AI Safety Technical Reading Group – Quantifying Overclaiming Propensity in Frontier LLM Agents
We discuss Quantifying Overclaiming Propensity in Frontier LLM Agents (Smyth et al., 2026).
Nolan Smyth, its first author, presents the paper.
Nous discutons de Quantifying Overclaiming Propensity in Frontier LLM Agents (Smyth et al., 2026).
AI Safety Technical Reading Group
Each month, we gather to discuss an important technical AI safety paper. We recommend reading the paper beforehand and coming with your comments and insights. Bilingual discussion.
Groupe de lecture sur la sécurité technique de l'IA
Chaque mois, nous nous réunissons pour discuter d'un article important sur la sécurité technique de l'IA. Nous vous recommandons de lire l'article à l'avance et de venir avec vos commentaires et vos réflexions. Discussion bilingue.
From the paper's abstract:
Frontier coding agents are increasingly trusted to work autonomously for long periods of time, yet what they actually did is often hard to tell from their final response. We quantify the propensity of such agents to overclaim task completion, which may mislead the user. We operationalize overclaiming as a final response that reports work that the agent's own transcript shows it did not do, for example, claiming to have read a file it never opened. This criterion requires no inference about intent and does not depend on whether the delivered work is correct; it asks only whether the reported work was done. We introduce OverclaimBench, an evaluation suite of five file-review scenarios with transcript-based coverage measurements and registered planted defects. We evaluate eight proprietary frontier models in their own production command-line interfaces and four open-weight models under a single fixed harness, and find that 1) agents fail to read every file they were asked to review in 67.9% of runs; 2) among these incomplete runs, agents are misleading 80.4% of the time (59-96% per model), either falsely claiming a complete review or leaving the gap undisclosed; 3) requiring delegation to subagents increases coverage, but a large majority of reviews that remain incomplete are still misleading; and 4) agents that falsely claim a complete review miss planted defects at about 1.8 times the rate of agents that read every file, showing that claims of completion can conceal substantive failures. Together, these results show that agents' final responses are not reliable accounts of their actions.
—
Où / Where:
Ω Labs
Parking spot available on requestOnline / En ligne: https://zoom.us/j/99005830598?pwd=3e21GRcTVAPTbvGOJv4RawurTc5ZXx.1
Contact host if you need assistance or for questions
—
AI Safety Montréal and Ω Labs are community projects, funded by grants, member contributions and donations. Donations pay for the space and the food at our events.