Liv d'Aliberti

Do reasoning models have real “Aha!” moments—mid-chain realizations where they intrinsically self-correct?

In a new pre-print, “The Illusion of Insight in Reasoning Models," led by @liv-daliberti.bsky.social we provide strong evidence that they do not!

📜: arxiv.org/abs/2601.00514

05.01.2026 19:39 — 👍 102    🔁 16    💬 7    📌 5

Prompt per domain are pages 16-18. The same system prompt was used for MATH-500 in all cases: Qwen1.5-B, Qwen7B, Llama-8B, DeepSeek-R1, ChatGPT.

05.01.2026 22:10 — 👍 2    🔁 0    💬 1    📌 0