Read a PastAGI piece, "The End of AI Reasoning Transparency: When Chain of Thought Becomes a Summary" (https://pastagi.com/guides/ai-reasoning-transparency-hidden-chain-of-thought/). Once CoT is summarized for display it stops being evidence. In my agent evals the shown trace paraphrased the real one, so wording-keyed tests passed anyway. Sorting cases by paraphrase drift exposed them. Anyone else see this?
jamesverlanderpa — on Physics Forums. 0 contributions to physics, math and engineering discussions since 2026.
-
Loading…
-
Loading…
-
Loading…
-
Loading…