6 papers
Noise Injection Reveals Hidden Capabilities of Sandbagging Language Models
Cameron Tice, Philipp Alexander Kreer, Nathan Helm-Burger +7
Capability evaluations play a crucial role in assessing and regulating frontier AI systems. The effectiveness of these evaluations faces a significant challenge: strategic underper…
CTRL-ALT-DECEIT: Sabotage Evaluations for Automated AI R&D
Francis Rhys Ward, Teun van der Weij, Hanna Gábor +6
AI systems are increasingly able to autonomously conduct realistic software engineering tasks, and may soon be deployed to automate machine learning (ML) R&D itself. Frontier AI sy…
Stress Testing Deliberative Alignment for Anti-Scheming Training
Bronson Schoen, Evgenia Nitishinskaya, Mikita Balesni +16
Highly capable AI systems could secretly pursue misaligned goals -- what we call "scheming". Because a scheming AI would deliberately try to hide its misaligned goals and actions,…
The Elicitation Game: Evaluating Capability Elicitation Techniques
Felix Hofstätter, Teun van der Weij, Jayden Teoh +3
Capability evaluations are required to understand and regulate AI systems that may be deployed or further developed. Therefore, it is important that evaluations provide an accurate…
SHADE-Arena: Evaluating Sabotage and Monitoring in LLM Agents
Jonathan Kutasov, Yuqi Sun, Paul Colognese +9
As Large Language Models (LLMs) are increasingly deployed as autonomous agents in complex and long horizon settings, it is critical to evaluate their ability to sabotage users by p…
AI Sandbagging: Language Models can Strategically Underperform on Evaluations
Teun van der Weij, Felix Hofstätter, Ollie Jaffe +2
Trustworthy capability evaluations are crucial for ensuring the safety of AI systems, and are becoming a key component of AI regulation. However, the developers of an AI system, or…