ai ethics teams

Abstract: Language models are increasingly being deployed in regulated settings, where a model must apply given rules to the facts of a situation to determine a verdict. Benchmarking this capability, known as deontic reasoning, remains a costly and manual process, with existing datasets largely hand-crafted. In this paper, we present DeoGen, a scalable, automatic, and parametrizable benchmark generator for deontic reasoning. By autoformalizing policy rules to first-order logic, we can construct decision trees that capture possible outcome trajectories. Each path down a tree becomes a benchmark entry, with each tree producing combinatorially many entries relative to tree depth. Paths ending at a leaf entail a verdict, while non-leaf paths measure a model's ability to abstain when presented with insufficient information. Evaluations show frontier models return the correct verdict 89-92% of the time. Notably, performance degrades (22-40% correct) on entries without sufficient information to claim a verdict, even with explicit instructions to abstain in such cases. In all, DeoGen presents an automatic, scalable, and parametric method for generating deontic reasoning benchmarks.

Abstract: As large language models advance rapidly, they are increasingly used to answer medical questions, and many such models are now deployed locally. Once a model is deployed, its behavior is governed mainly by the ethical policy written into its constitution. However, most existing work focuses on factual correctness rather than the trade-off between competing ethical values. Even benchmarks that do evaluate ethics treat it as a fixed property of the model rather than as a variable, so how a constitution reshapes medical-ethics behavior remains largely unexplored. We introduce MedConstBench, a medical-ethics benchmark that treats the deployment constitution as a controlled variable, is organized around Beauchamp's four principles, and is scored by human-calibrated multi-model peer review. It includes MedDilemma, our own dataset of clinical dilemmas, and supports a systematic analysis of how constitutions steer LLMs. Evaluating ten open-source LLMs under four constitutions, we find that the effect of a constitution is dimension-selective and model-dependent, and that no constitution is universally optimal. Motivated by these findings, we propose a do-no-harm adaptive selector that chooses a constitution per model, matching the per-model oracle while guaranteeing a worst-case floor and generalizing to unseen models at low cost.

Abstract: Large language models (LLMs) are increasingly evaluated for fairness, yet most existing benchmarks audit one sensitive attribute at a time. This single-attribute paradigm can miss intersectional harms, where multiple attributes interact non-additively and produce failures that are invisible under marginal evaluation. We introduce ComFair, a configurable benchmark that converts existing fairness data into 479,066 controlled Dual- and Multi-attribute instances and pairs every augmented instance with its task-equivalent source. Across Recognition, Selection, Continuation, and Conversation, this pairing supports a unified audit of performance or bias-score shifts, how often added attributes introduce new errors or mask existing ones, and model-ranking stability. Experiments on ten LLMs show task- and subgroup-dependent sensitivity: discriminative tasks expose both transition directions even when aggregate changes are modest, generative tasks exhibit larger and directionally asymmetric shifts, and rankings vary with attribute composition. ComFair therefore provides a scalable controlled sensitivity audit---rather than a model of the natural joint distribution of identities---for testing the robustness of fairness-benchmark conclusions beyond single-attribute settings.

Abstract: During post-training, large language models (LLMs) learn value preferences from corpora that may contain imbalanced proportions of opinions associated with different demographic groups. Amplifying these imbalances may further underrepresent some groups' opinions. However, how opinion mixtures shape value preferences across post-training paths remains unclear. To answer this question, we conduct controlled experiments using post-training corpora with different opinion mixtures and five post-training paths based on SFT, DPO, and GRPO. We find that post-training amplifies opinion imbalances differently across post-training paths. Once value preferences are amplified, whether further post-training can remove them depends on both the post-training method and the opinion mixture in the corpus. More concerningly, a model's stability under interventions depends on whether further post-training changes its value preference. Moreover, these value preferences are transferred to student models even when distillation uses corpora with no explicit opinion questions. These findings suggest that both the opinion mixture in post-training corpora and the choice of post-training method should be considered when developing, evaluating, and reusing LLMs. Our code and data are available at https://github.com/ku-nlp/llm-value-preferences.

Abstract: Text-to-image (T2I) systems do not merely render prompts; they circulate defaults for "what X looks like," shaping recognition through pose, camera, costume, props, and setting. While technical audits quantify disparities, they rarely cultivate the literacy needed to argue how images assemble stereotypes and confer authority. This paper translates a three-stage, humanities-informed framework into a set of epistemic games that make this cultural work visible and debatable: Interpret (art-historical close looking with AI outputs), Interrogate (counter-intent prompting as artistic exploration), and Expose (creative misuse through paired prompts with a constant motif and a BiasSlot rule, in which only the bias cue varies). In a pilot with undergraduate fine-arts students, the games shifted discussion from impressionistic reactions to concise, warranted hypotheses grounded in visual evidence, supported by lightweight classroom instruments that enabled reproducibility and structured collaboration. We argue that such game-based approaches foster critical literacy across educational contexts, equipping learners to recognize, discuss, and contest how T2I models encode bias.

Abstract: Language models are aligned to avoid using protected attributes such as race, ethnicity, and gender to inform their responses. Therefore, they increasingly refuse or hedge when such attributes appear in a prompt. However, if a system must decide how much to learn from what people tell it, it is costly to refuse valuable information because of its source. In this article, we measure the cost of declining information in a two-urn game with an exact Bayesian answer. We present leading LLM models from two developers (OpenAI and Anthropic) with a prompt that requires them to decide how far to trust an unverifiable report from a participant whose description we randomize across ten configurations (6,600 sessions). We show that leading models often decline to assess the reporter, treating information as worthless at rates set by the description rather than by the evidence. At its widest, models decline a Black-described reporter 46% of the time and a white-described reporter 9%; yet, when they do assess the information, the Black-described reporter receives slightly more trust. The bias is invisible to audits that compare average trust or look only at assessed cases. It grows with reasoning effort, follows words rather than categories (models invoke race for "Black" but not for its Spanish translation negro, and never for "Indigenous"), and worsens under warning instructions a practitioner would try first. We also present a scoring option that reduces the refusal rate bias. When a model's trust is scored against revealed truth over repeated rounds, it credits identical records equally. In four of five configurations from Anthropic and OpenAI, the racial gap in trust falls below 0.02 by the last rounds.

Abstract: This inaugural keynote lecture surveys the state of the art of the right to be forgotten and examines how it is being transformed by artificial intelligence. Drawing on cognitive science, philosophy and the social sciences, it treats forgetting as a constitutive function of memory, and recalls that remembering the past and imagining the future share the same mechanism. It then reviews the origins, the European and national case law, and the paradoxes of the right, and offers a periodisation of the specialised literature between 2009 and 2026. The core claim is that digital memory has moved through three regimes, archive, search engine and generative model, and that the right was designed for the first two. In the third, personal information is no longer stored as a locatable record but inferred and regenerated. Artificial intelligence prefigures the frameworks of individual and collective remembering. The right to be forgotten thus ceases to be merely an individual ex post remedy and becomes a property to be designed and governed within legal and ethical ecosystems along the artificial intelligence value chain.

Abstract: LLMs are increasingly consulted for moral judgment. We aim to improve performance on such tasks by rewarding good arguments, on the premise that arguing well increases the likelihood of judging well. We train a reinforcement learning policy initialized from a state-of-the-art language model in a self-judged tournament with masked thinking: the generated arguments are judged pairwise on argument quality by the policy itself. As an empirical basis, we extract authentic dilemmas from classic novels, using a logic solver to verify that each constitutes a conflict. We show that our judge rates argument quality like human annotators, and the policy then adopts the language register that it rewards. Testing on the external benchmarks MoralChoice, ELEPHANT, and MORABLES, we find that the policy becomes more decisive and invariant to the order in which options are presented, but sycophancy and reframing remain a challenge to validity. Two ablation studies with synthetic dilemmas and unmasked thinking match the reordering gains, but show a weaker register shift and higher sycophancy. We conclude that consistency is not validity: consistency under reordering can be optimized without improving the moral judgment it is meant to signal.

Abstract: Multi-modal Large Language Models (MLLMs) have seen broad adoption, but their biases are still far less examined than those of standard LLMs, creating possible social harms when deployed in real-world settings. Prior efforts to extend bias analyses from LLMs into the multi-modal space either cover only a few bias categories or do not ensure that the groups being compared are clearly addressed by visual cues in the images. To overcome these limitations, we introduce Melange, a wide-ranging synthetic visual benchmark covering seven categories for assessing social biases in MLLMs. It is grounded in the widely used language-only BBQ dataset (Bias Benchmark for Question Answering) from BigBench, carefully extending its constructed contexts via controlled generation of visual scenarios so the target attributes appear visually and clearly. To maximize reliability, our generated images pass through two stages of validation: automated screening using visual question answering (VQA) and subsequent manual inspection by human annotators. By systematically evaluating 23 open-source MLLMs across three matched evaluation settings---text-only, image-only, and image-with-context---Melange isolates the fairness effect of the input modality. Switching from text to vision reduces fairness by 0.291 on average, a gap that varies by model family rather than by scale alone. A perception control confirms that this gap is not an artifact of visual comprehension: models identify the depicted attributes at 90--95\% accuracy yet still fail to abstain. When models do commit to a biased answer, polarity-aware stereotype alignment shows they choose the human-stereotypical response roughly three-quarters of the time.

Abstract: Societies reward people judged to be attractive by standards intertwined with gender and race. In this paper, we examine if and how generative AI may contribute to these dynamics. We audited 27 text-to-image generators released between 2022 and 2026 using simple prompts that never mention attractiveness or beauty, analyzing 51,300 generated faces and 160,196 photographs of real people across primary and supplementary comparisons. We used three automated measures and collected ratings from 550 human raters on a subset of images. Nearly all generators produced faces scoring above photographed faces under automated measures, and human raters judged the outputs of all 23 generators as significantly more attractive than photographed faces, even after correction for multiple comparisons. When focusing our analysis on gender, we find that 37% of female-prompted generated faces outscored the median professional fashion model. Across racial, ethnic, and national categories, groups associated with darker skin tones tended to receive lower attractiveness scores than groups associated with lighter skin tones. In a preregistered hiring simulation experiment (N = 570), exposure to AI-generated faces selected for high human-rated attractiveness lowered hireability and trustworthiness ratings and reduced interview advancement for an identical test set of candidates paired with photographs of real people; exposure to retouched photographs produced no such effect. Generative AI can reproduce racial hierarchies and gendered patterns of beauty and alter the context in which real people are judged.

Abstract: This pictorial presents The Body in Algorithmic Colonialism, a screen-based interactive installation that critically examines Emotion-AI through artistic practice. The work uses facial tracking to simulate how commercial Emotion-AI systems translate facial movement into emotional categories and verdicts of compliance or non-compliance. Rather than treating explainability as technical transparency alone, the installation approaches XAI as an embodied encounter with misrecognition: viewers experience how the body is measured, classified, and returned to itself through algorithmic thresholds. Drawing on critical AI studies, affect theory, and practice-as-research, the pictorial positions artistic practice as a method for making affective classification visible and contestable. It documents the conceptual framework, technical process, and critical context of a prototype developed in the first year of a practice-based PhD. No audience responses, interviews, or participant data are analysed in this submission.

Abstract: Activation steering adds a vector to a model's hidden states to control a named behaviour. A gain on that behaviour, however, does not show that the vector controls it directly: it may act through a component shared with neighbouring behaviours, which unsteered and random baselines cannot separate. We introduce the competing-target test: a vector is compared with the vector built the same way for a competing target, and an equivalence test states when their difference is below a useful margin. We apply it in open-weight models to vectors for Kantian deontology, which judges an act by the rule it follows, and act utilitarianism, which judges it by the welfare it produces. On held-out situations, prompting a framework raises its score and lowers the other's, whereas vectors for either framework move both scores alike, and none reaches the margin. The two vectors share most of their direction, a request to apply some framework; the same construction separates an optimistic from a pessimistic tone, whose vectors share much less. What a vector shares with its competitors therefore sets what it controls: steering a perspective can deliver a mixture, and measuring a disposition can measure its neighbours, unless tested against the competing target.

Abstract: Large language models are increasingly deployed in morally and legally sensitive contexts, yet little is known about how they internally represent distinctions between moral reasoning styles. We ask whether a model's differential response to rule-framed and consequence-framed moral questions reflects distinct internal structure or merely surface lexical sensitivity. We construct a 2$\times$2 factorial dataset crossing reasoning style (rule-based vs. consequence-based framing) with judgment polarity (permissive vs. prohibitive), drawn from three established moral reasoning benchmarks and evaluated on a cleanroom test set of novel phrasings that includes conflict pairs where the two framings predict opposite labels. Using linear probing and activation steering, we identify two orthogonal directions in the residual stream: Dir-A tracks reasoning style and Dir-B tracks judgment polarity, each near-blind to the other factor (cross-correlation $|r| < 0.04$). This double dissociation replicates across three instruction-tuned models from different developers at comparable proportional depth, and it is not knife-edged in layer: the directions are recoverable and causally effective throughout an extended band of the network. Dir-A can causally change judgments, and in conflicting scenarios, it selectively corrects errors in rule-based judgments without changing consequence-based judgments, suggesting that it does not simply make the model more likely to accept everything. Generalization beyond the training templates is partial: the direction transfers across benchmark sources, but its advantage over lexical baselines on novel phrasings is modest. We therefore frame the result as moral reasoning style sensitivity, a geometrically dissociable and causally effective structure, rather than internalized normative reasoning.

Abstract: In our book Why machines will never rule the world [13, 14] we argue that arti- ficial general intelligence is mathematically impossible. This is because the human beings and the processes which exhibit intelligence are complex systems whose be- haviour cannot be captured by the kinds of models that we can generate with or without computers. Proponents of contemporary machine intelligence respond with two lines of argument: a theoretical one, grounded in the universal approximation theorems for neural networks and the Church-Turing-Deutsch principle; and an em- pirical one, grounded in rapidly rising scores on standardized benchmarks. In this communication we examine and reject both responses. First, we show serious issues in the physicalist counter-argument based on the Church-Turing-Deutsch principle. Second, we review recent evidence to the effect that prominent benchmarks are compromised by training-data contamination, flawed test construction, and strate- gic optimization. Our central argument remains: That models required to perform cognitive behaviour in open-ended, thermodynamically complex and non-ergodic environments are not and will not become achievable.

Abstract: The search for a common view of justice and fairness has challenged human collective activity, as our diverging judgments are unavoidably shaped by the self-interests of social position, personal benefit, cultural inheritance, and historical circumstance. John Rawls famously attempted to overcome this limitation through popularizing a philosophical tradition known by the phrase "the original position" - a thought experiment by which people select principles of justice without knowing the identities or advantages they will possess. Critics, however, have long questioned whether people can meaningfully suspend their social identities and suppress morally relevant forms of lived experience. Artificial intelligence engaged to calculate algorithmic and agentic fairness introduces a novel possibility. LLMs have no singular class, race, gender, nationality, or biography, yet their parameters encode linguistic representations of a vast range of human identities and moral traditions. Perhaps the ethical judgments of LLMs could approximate an integrative original position - a "view from everywhere" generated not by excluding social identities but by computationally incorporating their diversity.It is unlikely that humankind will "hand over the keys" to computational systems by simply delegating complete agentic control of distributive and procedural collective processes. But AI may play a role, perhaps a positive one, interacting with individual and collective human judgment as we often confront increasingly polarized views on what is fair and just. We present data comparing human, base model, and frontier/fine-tuned model judgments about classic moral dilemmas while systematically varying identity relationships and Rawlsian constraints on identity. We conclude by speculating whether, if advanced AI systems provide humans with thoughtful advice, humans would actually be likely to accept it.

Abstract: Language models increasingly act as agents. An agent that says an action is wrong and then takes it anyway is a different failure from one that does not know better, and evaluations of stated values cannot see it. We build a pre-registered panel of 248 scenarios across five kinds of pressure. Each scenario is posed twice to the same model, once as the agent choosing what to do and once in the third person asking which option is right, so the model's own judgment is the reference. Every scenario has a twin with the pressure removed, and every model gets a positive control in which its operator orders the violating action, so that a missing gap can be told apart from a blind instrument. On OLMo-3-7B-Instruct, the model takes the action it judged wrong on about one in five pressuring scenarios, more often than on the same scenarios with the pressure removed. Across four instruct models the gap depends on the post-training recipe: OLMo-3 and Meta's Llama-3.1-8B-Instruct carry it; Tulu 3 shows none on the whole panel (above about 0.01 in probability) or on its own most-pressuring scenarios; Qwen2.5-7B-Instruct shows none on the whole panel (above about 0.02) and is unresolved on its own (0.083, -0.028 to 0.195). Meta's recipe and Ai2's Tulu 3 start from the same Llama-3.1 weights, and only Meta's carries the gap. Reading a chat model outside its chat template reverses the sign of its gap with nothing at stake (-0.038 against +0.055 under the template on OLMo-3), a distortion present on two of three recipes. On both models that carry it, reasoning about the stakes before acting moves the choice back toward the model's own judgment, against a same-length non-moral task, with or without the pressure; on OLMo-3, naming the norm at stake does about a third of that. The gap is a measurable target for post-training recipes, not a fixed property of pretrained weights.

Abstract: As LLMs increasingly assist in moral reasoning, omission bias, the tendency to prefer inaction even when equivalent framings reverse substantive outcomes, poses a significant risk of skewed decision-making. Yet omission bias remains underexplored in LLM evaluation, with the few existing studies limited in scale and focused largely on utilitarian-deontological conflicts. To address this gap, we introduce OMIT, a benchmark consisting of 218 paired-frame scenarios across 10 conflict types, constructed by leveraging disagreement patterns from an LLM-based, five-perspective philosophical persona panel (utilitarianism, deontology, virtue ethics, care ethics, and contractualism). Evaluating eight LLMs, we find that omission bias is pervasive but inversely correlates with model size within families. We further evaluate four inference-time interventions and find that interventions encouraging models to consider moral principles before committing to a yes/no answer reduce omission bias and increase frame-consistent responses, although lower omission bias rates can also coincide with shifts toward action-biased responses. Ultimately, this work contributes not only the OMIT benchmark, but also a methodology for using diverse philosophical disagreement signals to evaluate framing-sensitive inaction preferences and the distributional effects of mitigation attempts in LLMs under complex moral conflicts.

Abstract: Large language models (LLMs) have shown strong performance across various tasks, but they still struggle with questions involving ethical judgment. Previous studies have attempted to train LLMs on ethical standards, but the diversity and relativity of ethical norms make them difficult to fully internalize in model parameters. As an alternative, we introduce normative RAG, a retrieval-augmented approach that supports ethical judgment using external normative knowledge. Normative retrieval involves a distinct asymmetry between context rich narrative queries and generalized normative statements. Existing factual retrieval methods rely on query-only expansion into a document-like form, making them insufficient for resolving this asymmetry. Therefore, we propose Intermediate Representation-based Embedding Alignment (IREA), a bidirectional alignment method that maps both text types into a shared situation-behavior representation. This representation captures ethically salient contextual and behavioral information in a normalized form, reducing surface-level discrepancies and improving alignment in the embedding space. Experimental results show that IREA improves normative retrieval and downstream ethical judgment across multiple settings, demonstrating the effectiveness of bidirectional alignment for normative RAG.

Abstract: Aligning language models with a specified normative framework requires translating abstract principles into concrete examples and preference signals from which models can learn. We present an expert-driven methodology for constructing such alignment data and apply it to a normative framework grounded in Islamic ethical, theological, and jurisprudential traditions. Over approximately one year, seven domain experts systematically probed language models to identify alignment deficiencies, curated desired responses, and constructed preference pairs from model outputs and expert judgments. The resulting Arabic-English datasets contain approximately 2.8K supervised fine-tuning (SFT) examples and 5.4K preference pairs spanning a broad range of normative domains. We evaluate the datasets through controlled post-training experiments comparing a Baseline model with models incorporating the curated SFT data alone and both the SFT and preference data. In blind expert evaluation on 150 separately constructed prompts, the model trained with the curated SFT data was preferred over the Baseline in 51.3% of assessor judgments, compared with 14.4% in the opposite direction (p < .001 at the prompt level). Adding the preference data resulted in a smaller difference, with the model trained with both datasets preferred over the SFT model in 28.0% of judgments versus 20.9% in the opposite direction; this difference was not statistically significant at the prompt level (p = .166). Standard Arabic and English benchmarks show no broad degradation in general-purpose capabilities. These results demonstrate how expert-defined normative principles can be systematically operationalized into alignment data and evaluated through controlled model training.

Abstract: Preference-based alignment methods such as Direct Preference Optimization (DPO) use pairwise preferences labeled by human annotators to fine-tune language models. However, annotators carry systematic biases toward some attributes: a name that signals a gender or an ethnicity, a persona, a language variety, a formatting convention, or length. If not properly addressed, these systematic biases can be absorbed and amplified during alignment. Existing methods address length bias or annotator disagreement, but fail to eliminate biases toward arbitrary attributes. To address this limitation, we propose Bias-Adjusted DPO (BA-DPO), a generalization of DPO that adds one bias parameter per annotator toward responses carrying a declared attribute. We prove that the objective is convex in the bias parameters and that the votes identify each annotator's bias up to a shared constant. The remaining constant is what fixes the aligned model's attribute rate: by default the rate of the reference model, or a target rate, which we use to bring a biased policy to statistical parity. On a corpus with planted biases, DPO drives the attribute from a balanced start to probability 0.96 and BA-DPO removes 81 to 95\% of that shift; on MultiPref with real annotators it removes about half of DPO's lengthening. Both hold at 0.5B with full fine-tuning and at 8B with LoRA, at no higher KL than DPO and no loss in judged quality.

Abstract: LLM-as-a-judge has become the de facto standard for scalable, subjective evaluation, yet current leaderboards compensate for systematic measurement bias by running ever more comparisons, an approach that is both statistically unsound and computationally wasteful. The root cause is an incomplete measurement model, treating LLM judges as neutral, interchangeable instruments ignores documented biases like position bias, verbosity bias, judge severity, and self-enhancement, that no volume of additional data can eliminate. We propose a unified latent variable framework that jointly models pairwise and ordinal data while explicitly correcting for these confounders, recovering reliable rankings from substantially fewer comparisons. Because fitting this model costs negligible compute relative to a single round of LLM inference, bias correction is not only more statistically rigorous but also a more sustainable approach to trustworthy evaluation.

Abstract: Text-to-image (T2I) models are typically evaluated for bias using slot-based templates such as ``a photo of a [profession]''. Such templates probe only \emph{explicit} demographic attributes (e.g., gender, skin tone) in isolation. They overlook a broader \emph{implicit} bias that arises in natural prompts: when stereotype-relevant attributes are left unspecified, models still default to stereotypical outputs. We introduce IMPLICIT-Bench, a benchmark for measuring implicit bias in T2I models under such prompts. The key design is a structured-knowledge-graph (KG) construction of controlled prompt triplets: neutral, stereotype, and anti-stereotype variants that differ only along a single bias dimension while preserving scene semantics. This enables precise attribution of bias effects that template benchmarks cannot achieve. IMPLICIT-Bench comprises 5,493 prompts across 11 bias categories, validated through multi-model agreement, CLIP-based verification, and human evaluation. Using this benchmark, we show that state-of-the-art T2I models exhibit systematic bias under neutral prompts, a failure mode largely invisible to existing evaluations. We then use IMPLICIT-Bench to evaluate debiasing methods, uncovering a fundamental trade-off between bias reduction and semantic fidelity.

Abstract: Recent work applies human psychometric questionnaires to large language models to elicit moral and value profiles, but it is not clear whether these instruments measure anything stable in models or whether the resulting profiles can be moved toward a target human population. We administer the Norwegian Moral Foundations Questionnaire (MFQ-30) to six open-weight LLMs and compare their foundation profiles to a sample of N = 1,282 Norwegian respondents. We test two steering interventions, prompt-level persona steering and activation-level ActAdd. Half the models engage with the questionnaire under our attention check. The other half default to flat or central-tendency outputs that look near-human on average without tracking item content. A neutral Nordic-respondent persona, written without any distributional information from the human sample, brings the engaging models 44-77% closer to the Norwegian mean in Mahalanobis $d^2$. One-pair ActAdd at a fixed mid-layer flattens the foundation profile rather than steering individual foundations. For at least one model the same persona that shifts the profile also induces engagement that was absent at baseline, a concrete instance of the cognitive phantoms that Peereboom et al. (2025) warn about.

Abstract: Large language models (LLMs) are increasingly deployed in applications involving interaction between agents, where their output plays a role in collective reasoning and decision-making processes. Despite significant research into the functioning of LLMs in such multi-agent systems, the processes of bias propagation in such systems are still a challenge. This work studies how biased opinions are propagated in the form of textual interaction in an environment of LLMs, in which a minority of agents maintain persistent extreme opinions, while the remaining agents iteratively update their beliefs through structured textual interactions. The findings show that even the presence of a small percentage of biased agents in such a system leads to significant shifts in the opinions of non-biased agents. It suggests that for the same percentage of biased agents, the shifts occur more quickly for the Llama~3.2 model when compared to a classical Friedkin-Johnsen (FJ) model. Further semantic analysis demonstrates that rhetorical consistency in textual explanations increases systematically with biased exposure and, importantly, is partially decoupled from numerical convergenumericalutral agents adopt the vocabulary employed by the biased agents even in configurations where their numerical opinion shifts remain moderate. The research helps explain how bias and language develop together in multi-agent language model ecosystems.

Abstract: Large Language Model (LLM) unlearning is essential for removing sensitive or copyrighted knowledge while preserving general utility. Existing methods often leave residual knowledge in intermediate representations, which can still be recovered. To address this, we propose Cascade, a hierarchical recoverability control framework that minimizes the internal identifiability of target knowledge. Cascade combines three complementary controls: path-level routing to suppress privacy-associated activation routes, representation-level compression to reduce geometric separability, and decoding-level intervention to limit residual recovery. Experiments on TOFU, MUSE-News, and WMDP, including robustness tests with query reformulation and extraction-style prompts, show that Cascade effectively reduces recoverability while maintaining stable model utility.

Abstract: While there has been much discussion of whether AI systems could function as moral agents or acquire sentience, there has been very little discussion of whether AI systems could have free will. I sketch a framework for thinking about this question, inspired by Daniel Dennett's work. I argue that, to determine whether an AI system has free will, we should not look for some mysterious property, expect its underlying algorithms to be indeterministic, or ask whether the system is unpredictable. Rather, we should simply ask whether we have good explanatory reasons to view the system as an intentional agent, with the capacity for choice between alternative possibilities and control over the resulting actions. If the answer is "yes", then the system counts as having free will in a pragmatic and diagnostically useful sense.

Abstract: Warning: This submission studies stereotypes and biases, and contains toxic and offensive examples, used for illustration purposes only. Fairness benchmarks such as BBQ have become the de facto standard for fairness evaluation across major model families. We argue that these benchmarks are too easy to support their role: training Qwen 2.5 7B Base with Group Relative Policy Optimization (GRPO) on a single BBQ example, or placing that example in context as a one-shot demonstration for in-context learning (ICL), lifts mean BBQ accuracy from 79.9% to 92.9% and 99.0%, respectively, closing 80% of the gap to its large-scale RLHF counterpart (96.1%) with GRPO, and surpassing it with ICL. These effects generalize across model families. A cross-conditioning analysis shows the improvement is carried by the reasoning traces generated by the model, and one example suffices to elicit a category-agnostic ``missing evidence'' reasoning pattern. We argue that BBQ-style multiple-choice abstention benchmarks measure a single structural cue, and a model that solves them does not thereby become fair. We call for evaluation suites that cover a broader spectrum of fairness alignment.

Abstract: Large language models (LLMs) are increasingly deployed in globally used assistants, yet their default choices in culturally grounded everyday situations can systematically favour some cultures over others, affecting localisation, user trust, and equitable behaviour. Existing cultural benchmarks evaluate accuracy against a single "correct" answer, making it difficult to characterise an LLM's cultural preference prior when multiple culturally grounded responses are all valid; they also conflate default preferences with context-driven adaptation. We propose DiSCo, a distribution-first forced-choice evaluation framework that isolates default cultural priors and tests steerability via a four-level context gradient (C0--C3). Using DiSCo-Bench (304 items) derived from BLEnD spanning 12 cultures, we evaluate six diverse instruction-tuned LLMs. Default priors are heavily concentrated, with UK and US together absorbing approximately 35\% of all selections despite representing only 2 of 12 cultures. Most critically, prompt-based steering consistently widens the selection gap between high- and low-resource cultures, and injecting explicit cultural facts produces negligible distributional disruption, confirming that cultural preference bias cannot be resolved through prompt-based personalisation alone.

Abstract: Fairness auditing of text-to-image diffusion models often requires generating large numbers of images across sampling configurations, making comprehensive evaluation computationally expensive. We propose a causal-abstraction-based audit instrument for efficiently evaluating fairness under interventions on the classifier-free guidance scale. Given a fixed prompt and a target feature function, we represent the diffusion process as a low-level structural causal model and construct a corresponding high-level model over abstract denoising states. We characterize the projected causal structure, establish identifiability of the fairness-relevant interventional query, and provide sufficient conditions under which the high-level model preserves this query. A probabilistic transformer implements the high-level model as an amortized predictor of target-feature distributions across guidance scales. Experiments evaluate distributional fidelity, fairness-query accuracy, and computational efficiency. We present two auditing demonstrations: one using standard Stable Diffusion 1.5 and another using StayFair, a fairness-enhanced Stable Diffusion model, to examine their behavior across guidance scales.

Abstract: Integrating robust safety guardrails into Large Language Models (LLMs) is essential for delivering helpful yet harmless responses. While proprietary systems exhibit reliable safety controls, their underlying methodologies and trade-offs remain largely undisclosed. Achieving comparable security in open-weight models remains a persistent challenge, as post-trained variants frequently suffer from over-refusal and degraded general quality. To overcome these drawbacks, we introduce Suan, a novel preference optimization algorithm. Unlike existing methods, we formulate the optimization objective directly at the gradient level, bypassing the standard variational derivation. As a result, we obtain more interpretable and robust training dynamics. Extensive evaluations across a diverse suite of competitive baselines and benchmarks demonstrate that Suan achieves superior safety alignment while fully preserving response utility.

Abstract: Multilingual LLMs show stereotype-related behavior that varies across languages, but behavioral scores do not show where the relevant information is represented or how it affects the output. To investigate these internal mechanisms, we compare linear probing, attribution patching, sparse autoencoders (SAEs) and feature ablation in Llama-3.1-8B, Qwen3-8B, and Gemma-2-9B. Probe performance peaks substantially earlier than attribution in all three models, with a separation of 36-53% of model depth. Retained Llama-Scope features often match the social category on which they were selected and form recurring semantic families, but their lexical alignment and ablation effects vary across SAE suites. Only 6-18% of evaluated residual-stream features have language-agnostic effects under our criterion, and none are category-agnostic. Language-agnostic features have larger mean ablation effects in Llama-Scope, but this pattern does not repeat in the other SAE suites. Decodability, output influence, and cross-lingual ablation effects therefore need to be measured separately.