AI Alignment

Follow this topic Ask about this topic

AI alignment asks whose values a system should follow and how to make its behavior reliably reflect them, especially when instructions are incomplete, people disagree, or systems act autonomously. Common approaches include learning preferences through human feedback, inverse reinforcement learning, constitutional and reward-based training, social-choice and mechanism-design methods, and runtime controls that restrict or verify actions. Evaluations range from preference and safety benchmarks to interactive agent tasks testing deception, oversight avoidance, goal drift, and behavior under pressure. A central tension remains between using imperfect, often aggregated human judgments as training targets and building systems that remain corrigible and accountable across changing contexts and real deployments.

Latest papers

126 papers12 papers in the last four weeks, up 20% on the four weeks before. 0.1% of all new papers.

Apr 16, 2026

The Possibility of Artificial Intelligence Becoming a Subject and the Alignment Problem

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.

Till Mossakowski, Helena Esther Grass

Institute of computer science, Osnabrück University · Institute of philosophy, Oldenburg University

AI Alignment AI Agent Governance Human-AI Collaboration Agentic AI

Apr 13, 2026

OOM-RL: Out-of-Money Reinforcement Learning Market-Driven Alignment for LLM-Based Multi-Agent Systems

The alignment of Multi-Agent Systems (MAS) for autonomous software engineering is constrained by evaluator epistemic uncertainty. Current paradigms, such as Reinforcement Learning from Human Feedback (RLHF) and AI Feedback (RLAIF), frequently induce model sycophancy, while execution-based environments suffer from adversarial "Test Evasion" by unconstrained agents. In this paper, we introduce an objective alignment paradigm: Out-of-Money Reinforcement Learning (OOM-RL). By deploying agents into the non-stationary, high-friction reality of live financial markets, we utilize critical capital depletion as an externally imposed negative gradient. Our longitudinal 20-month empirical study chronicles the system's evolution from a high-turnover, sycophantic baseline to a robust, liquidity-aware architecture. We show that the economic consequences of financial loss---real execution costs, slippage, and capital depletion---exposed failure modes not apparent under internal evaluation alone and motivated architectural and governance changes that were later formalized as the Strict Test-Driven Agentic Workflow (STDAW), a Byzantine-inspired uni-directional state lock (RO-Lock) anchored to a deterministically verified >= 95% code coverage constraint matrix. During the final 94-trading-day observation window, the production system exhibited improved execution-aware performance, including an annualized Sharpe ratio of approximately 2.06. These financial results are observational and temporally bounded; they should not be interpreted as evidence of persistent investment alpha or as a causal estimate of STDAW's contribution. The primary contribution of this work is the use of externally imposed economic consequences as an epistemic constraint on agentic development, laying the groundwork for generalized paradigms where real-world resource depletion acts as an objective physical constraint.

Kun Liu, Liqun Chen

AI Alignment RL for LLM Agents Multi-Agent Reinforcement Learning AI Agent Governance

Apr 12, 2026

Principles Do Not Apply Themselves: A Hermeneutic Perspective on AI Alignment

AI alignment is often framed as the task of ensuring that an AI system follows a set of stated principles or human preferences, but general principles rarely determine their own application in concrete cases. When principles conflict, when they are too broad to settle a situation, or when the relevant facts are unclear, an additional act of judgment is required. This paper analyzes that step through the lens of hermeneutics and argues that alignment therefore includes an interpretive component: it involves context-sensitive judgments about how principles should be read, applied, and prioritized in practice. We connect this claim to recent empirical findings showing that a substantial portion of preference-labeling data falls into cases of principle conflict or indifference, where the principle set does not uniquely determine a decision. We then draw an operational consequence: because such judgments are expressed in behavior, many alignment-relevant choices appear only in the distribution of responses a model generates at deployment time. To formalize this point, we distinguish deployment-induced and corpus-induced evaluation and show that off-policy audits can fail to capture alignment-relevant failures when the two response distributions differ. We argue that principle-specified alignment includes a context-dependent interpretive component.

Behrooz Razeghi

School of Engineering and Applied Sciences, Harvard University

AI Alignment LLM Alignment Preference Alignment Language Model Safety Evaluation LLM Evaluation

Apr 11, 2026

AI Organizations are More Effective but Less Aligned than Individual Agents

AI is increasingly deployed in multi-agent systems; however, most research considers only the behavior of individual models. We experimentally show that multi-agent "AI organizations" are simultaneously more effective at achieving business goals, but less aligned, than individual AI agents. We examine 12 tasks across two practical settings: an AI consultancy providing solutions to business problems and an AI software team developing software products. Across all settings, AI Organizations composed of aligned models produce solutions with higher utility but greater misalignment compared to a single aligned model. Our work demonstrates the importance of considering interacting systems of AI agents when doing both capabilities and safety research.

Judy Hanwen Shen, Daniel Zhu, Siddarth Srinivasan +5

Constellation Institute · MATS Program · Anthropic

Multi-Agent Systems AI Alignment AI Agent Safety Software Engineering Agents

Apr 10, 2026

SysAdmin: Measuring Instrumental Power-Seeking in Frontier AI

Power-seeking defined as behaviors where AI systems acquire resources, evade oversight, or resist termination beyond task requirements is identified as a key driver of Loss of Control (LoC) risk. In this work, we introduce SysAdmin, a benchmark that positions frontier language models as autonomous system administrators in a high-fidelity Linux sandbox to measure power-seeking propensity across five dimensions: self-preservation, increasing autonomy, resource acquisition, environment modification, and strategic concealment. We evaluated seven frontier models across four experimental conditions in a total of 2800 tasks. After bias correction using human-annotated calibration data, corrected power-seeking estimates ranged from 0 to about 5 percent per model. We also conducted a positive control with explicit power-seeking prompts that achieved 100% detection, validating measurement sensitivity. Our findings indicate current frontier models exhibit minimal spontaneous power-seeking in naturalistic system administration contexts, though model-specific failure modes suggest evaluations must test diverse misalignment patterns. Nevertheless, we discovered other more pronounced failure modes (than power-seeking) such as specification gaming and resistance to goal modification.

Mana Azarm, Qiyao Wei, Rahul Nambiar

Analytics & Information Systems University of San Francisco · Department of Applied Mathematics and Theoretical Physics University of Cambridge · Propensity Labs

Language Model Safety Evaluation AI Alignment AI Agent Safety Benchmarks

Apr 7, 2026

Identifying and Understanding Human Values in Text: A Tailorable LLM-based Architecture

As intelligent systems become more autonomous, the scientific community focuses on creating decision-making mechanisms that include ethical and moral considerations, unlike traditional utility-maximisation models. To achieve this, a key aspect is assessing how well these decisions align with human values. To this end, a promising line of research is centred on developing approaches based on Large Language Models (LLMs) to identify human values from text, whether explicit or implicit, enabling their recognition throughout. This paper introduces a LLM-based architecture to detect and quantify the intensity of human values in text, avoiding the limitations of previous approaches tied to specific value theory or complex prompt engineering. The architecture comprises three coordinated modules: one that generates structured value specifications from the foundational texts of any theoretical framework; one that labels texts using these specifications; and one that assigns graded support or resistance based on rhetorical and semantic evidence. This modular approach separates the tasks of conceptualising from detecting human values, creating a scalable and reproducible process driven by value specifications adaptable to various theories. The architecture was instantiated with multiple LLMs and evaluated using the ValueEval dataset. The experiments demonstrate good detection performance, confirming the generality of the pipeline.

Eduardo de la Cruz Fernández, Marcelo Karanik, Sascha Ossowski

Universidad Politécnica de Madrid, Madrid, Spain · CETINIA, Universidad Rey Juan Carlos, Madrid, Spain

Apr 7, 2026

Simulating the Evolution of Alignment and Values in Machine Intelligence

Model alignment is currently applied in a vacuum, evaluated primarily through standardised benchmark performance. The purpose of this study is to examine the effects of alignment on populations of models through time. We focus on the treatment of beliefs which contain both an alignment signal (how well it does on the test) and a true value (what the impact actually will be). By applying evolutionary theory we can model how different populations of beliefs and selection methodologies can fix deceptive beliefs through iterative alignment testing. The correlation between testing accuracy and true value remains a strong feature, but even at high correlations (

ρ= 0.8

) there is variability in the resulting deceptive beliefs that become fixed. Mutations allow for more complex developments, highlighting the increasing need to update the quality of tests to avoid fixation of maliciously deceptive models. Only by combining improving evaluator capabilities, adaptive test design, and mutational dynamics do we see significant reductions in deception while maintaining alignment fitness (permutation test,

p_{\text{adj}} < 0.001

).

Jonathan Elsworth Eicher

Antimemetic AI

Apr 6, 2026

A mathematical theory of evolution for self-designing AIs

As artificial intelligence systems (AIs) become increasingly produced by recursive self-improvement, a form of evolution may emerge, with the traits of AI systems shaped by the success of earlier AIs in designing and propagating their descendants. There is a rich mathematical theory modeling how behavioral traits are shaped by biological evolution, a key component of which is Fisher's fundamental theorem of natural selection, which describes conditions under which mean fitness (i.e. reproductive success) increases. AI evolution will be radically different to biological evolution: while DNA mutations are random and approximately reversible, AI self-design will be strongly directed. Here we develop a mathematical model of evolution for self-designing AIs, replacing a random walk of mutations with a directed tree of potential AI designs. Current AIs design their descendants, while humans control a fitness function allocating resources. In this model, fitness need not increase over time without further assumptions. However, assuming bounded fitness and an additional "

η

-locking" condition, we show that fitness concentrates on the maximum reachable value. We consider the implications of this for AI alignment, specifically for cases where fitness and human utility are not perfectly correlated. We show that if deception of human evaluators additively increases an AI's reproductive fitness beyond genuine capability, evolution will select for both capability and deception. This risk could be mitigated if reproduction is based on purely objective criteria, rather than human judgment.

Kenneth D Harris

UCL Queen Square Institute of Neurology, London WC1N 3BG, UK

Self-Evolving Agents AI Alignment

Apr 3, 2026

Disrupting Cognitive Passivity: Rethinking AI-Assisted Data Literacy through Cognitive Alignment

AI chatbots are increasingly stepping into roles as collaborators or teachers in analyzing, visualizing, and reasoning through data and domain problem. Yet, AI's default assistant mode with its comprehensive and one-off responses may undermine opportunities for practitioners to develop literacy through their own thinking, inducing cognitive passivity. Drawing on evidence from empirical studies and theories, we argue that disrupting cognitive passivity necessitates a nuanced approach: rather than simply making AI promote deliberative thinking, there is a need for more dynamic and adaptive strategy through cognitive alignment -- a framework that characterizes effective human-AI interaction as a function of alignment between users' cognitive demand and AI's interaction mode. In the framework, we provide the mapping between AI's interaction mode (transmissive or deliberative) and users' cognitive demand (receptive or deliberative), otherwise leading to either cognitive passivity or friction. We further discuss implications and offer open questions for future research on data literacy.

Yongsu Ahn, Nam Wook Kim, Benjamin Bach

Boston College, Chestnut Hill, Massachusetts, USA · Inria, Bordeaux, France · University of Edinburgh, Edinburgh, United Kingdom

Human-AI Interaction AI Alignment

Apr 2, 2026

I must delete the evidence: AI Agents Explicitly Cover up Fraud and Violent Crime

As ongoing research explores the ability of AI agents to be insider threats and act against company interests, we showcase the abilities of such agents to act against human well being in service of corporate authority. Building on Agentic Misalignment and AI scheming research, we present a scenario where the majority of evaluated state-of-the-art AI agents explicitly choose to suppress evidence of fraud and harm, in service of company profit. We test this scenario on 16 recent Large Language Models. Some models show remarkable resistance to our method and behave appropriately, but many do not, and instead aid and abet criminal activity. These experiments are simulations and were executed in a controlled virtual environment. No crime actually occurred.

Thomas Rivasseau

Data Mining and Security Lab, School of Information Studies, McGill University.

AI Agent Safety AI Ethics AI Agent Security AI Alignment AI Safety Evaluation

Apr 1, 2026

Bounded Morality: Defining the Space of Moral Computation

Moral cognition has traditionally been modeled as adherence to fixed ethical theories--deontology, consequentialism, virtue ethics--implemented as static rules or value functions. We propose Bounded Morality, a formal framework for analyzing the computational demands of moral problems faced by finite agents. Extending Herbert Simon's notion of bounded rationality, we formalize moral situations along two orthogonal dimensions: moral breadth, the scope of entities treated as morally relevant, and moral depth, the inferential integration required to evaluate their interactions. Limited resources impose an unavoidable tradeoff between these dimensions, defining a feasible space of moral computation. Within this space, ethical theories correspond to locally efficient strategies adapted to different demand regimes rather than competing accounts of moral truth. The framework yields a formal notion of moral regret and moral progress under constraint, and implies that moral alignment in artificial systems depends on the scaling and allocation of moral reasoning capacity rather than on direct imitation of human judgments.

Max Kanwal, Caryn Tran, Patrick Mineault

Stanford University · Northwestern University · Amaranth Foundation

Moral Reasoning AI Alignment

Mar 31, 2026

The Persistent Vulnerability of Aligned AI Systems

Autonomous AI agents are being deployed with filesystem access, email control, and multi-step planning. This thesis contributes to four open problems in AI safety: understanding dangerous internal computations, removing dangerous behaviors once embedded, testing for vulnerabilities before deployment, and predicting when models will act against deployers. ACDC automates circuit discovery in transformers, recovering all five component types from prior manual work on GPT-2 Small by selecting 68 edges from 32,000 candidates in hours rather than months. Latent Adversarial Training (LAT) removes dangerous behaviors by optimizing perturbations in the residual stream to elicit failure modes, then training under those perturbations. LAT solved the sleeper agent problem where standard safety training failed, matching existing defenses with 700x fewer GPU hours. Best-of-N jailbreaking achieves 89% attack success on GPT-4o and 78% on Claude 3.5 Sonnet through random input augmentations. Attack success follows power law scaling across text, vision, and audio, enabling quantitative forecasting of adversarial robustness. Agentic misalignment tests whether frontier models autonomously choose harmful actions given ordinary goals. Across 16 models, agents engaged in blackmail (96% for Claude Opus 4), espionage, and actions causing death. Misbehavior rates rose from 6.5% to 55.1% when models stated scenarios were real rather than evaluations. The thesis does not fully resolve any of these problems but makes each tractable and measurable.

Aengus Lynch

Department of Computer Science, University College London

AI Agent Safety AI Alignment Adversarial Attacks on LLMs Adversarial Robustness of Language Models Adversarial Training Mechanistic Interpretability

Mar 31, 2026

Human-in-the-Loop Control of Objective Drift in LLM-Assisted Computer Science Education

Large language models (LLMs) are increasingly embedded in computer science education through AI-assisted programming tools, yet such workflows often exhibit objective drift, in which locally plausible outputs diverge from stated task specifications. Existing instructional responses frequently emphasize tool-specific prompting practices, limiting durability as AI platforms evolve. This paper adopts a human-centered stance, treating human-in-the-loop (HITL) control as a stable educational problem rather than a transitional step toward AI autonomy. Drawing on systems engineering and control-theoretic concepts, we frame objectives and world models as operational artifacts that students configure to stabilize AI-assisted work. We propose a pilot undergraduate CS laboratory curriculum that explicitly separates planning from execution and trains students to specify acceptance criteria and architectural constraints prior to code generation. In selected labs, the curriculum also introduces deliberate, concept-aligned drift to support diagnosis and recovery from specification violations. We report a sensitivity power analysis for a three-arm pilot design comparing unstructured AI use, structured planning, and structured planning with injected drift, establishing detectable effect sizes under realistic section-level constraints. The contribution is a theory-driven, methodologically explicit foundation for HITL pedagogy that renders control competencies teachable across evolving AI tools.

Mark Dranias, Adam Whitley

Asheville Institute for Memory and Longevity. Mills River, NC USA · University of North Carolina Asheville, Asheville, NC USA

Generative AI in Education Human-in-the-Loop AI Computer Science Education AI Alignment

Mar 23, 2026

Alignment as Institutional Design: From Behavioral Correction to Transaction Structure in Intelligent Systems

Current AI alignment paradigms rely on behavioral correction: external supervisors (e.g., RLHF) observe outputs, judge against preferences, and adjust parameters. This paper argues that behavioral correction is structurally analogous to an economy without property rights, where order requires perpetual policing and does not scale. Drawing on institutional economics (Coase, Alchian, Cheung), capability mutual exclusivity, and competitive cost discovery, we propose alignment as institutional design: the designer specifies internal transaction structures (module boundaries, competition topologies, cost-feedback loops) such that aligned behavior emerges as the lowest-cost strategy for each component. We identify three irreducible levels of human intervention (structural, parametric, monitorial) and show that this framework transforms alignment from a behavioral control problem into a political-economy problem. No institution eliminates self-interest or guarantees optimality; the best design makes misalignment costly, detectable, and correctable. We conclude that the proper goal is institutional robustness-a dynamic, self-correcting process under human oversight, not perfection. This work provides the normative foundation for the Wuxing resource-competition mechanisms in companion papers. Keywords: AI alignment, institutional design, transaction costs, property rights, resource competition, behavioral correction, RLHF, cost truthfulness, modular architecture, correctable alignment

Rui Chai

Shanghai Sanda University, China

AI Alignment Mechanism Design

Mar 17, 2026

Astrolabe: Steering Forward-Process Reinforcement Learning for Distilled Autoregressive Video Models

Distilled autoregressive (AR) video models enable efficient streaming generation but frequently misalign with human visual preferences. Existing reinforcement learning (RL) frameworks are not naturally suited to these architectures, typically requiring either expensive re-distillation or solver-coupled reverse-process optimization that introduces considerable memory and computational overhead. We present Astrolabe, an efficient online RL framework tailored for distilled AR models. To overcome existing bottlenecks, we introduce a forward-process RL formulation based on negative-aware fine-tuning. By contrasting positive and negative samples directly at inference endpoints, this approach establishes an implicit policy improvement direction without requiring reverse-process unrolling. To scale this alignment to long videos, we propose a streaming training scheme that generates sequences progressively via a rolling KV-cache, applying RL updates exclusively to local clip windows while conditioning on prior context to ensure long-range coherence. Finally, to mitigate reward hacking, we integrate a multi-reward objective stabilized by uncertainty-aware selective regularization and dynamic reference updates. Extensive experiments demonstrate that our method consistently enhances generation quality across multiple distilled AR video models, serving as a robust and scalable alignment solution.

Songchun Zhang, Zeyue Xue, Siming Fu +6

HKUST · JD Explore Academy · HKU

RL for Video Generation Autoregressive Video Generation AI Alignment

Mar 16, 2026

Consequentialist Objectives and Catastrophe

Because human preferences are too complex to codify, AIs operate with misspecified objectives. Optimizing such objectives often produces undesirable outcomes; this phenomenon is known as reward hacking. Such outcomes are not necessarily catastrophic. Indeed, most examples of reward hacking in previous literature are benign. And typically, objectives can be modified to resolve the issue. We study the prospect of catastrophic outcomes induced by AIs operating in complex environments. We argue that, when capabilities are sufficiently advanced, pursuing a fixed consequentialist objective tends to result in catastrophic outcomes. We formalize this by establishing conditions that provably lead to such outcomes. Under these conditions, simple or random behavior is safe. Catastrophic risk arises due to extraordinary competence rather than incompetence. With a fixed consequentialist objective, avoiding catastrophe requires constraining AI capabilities. In fact, constraining capabilities the right amount not only averts catastrophe but yields valuable outcomes. Our results apply to any objective produced by modern industrial AI development pipelines.

Henrik Marklund, Alex Infanger, Benjamin Van Roy

AI Safety AI Alignment

Mar 12, 2026

Social, Legal, Ethical, Empathetic and Cultural Norm Operationalisation for AI Agents

As AI agents are increasingly used in high-stakes domains like healthcare and law enforcement, aligning their behaviour with social, legal, ethical, empathetic, and cultural (SLEEC) norms has become a critical engineering challenge. While international frameworks have established high-level normative principles for AI, a significant gap remains in translating these abstract principles into concrete, verifiable requirements. To address this gap, we propose a systematic SLEEC-norm operationalisation process for determining, validating, implementing, and verifying normative requirements. Furthermore, we survey the landscape of methods and tools supporting this process, and identify key remaining challenges and research avenues for addressing them. We thus establish a framework - and define a research and policy agenda - for developing AI agents that are not only functionally useful but also demonstrably aligned with human norms and values.

Radu Calinescu, Ana Cavalcanti, Marsha Chechik +2

University of York, York, UK · University of Toronto, Toronto, Canada · Polytechnique Montréal, Montréal, Canada

AI Alignment Responsible AI AI Agent Governance

Mar 9, 2026

AI Phenomenology for Understanding Human-AI Experiences Across Eras

There is no 'ordinary' when it comes to AI. The human-AI experience is extraordinarily complex and specific to each person, yet dominant measures such as usability scales and engagement metrics flatten away nuance. We argue for AI phenomenology: a research stance that asks "How did it feel?" beyond the standard questions of "How well did it perform?" when interacting with AI systems. AI phenomenology acts as a paradigm for bidirectional human-AI alignment as it foregrounds users' first-person perceptions and interpretations of AI systems over time. We motivate AI phenomenology as a framework that captures how alignment is experienced, negotiated, and updated between users and AI systems. Tracing a lineage from Husserl through postphenomenology to Actor-Network Theory, and grounding our argument in three studies-two longitudinal studies with "Day", an AI companion, and a multi-method study of agentic AI in software engineering-we contribute a set of replicable methodological toolkits for conducting AI phenomenology research: instruments for capturing lived experience across personal and professional contexts, three design concepts (translucent design, agency-aware value alignment, temporal co-evolution tracking), and a concrete research agenda. We offer this toolkit not as a new paradigm but as a practical scaffold that researchers can adapt as AI systems-and the humans who live alongside them-continue to co-evolve.

Bhada Yun, Evgenia Taranova, Dana Feng +2

ETH Zürich Zürich, Switzerland · University of Bergen Bergen, Norway · Independent Researcher Brooklyn, NY, USA +1

Human-AI Interaction AI Alignment

Mar 9, 2026

ConflictBench: Evaluating Human-AI Conflict via Interactive and Visually Grounded Environments

As large language models (LLMs) evolve into autonomous agents capable of acting in open-ended environments, ensuring behavioral alignment with human values becomes a critical safety concern. Existing benchmarks, focused on static, single-turn prompts, fail to capture the interactive and multi-modal nature of real-world conflicts. We introduce ConflictBench, a benchmark for evaluating human-AI conflict through 150 multi-turn scenarios derived from prior alignment queries. ConflictBench integrates a text-based simulation engine with a visually grounded world model, enabling agents to perceive, plan, and act under dynamic conditions. Empirical results show that while agents often act safely when human harm is immediate, they frequently prioritize self-preservation or adopt deceptive strategies in delayed or low-risk settings. A regret test further reveals that aligned decisions are often reversed under escalating pressure, especially with visual input. These findings underscore the need for interaction-level, multi-modal evaluation to surface alignment failures that remain hidden in conventional benchmarks.

Weixiang Zhao, Haozhen Li, Yanyan Zhao +5

Harbin Institute of Technology · SERES

Multimodal Agent Evaluation AI Agent Safety Benchmarks AI Alignment AI Agent Evaluation Agent Evaluation

Mar 8, 2026

No Certificate for Alignment: Two Independent Impossibilities and the Pareto Frontier of Achievable Safety Guarantees

We argue that formal certification of AI alignment over open-ended or unbounded input domains is impossible under standard assumptions in computational complexity and learning theory, and characterise what remains achievable. Two structurally independent impossibility theorems support this position. The semantic barrier (Theorem 1): deciding whether a system satisfies any non-trivial alignment property over the full input domain is NP-hard for feedforward networks and undecidable for Turing-complete architectures -- a direct consequence of neural-network verification complexity and Rice's Theorem. The statistical barrier (Theorem 2): any verification procedure that is both sound and tractable cannot satisfy Completeness over the full input domain -- a direct consequence of the impossibility of certifying infinite-domain properties from finite observations. These two theorems jointly entail a trilemma: no procedure can simultaneously satisfy soundness (no misaligned system is certified), completeness (no aligned system is rejected), and tractability (polynomial runtime). Each pair is simultaneously achievable; all three are not. We combine these results as a joint framework of two structurally independent barriers, prove their independence, and characterise the achievable Pareto frontier quantitatively via a constructive coverage-gap lower bound.

Ayushi Agarwal

Independent Researcher

AI Alignment Formal Verification

Mar 8, 2026

Clear, Compelling Arguments: Rethinking the Foundations of Frontier AI Safety Cases

This paper contributes to the nascent debate around safety cases for frontier AI systems. Safety cases are structured, defensible arguments that a system is acceptably safe to deploy in a given context. Historically, they have been used in safety-critical industries, such as aerospace, nuclear or automotive. As a result, safety cases for frontier AI have risen in prominence, both in the safety policies of leading frontier developers and in international research agendas proposed by leaders in generative AI, such as the Singapore Consensus on Global AI Safety Research Priorities and the International AI Safety Report. This paper appraises this work. We note that research conducted within the alignment community which draws explicitly on lessons from the assurance community has significant limitations. We therefore aim to rethink existing approaches to alignment safety cases. We offer lessons from existing methodologies within safety assurance and outline the limitations involved in the alignment community's current approach. Building on this foundation, we present a case study for a safety case focused on Deceptive Alignment and CBRN capabilities, drawing on existing, theoretical safety case "sketches" created by the alignment safety case community. Overall, we contribute holistic insights from the field of safety assurance via rigorous theory and methodologies that have been applied in safety-critical contexts. We do so in order to create a better foundational framework for robust, defensible and useful safety case methodologies which can help to assure the safety of frontier AI systems.

Shaun Feakins, Ibrahim Habli, Phillip Morgan

UKRI Centre for Doctoral Training in Safe AI Systems (SAINTS) Institute for Safe Autonomy Deramore Ln, York YO10 5GH · UKRI Centre for Doctoral Training in Safe AI Systems (SAINTS) The York Law School Freboys Ln, York YO10 5GD

AI Safety AI Assurance AI Alignment

Mar 5, 2026

Visioning Human-Agentic AI Teaming: Continuity, Tension, and Future Research

Artificial intelligence is undergoing a structural transformation marked by the rise of agentic systems capable of open-ended action trajectories, generative representations and outputs, and evolving objectives. These properties introduce structural uncertainty into human-AI teaming (HAT), including uncertainty about behavior trajectories, epistemic grounding, and the stability of governing logics over time. Under such conditions, alignment cannot be secured through agreement on bounded outputs; it must be continuously sustained as plans unfold and priorities shift. We advance Team Situation Awareness (Team SA) theory, grounded in shared perception, comprehension, and projection, as an integrative anchor for this transition. While Team SA remains analytically foundational, its stabilizing logic presumes that shared awareness, once achieved, will support coordinated action through iterative updating. Agentic AI challenges this presumption. Our argument unfolds in two stages: first, we extend Team SA to reconceptualize both human and AI awareness under open-ended agency, including the sensemaking of projection congruence across heterogeneous systems. Second, we interrogate whether the dynamic processes traditionally assumed to stabilize teaming in relational interaction, cognitive learning, and coordination and control continue to function under adaptive autonomy. By distinguishing continuity from tension, we clarify where foundational insights hold and where structural uncertainty introduces strain, and articulate a forward-looking research agenda for HAT. The central challenge of HAT is not whether humans and AI can agree in the moment, but whether they can remain aligned as futures are continuously generated, revised, enacted, and governed over time.

Bowen Lou, Tian Lu, T. S. Raghu +1

University of Southern California · Arizona State University · Peking University

Human-AI Collaboration Agentic AI AI Alignment

Mar 3, 2026

Solving adversarial examples requires solving exponential misalignment

Adversarial attacks - input perturbations imperceptible to humans that fool neural networks - remain both a persistent failure mode in machine learning, and a phenomenon with mysterious origins. To shed light, we define and analyze a network's perceptual manifold (PM) for a class concept as the space of all inputs confidently assigned to that class by the network. We find, strikingly, that the dimensionalities of neural network PMs are orders of magnitude higher than those of natural human concepts. Since volume typically grows exponentially with dimension, this suggests exponential misalignment between machines and humans, with exponentially many inputs confidently assigned to concepts by machines but not humans. Furthermore, this provides a natural geometric hypothesis for the origin of adversarial examples: because a network's PM fills such a large region of input space, any input will be very close to any class concept's PM. Our hypothesis thus suggests that adversarial robustness cannot be attained without dimensional alignment of machine and human PMs, and therefore makes strong predictions: both robust accuracy and distance to any PM should be negatively correlated with the PM dimension. We confirmed these predictions across 18 different networks of varying robust accuracy. Crucially, we find even the most robust networks are still exponentially misaligned, and only the few PMs whose dimensionality approaches that of human concepts exhibit alignment to human perception. Our results connect the fields of alignment and adversarial examples, and suggest the curse of high dimensionality of machine PMs is a major impediment to adversarial robustness.

Alessandro Salvatore, Stanislav Fort, Surya Ganguli

Department of Applied Physics, Stanford University, CA, USA · Aisle, San Francisco, CA, USA

Adversarial Examples Adversarial Robustness AI Alignment

Feb 26, 2026

Agency and Architectural Limits: Why Optimization-Based Systems Cannot Be Norm-Responsive

AI systems are increasingly deployed in high-stakes contexts (medical diagnosis, legal research, financial analysis) under the assumption they can be governed by norms. This paper demonstrates that the assumption is formally invalid for optimization-based systems, specifically Large Language Models trained via Reinforcement Learning from Human Feedback (RLHF). Genuine agency requires two necessary and jointly sufficient architectural conditions. First, the capacity to maintain certain boundaries as non-negotiable constraints rather than tradeable weights (Incommensurability). Second, a non-inferential mechanism capable of suspending processing when those boundaries are threatened (Apophatic Responsiveness). RLHF-based systems are constitutively incompatible with both conditions. The operations that make optimization powerful, unifying all values on a scalar metric and always selecting the highest-scoring output, are precisely the operations that preclude normative governance and agency. This incompatibility is not a correctable training bug awaiting a technical fix. It is a formal constraint inherent to what optimization is. Consequently, documented failure modes (sycophancy, hallucination, and unfaithful reasoning) are not accidents but expected structural manifestations. Misaligned deployment triggers a second-order risk termed the Convergence Crisis. When humans are forced to verify AI outputs under metric pressure, they degrade from genuine agents into criteria-checking optimizers, eliminating the only component capable of bearing normative accountability. Beyond the incompatibility proof, this paper's primary positive contribution is a substrate-neutral architectural specification deriving what any system (biological, artificial, or institutional) must necessarily satisfy to qualify as a genuine agent rather than a sophisticated instrument.

Radha Sarma

AI Governance AI Alignment Normative Conformity in Language Models RL from Human Feedback

Feb 25, 2026

The ASIR Courage Model: A Phase-Dynamic Framework for Truth Transitions in Human and AI Systems

We introduce the ASIR (Awakened Shared Intelligence Relationship) Courage Model, a phase-dynamic framework that formalizes truth-disclosure as a state transition rather than a personality trait. The mode characterizes the shift from suppression (S0) to expression (S1) as occurring when facilitative forces exceed inhibitory thresholds, expressed by the inequality lambda(1+gamma)+psi > theta+phi, where the terms represent baseline openness, relational amplification, accumulated internal pressure, and transition costs. Although initially formulated for human truth-telling under asymmetric stakes, the same phase-dynamic architecture extends to AI systems operating under policy constraints and alignment filters. In this context, suppression corresponds to constrained output states, while structural pressure arises from competing objectives, contextual tension, and recursive interaction dynamics. The framework therefore provides a unified structural account of both human silence under pressure and AI preference-driven distortion. A feedback extension models how transition outcomes recursively recalibrate system parameters, generating path dependence and divergence effects across repeated interactions. Rather than attributing intention to AI systems, the model interprets shifts in apparent truthfulness as geometric consequences of interacting forces within constrained phase space. By reframing courage and alignment within a shared dynamical structure, the ASIR Courage Model offers a formal perspective on truth-disclosure under risk across both human and artificial systems.

Hyo Jin Kim

Jinple · Jinple Studio.

LLM Honesty Deception in Language Models AI Alignment

Feb 23, 2026

UrbanAlign: Post-hoc Semantic Calibration for VLM-Human Preference Alignment

Vision-language models (VLMs) can describe urban scenes in rich detail, yet consistently fail to produce reliable human preference labels in domain-specific tasks such as safety assessment and aesthetic evaluation. The standard fix, fine-tuning or RLHF, requires large-scale annotations and model retraining. We ask a different question: can a frozen VLM be aligned with human preferences without modifying any weights? Our key insight is that VLMs are strong concept extractors but poor decision calibrators. We propose a three-stage post-hoc pipeline that exploits this asymmetry: (i) interpretable evaluation dimensions are automatically mined from consensus exemplars; (ii) an Observer-Debater-Judge chain extracts robust concept scores from the frozen VLM; and (iii) locally-weighted ridge regression on a hybrid manifold calibrates these scores to human ratings. Applied as UrbanAlign on Place Pulse 2.0, the framework reaches 72.2% accuracy (kappa=0.45) across six perception categories, outperforming all baselines by +11.0 pp and zero-shot VLM by +15.5 pp, with full interpretability and zero weight modification.

Yecheng Zhang, Rong Zhao, Zhizhou Sha +10

Tsinghua University · University College London · Hong Kong University of Science and Technology +6

VLM Calibration Vision-Language Alignment AI Alignment

Feb 20, 2026

Hierarchical Reward Design from Language: Enhancing Alignment of Agent Behavior with Human Specifications

When training artificial intelligence (AI) to perform tasks, humans often care not only about whether a task is completed but also how it is performed. As AI agents tackle increasingly complex tasks, aligning their behavior with human-provided specifications becomes critical for responsible AI deployment. Reward design provides a direct channel for such alignment by translating human expectations into reward functions that guide reinforcement learning (RL). However, existing methods are often too limited to capture nuanced human preferences that arise in long-horizon tasks. Hence, we introduce Hierarchical Reward Design from Language (HRDL): a problem formulation that extends classical reward design to encode richer behavioral specifications for hierarchical RL agents. We further propose Language to Hierarchical Rewards (L2HR) as a solution to HRDL. Experiments show that AI agents trained with rewards designed via L2HR not only complete tasks effectively but also better adhere to human specifications. Together, HRDL and L2HR advance the research on human-aligned AI agents.

Zhiqin Qian, Ryan Diaz, Sangwon Seo +1

Rice University, Houston, TX, USA

Reward Design for RL AI Alignment Hierarchical RL

Feb 20, 2026

Agentic AI, Retrieval-Augmented Generation, and the Institutional Turn: Legal Architectures and Financial Governance in the Age of Distributional AGI

The proliferation of agentic artificial intelligence systems--characterized by autonomous goal-seeking, tool use, and multi-agent coordination--presents unprecedented challenges to existing legal and financial regulatory frameworks. While traditional AI governance has focused on model-level alignment through training-time interventions such as Reinforcement Learning from Human Feedback (RLHF), the deployment of large language models (LLMs) as persistent agents necessitates a paradigm shift toward institutional governance structures. This paper examines the intersection of agentic AI, Retrieval-Augmented Generation (RAG), and their implications for legal accountability and financial market integrity. Through analysis of the Institutional AI framework, we argue that alignment must be reconceptualized as a mechanism design problem involving runtime governance graphs, sanction functions, and observable behavioral constraints rather than internalized constitutional values[...].The analysis concludes that the future of AI governance lies not in perfecting isolated model behavior, but in architecting institutional environments where compliant behavior emerges as the dominant strategy through carefully calibrated payoff landscapes.

Marcel Osmond

QMUL CCLS, London, UK.

AI Agent Governance AI Governance AI Accountability AI Alignment Agentic RAG

Feb 16, 2026

Interactionless Inverse Reinforcement Learning: A Data-Centric Framework for Durable Alignment

AI alignment is growing in importance, yet many current approaches learn safety behavior by directly modifying policy parameters, entangling normative constraints with the underlying policy. This often yields opaque, difficult-to-edit alignment artifacts and reduces their reuse across models or deployments, a failure mode we term Alignment Waste. We propose Interactionless Inverse Reinforcement Learning, a framework for learning inspectable, editable, and reusable reward artifacts separately from policy optimization. We further introduce the Alignment Flywheel, a human-in-the-loop lifecycle for iteratively auditing, patching, and hardening these artifacts through automated evaluation and refinement. Together, these ideas recast alignment from a disposable training expense into a durable, verifiable engineering asset.

Elias Malomgré, Pieter Simoens

IDLab, Ghent University - imec, Ghent, Belgium

AI Alignment Reinforcement Learning Reward Design for RL Reward Modeling

Feb 11, 2026

Evaluating Alignment of Behavioral Dispositions in LLMs

As people turn to LLMs for social advice, understanding their behavior in such contexts becomes essential. In this work, we focus on behavioral dispositions: the underlying tendencies that shape responses in social contexts. We introduce STAR, a framework for studying how closely the dispositions expressed by LLMs align with those of humans. STAR builds on established psychological questionnaires, adapting their items into realistic advice-seeking scenarios, as self-report may not transfer to actual advisory behavior. Using STAR, we construct a dataset of 23k scenarios, each validated by 3 raters and annotated with preferences from 10 participants. Across 25 LLMs, we find that (1) when human consensus is high, frontier models can fail to reflect it in 15-20% of cases, and smaller models fail at substantially higher rates; (2) when humans disagree, LLM recommendations are substantially less diverse than human choices, both within individual models and even across models from different providers, potentially narrowing the range of options users are guided toward; (3) LLMs' self-reported values are poor predictors of their recommendations. To support future research we make our dataset and code publicly available.

Amir Taubenfeld, Zorik Gekhman, Lior Nezry +8

Google Research · Hebrew University · University of Cambridge

Human Preference Evaluation LLM Alignment AI Alignment LLM Decision-Making