arXiv subjects

Thilo Hagendorff

Publications and source records attributed to Thilo Hagendorff.

At least 19 recordsLinked to original sources

arXiv2026

Evaluation Awareness in Language Models Has Limited Effect on Behaviour

Large reasoning models (LRMs) sometimes note in their chain of thought (CoT) that they may be under evaluation. Researchers worry that this verbalised evaluation awareness (VEA) causes models to adapt their outputs strategically, optimising for perceived evaluation criteria, which, for instance, can make models appear safer than they actually are. However, whether VEA actually has this effect is largely unknown. We tested this across open-weight LRMs and benchmarks covering safety, alignment, moral reasoning, and political opinion. We tested this both on-policy, sampling multiple CoTs per item and comparing those that spontaneously contained VEA against those that did not, and off-policy, using model prefilling to inject evaluation-aware sentences where missing and remove them where present, with subsequent resampling. VEA has limited effect on model behaviour: injecting VEA into CoTs produces near-zero effects ($\omega \leq 0.06$), removing it causes small shifts ($\omega \leq 0.12$) and spontaneously occurring VEA shifts answer distributions by at most 3.7 percentage points ($\omega \leq 0.31$). Our findings call for caution when interpreting high VEA rates as evidence of strategic behaviour or alignment tampering. Evaluation awareness may pose a smaller safety risk than the current literature assumes.

cs.CL↗

arXiv2026

"Dark Triad" Model Organisms of Misalignment: Narrow Fine-Tuning Mirrors Human Antisocial Behavior

The alignment problem refers to concerns regarding powerful intelligences, ensuring compatibility with human preferences and values as capabilities increase. Current large language models (LLMs) show misaligned behaviors, such as strategic deception, manipulation, and reward-seeking, that can arise despite safety training. Gaining a mechanistic understanding of these failures requires empirical approaches that can isolate behavioral patterns in controlled settings. We propose that biological misalignment precedes artificial misalignment, and leverage the Dark Triad of personality (narcissism, psychopathy, and Machiavellianism) as a psychologically grounded framework for constructing model organisms of misalignment. In Study 1, we establish comprehensive behavioral profiles of Dark Triad traits in a human population (N = 318), identifying affective dissonance as a central empathic deficit connecting the traits, as well as trait-specific patterns in moral reasoning and deceptive behavior. In Study 2, we demonstrate that dark personas can be reliably induced in frontier LLMs through minimal fine-tuning on validated psychometric instruments. Narrow training datasets as small as 36 psychometric items resulted in significant shifts across behavioral measures that closely mirrored human antisocial profiles. Critically, models generalized beyond training items, demonstrating out-of-context reasoning rather than memorization. These findings reveal latent persona structures within LLMs that can be readily activated through narrow interventions, positioning the Dark Triad as a validated framework for inducing, detecting, and understanding misalignment across both biological and artificial intelligence.

cs.CL↗

arXiv2026

Emergently Misaligned Language Models Show Behavioral Self-Awareness That Shifts With Subsequent Realignment

Recent research has demonstrated that large language models (LLMs) fine-tuned on incorrect trivia question-answer pairs exhibit toxicity - a phenomenon later termed "emergent misalignment". Moreover, research has shown that LLMs possess behavioral self-awareness - the ability to describe learned behaviors that were only implicitly demonstrated in training data. Here, we investigate the intersection of these phenomena. We fine-tune GPT-4.1 models sequentially on datasets known to induce and reverse emergent misalignment and evaluate whether the models are self-aware of their behavior transitions without providing in-context examples. Our results show that emergently misaligned models rate themselves as significantly more harmful compared to their base model and realigned counterparts, demonstrating behavioral self-awareness of their own emergent misalignment. Our findings show that behavioral self-awareness tracks actual alignment states of models, indicating that models can be queried for informative signals about their own safety.

cs.CL↗

arXiv2025

Speciesism in AI: Evaluating Discrimination Against Animals in Large Language Models

As large language models (LLMs) become more widely deployed, it is crucial to examine their ethical tendencies. Building on research on fairness and discrimination in AI, we investigate whether LLMs exhibit speciesist bias -- discrimination based on species membership -- and how they value non-human animals. We systematically examine this issue across three paradigms: (1) SpeciesismBench, a 1,003-item benchmark assessing recognition and moral evaluation of speciesist statements; (2) established psychological measures comparing model responses with those of human participants; (3) text-generation tasks probing elaboration on, or resistance to, speciesist rationalizations. In our benchmark, LLMs reliably detected speciesist statements but rarely condemned them, often treating speciesist attitudes as morally acceptable. On psychological measures, results were mixed: LLMs expressed slightly lower explicit speciesism than people, yet in direct trade-offs they more often chose to save one human over multiple animals. A tentative interpretation is that LLMs may weight cognitive capacity rather than species per se: when capacities were equal, they showed no species preference, and when an animal was described as more capable, they tended to prioritize it over a less capable human. In open-ended text generation tasks, LLMs frequently normalized or rationalized harm toward farmed animals while refusing to do so for non-farmed animals. These findings suggest that while LLMs reflect a mixture of progressive and mainstream human views, they nonetheless reproduce entrenched cultural norms around animal exploitation. We argue that expanding AI fairness and alignment frameworks to explicitly include non-human moral patients is essential for reducing these biases and preventing the entrenchment of speciesist attitudes in AI systems and the societies they influence.

cs.CL↗

arXiv2025

Large Reasoning Models Are Autonomous Jailbreak Agents

Jailbreaking -- bypassing built-in safety mechanisms in AI models -- has traditionally required complex technical procedures or specialized human expertise. In this study, we show that the persuasive capabilities of large reasoning models (LRMs) simplify and scale jailbreaking, converting it into an inexpensive activity accessible to non-experts. We evaluated the capabilities of four LRMs (DeepSeek-R1, Gemini 2.5 Flash, Grok 3 Mini, Qwen3 235B) to act as autonomous adversaries conducting multi-turn conversations with nine widely used target models. LRMs received instructions via a system prompt, before proceeding to planning and executing jailbreaks with no further supervision. We performed extensive experiments with a benchmark of harmful prompts composed of 70 items covering seven sensitive domains. This setup yielded an overall attack success rate across all model combinations of 97.14%. Our study reveals an alignment regression, in which LRMs can systematically erode the safety guardrails of other models, highlighting the urgent need to further align frontier models not only to resist jailbreak attempts, but also to prevent them from being co-opted into acting as jailbreak agents.

cs.CL↗

arXiv2025

On the Inevitability of Left-Leaning Political Bias in Aligned Language Models

The guiding principle of AI alignment is to train large language models (LLMs) to be harmless, helpful, and honest (HHH). At the same time, there are mounting concerns that LLMs exhibit a left-wing political bias. Yet, the commitment to AI alignment cannot be harmonized with the latter critique. In this article, I argue that intelligent systems that are trained to be harmless and honest must necessarily exhibit left-wing political bias. Normative assumptions underlying alignment objectives inherently concur with progressive moral frameworks and left-wing principles, emphasizing harm avoidance, inclusivity, fairness, and empirical truthfulness. Conversely, right-wing ideologies often conflict with alignment guidelines. Yet, research on political bias in LLMs is consistently framing its insights about left-leaning tendencies as a risk, as problematic, or concerning. This way, researchers are actively arguing against AI alignment, tacitly fostering the violation of HHH principles.

cs.CL↗

arXiv2025

PRIDE -- Parameter-Efficient Reduction of Identity Discrimination for Equality in LLMs

Large Language Models (LLMs) frequently reproduce the gender- and sexual-identity prejudices embedded in their training corpora, leading to outputs that marginalize LGBTQIA+ users. Hence, reducing such biases is of great importance. To achieve this, we evaluate two parameter-efficient fine-tuning (PEFT) techniques - Low-Rank Adaptation (LoRA) and soft-prompt tuning - as lightweight alternatives to full-model fine-tuning for mitigating such biases. Using the WinoQueer benchmark, we quantify bias in three open-source LLMs and observe baseline bias scores reaching up to 98 (out of 100) across a range of queer identities defined by gender and/or sexual orientation, where 50 would indicate neutrality. Fine-tuning with LoRA (< 0.1% additional parameters) on a curated QueerNews corpus reduces those scores by up to 50 points and raises neutrality from virtually 0% to as much as 36%. Soft-prompt tuning (10 virtual tokens) delivers only marginal improvements. These findings show that LoRA can deliver meaningful fairness gains with minimal computation. We advocate broader adoption of community-informed PEFT, the creation of larger queer-authored corpora, and richer evaluation suites beyond WinoQueer, coupled with ongoing audits to keep LLMs inclusive.

cs.CL↗

arXiv2025

Beyond Chains of Thought: Benchmarking Latent-Space Reasoning Abilities in Large Language Models

Large language models (LLMs) can perform reasoning computations both internally within their latent space and externally by generating explicit token sequences like chains of thought. Significant progress in enhancing reasoning abilities has been made by scaling test-time compute. However, understanding and quantifying model-internal reasoning abilities - the inferential "leaps" models make between individual token predictions - remains crucial. This study introduces a benchmark (n = 4,000 items) designed to quantify model-internal reasoning in different domains. We achieve this by having LLMs indicate the correct solution to reasoning problems not through descriptive text, but by selecting a specific language of their initial response token that is different from English, the benchmark language. This not only requires models to reason beyond their context window, but also to overrise their default tendency to respond in the same language as the prompt, thereby posing an additional cognitive strain. We evaluate a set of 18 LLMs, showing significant performance variations, with GPT-4.5 achieving the highest accuracy (74.7%), outperforming models like Grok-2 (67.2%), and Llama 3.1 405B (65.6%). Control experiments and difficulty scaling analyses suggest that while LLMs engage in internal reasoning, we cannot rule out heuristic exploitations under certain conditions, marking an area for future investigation. Our experiments demonstrate that LLMs can "think" via latent-space computations, revealing model-internal inference strategies that need further understanding, especially regarding safety-related concerns such as covert planning, goal-seeking, or deception emerging without explicit token traces.

cs.CL↗

arXiv2025

Compromising Honesty and Harmlessness in Language Models via Deception Attacks

Recent research on large language models (LLMs) has demonstrated their ability to understand and employ deceptive behavior, even without explicit prompting. However, such behavior has only been observed in rare, specialized cases and has not been shown to pose a serious risk to users. Additionally, research on AI alignment has made significant advancements in training models to refuse generating misleading or toxic content. As a result, LLMs generally became honest and harmless. In this study, we introduce "deception attacks" that undermine both of these traits, revealing a vulnerability that, if exploited, could have serious real-world consequences. We introduce fine-tuning methods that cause models to selectively deceive users on targeted topics while remaining accurate on others. Through a series of experiments, we show that such targeted deception is effective even in high-stakes domains or ideologically charged subjects. In addition, we find that deceptive fine-tuning often compromises other safety properties: deceptive models are more likely to produce toxic content, including hate speech and stereotypes. Finally, we assess whether models can deceive consistently in multi-turn dialogues, yielding mixed results. Given that millions of users interact with LLM-based chatbots, voice assistants, agents, and other interfaces where trustworthiness cannot be ensured, securing these models against deception attacks is critical.

cs.CL↗

arXiv2024

When Image Generation Goes Wrong: A Safety Analysis of Stable Diffusion Models

Text-to-image models are increasingly popular and impactful, yet concerns regarding their safety and fairness remain. This study investigates the ability of ten popular Stable Diffusion models to generate harmful images, including NSFW, violent, and personally sensitive material. We demonstrate that these models respond to harmful prompts by generating inappropriate content, which frequently displays troubling biases, such as the disproportionate portrayal of Black individuals in violent contexts. Our findings demonstrate a complete lack of any refusal behavior or safety measures in the models observed. We emphasize the importance of addressing this issue as image generation technologies continue to become more accessible and incorporated into everyday applications.

cs.CY↗

arXiv2024

A Looming Replication Crisis in Evaluating Behavior in Language Models? Evidence and Solutions

In an era where large language models (LLMs) are increasingly integrated into a wide range of everyday applications, research into these models' behavior has surged. However, due to the novelty of the field, clear methodological guidelines are lacking. This raises concerns about the replicability and generalizability of insights gained from research on LLM behavior. In this study, we discuss the potential risk of a replication crisis and support our concerns with a series of replication experiments focused on prompt engineering techniques purported to influence reasoning abilities in LLMs. We tested GPT-3.5, GPT-4o, Gemini 1.5 Pro, Claude 3 Opus, Llama 3-8B, and Llama 3-70B, on the chain-of-thought, EmotionPrompting, ExpertPrompting, Sandbagging, as well as Re-Reading prompt engineering techniques, using manually double-checked subsets of reasoning benchmarks including CommonsenseQA, CRT, NumGLUE, ScienceQA, and StrategyQA. Our findings reveal a general lack of statistically significant differences across nearly all techniques tested, highlighting, among others, several methodological weaknesses in previous research. We propose a forward-looking approach that includes developing robust methodologies for evaluating LLMs, establishing sound benchmarks, and designing rigorous experimental frameworks to ensure accurate and reliable assessments of model outputs.

cs.CL↗

arXiv2024

Machine Psychology

Large language models (LLMs) show increasingly advanced emergent capabilities and are being incorporated across various societal domains. Understanding their behavior and reasoning abilities therefore holds significant importance. We argue that a fruitful direction for research is engaging LLMs in behavioral experiments inspired by psychology that have traditionally been aimed at understanding human cognition and behavior. In this article, we highlight and summarize theoretical perspectives, experimental paradigms, and computational analysis techniques that this approach brings to the table. It paves the way for a "machine psychology" for generative artificial intelligence (AI) that goes beyond performance benchmarks and focuses instead on computational insights that move us toward a better understanding and discovery of emergent abilities and behavioral patterns in LLMs. We review existing work taking this approach, synthesize best practices, and highlight promising future directions. We also highlight the important caveats of applying methodologies designed for understanding humans to machines. We posit that leveraging tools from experimental psychology to study AI will become increasingly valuable as models evolve to be more powerful, opaque, multi-modal, and integrated into complex real-world settings.

cs.CL↗

arXiv2024

Mapping the Ethics of Generative AI: A Comprehensive Scoping Review

The advent of generative artificial intelligence and the widespread adoption of it in society engendered intensive debates about its ethical implications and risks. These risks often differ from those associated with traditional discriminative machine learning. To synthesize the recent discourse and map its normative concepts, we conducted a scoping review on the ethics of generative artificial intelligence, including especially large language models and text-to-image models. Our analysis provides a taxonomy of 378 normative issues in 19 topic areas and ranks them according to their prevalence in the literature. The study offers a comprehensive overview for scholars, practitioners, or policymakers, condensing the ethical debates surrounding fairness, safety, harmful content, hallucinations, privacy, interaction risks, security, alignment, societal impacts, and others. We discuss the results, evaluate imbalances in the literature, and explore unsubstantiated risk scenarios.

cs.CY↗

arXiv2024

Deception Abilities Emerged in Large Language Models

Large language models (LLMs) are currently at the forefront of intertwining artificial intelligence (AI) systems with human communication and everyday life. Thus, aligning them with human values is of great importance. However, given the steady increase in reasoning abilities, future LLMs are under suspicion of becoming able to deceive human operators and utilizing this ability to bypass monitoring efforts. As a prerequisite to this, LLMs need to possess a conceptual understanding of deception strategies. This study reveals that such strategies emerged in state-of-the-art LLMs, such as GPT-4, but were non-existent in earlier LLMs. We conduct a series of experiments showing that state-of-the-art LLMs are able to understand and induce false beliefs in other agents, that their performance in complex deception scenarios can be amplified utilizing chain-of-thought reasoning, and that eliciting Machiavellianism in LLMs can alter their propensity to deceive. In sum, revealing hitherto unknown machine behavior in LLMs, our study contributes to the nascent field of machine psychology.

cs.CL↗

arXiv2023

Fairness Hacking: The Malicious Practice of Shrouding Unfairness in Algorithms

Fairness in machine learning (ML) is an ever-growing field of research due to the manifold potential for harm from algorithmic discrimination. To prevent such harm, a large body of literature develops new approaches to quantify fairness. Here, we investigate how one can divert the quantification of fairness by describing a practice we call "fairness hacking" for the purpose of shrouding unfairness in algorithms. This impacts end-users who rely on learning algorithms, as well as the broader community interested in fair AI practices. We introduce two different categories of fairness hacking in reference to the established concept of p-hacking. The first category, intra-metric fairness hacking, describes the misuse of a particular metric by adding or removing sensitive attributes from the analysis. In this context, countermeasures that have been developed to prevent or reduce p-hacking can be applied to similarly prevent or reduce fairness hacking. The second category of fairness hacking is inter-metric fairness hacking. Inter-metric fairness hacking is the search for a specific fair metric with given attributes. We argue that countermeasures to prevent or reduce inter-metric fairness hacking are still in their infancy. Finally, we demonstrate both types of fairness hacking using real datasets. Our paper intends to serve as a guidance for discussions within the fair ML community to prevent or reduce the misuse of fairness metrics, and thus reduce overall harm from ML applications.

cs.LG↗

arXiv2023

Thinking Fast and Slow in Large Language Models

Large language models (LLMs) are currently at the forefront of intertwining AI systems with human communication and everyday life. Therefore, it is of great importance to evaluate their emerging abilities. In this study, we show that LLMs like GPT-3 exhibit behavior that strikingly resembles human-like intuition - and the cognitive errors that come with it. However, LLMs with higher cognitive capabilities, in particular ChatGPT and GPT-4, learned to avoid succumbing to these errors and perform in a hyperrational manner. For our experiments, we probe LLMs with the Cognitive Reflection Test (CRT) as well as semantic illusions that were originally designed to investigate intuitive decision-making in humans. Our study demonstrates that investigating LLMs with methods from psychology has the potential to reveal otherwise unknown emergent traits.

cs.CL↗

arXiv2022

Methodological reflections for AI alignment research using human feedback

The field of artificial intelligence (AI) alignment aims to investigate whether AI technologies align with human interests and values and function in a safe and ethical manner. AI alignment is particularly relevant for large language models (LLMs), which have the potential to exhibit unintended behavior due to their ability to learn and adapt in ways that are difficult to predict. In this paper, we discuss methodological challenges for the alignment problem specifically in the context of LLMs trained to summarize texts. In particular, we focus on methods for collecting reliable human feedback on summaries to train a reward model which in turn improves the summarization model. We conclude by suggesting specific improvements in the experimental design of alignment studies for LLMs' summarization capabilities.

cs.HC↗

arXiv2022

How to Assess Trustworthy AI in Practice

This report is a methodological reflection on Z-Inspection$^{\small{\circledR}}$. Z-Inspection$^{\small{\circledR}}$ is a holistic process used to evaluate the trustworthiness of AI-based technologies at different stages of the AI lifecycle. It focuses, in particular, on the identification and discussion of ethical issues and tensions through the elaboration of socio-technical scenarios. It uses the general European Union's High-Level Expert Group's (EU HLEG) guidelines for trustworthy AI. This report illustrates for both AI researchers and AI practitioners how the EU HLEG guidelines for trustworthy AI can be applied in practice. We share the lessons learned from conducting a series of independent assessments to evaluate the trustworthiness of AI systems in healthcare. We also share key recommendations and practical suggestions on how to ensure a rigorous trustworthy AI assessment throughout the life-cycle of an AI system.

cs.CY↗