How to Steal Reasoning Without Reasoning Traces

Tingwei Zhang
Cornell Tech, Computer Science

Time: 2026-11-04, 12:00 - 13:00 ET
Location: Rice 540 and Zoom

Abstract Reasoning traces are increasingly valuable training data, motivating frontier models to hide their internal chain-of-thought and expose only summaries or final answers. In this talk, I will introduce Trace Inversion, an attack that asks whether reasoning capability can still be recovered when the hidden trace itself is completely inaccessible. Using traces from a small open-source surrogate, we train an inverter that reconstructs useful reasoning supervision from a stronger victim’s observable outputs. These synthetic traces substantially improve downstream student reasoning, including when attacking black-box commercial models and when no reasoning summary is available. Surprisingly, synthetic traces can sometimes outperform the victim’s original traces as training data. Together, these results suggest that hiding reasoning traces does not necessarily hide reasoning capability.

Bio: Tingwei Zhang is a fouth-year PhD student in Computer Science at Cornell Tech, advised by Professor Vitaly Shmatikov. His research focuses on security and privacy challenges in machine learning technologies, particularly in real-world scenarios and under adversarial conditions, with the goal of developing secure, ethical, and privacy-preserving AI systems.