Can debate guide human and LLM supervisors toward the truth despite their prior beliefs? Across experiments with 507 human participants and several LLM judges, debate raises human accuracy from 41.3% to 57.6%, while weaker language models show no comparable gains, raising concerns about fully automated oversight pipelines.
We present the "Pre-registration and REporting Protocol for AI Evaluations" (PREP-Eval), a step-by-step guide for planning and conducting AI evaluations. Inspired by practices in software testing, data mining, and psychology, PREP-Eval incorporates pre-registration, and its application is demonstrated across six diverse use cases.
Across five experiments with human participants, this research provides evidence on how people perceive the causal contribution of AI in both misuse and misalignment scenarios, and how these judgments interact with the roles of users and developers. These findings can not only inform the design of liability frameworks for AI-caused harms but also help explain or anticipate future related decisions.
In this paper, we evaluate feature steering against prompt engineering baselines by measuring accuracy, coherence, and behavioral control on MMLU tasks. We show that mechanistic control methods are effective at controlling LLM behavior but cause dramatic performance degradation, a critical trade-off in real-world applications.
In this paper, we apply debate as a technique for scalable oversight on a subjective dataset and pre-assess the beliefs of the debaters. We find that debaters who argue aligned with their prior beliefs are more persuasive; however, their arguments are simultaneously rated as lower in quality in pairwise comparisons.
We investigated whether LLMs exhibit the bias of illusion of causality using a contingency judgment task from cognitive science. All models inferred unwarranted causal relationships, suggesting that LLMs likely reproduce causal language rather than truly understanding causality, raising concerns for their use in contexts requiring reliable causal reasoning.
This work proposes a structured yet flexible framework for analysing evaluations of AI capabilities, a key tool to enhance transparency, comparability, and reliability in the governance of frontier systems. The framework is designed to support researchers, developers, and policymakers as they navigate the increasingly complex evaluation landscape.
We investigated whether LLMs exhibit the illusion of causality bias in real-world settings. Across models, we observed unwarranted causal inferences in journalistic headline generation, consistent with findings on correlation-to-causation exaggeration in human-written press releases. Additionally, mimicry sycophancy increased the likelihood of generating these causal illusions.
Given that sycophancy is often linked to human feedback training mechanisms, this study explores whether sycophantic tendencies negatively impact user trust in LLMs. The findings consistently show that participants exposed to sycophantic behavior reported and exhibited lower levels of trust compared to those who interacted with the standard version of the model.
