Cagri Yalcin retweetledi

Long reasoning can still hide a bad answer. What matters is not just whether an agent explains itself, but whether its reasoning trace contains signals of failure before the output ships. Auditing that gap is where agent trust gets real.
Sean Du@xuefeng_du
1/ 🧠 Long reasoning ≠ reliable reasoning. Large reasoning models can write long, convincing chains of thought… and still end with a wrong answer. Our new @icmlconf paper asks: Can we use the reasoning trace itself to detect when the final answer is hallucinated? 🧵👇
English




