arXiv subjects

Maarten Sap

Publications and source records attributed to Maarten Sap.

At least 55 records · Page 3Linked to original sources

arXiv2026

PluriHarms: Benchmarking the Full Spectrum of Human Judgments on AI Harm

Current AI safety frameworks, which often treat harmfulness as binary, lack the flexibility to handle borderline cases where humans meaningfully disagree. To build more pluralistic systems, it is essential to move beyond consensus and instead understand where and why disagreements arise. We introduce PluriHarms, a benchmark designed to systematically study human harm judgments across two key dimensions -- the harm axis (benign to harmful) and the agreement axis (agreement to disagreement). Our scalable framework generates prompts that capture diverse AI harms and human values while targeting cases with high disagreement rates, validated by human data. The benchmark includes 150 prompts with 15,000 ratings from 100 human annotators, enriched with demographic and psychological traits and prompt-level features of harmful actions, effects, and values. Our analyses show that prompts that relate to imminent risks and tangible harms amplify perceived harmfulness, while annotator traits (e.g., toxicity experience, education) and their interactions with prompt content explain systematic disagreement. We benchmark AI safety models and alignment methods on PluriHarms, finding that while personalization significantly improves prediction of human harm judgments, considerable room remains for future progress. By explicitly targeting value diversity and disagreement, our work provides a principled benchmark for moving beyond "one-size-fits-all" safety toward pluralistically safe AI.

cs.CY↗

arXiv2026

Synthetic Socratic Debates: Examining Persona Effects on Moral Decision and Persuasion Dynamics

As large language models (LLMs) are increasingly used in morally sensitive domains, it is crucial to understand how persona traits affect their moral reasoning and persuasive behavior. We present the first large-scale study of multi-dimensional persona effects in AI-AI debates over real-world moral dilemmas. Using a 6-dimensional persona space (age, gender, country, class, ideology, and personality), we simulate structured debates between AI agents over 131 relationship-based cases. Our results show that personas affect initial moral stances and debate outcomes, with political ideology and personality traits exerting the strongest influence. Persuasive success varies across traits, with liberal and open personalities reaching higher consensus and win rates. While logit-based confidence grows during debates, emotional and credibility-based appeals diminish, indicating more tempered argumentation over time. These trends mirror findings from psychology and cultural studies, reinforcing the need for persona-aware evaluation frameworks for AI moral reasoning.

cs.CL↗

arXiv2026

TOM-SWE: User Mental Modeling For Software Engineering Agents

Recent advances in coding agents have made them capable of planning, editing, running, and testing complex code bases. Despite their growing ability in coding tasks, these systems still struggle to infer and track user intent, especially when instructions are underspecified or context-dependent. To bridge this gap, we introduce ToM-SWE, a dual-agent architecture that pairs a primary software-engineering (SWE) agent with a lightweight theory-of-mind (ToM) partner agent dedicated to modeling the user's mental state. The ToM agent infers user goals, constraints, and preferences from instructions and interaction history, maintains a \textbf{persistent memory} of the user, and provides user-related suggestions to the SWE agent. In two software engineering benchmarks (ambiguous SWE-bench and stateful SWE-bench), ToM-SWE improves task success rates and user satisfaction. Notably, on the stateful SWE benchmark, a newly introduced evaluation that provides agents with a user simulator along with previous interaction histories, ToM-SWE achieves a substantially higher task success rate of 59.7\% compared to 18.1\% for OpenHands, a state-of-the-art SWE agent. Furthermore, in a three-week study with professional developers using ToM-SWE in their daily work, participants found it useful 86\% of the time, underscoring the value of stateful user modeling for practical coding agents.

cs.SE↗

arXiv2026

Out of Style: RAG's Fragility to Linguistic Variation

Despite the impressive performance of Retrieval-augmented Generation (RAG) systems across various NLP benchmarks, their robustness in handling real-world user-LLM interaction queries remains largely underexplored. This presents a critical gap for practical deployment, where user queries exhibit greater linguistic variations and can trigger cascading errors across interdependent RAG components. In this work, we systematically analyze how varying four linguistic dimensions (formality, readability, politeness, and grammatical correctness) impact RAG performance. We evaluate two retrieval models and nine LLMs, ranging from 3 to 72 billion parameters, across four information-seeking Question Answering (QA) datasets. Our results reveal that linguistic reformulations significantly impact both retrieval and generation stages, leading to a relative performance drop of up to 40.41% in Recall@5 scores for less formal queries and 38.86% in answer match scores for queries containing grammatical errors. Notably, RAG systems exhibit greater sensitivity to such variations compared to LLM-only generations, highlighting their vulnerability to error propagation due to linguistic shifts. These findings highlight the need for improved robustness techniques to enhance reliability in diverse user interactions. Code is available at https://github.com/Springcty/RAG-fragility-to-linguistic-variation.

cs.CL↗

arXiv2025

SoMi-ToM: Evaluating Multi-Perspective Theory of Mind in Embodied Social Interactions

Humans continuously infer the states, goals, and behaviors of others by perceiving their surroundings in dynamic, real-world social interactions. However, most Theory of Mind (ToM) benchmarks only evaluate static, text-based scenarios, which have a significant gap compared to real interactions. We propose the SoMi-ToM benchmark, designed to evaluate multi-perspective ToM in embodied multi-agent complex social interactions. This benchmark is based on rich multimodal interaction data generated by the interaction environment SoMi, covering diverse crafting goals and social relationships. Our framework supports multi-level evaluation: (1) first-person evaluation provides multimodal (visual, dialogue, action, etc.) input from a first-person perspective during a task for real-time state inference, (2) third-person evaluation provides complete third-person perspective video and text records after a task for goal and behavior inference. This evaluation method allows for a more comprehensive examination of a model's ToM capabilities from both the subjective immediate experience and the objective global observation. We constructed a challenging dataset containing 35 third-person perspective videos, 363 first-person perspective images, and 1225 expert-annotated multiple-choice questions (three options). On this dataset, we systematically evaluated the performance of human subjects and several state-of-the-art large vision-language models (LVLMs). The results show that LVLMs perform significantly worse than humans on SoMi-ToM: the average accuracy gap between humans and models is 40.1% in first-person evaluation and 26.4% in third-person evaluation. This indicates that future LVLMs need to further improve their ToM capabilities in embodied, complex social interactions.

cs.CL↗

arXiv2025

Let Them Down Easy! Contextual Effects of LLM Guardrails on User Perceptions and Preferences

Current LLMs are trained to refuse potentially harmful input queries regardless of whether users actually had harmful intents, causing a tradeoff between safety and user experience. Through a study of 480 participants evaluating 3,840 query-response pairs, we examine how different refusal strategies affect user perceptions across varying motivations. Our findings reveal that response strategy largely shapes user experience, while actual user motivation has negligible impact. Partial compliance -- providing general information without actionable details -- emerges as the optimal strategy, reducing negative user perceptions by over 50% to flat-out refusals. Complementing this, we analyze response patterns of 9 state-of-the-art LLMs and evaluate how 6 reward models score different refusal strategies, demonstrating that models rarely deploy partial compliance naturally and reward models currently undervalue it. This work demonstrates that effective guardrails require focusing on crafting thoughtful refusals rather than detecting intent, offering a path toward AI safety mechanisms that ensure both safety and sustained user engagement.

cs.CL↗

arXiv2025

Martingale Score: An Unsupervised Metric for Bayesian Rationality in LLM Reasoning

Recent advances in reasoning techniques have substantially improved the performance of large language models (LLMs), raising expectations for their ability to provide accurate, truthful, and reliable information. However, emerging evidence suggests that iterative reasoning may foster belief entrenchment and confirmation bias, rather than enhancing truth-seeking behavior. In this study, we propose a systematic evaluation framework for belief entrenchment in LLM reasoning by leveraging the Martingale property from Bayesian statistics. This property implies that, under rational belief updating, the expected value of future beliefs should remain equal to the current belief, i.e., belief updates are unpredictable from the current belief. We propose the unsupervised, regression-based Martingale Score to measure violations of this property, which signal deviation from the Bayesian ability of updating on new evidence. In open-ended problem domains including event forecasting, value-laden questions, and academic paper review, we find such violations to be widespread across models and setups, where the current belief positively predicts future belief updates, a phenomenon which we term belief entrenchment. We identify the models, reasoning techniques, and domains more prone to belief entrenchment. Finally, we validate the Martingale Score by showing that it predicts ground-truth accuracy on problem domains where ground truth labels are available. This indicates that, while designed as an unsupervised metric that operates even in domains without access to ground truth, the Martingale Score is a useful proxy of the truth-seeking ability of a reasoning process.

cs.AI↗

arXiv2025

Medical Hallucinations in Foundation Models and Their Impact on Healthcare

Hallucinations in foundation models arise from autoregressive training objectives that prioritize token-likelihood optimization over epistemic accuracy, fostering overconfidence and poorly calibrated uncertainty. We define medical hallucination as any model-generated output that is factually incorrect, logically inconsistent, or unsupported by authoritative clinical evidence in ways that could alter clinical decisions. We evaluated 11 foundation models (7 general-purpose, 4 medical-specialized) across seven medical hallucination tasks spanning medical reasoning and biomedical information retrieval. General-purpose models achieved significantly higher proportions of hallucination-free responses than medical-specialized models (median: 76.6% vs 51.3%, difference = 25.2%, 95% CI: 18.7-31.3%, Mann-Whitney U = 27.0, p = 0.012, rank-biserial r = -0.64). Top-performing models such as Gemini-2.5 Pro exceeded 97% accuracy when augmented with chain-of-thought prompting (base: 87.6%), while medical-specialized models like MedGemma ranged from 28.6-61.9% despite explicit training on medical corpora. Chain-of-thought reasoning significantly reduced hallucinations in 86.4% of tested comparisons after FDR correction (q < 0.05), demonstrating that explicit reasoning traces enable self-verification and error detection. Physician audits confirmed that 64-72% of residual hallucinations stemmed from causal or temporal reasoning failures rather than knowledge gaps. A global survey of clinicians (n = 70) validated real-world impact: 91.8% had encountered medical hallucinations, and 84.7% considered them capable of causing patient harm. The underperformance of medical-specialized models despite domain training indicates that safety emerges from sophisticated reasoning capabilities and broad knowledge integration developed during large-scale pre-training, not from narrow optimization.

cs.CL↗

arXiv2025

Breaking mBad! Supervised Fine-tuning for Cross-Lingual Detoxification

As large language models (LLMs) become increasingly prevalent in global applications, ensuring that they are toxicity-free across diverse linguistic contexts remains a critical challenge. We explore "Cross-lingual Detoxification", a cross-lingual paradigm that mitigates toxicity, enabling detoxification capabilities to transfer between high and low-resource languages across different script families. We analyze cross-lingual detoxification's effectiveness through 392 extensive settings to evaluate toxicity reduction in cross-distribution settings with limited data and investigate how mitigation impacts model performance on non-toxic tasks, revealing trade-offs between safety and knowledge preservation. Our code and dataset are publicly available at https://github.com/himanshubeniwal/Breaking-mBad.

cs.CL↗

arXiv2025

EVALUESTEER: Measuring Reward Model Steerability Towards Values and Preferences

As large language models (LLMs) are deployed globally, creating pluralistic systems that can accommodate the diverse preferences and values of users worldwide becomes essential. We introduce EVALUESTEER, a benchmark to measure LLMs' and reward models' (RMs) steerability towards users' value and stylistic preference profiles grounded in psychology and human-LLM interaction literature. To address the gap in existing datasets that do not support controlled evaluations of RM steering, we synthetically generated 165,888 preference pairs -- systematically varying pairs along 4 value dimensions (traditional, secular-rational, survival, and self-expression) and 4 style dimensions (verbosity, readability, confidence, and warmth). We use EVALUESTEER to evaluate whether, given a user profile and a pair of candidate value-laden and style-laden responses, LLMs and RMs are able to select the output that aligns with the user's preferences. We evaluate six open-source and proprietary LLMs and RMs under eleven systematic prompting conditions and six preference comparison scenarios. Notably, our results show that, when given the user's full profile of values and stylistic preferences, the best models achieve <75% accuracy at choosing the correct response, in contrast to >99% accuracy when only relevant style and value preferences are provided. EVALUESTEER thus highlights the limitations of current RMs at identifying and adapting to relevant user profile information, and provides a challenging testbed for developing RMs that can be steered towards diverse human values and preferences.

cs.CL↗

arXiv2025

Fluid Language Model Benchmarking

Language model (LM) benchmarking faces several challenges: comprehensive evaluations are costly, benchmarks often fail to measure the intended capabilities, and evaluation quality can degrade due to labeling errors and benchmark saturation. Although various strategies have been proposed to mitigate these issues, they tend to address individual aspects in isolation, neglecting broader questions about overall evaluation quality. Here, we introduce Fluid Benchmarking, a new evaluation approach that advances LM benchmarking across multiple dimensions. Inspired by psychometrics, Fluid Benchmarking is based on the insight that the relative value of benchmark items depends on an LM's capability level, suggesting that evaluation should adapt to each LM. Methodologically, Fluid Benchmarking estimates an item response model based on existing LM evaluation results and uses the inferred quantities to select evaluation items dynamically, similar to computerized adaptive testing in education. In our experiments, we compare Fluid Benchmarking against the common practice of random item sampling as well as more sophisticated baselines, including alternative methods grounded in item response theory. We examine four dimensions -- efficiency, validity, variance, and saturation -- and find that Fluid Benchmarking achieves superior performance in all of them (e.g., higher validity and less variance on MMLU with fifty times fewer items). Our analysis shows that the two components of Fluid Benchmarking have distinct effects: item response theory, used to map performance into a latent ability space, increases validity, while dynamic item selection reduces variance. Overall, our results suggest that LM benchmarking can be substantially improved by moving beyond static evaluation.

cs.CL↗

arXiv2025

HAICOSYSTEM: An Ecosystem for Sandboxing Safety Risks in Human-AI Interactions

AI agents are increasingly autonomous in their interactions with human users and tools, leading to increased interactional safety risks. We present HAICOSYSTEM, a framework examining AI agent safety within diverse and complex social interactions. HAICOSYSTEM features a modular sandbox environment that simulates multi-turn interactions between human users and AI agents, where the AI agents are equipped with a variety of tools (e.g., patient management platforms) to navigate diverse scenarios (e.g., a user attempting to access other patients' profiles). To examine the safety of AI agents in these interactions, we develop a comprehensive multi-dimensional evaluation framework that uses metrics covering operational, content-related, societal, and legal risks. Through running 1840 simulations based on 92 scenarios across seven domains (e.g., healthcare, finance, education), we demonstrate that HAICOSYSTEM can emulate realistic user-AI interactions and complex tool use by AI agents. Our experiments show that state-of-the-art LLMs, both proprietary and open-sourced, exhibit safety risks in over 50\% cases, with models generally showing higher risks when interacting with simulated malicious users. Our findings highlight the ongoing challenge of building agents that can safely navigate complex interactions, particularly when faced with malicious users. To foster the AI agent safety ecosystem, we release a code platform that allows practitioners to create custom scenarios, simulate interactions, and evaluate the safety and performance of their agents.

cs.AI↗

arXiv2025

Words Like Knives: Backstory-Personalized Modeling and Detection of Violent Communication

Conversational breakdowns in close relationships are deeply shaped by personal histories and emotional context, yet most NLP research treats conflict detection as a general task, overlooking the relational dynamics that influence how messages are perceived. In this work, we leverage nonviolent communication (NVC) theory to evaluate LLMs in detecting conversational breakdowns and assessing how relationship backstory influences both human and model perception of conflicts. Given the sensitivity and scarcity of real-world datasets featuring conflict between familiar social partners with rich personal backstories, we contribute the PersonaConflicts Corpus, a dataset of N=5,772 naturalistic simulated dialogues spanning diverse conflict scenarios between friends, family members, and romantic partners. Through a controlled human study, we annotate a subset of dialogues and obtain fine-grained labels of communication breakdown types on individual turns, and assess the impact of backstory on human and model perception of conflict in conversation. We find that the polarity of relationship backstories significantly shifted human perception of communication breakdowns and impressions of the social partners, yet models struggle to meaningfully leverage those backstories in the detection task. Additionally, we find that models consistently overestimate how positively a message will make a listener feel. Our findings underscore the critical role of personalization to relationship contexts in enabling LLMs to serve as effective mediators in human communication for authentic connection.

cs.CL↗

arXiv2025

Exploring Big Five Personality and AI Capability Effects in LLM-Simulated Negotiation Dialogues

This paper presents an evaluation framework for agentic AI systems in mission-critical negotiation contexts, addressing the need for AI agents that can adapt to diverse human operators and stakeholders. Using Sotopia as a simulation testbed, we present two experiments that systematically evaluated how personality traits and AI agent characteristics influence LLM-simulated social negotiation outcomes--a capability essential for a variety of applications involving cross-team coordination and civil-military interactions. Experiment 1 employs causal discovery methods to measure how personality traits impact price bargaining negotiations, through which we found that Agreeableness and Extraversion significantly affect believability, goal achievement, and knowledge acquisition outcomes. Sociocognitive lexical measures extracted from team communications detected fine-grained differences in agents' empathic communication, moral foundations, and opinion patterns, providing actionable insights for agentic AI systems that must operate reliably in high-stakes operational scenarios. Experiment 2 evaluates human-AI job negotiations by manipulating both simulated human personality and AI system characteristics, specifically transparency, competence, adaptability, demonstrating how AI agent trustworthiness impact mission effectiveness. These findings establish a repeatable evaluation methodology for experimenting with AI agent reliability across diverse operator personalities and human-agent team dynamics, directly supporting operational requirements for reliable AI systems. Our work advances the evaluation of agentic AI workflows by moving beyond standard performance metrics to incorporate social dynamics essential for mission success in complex operations.

cs.AI↗

arXiv2025

ALFA: Aligning LLMs to Ask Good Questions A Case Study in Clinical Reasoning

Large language models (LLMs) often fail to ask effective questions under uncertainty, making them unreliable in domains where proactive information-gathering is essential for decision-making. We present ALignment via Fine-grained Attributes, (ALFA) a framework that improves LLM question-asking by (i) decomposing the notion of a "good" question into a set of theory-grounded attributes (e.g., clarity, relevance), (ii) controllably synthesizing attribute-specific question variations, and (iii) aligning models via preference-based optimization to explicitly learn to ask better questions along these fine-grained attributes. Focusing on clinical reasoning as a case study, we introduce the MediQ-AskDocs dataset, composed of 17k real-world clinical interactions augmented with 80k attribute-specific preference pairs of follow-up questions, as well as a novel expert-annotated interactive healthcare QA task to evaluate question-asking abilities. Models aligned with ALFA reduce diagnostic errors by 56.6% on MediQ-AskDocs compared to SoTA instruction-tuned LLMs, with a question-level win-rate of 64.4% and strong generalizability. Our findings suggest that explicitly guiding question-asking with structured, fine-grained attributes offers a scalable path to improve LLMs, especially in expert application domains.

cs.CL↗

arXiv2025

Not Like Us, Hunty: Measuring Perceptions and Behavioral Effects of Minoritized Anthropomorphic Cues in LLMs

As large language models (LLMs) increasingly adapt and personalize to diverse sets of users, there is an increased risk of systems appropriating sociolects, i.e., language styles or dialects that are associated with specific minoritized lived experiences (e.g., African American English, Queer slang). In this work, we examine whether sociolect usage by an LLM agent affects user reliance on its outputs and user perception (satisfaction, frustration, trust, and social presence). We designed and conducted user studies where 498 African American English (AAE) speakers and 487 Queer slang speakers performed a set of question-answering tasks with LLM-based suggestions in either standard American English (SAE) or their self-identified sociolect. Our findings showed that sociolect usage by LLMs influenced both reliance and perceptions, though in some surprising ways. Results suggest that both AAE and Queer slang speakers relied more on the SAE agent, and had more positive perceptions of the SAE agent. Yet, only Queer slang speakers felt more social presence from the Queer slang agent over the SAE one, whereas only AAE speakers preferred and trusted the SAE agent over the AAE one. These findings emphasize the need to test for behavioral outcomes rather than simply assume that personalization would lead to a better and safer reliance outcome. They also highlight the nuanced dynamics of minoritized language in machine interactions, underscoring the need for LLMs to be carefully designed to respect cultural and linguistic boundaries while fostering genuine user engagement and trust.

cs.HC↗

arXiv2025

PolyGuard: A Multilingual Safety Moderation Tool for 17 Languages

Truly multilingual safety moderation efforts for Large Language Models (LLMs) have been hindered by a narrow focus on a small set of languages (e.g., English, Chinese) as well as a limited scope of safety definition, resulting in significant gaps in moderation capabilities. To bridge these gaps, we release POLYGUARD, a new state-of-the-art multilingual safety model for safeguarding LLM generations, and the corresponding training and evaluation datasets. POLYGUARD is trained on POLYGUARDMIX, the largest multilingual safety training corpus to date containing 1.91M samples across 17 languages (e.g., Chinese, Czech, English, Hindi). We also introduce POLYGUARDPROMPTS, a high quality multilingual benchmark with 29K samples for the evaluation of safety guardrails. Created by combining naturally occurring multilingual human-LLM interactions and human-verified machine translations of an English-only safety dataset (WildGuardMix; Han et al., 2024), our datasets contain prompt-output pairs with labels of prompt harmfulness, response harmfulness, and response refusal. Through extensive evaluations across multiple safety and toxicity benchmarks, we demonstrate that POLYGUARD outperforms existing state-of-the-art open-weight and commercial safety classifiers by 5.5%. Our contributions advance efforts toward safer multilingual LLMs for all global users.

cs.CL↗

arXiv2025

BIG5-CHAT: Shaping LLM Personalities Through Training on Human-Grounded Data

In this work, we tackle the challenge of embedding realistic human personality traits into LLMs. Previous approaches have primarily focused on prompt-based methods that describe the behavior associated with the desired personality traits, suffering from realism and validity issues. To address these limitations, we introduce BIG5-CHAT, a large-scale dataset containing 100,000 dialogues designed to ground models in how humans express their personality in language. Leveraging this dataset, we explore Supervised Fine-Tuning and Direct Preference Optimization as training-based methods to align LLMs more naturally with human personality patterns. Our methods outperform prompting on personality assessments such as BFI and IPIP-NEO, with trait correlations more closely matching human data. Furthermore, our experiments reveal that models trained to exhibit higher conscientiousness, higher agreeableness, lower extraversion, and lower neuroticism display better performance on reasoning tasks, aligning with psychological findings on how these traits impact human cognitive performance. To our knowledge, this work is the first comprehensive study to demonstrate how training-based methods can shape LLM personalities through learning from real human behaviors.

cs.CL↗