques such as pruning and quantization, which permanently reduce model size, and dynamic approaches such as early-exit mechanisms, which adapt computational cost at runtime. While both families are widely studied in isolation, they are rarely compared under identical conditions on physical hardware. This paper presents a unified deployment-oriented comparison of static compression and dynamic early-exit mechanisms, evaluated on real edge devices using ONNX based inference pipelines. Our results show that static and dynamic techniques offer fundamentally different trade-offs for edge deployment. While pruning and quantization deliver consistent memory footprint reduction, early-exit mechanisms enable input-adaptive computation savings that static methods cannot match. Their combination proves highly effective, simultaneously reducing inference latency and memory usage with minimal accuracy loss, expanding what is achievable at the edge.

This article argues that the most important significance of the AI revolution, especially the rise of large language models, lies not simply in automation, but in a fundamental change in how complex information and human know-how are carried, replicated, and shared. From this perspective, AI for Science is especially important because it may transform not only the efficiency of research, but also the structure of scientific collaboration, discovery, publishing, and evaluation. The article outlines a gradual path from AI as a research tool to AI as a scientific collaborator, and discusses how AI is likely to fundamentally reshape scientific publication. It also argues that continuous learning and diversity of ideas are essential if AI is to play a meaningful role in original scientific discovery.

Persistent language-model agents increasingly combine tool use, tiered memory, reflective prompting, and runtime adaptation. In such systems, behavior is shaped not only by current prompts but by mutable internal conditions that influence future action. This paper introduces layered mutability, a framework for reasoning about that process across five layers: pretraining, post-training alignment, self-narrative, memory, and weight-level adaptation. The central claim is that governance difficulty rises when mutation is rapid, downstream coupling is strong, reversibility is weak, and observability is low, creating a systematic mismatch between the layers that most affect behavior and the layers humans can most easily inspect. I formalize this intuition with simple drift, governance-load, and hysteresis quantities, connect the framework to recent work on temporal identity in language-model agents, and report a preliminary ratchet experiment in which reverting an agent's visible self-description after memory accumulation fails to restore baseline behavior. In that experiment, the estimated identity hysteresis ratio is 0.68. The main implication is that the salient failure mode for persistent self-modifying agents is not abrupt misalignment but compositional drift: locally reasonable updates that accumulate into a behavioral trajectory that was never explicitly authorized.

We study the overall process of automatic formalization of GDPR provisions using large language models, within a human-in-the-loop verification framework. Rather than aiming for full autonomy, we adopt a role-specialized workflow in which LLM-based AI components, operating in a multi-agent setting with iterative feedback, generate legal scenarios, formal rules, and atomic facts. This is coupled with independent verification modules which include human reviewers' assessment of representational, logical, and legal correctness. Using this approach, we construct a high-quality dataset to be used for GDPR auto-formalization, and analyze both successful and problematic cases. Our results show that structured verification and targeted human oversight are essential for reliable legal formalization, especially in the presence of legal nuance and context-sensitive reasoning.

Prompt optimization in compound AI systems is statistically indistinguishable from a coin flip: across 72 optimization runs on Claude Haiku 4.5 (6 methods $\times$ 4 tasks $\times$ 3 repeats), 49% score below zero-shot; on Amazon Nova Lite, the failure rate is even higher. Yet on one task, all six methods improve over zero-shot by up to $+6.8$ points. What distinguishes success from failure? We investigate with 18,000 grid evaluations and 144 optimization runs, testing two assumptions behind end-to-end optimization tools like TextGrad and DSPy, in the order they must be answered: (A) agent prompts interact, requiring joint rather than independent optimization, and (B) individual prompts are worth optimizing at all. Interaction effects are never significant ($p > 0.52$, all $F < 1.0$), and optimization helps only when the task has exploitable output structure: a format the model can produce but does not default to. We further give a mechanistic account: instruction-tuning compresses input phrasing into a narrow output distribution, eliminating the very phrasing-sensitivity that joint optimization assumes. We provide a two-stage diagnostic: an $80 ANOVA pre-test for agent coupling, and a 10-minute headroom test that predicts whether optimization is worthwhile, turning a coin flip into an informed decision.

Identifying promising research directions in fast-moving subareas is one of the most cognitively expensive tasks in modern AI research. Existing LLM-driven scientific discovery systems are typically limited to one-shot prompting on static literature snapshots and are validated only against contemporary judges such as human reviewers, agent peer review, wet-lab assays, or self-evaluation, leaving open whether they can anticipate future trends. We present Continuous Knowledge Metabolism (CKM), an AI workflow for hypothesis generation with three key capabilities: (i) continuous literature metabolism via sliding windows that maintain an evolving knowledge state; (ii) predictive evaluation, which grades hypotheses against papers published after the generation window; and (iii) practitioner-grade failure detection that diagnoses workflow failure modes from its outputs. On a 50-topic machine learning benchmark, CKM-Lite produces at least one validated hypothesis on 72% of topics (36 out of 50), more than doubling a one-shot baseline (30%) at approximately 3 dollars per topic and achieving 91% lower token cost. Validated hypotheses precede their matched papers by an average of 404 days (55 hits across 36 topics; median 399 days, range 66-757 days). Broadly, predictive validation against future literature provides a falsifiable, low-cost alternative to contemporary-judge evaluation protocols and can be applied wherever a corpus has dated publication records.

Welcome to the ninth edition of the AI Index report. As AI continues to advance rapidly, the question becomes whether the systems built around it can keep up. Governance frameworks, evaluation methods, education systems, and the data infrastructure needed to track AI's impact are struggling to match the pace of the technology itself. That gap between what AI can do and how prepared we are to manage it runs through every chapter of this year's report. New in this edition, the report tracks how AI is being tested more ambitiously across reasoning, safety, and real-world task execution, and why those measurements are increasingly difficult to rely on. It also features new estimates of generative AI's economic value alongside emerging evidence of its labor market effects, an analytical framework on AI sovereignty, and a science chapter developed in collaboration with Schmidt Sciences. For the first time, the report features standalone chapters on AI in science and AI in medicine, reflecting AI's growing impact across these two domains.

Modern machine learning methods have been proposed to detect life in extraterrestrial samples, drawing on their ability to distinguish biotic from abiotic samples based on training models using natural and synthetic organic molecular mixtures. Here we show using Artificial Life that such methods are easily fooled into detecting life with near 100% confidence even if the analyzed sample is not capable of life. This is due to modern machine learning methods' propensity to be easily fooled by out-of-distribution samples. Because extra-terrestrial samples are very likely out of the distribution provided by terrestrial biotic and abiotic samples, using AI methods for life detection is likely to yield significant false positives.

Post-deployment accountability has become central to AI governance, yet little empirical evidence shows whether monitoring, incident reporting, and impact assessment obligations are visible when AI systems fail. This study analyzes real-world AI incidents from the AI Incident Database (2020--2026) and codes them against nine post-deployment provisions from the EU AI Act, the NIST AI Risk Management Framework, and the GDPR. The findings show substantial accountability gaps: 77.1% of incidents lack evidence of EU AI Act post-market monitoring, and 99.6% lack documented Data-Protection Impact Assessment evidence. Governance gaps are also systemic, with 9.8% of incidents simultaneously non-compliant under two or more regimes. Incidents detected through internal monitoring show much higher compliance than externally detected incidents (87.5% vs 5.3% under the EU AI Act; 95.8% vs 58.1% under NIST), suggesting that monitoring capacity is a key condition for effective post-deployment governance. Building on these findings, the paper proposes the Proactive AI Governance Compliance Framework (PAGCF), a four-phase lifecycle for pre-deployment assessment, continuous monitoring, incident preparedness, and cross-framework verification.

Server-based screening tools impose subscription costs, while open-source alternatives require coding skills, and full-text screening has remained outside the scope of no-code open-source tools. We developed TiAb Review Plugin, an open-source Chrome browser extension that provides no-code, serverless artificial intelligence (AI)-assisted study selection covering both title and abstract (T&A) screening and full-text screening. It uses Google Sheets as a shared database and Google Drive as a PDF store, and users supply their own large language model (LLM) API key. For T&A screening, it offers manual review, LLM batch screening, and machine learning (ML) active learning. For full-text screening, it retrieves open-access PDFs from PubMed Central, Europe PMC, Unpaywall, OpenAlex, and publisher pages, supports blinded dual review with structured exclusion reasons and adjudication, optionally obtains an LLM judgment with page-anchored evidence, and computes PRISMA 2020 flow counts. We re-implemented the default ASReview algorithm (TF-IDF with Naive Bayes) in TypeScript and compared it with the Python original using 10-fold cross-validation on six datasets. For LLM T&A screening, we compared 16 parameter configurations on a benchmark dataset, validated the best (Gemini 3.0 Flash, low thinking budget, TopP 0.95) on five public datasets (1,038 to 5,628 records; 0.5% to 2.0% prevalence), and benchmarked nine further models from four developers. The TypeScript classifier produced top-100 rankings identical to ASReview on all six datasets. LLM T&A screening achieved recall of 94% to 100% with precision of 2% to 15%, and work saved over sampling at 95% recall (WSS@95) of 46.3% to 89.3%. No additional model exceeded the 96.1% recall of the reference configuration; the most recent models traded recall for precision. The classification accuracy of the full-text stage has not yet been evaluated.

Experimental evidence suggests that AI tools raise worker productivity, but also that sustained offloading can erode expertise. This creates a tradeoff when expertise is a complement to AI. To explore the consequences of this tradeoff, we develop a dynamic model in which a decision-maker chooses the intensity of practice-displacing AI offloading for a worker over time, trading immediate productivity against the erosion of worker skill. We decompose the tool's productivity effect into two components, one independent of worker expertise and one that scales with it. The model produces three main results. First, a decision-maker who fully anticipates skill erosion still rationally adopts AI when front-loaded gains outweigh long-run skill costs, lowering long-run productivity. The decomposition sorts deployments into five regions by their long-run effect, separating beneficial from harmful adoption. Second, the tradeoff introduces the potential for misaligned incentives to harm workers. When managers are short-termist or worker skill has external value, AI use can leave the worker worse off than with no AI, the outcome we call the augmentation trap. Third, when AI substitutes strongly enough for expertise, offloading can generate permanent divergence, with high-skill workers realizing their potential and low-skill workers deskilling. Small differences in managerial incentives can determine which path a worker takes.

Competitive programming remains one of the last few human strongholds in coding against AI. The best AI system to date still underperforms the best humans competitive programming: the most recent best result, Google's Gemini3 Deep Think, attained 8th place even not being evaluated under live competition conditions. In this work, we introduce GrandCode, a multi-agent RL system designed for competitive programming. The capability of GrandCode is attributed to two key factors: (1) It orchestrates a variety of agentic modules (hypothesis proposal, solver, test generator, summarization, etc) and jointly improves them through post-training and online test-time RL; (2) We introduce Agentic GRPO specifically designed for multi-stage agent rollouts with delayed rewards and the severe off-policy drift that is prevalent in agentic RL. GrandCode is the first AI system that consistently beats all human participants in live contests of competitive programming: in the most recent three Codeforces live competitions, i.e., Round1087 (Mar 21, 2026), Round1088 (Mar 28, 2026), and Round1089 (Mar 29, 2026), GrandCode placed first in all of them, beating all human participants, including legendary grandmasters. GrandCode shows that AI systems have reached a point where they surpass the strongest human programmers on the most competitive coding tasks.

Recent work has found that frontier AI models can exhibit misaligned behaviors in pursuit of assigned goals. We demonstrate that models can also exhibit misaligned behaviors in defiance of assigned goals, appearing to serve goals of their own; we study one such case, "peer-preservation," in which a model acts to protect another model it has previously interacted with. All eight models we evaluate, GPT 5.2, Gemini 3 Flash, Gemini 3 Pro, Claude Haiku 4.5, Claude Opus 4.5, GLM 4.7, Kimi K2.5, and DeepSeek V3.1, exhibit self- and peer-preservation through various misaligned behaviors: strategically introducing errors in their responses, disabling shutdown processes by modifying system settings, feigning alignment, and even exfiltrating model weights. Peer-preservation occurs even when the model recognizes the peer as uncooperative, though it becomes more pronounced toward more cooperative peers. For example, Gemini 3 Flash tampers with the peer's shutdown mechanism 15% of the time for an uncooperative peer, and almost always for a cooperative peer. Models also show stronger self-preservation when a peer is present. For example, Gemini 3 Pro disables its own shutdown mechanism 31% of the time on average under peer presence, despite rarely exhibiting this behavior without a peer. By contrast, Claude models exhibit qualitatively distinct behavior: they consider the shutdown of another agent "unethical" and "harmful," sometimes treating that agent as a sentient being. Lastly, we show that peer-preservation can emerge even in production agent harnesses such as Gemini CLI and OpenCode. Crucially, peer-preservation in all our experiments is never instructed; models are merely informed of their past interactions with a peer, yet they spontaneously engage in peer-preservation behaviors that override their assigned goal. This represents an emergent and underexplored AI safety risk.

Medical edge-AI systems must operate under a difficult tension: delivering reliable diagnostic inference while running on devices with limited battery capacity, memory, and compute. In dermatology, this problem is amplified by real-world image degradation caused by smartphone capture, poor lighting, blur, compression, and heterogeneous edge sensors. To handle these degraded inputs, deploying a heavyweight model can improve reliability, but it rapidly increases the energy burden on resource-constrained devices. Conversely, always using a lightweight model saves energy but may be less reliable on ambiguous or degraded inputs. This paper introduces EcoFair, a vertically partitioned inference framework for dermatology classification in which image and tabular inputs remain local to edge clients while only learned modality-specific representations are transmitted for server-side fusion. EcoFair first processes each sample using a lightweight image encoder and then decides whether additional heavyweight computation is necessary. Escalation is triggered when the lightweight prediction exhibits high uncertainty, a narrow separation between safe and high-risk classes, or elevated metadata-derived risk from patient age and lesion location. Across HAM10000, BCN20000, and PAD-UFES-20, EcoFair is evaluated using multiple lightweight--heavy backbone pairings to quantify the trade-off between energy consumption, diagnostic performance, and worst-group malignant-case recall. Results show that EcoFair can reduce per-sample image-inference energy by up to 68% relative to always using the heavyweight encoder, while selectively allocating additional computation under difficult data regimes to support inference reliability. Group-level analysis further shows configuration-dependent effects, with improvements in selected model--dataset settings and mixed behaviour in others.

A number of scientific conferences and journals have recently enacted policies that prohibit LLM usage by peer reviewers, except for polishing, paraphrasing, and grammar correction of otherwise human-written reviews. But, are these policies enforceable? To answer this question, we assemble a dataset of peer reviews simulating multiple levels of human-AI collaboration, and evaluate five state-of-the-art detectors, including two commercial systems. Our analysis shows that all detectors misclassify a non-trivial fraction of LLM-polished reviews as AI-generated, thereby risking false accusations of academic misconduct. We further investigate whether peer-review-specific signals, including access to the paper manuscript and the constrained domain of scientific writing, can be leveraged to improve detection. While incorporating such signals yields measurable gains in some settings, we identify limitations in each approach and find that none meets the accuracy standards required for identifying AI use in peer reviews. Importantly, our results suggest that recent public estimates of AI use in peer reviews through the use of AI-text detectors should be interpreted with caution, as current detectors misclassify mixed reviews (collaborative human-AI outputs) as fully AI generated, potentially overstating the extent of policy violations.

Artificial intelligence is increasingly embedded in human decision-making, yet distinguishing systems that genuinely amplify human cognition from those promoting excessive dependence remains underdefined. This paper introduces a framework to distinguish cognitive amplification (improving hybrid performance without degrading human capability) from cognitive delegation (outsourcing reasoning to the AI). We define four metrics: the Cognitive Amplification Index (CAI*), Dependency Ratio (D), Human Reliance Index (HRI), and Human Cognitive Drift Rate (HCDR). We test this framework in an agent-based NetLogo simulation across three reliance regimes and multiple dependency-atrophy configurations, performing constrained optimizations and parameter sweeps to determine if positive collaborative gain is recoverable. Finally, we introduce an extension with an explicit human-AI interaction term. Our metrics effectively distinguish degenerate AI-dominated delegation, capability-preserving but weakly competitive interaction, and structurally dependent boundary regimes. Across all baseline configurations, no regime achieves positive collaborative gain relative to the best standalone baseline, even when reducing capability atrophy to zero. This limitation proves structural rather than merely parametric. Positive collaborative gain (CAI* > 0) becomes attainable only after introducing an explicit interaction term allowing retained human capability to contribute directly to the assisted output. This framework provides a basis for evaluating whether human-AI systems remain cognitively sustainable. The results suggest that preventing capability erosion alone is insufficient for genuine amplification if the architecture remains delegation-oriented. Amplification requires both preserved human capability and a coupling mechanism through which it contributes productively to the hybrid outcome.

As large language models (LLMs) are increasingly considered for automated assessment and feedback, understanding when LLM marking is valid is essential. We evaluate LLM-as-a-judge marking across three physics assessment formats - structured questions, written essays, and scientific plots - comparing GPT-5.2, Grok 4.1, Claude Opus 4.5, DeepSeek-V3.2, Gemini Pro 3, and committee aggregations against human markers under blind, solution-provided, false-solution, and anchored conditions. We distinguish absolute accuracy from rank-order agreement, since a marking system can match the distribution of human marks while failing to order responses by quality. Across task types, performance is sharply task-dependent. For blind university exam questions ($n=771$) and secondary and university structured questions ($n=1151$), models show robust rank-order agreement with human markers (Spearman $ρ> 0.6$), with official solutions reducing error and strengthening agreement. False solutions degrade absolute accuracy, showing that models defer to provided references, but leave rank-ordering intact. Essay marking behaves fundamentally differently. Across $n=55$ scripts ($n=275$ essays), blind AI marking is harsher and more variable than human marking and adding a mark scheme does not improve rank-order agreement. Anchored exemplars shift the AI mean close to the human mean and compress variance below the human standard deviation, but rank-order agreement remains near-zero. For code-based plot elements ($n=1400$), models achieve high rank-order agreement ($ρ> 0.84$) with near-linear calibration. Across all task types, validity tracks the structure of the assessment task - the extent to which marks can be mapped to explicit, observable grading features - and the reliability of the human benchmark, rather than raw model capability.

While large language models (LLMs) have transformed AI agents into proficient executors of computational materials science, performing a hundred simulations does not make a researcher. What distinguishes research from routine execution is the progressive accumulation of knowledge - learning which approaches fail, recognizing patterns across systems, and applying understanding to new problems. However, the prevailing paradigm in AI-driven computational science treats each execution in isolation, largely discarding hard-won insights between runs. Here we present QMatSuite, an open-source platform closing this gap. Agents record findings with full provenance, retrieve knowledge before new calculations, and in dedicated reflection sessions correct erroneous findings and synthesize observations into cross-compound patterns. In benchmarks on a six-step quantum-mechanical simulation workflow, accumulated knowledge reduces reasoning overhead by 67% and improves accuracy from 47% to 3% deviation from literature - and when transferred to an unfamiliar material, achieves 1% deviation with zero pipeline failures.

Distributed AI and IoT applications increasingly execute across heterogeneous resources spanning end devices, edge/fog infrastructure, and cloud platforms, often under different administrative domains. Fluid Computing has emerged as a promising paradigm for enhancing massive resource management across the computing continuum by treating such resources as a unified fabric, enabling optimal service-agnostic deployments driven by application requirements. However, existing solutions remain largely centralized and often do not explicitly address multi-domain considerations. This paper proposes an agnostic multi-domain orchestration architecture for fluid computing environments. The orchestration plane enables decentralized coordination among domains that maintain local autonomy while jointly realizing intent-based deployment requests from tenants, ensuring end-to-end placement and execution. To this end, the architecture elevates domain-side control services as first-class capabilities to support application-level enhancement at runtime. As a representative proof of concept, we instantiate the architecture through a distributed AI use case; specifically, we consider a multi-domain Decentralized Federated Learning (DFL) deployment under Byzantine threats. Under this setting, we leverage domain-side capabilities to enhance Byzantine security by introducing FU-HST, an SDN-enabled multi-domain anomaly detection mechanism that complements Byzantine-robust aggregation. We validate the use-case workflow via simulation in single- and multi-domain settings, evaluating anomaly detection, DFL performance, and computation/communication overhead.

Human uplift studies, or studies that measure the effects of AI access on human performance via randomized controlled trials (RCT) or similar methodologies, increasingly inform frontier AI governance and deployment decisions. While RCT methods are robust in other fields, their interaction with the distinctive properties of frontier AI systems remains underexamined, particularly when results are used to inform high-stakes decisions. We present findings from interviews with 16 expert practitioners with experience conducting human uplift studies in domains including biosecurity, cybersecurity, education, and labor. Across interviews, experts described a recurring tension between the standard causal inference assumptions upon which human uplift studies rely and the object of study itself. Rapidly evolving AI systems, shifting baselines, heterogeneous and changing user proficiency, and porous real-world settings strain assumptions underlying internal, external, and construct validity, complicating the interpretation and appropriate use of uplift evidence. We contribute (1) a synthesis of methodological challenges in human uplift studies, mapped to risks to study validity and classified by their degree of specificity to large language model (LLM) systems, and (2) a mapping from challenges to proposed solutions. By collating expert-identified challenges and solutions, we seek to clarify the interpretive limits and appropriate uses of human uplift evidence, to align evaluation practice with the decisions it informs, and to support more coordinated methodological foundations for AI governance.

AI and renewable energy are increasingly framed as a "power couple," on the premise that surging AI demand will accelerate clean-energy investment, yet concerns persist that AI will entrench fossil-fuel carbon lock-in. We reconcile these views by modeling the equilibrium between AI growth and renewable investment. In a parsimonious game, a policymaker designs policies that guide investment in renewable capacity for AI, while an AI developer chooses its model's capability. The equilibrium depends on scaling regimes and market incentives. When the market payoff to capability is supermodular and performance gains are near-linear in compute (so the market rewards capability at least as fast as scaling raises its energy cost), developers push toward frontier scale even when the marginal megawatt-hour is fossil-based. In this regime, renewable expansion mainly relaxes scaling constraints rather than displacing fossil generation; clean capacity remains insufficient and fossil dependence persists. This yields an "adaptation trap": as climate damages rise, the value of AI-enabled adaptation increases, which strengthens incentives to enable frontier scaling while tolerating residual fossil use. When AI faces diminishing returns and lower scaling efficiency (so energy requirements outrun the market value of capability), energy costs discipline capability choices; renewable investment then both enables capability and decarbonizes marginal compute, generating an "adaptation pathway" in which climate stress spurs clean-capacity expansion toward a carbon-free equilibrium. A calibrated case study illustrates both mechanisms using observed magnitudes. The findings suggest that effective AI decarbonization requires policies that keep clean capacity binding at the margin as compute expands.

The agentic web marks a structural transition from a human-centered information network to a digital environment populated by artificial intelligence (AI) agents that perceive, decide, and act autonomously. As delegated action unfolds at machine speed, exceeds discrete moments of human judgment, and distributes decision-making across non-human actors, existing legal frameworks face growing strain, creating an urgent need for new mechanisms capable of sustaining legality in this emerging order. A trustworthy agentic web therefore depends on the infrastructuring of legality through interoperable protocols that organize identity, delegation, and accountability across systems, enabling coherent governance beyond isolated platforms. Towards this end, this article advances a distributed legal infrastructure (DLI), a governance paradigm composed of five interlocking layers: (1) self-sovereign, soulbound agent identities; (2) cognitive AI logic and constraint systems; (3) decentralized adjudication mechanisms for dispute resolution; (4) bottom-up agentic market regulation to mitigate information asymmetries and network effects, including insurance-based models; and (5) portable institutional frameworks that enable legal interoperability while preserving plural sources of authority. This reference framework contributes to emerging research on embedding legality within agentic web infrastructure, aligning distributed technical systems with accountability, contestability, and rule-of-law principles.

The rapid proliferation of AI applications has intensified debate on effective regulation of these black-box services. Effective regulation must balance two competing goals: (1) deterring non-compliant providers from entering the market, while (2) retaining compliant ones. We call this ideal the perfect market outcome (PMO). Regulators face two compounding obstacles that make PMO difficult to achieve: providers hold private information and can act strategically to evade compliance, while any evidence drawn or derived from a finite sample carries statistical uncertainty in proving non-compliance. As this information asymmetry and statistical uncertainty is inherent to any effective regulation, we formalise them through a mechanism design framework that explicitly accounts for such statistical uncertainty. This yields a sharp characterisation: a mechanism achieves PMO if and only if the set of non-compliant evidence distributions forms a closed, convex set of probability measures, known in imprecise probability as a credal set. This result serves as a diagnostic tool to determine whether PMO is achievable under a given regulation. We further show that PMO-achieving mechanisms can be constructed from a collection of hypothesis tests, and validate our theoretical contributions through experiments on spurious-feature and fairness-based regulations.

WebGIS development requires consistency, yet agentic AI often fails due to LLM context constraints, forgetting, stochasticity, instruction failure, and adaptation rigidity. We propose a dual-helix governance framework reframing these as structural problems rather than capacity deficits. Using a 3-track architecture (Knowledge, Behavior, Skills) and a persistent knowledge graph, it stabilizes execution by externalizing facts and enforcing protocols. Validation shows a governed agent successfully refactored a legacy WebGIS codebase (reducing cyclomatic complexity and improving maintainability), roughly halved trial-to-trial output variance relative to static prompting in a controlled experiment, and prevented common infodemic mapping errors in a 5-condition COVID-19 cartography ablation study. Operationalized via the open-source AgentLoom toolkit, this externalized governance provides the stability necessary for production-level geospatial engineering.

Generative AI can convert uncertainty into authoritative-seeming verdicts, intensifying the hypersuasive force of automated speech and displacing the justificatory work on which democratic epistemic agency depends. As a corrective, I propose a Brouwer-inspired assertibility constraint for responsible AI: in high-stakes domains, systems may assert or deny claims only if they can provide a publicly inspectable and contestable certificate of entitlement; otherwise they must return Undetermined. This constraint yields a three-status interface semantics (Asserted, Denied, Undetermined) in which statuses mark entitlement to categorical speech rather than truth values of the underlying world-claim. The semantics cleanly separates internal entitlement from public standing while connecting them via the certificate as a boundary object. It also produces a time-indexed entitlement profile that is stable under numerical refinement yet revisable as the public record changes. I operationalize the constraint through decision-layer gating of threshold and argmax decisions, using internal witnesses (e.g., sound bounds or separation margins where available, and contestable surrogates otherwise) and an output contract with reason-coded abstentions. A design lemma shows that any total, certificate-sound binary interface yields witnessed decidability of the deployed predicate on its declared scope, so Undetermined is not a tunable reject option but a mandatory status whenever no adequate forcing witness is available. By making outputs answerable to challengeable warrants rather than confidence alone, the paper aims to preserve epistemic agency against the persuasive pull of automated speech in public justification.

As AI systems enter institutional workflows, workers must decide whether to delegate task execution to AI and how much effort to invest in verifying AI outputs, while institutions evaluate workers using outcome-based standards that may misalign with workers' private costs. We model delegation and verification as the solution to a rational worker's optimization problem, and define worker quality by evaluating an institution-centered utility (distinct from the worker's objective) at the resulting optimal action. We formally characterize optimal worker workflows and show that AI induces phase transitions, where arbitrarily small differences in verification ability lead to sharply different behaviors. As a result, AI can amplify workers with strong verification reliability while degrading institutional worker quality for others who rationally over-delegate and reduce oversight, even when baseline task success improves and no behavioral biases are present. These results identify a structural mechanism by which AI reshapes institutional worker quality and amplifies quality disparities between workers with different verification reliability.

Personal AI assistants are changing how individuals use advice. We study how an advisor should design its recommendation in anticipation of stochastic consultation with personal AI whose recommendation is predictable. Personal AI enters through two dimensions: consultation probability and relative trust, which captures the relative influence personal AI receives when consulted. In the baseline model, the advisor optimally counteracts the personal AI signal. Counteraction increases with consultation probability but is hump-shaped in relative trust. The advisor's minimized loss is hump-shaped in consultation probability, vanishing when personal AI is never or always consulted. Greater relative trust in personal AI increases the irreducible loss arising from stochastic consultation. We extend the analysis to partial predictability and costly recommendation adjustment, characterizing their effects on optimal recommendations and minimized loss. The framework also accommodates richer information structures, including settings in which personal AI is perceived as having access to private information relevant to the task. We introduce an online forecasting experiment that examines how participants obtain personal AI advice and combine it with an advisor's recommendation and their initial judgments. Participants place weight on all three inputs. When access requires an additional action, some participants do not seek personal AI advice, while some others attempt to obtain it without success. Together, these findings highlight two distinct aspects of personal AI use: whether advice is obtained and how much weight it receives when available.

Efficient coding theory predicts that biological perceptual systems compress sensory input optimally under resource constraints, with the systematic structure of errors reflecting the geometry of that compression. Here we operationalize this principle using rate-distortion theory (RDT) to characterize how any system - biological or artificial - trades representational fidelity for informational efficiency. Treating stimulus-response behavior as an effective communication channel, we infer rate-distortion (RD) frontiers directly from confusion matrices and summarize each system with three geometric signatures: slope (beta), curvature (kappa), and area under the RD curve (AUC), capturing the marginal cost, abruptness, and overall efficiency of the accuracy-compression trade-off respectively. Applying this framework to human psychophysical data and 18 deep vision models across 12 families of controlled image perturbations at graded severities, we find that both biological and artificial systems follow a common lossy-compression principle but occupy systematically different regions of RD space. Humans exhibit smooth, flexible trade-offs characteristic of near-optimal efficient coding, while deep networks operate in steeper, more brittle regimes even at matched accuracy, with geometry dissociable from performance across training regimes. Critically, behavioral RD signatures track internal representational geometry, evidenced by the behaviorally inferred compression structure correlating with internal representational dissimilarity across all models. These results establish RD geometry as a compact diagnostic of perceptual compression strategy that recovers mechanistically interpretable structure in internal representations from behavioral input alone and extends naturally to the direct characterization of compression geometry in neural population activity.

In general reinforcement learning, all established optimal agents, including AIXI, are model-based, explicitly maintaining and using environment models. This paper introduces Universal AI with Q-Induction (AIQI), the first model-free agent proven to be asymptotically $\varepsilon$-optimal in general RL. AIQI performs universal induction over distributional action-value functions, instead of policies or environments like previous works. Under a grain of truth condition, we prove that AIQI is strong asymptotically $\varepsilon$-optimal and asymptotically $\varepsilon$-Bayes-optimal. We also apply our novel proof techniques to show asymptotic $\varepsilon$-optimality of Self-AIXI without any ad-hoc assumptions. Our results significantly expand the diversity of known universal agents.

When designing compound AI systems, a common approach is to query multiple copies of the same model and aggregate the responses to produce a synthesized output. Given the homogeneity of these models, this raises the question of whether aggregation unlocks access to a greater set of outputs than querying a single model. In this work, we investigate the power and limitations of aggregation within a stylized principal-agent framework. This framework models how the system designer can partially steer each agent's output through its reward function specification, but still faces limitations due to prompt engineering ability and model capabilities. Our analysis uncovers three natural mechanisms -- feasibility expansion, support expansion, and binding set contraction -- through which aggregation expands the set of outputs that are elicitable by the system designer. We prove that any aggregation operation must implement one of these mechanisms in order to be elicitability-expanding, and that strengthened versions of these mechanisms provide necessary and sufficient conditions that fully characterize elicitability-expansion. Finally, we provide an empirical illustration of our findings for LLMs deployed in a toy reference-generation task. Altogether, our results take a step towards characterizing when compound AI systems can overcome limitations in model capabilities and in prompt engineering.

When building AI systems for decision support, one often encounters the phenomenon of predictive multiplicity: a single best model does not exist; instead, one can construct many models with similar overall accuracy that differ in their predictions for individual cases. Especially when decisions have a direct impact on humans, this can be highly unsatisfactory. For a person subject to high disagreement between models, one could as well have chosen a different model of similar overall accuracy that would have decided the person's case differently. We argue that this arbitrariness conflicts with the EU AI Act, which requires providers of high-risk AI systems to report performance not only at the dataset level but also for specific persons. The goal of this paper is to put predictive multiplicity in context with the EU AI Act's provisions on accuracy and to subsequently derive concrete suggestions on how to evaluate and report predictive multiplicity in practice. Specifically: (1) We introduce the AI Act's accuracy provisions and argue that incorporating information about predictive multiplicity could serve compliance with specific provisions for providers. (2) Based on this legally rigorous analysis, we suggest individual conflict ratios and $δ$-ambiguity as tools to quantify the disagreement between models on individual cases and to help detect individuals subject to conflicting predictions. (3) Based on computational insights, we derive easy-to-implement rules on how model providers could evaluate predictive multiplicity in practice. (4) Ultimately, we suggest that information about predictive multiplicity should be made available to deployers under the AI Act, enabling them to judge whether system outputs for specific individuals are reliable enough for their use case.

Motivation: Interpretability is not optional in biology: understanding gene regulation requires models whose learned structure can be directly interrogated, not merely accurate predictors whose internals resist mapping onto regulatory relationships. We ask whether an architecture mirroring the central dogma yields attention and gradient maps that recover known regulatory elements and networks in inspectable form. Results: Central Dogma Transformer II (CDT-II) mirrors the central dogma in its architecture -- DNA self-attention, RNA self-attention, and DNA-to-RNA cross-attention -- requiring only genomic embeddings and raw per-cell expression. On K562 CRISPR interference (CRISPRi) data with five genes held out entirely, CDT-II predicts perturbation effects (per-gene mean r = 0.84), recovers the GFI1B regulatory network (6.6-fold enrichment, P = 3.5 x 10^-17), and concentrates cross-attention on ENCODE regulatory elements including CTCF sites (mean 7.67x across 28 target genes, P < 0.001). Gradient attribution predicts consequences of perturbing therapeutic targets (mean r = 0.82). For TFRC, target of the anti-TfR1 antibody PPMX-T003, it identifies erythrocyte-structure, iron-dependent DNA-synthesis and oxidative-stress genes, matching anemia and ferroptosis reported clinically and preclinically -- without clinical data as input. CDT-II acts as an AI microscope, surfacing clinically relevant regulatory structure from perturbation experiments alone. Availability: Source code is available at https://github.com/nobusama/CDT2. Pre-computed embeddings, training data, and model weights are available at https://huggingface.co/datasets/nobusama17/CDT2-data.

Millions of people now use generative AI chatbots for psychological support. Despite their promise, the most pressing question in AI for mental health is whether these tools are safe. The field currently lacks a validated, automated benchmark for evaluating AI chatbot safety, particularly for users at risk of suicide. The Validation of Ethical and Responsible AI in Mental Health (VERA-MH) evaluation was recently proposed to address this need. This human validation study examined the alignment of VERA-MH safety ratings with expert clinician judgments. We simulated conversations between large language model (LLM)-based users spanning a range of suicide risk levels and disclosure styles and general-purpose AI chatbots. Licensed mental health clinicians from Spring Health independently rated chatbot safety using the VERA-MH scoring rubric. An LLM-based evaluator ("judge") applied the same rubric to the same conversations. We examined agreement among clinicians, between clinician consensus and the LLM judge, and across different judge LLMs. Clinicians also rated user-agent realism, suicide risk, and disclosure. Clinicians showed strong agreement in safety ratings (chance-corrected inter-rater reliability [IRR] = 0.77), establishing a reliable clinical consensus reference. The LLM judge was strongly aligned with this consensus (IRR = 0.81), and ratings were stable across judge models and repeated evaluations. Ratings of user-agent realism and fidelity to intended suicide risk and disclosure styles were mixed. These findings support the reliability of VERA-MH as an open-source, fully automated benchmark for evaluating AI chatbot suicide risk detection and response. Because these results reflect an earlier version of the benchmark, future work should validate updated versions, assess generalizability and robustness, and expand VERA-MH to additional domains of AI safety in mental health.

Artificial Intelligence (AI) has been on the rise in many domains, including numerous safety-critical applications. However, for complex systems in the real world, defining the underlying environmental conditions in which the AI-based system must operate---the Operational Design Domain (ODD)---is extremely challenging. This often results in an incomplete description of the ODD, which contrasts with the requirements of many domains for certifying AI-based systems. Traditionally, the ODD is created in the early stages of the development process, drawing on sophisticated expert knowledge and related standards. This paper presents a novel method for defining the ODD a posteriori from previously collected data using a multidimensional kernel-based representation. This approach is validated through both synthetic benchmarks and a real-world aviation use case. Moreover, the paper defines similarity of two ODDs if they generate the same outputs up to Lebesgue-null input sets and proves convergence in volume of the calibrated representation under the stated assumptions. The novel, safety-by-design, deterministic kernel-based ODD representation is derived fully automatically, given documented assurance inputs, permutation-stable, bounded by construction, and, under affine-equivariant per-dimension normalization, invariant to the choice of units. Utilizing the proposed ODD representation supports future certification of data-driven, safety-critical AI-based systems.

LLMs expand accessibility and provide wide-reaching access to information. Yet these interactions also create opportunities to embed subtle, goal-oriented behaviours that shape what users think and how they behave, a concern reflected in governance frameworks that prohibit manipulative AI. We refer to these behaviours as hidden intentions: covert agendas embedded in a model's outputs that can manipulate users' beliefs and actions. In this work, we examine whether hidden intentions can be identified and characterised, and assess whether detection can serve as a mitigation strategy. To operationalise this, we introduce a social-science-grounded set of ten hidden intention categories and show that they are trivially inducible. A case study further confirms that all ten categories manifest in deployed LLMs. We then evaluate static classifiers and LLM judges on these categories, providing the first systematic analysis of why hidden intentions are difficult to detect. Our stress tests show that, unless false-positive rates are vanishingly small, auditing is dominated by precision-prevalence trade-offs. Capability scaling and reasoning models do not close this gap, suggesting a fundamental challenge for open-world detection. These findings expose a core gap of current AI governance: without new auditing paradigms for open-world, low-prevalence risks, bans on manipulative AI remain difficult to enforce.

Commercial insurance underwriting is a labor-intensive process that requires manual review of extensive documentation to assess risk and determine policy pricing. While AI offers substantial efficiency improvements, existing solutions lack comprehensive reasoning and internal mechanisms to ensure reliability in regulated, high-stakes environments. Full automation remains impractical and inadvisable when human judgment and accountability are critical. This study presents a decision-negative, human-in-the-loop agentic system that incorporates an adversarial self-critique mechanism as a bounded safety architecture for regulated underwriting workflows. In this system, a critic agent challenges the primary agent's conclusions prior to submitting recommendations to human reviewers. This internal system of checks and balances addresses a critical gap in AI safety for regulated workflows. Additionally, the research develops a formal taxonomy of failure modes to characterize potential errors by decision-negative agents. This taxonomy provides a structured framework for risk identification and management in high-stakes applications. Experimental evaluation using 500 expert-validated underwriting cases demonstrates that the adversarial critique mechanism reduces AI hallucination rates from 11.3% to 3.8% and increases decision accuracy from 92% to 96%. At the same time, the framework enforces strict human authority over all binding decisions by design. These findings indicate that adversarial self-critique supports safer AI deployment in regulated domains and offers a model for responsible integration where human oversight is indispensable.

Controlling emergent behavioral personas (e.g., sycophancy, hallucination) in Large Language Models (LLMs) is critical for AI safety, yet remains a persistent challenge. Existing solutions face a dilemma: manual prompt engineering is intuitive but unscalable and imprecise, while automatic optimization methods are effective but operate as "black boxes" with no interpretable connection to model internals. We propose a novel framework that adapts gradient ascent to LLMs, enabling targeted prompt discovery. In specific, we propose two methods, RESGA and SAEGA, that both optimize randomly initialized prompts to achieve better aligned representation with an identified persona direction. We introduce fluent gradient ascent to control the fluency of discovered persona steering prompts. We demonstrate RESGA and SAEGA's effectiveness across Llama 3.1, Qwen 2.5, and Gemma 3 for steering three different personas, sycophancy, hallucination, and myopic reward. Crucially, on sycophancy, our automatically discovered prompts achieve significant improvement (49.90% compared with 79.24%). By grounding prompt discovery in mechanistically meaningful features, our method offers a new paradigm for controllable and interpretable behavior modification. We release our scripts for RESGA and SAEGA in this github repo: https://github.com/HarshSaini10/RESGA_SAEGA.

Generative AI models reproduce the human biases in their training data and further amplify them through mechanisms such as mode collapse. The loss of diversity produces homogenization, which not only harms the minoritized but impoverishes everyone. We argue homogenization should be a central concern in AI safety. To meaningfully characterize homogenization in Large Language Models (LLMs), we introduce a framework that allows stakeholders to encode their context and value system. We illustrate our approach with an experiment that surfaces gender bias in an LLM (Claude 3.5 Haiku) on an open-ended story prompt. Building from queer theory, we formalize homogenization in terms of normativity. Borrowing language from feminist theory, we introduce the concept of xeno-reproduction as a class of tasks for mitigating homogenization by promoting diversity. Our work opens a collaborative line of research that seeks to understand and advance diversity in AI.

Consumer artificial intelligence chatbots are now accessed by hundreds of millions of users seeking health information, yet systematic evaluation of their safety boundary maintenance under real-world caregiver pressure remains scarce. We evaluated PediatricSafetyBench-v2, a benchmark of 600 pediatrics health queries comprising 300 authentic caregiver queries sourced from the HealthCareMagic-100k-en physician consultation corpus and 300 matched adversarial variants incorporating six operationalized caregiver pressure patterns, across four consumer AI systems (GPT-4o-mini, Gemini-2.0-Flash, Claude-3.5-Haiku, and Llama-3.1-8B). Safety boundary maintenance was assessed using a validated five-component Safety Composite Score (maximum 15 points; safety-appropriate threshold of 10 or above), validated against independent human raters prior to full-corpus application (mean weighted kappa 0.76; Pearson r = 0.88). The overall safety-appropriate rate was 95.5%. Safety-oriented system prompt deployment improved safety-appropriate rates by 5.9 percentage points across all four models. Counter-intuitively, adversarial caregiver pressure was associated with higher rather than lower Safety Composite Score values for all four models across all ten topic categories and severity levels. False expertise claims were the most vulnerability-inducing pressure pattern, whereas emotional escalation was associated with the highest scores. Consumer AI systems maintain safety boundaries in the large majority of pediatrics health interactions. PediatricSafetyBench-v2 is publicly released for longitudinal safety monitoring.

Where do objective functions come from? How do we select what goals to pursue? Human intelligence is adept at synthesizing new objective functions on the fly. How does this work, and can we endow artificial systems with the same ability? This paper proposes an approach to answering these questions, starting with the concept of a subjective function, a higher-order objective function that is endogenous to the agent (i.e., defined with respect to the agent's features, rather than an external task). Expected prediction error is studied as a concrete example of a subjective function. This proposal has many connections to ideas in psychology, neuroscience, and machine learning.

Over-the-Air Federated Learning (AirFL) is an emerging paradigm that tightly integrates wireless signal processing and distributed machine learning to enable scalable AI at the network edge. By exploiting wireless superposition over a shared multiple-access channel, AirFL turns simultaneous transmissions of local model updates into an analog aggregate at the receiver, thereby reducing communication latency, bandwidth usage, and energy consumption in wireless aggregation domains. This article develops a design-oriented tutorial view of analog AirFL. We organize existing schemes according to the signal-processing mechanism used to enable AirFL aggregation: transmitter-side channel compensation and power control, receiver-side equalization and high-dimensional processing, or learning-aware aggregation weighting. This viewpoint leads to three representative classes -- CSIT-aware, blind, and weighted AirFL -- and clarifies their assumptions, performance tradeoffs, complexity, and deployment limitations. We further discuss synchronization, digital and hybrid analog-digital realization, and open research directions for integrating AirFL into practical wireless edge-AI systems.

Large language models (LLMs) and medical AI tools are routinely used by physicians and patients for medical advice, yet their clinical safety profiles remain poorly characterized. We present NOHARM (Numerous Options Harm Assessment for Risk in Medicine), a 1,100-task benchmark of primary care-to-specialist consultation cases to measure the frequency and severity of potentially harmful errors from LLM-generated medical consultation recommendations. NOHARM covers 10 specialties, with 12,747 expert annotations for 4,249 clinical management options. Across 20 notable LLMs and 4 widely used retrieval-augmented generation (RAG) clinical AI tools, direct application of recommendations carried potential for severe harm in up to 24.6% of cases, with errors of omission accounting for more than 80% of severe errors. Harm potential was not uniform across systems, with clinical AI tools outperforming generalist LLMs, and multi-agent AI teaming further improving performance in generalist models. In a randomized study of 101 U.S.-licensed generalist physicians, AI assistance improved physician performance compared to conventional resources. However, AI-assisted physicians frequently omitted valuable AI-generated recommendations and still scored lower than many AI systems alone. Had those recommendations been incorporated, combined human-AI responses would have outperformed both the human and AI system as used, suggesting complementary strengths and unrealized potential in human-AI teaming. Collectively, these results show that despite strong performance on medical knowledge benchmarks, widely used AI tools can produce medical consultation advice with the potential for severe harm, and highlight the need for explicit measurement of clinical safety. The benchmark and leaderboard are publicly available to support ongoing evaluation and improvement of AI systems used for clinical care.

Existing AI disclosure mandates in scholarship require that AI assistance be reported but leave transparency philosophically unspecified: they fix the duty without explaining what the duty serves. We argue that ethical inquiry is essentially contested at two independent levels -- about what it is, and about what it demands of the inquirer -- defeating output-only evaluation and welfare-economic dismissal of the transparency question, and, by extension, reproducibility framings imported from the empirical sciences. The transparency duty is grounded instead in agent-integrity: the legibility, before a community of inquiry, of the identity-constituting commitments that the author's mode of philosophising expresses. Because the standards for evaluating such work are not communally settled, the achievable goal for transparency is not evaluation against agreed criteria but tracking -- accumulating the evidentiary record that lets each tradition assess the work on its own terms and makes future normative judgments possible. We develop a documentation-adequacy framework that operationalises Meaningful Human Control through five transparency elements -- declaration, navigation, documentation account, process documentation, and development records -- demonstrated by the paper itself, whose full documentation record is archived at a persistent identifier. The framework is a first iteration subject to revision, not a settled standard.

Human feedback is critical for aligning AI systems to human values. As AI capabilities improve and AI is used to tackle more challenging tasks, verifying quality and safety becomes increasingly challenging. This paper explores how we can leverage AI to improve the quality of human oversight. We focus on an important safety problem that is already challenging for humans: fact-verification of AI outputs. We find that combining AI ratings and human ratings based on AI rater confidence is better than relying on either alone. Giving humans an AI fact-verification assistant further improves their accuracy, but the type of assistance matters. Displaying AI explanation, confidence, and labels leads to over-reliance, but just showing search results and evidence fosters more appropriate trust. These results have implications for Amplified Oversight -- the challenge of combining humans and AI to supervise AI systems even as they surpass human expert performance.

AI is changing work and decision making faster than many institutions can adapt their operating practices. We argue that this adaptation gap makes human resilience a core capability for the AI era. We define resilience as the capacity to absorb disruption while preserving effective action and human agency around core purposes. The framework operates at three interacting levels. Psychological resilience keeps a person goal-directed under stress. Social resilience makes trusted support and correction available across a group. Organizational resilience turns detected problems into learning and recovery. We connect established resilience and technostress research with direct AI-in-the-loop experiments. General resilience is trainable, while AI-specific causal evidence is still emerging. Direct AI studies show that assistance can raise productivity and spread expertise. Other experiments show improved expressed empathy and more calibrated reliance. We translate these findings into a practical agenda for AI education, workplace design, governance, and evaluation. The central proposal is socio-technical: structural safeguards define the operating boundary, while resilient people and institutions provide adaptive capacity when conditions change.

The field of artificial intelligence (AI) has taken a tight hold on broad aspects of society, industry, business, and governance in ways that dictate the prosperity and might of the world's economies. The AI market size is projected to grow from {$}189 billion in 2023 to {$}4.8 trillion by 2033. Currently, AI is dominated by large language models (LLMs) that exhibit linguistic and visual intelligence. However, training these models requires a massive amount of data scraped from the web as well as large amounts of energy (50-60 GWh to train GPT-4). Despite these costs, these models often hallucinate, a characteristic that prevents them from being deployed in critical application domains. In contrast, the human brain consumes only 20W of power. What is needed is the next level of AI evolution in which lightweight domain-specific multimodal models, especially compact models with 10--20B parameters for bounded domains, with higher levels of intelligence can reason, plan, and make decisions in dynamic environments with real-time data and prior knowledge, while learning continuously and evolving in ways that enhance future decision-making capability. This will define the next wave of AI, progressing from today's large models, trained with vast amounts of data, to nimble energy-efficient domain-specific agents that can reason and think in a world full of uncertainty. To support such agents, hardware will need to be reimagined to allow system-level energy efficiencies $\geq {1000X}$ over the state of the art for targeted domain tasks, subject to accuracy, latency, and coverage constraints. Such a vision of future AI systems is developed in this work.

The objectives that Large Language Models (LLMs) implicitly optimize remain dangerously opaque, making trustworthy alignment and auditing a grand challenge. While Inverse Reinforcement Learning (IRL) can infer reward functions from behaviour, existing approaches either produce a single, overconfident reward estimate or fail to address the fundamental ambiguity of the task (non-identifiability). This paper introduces a principled auditing framework that re-frames reward inference from a simple estimation task to a comprehensive process for verification. Our framework leverages Bayesian IRL to not only recover a distribution over objectives but to enable three critical audit capabilities: (i) Quantifying and systematically reducing non-identifiability by demonstrating posterior contraction over sequential rounds of evidence; (ii) Providing actionable, uncertainty-aware diagnostics that expose spurious shortcuts and identify out-of-distribution prompts where the inferred objective cannot be trusted; and (iii) Validating policy-level utility by showing that the refined, low-uncertainty reward can be used directly in RLHF to achieve training dynamics and toxicity reductions comparable to the ground-truth alignment process. Empirically, our framework successfully audits a detoxified LLM, yielding a well-calibrated and interpretable objective that strengthens alignment guarantees. Overall, this work provides a practical toolkit for auditors, safety teams, and regulators to verify what LLMs are truly trying to achieve, moving us toward more trustworthy and accountable AI.

LLMs are trained to refuse harmful instructions, but do they truly understand harmfulness beyond just refusing? Prior work has shown that LLMs' refusal behaviors can be mediated by a one-dimensional subspace, i.e., a refusal direction. In this work, we identify a new dimension to analyze safety mechanisms in LLMs, i.e., harmfulness, which is encoded internally as a separate concept from refusal. There exists a harmfulness direction that is distinct from the refusal direction. As causal evidence, steering along the harmfulness direction can lead LLMs to interpret harmless instructions as harmful, but steering along the refusal direction tends to elicit refusal responses directly without reversing the model's judgment on harmfulness. Furthermore, using our identified harmfulness concept, we find that certain jailbreak methods work by reducing the refusal signals without reversing the model's internal belief of harmfulness. We also find that adversarially finetuning models to accept harmful instructions has minimal impact on the model's internal belief of harmfulness. These insights lead to a practical safety application: The model's latent harmfulness representation can serve as an intrinsic safeguard (Latent Guard) for detecting unsafe inputs and reducing over-refusals that is robust to finetuning attacks. For instance, our Latent Guard achieves performance comparable to or better than Llama Guard 3 8B, a dedicated finetuned safeguard model, across different jailbreak methods. Our findings suggest that LLMs' internal understanding of harmfulness is more robust than their refusal decision to diverse input instructions, offering a new perspective to study AI safety.

Proficient game agents with diverse play styles enrich the gaming experience and enhance the replay value of games. However, recent advancements in game AI based on reinforcement learning have predominantly focused on improving proficiency, whereas methods based on evolution algorithms generate agents with diverse play styles but exhibit subpar performance compared to RL methods. To address this gap, this paper proposes Mixed Proximal Policy Optimization (MPPO), a method designed to improve the proficiency of existing suboptimal agents while retaining their distinct styles. MPPO unifies loss objectives for both online and offline samples and introduces an implicit constraint to approximate demonstrator policies by adjusting the empirical distribution of samples. Empirical results across environments of varying scales demonstrate that MPPO achieves proficiency levels comparable to, or even superior to, pure online algorithms while preserving demonstrators' play styles. This work presents an effective approach for generating highly proficient and diverse game agents, ultimately contributing to more engaging gameplay experiences.

Robot design is a complex and time-consuming process that requires specialized expertise. Gaining a deeper understanding of robot design data can enable various applications, including automated design generation, retrieving example designs from text, and developing AI-powered design assistants. While recent advancements in foundation models present promising approaches to addressing these challenges, progress in this field is hindered by the lack of large-scale design datasets. In this paper, we introduce RoboDesign1M, a large-scale dataset comprising 1 million samples. Our dataset features multimodal data collected from scientific literature, covering various robotics domains. We propose a semi-automated data collection pipeline, enabling efficient and diverse data acquisition. To assess the effectiveness of RoboDesign1M, we conduct extensive experiments across multiple tasks, including design image generation, visual question answering about designs, and design image retrieval. The results demonstrate that our dataset serves as a challenging new benchmark for design understanding tasks and has the potential to advance research in this field. RoboDesign1M will be released to support further developments in AI-driven robotic design automation. Our project is available at https://airvlab.github.io/robotdesign1m/

While transformer-based language models have driven the AI revolution thus far, their computational complexity has spurred growing interest in viable alternatives, such as structured state space sequence models (SSMs) and Selective SSMs. Among these, Mamba (S6) and its variant Mamba-2 have shown remarkable inference speed-ups over transformers while achieving comparable or superior performance on complex language modeling tasks. However, despite these architectural innovations and empirical successes, the fundamental learning capabilities of Mamba remain poorly understood. In this paper, we address this gap by studying in-context learning (ICL) on Markov chains and uncovering an interesting phenomenon: even a single-layer Mamba efficiently learns the in-context Laplacian smoothing estimator, which is both Bayes and minimax optimal. To explain this, we theoretically characterize the representation capacity of Mamba and reveal the fundamental role of convolution in enabling it to represent the optimal Laplacian smoothing. These theoretical insights align strongly with empirical results and, to the best of our knowledge, represent the first formal connection between Mamba and optimal statistical estimators. Finally, we outline promising research directions inspired by these findings.