Berkeley Talk: LLM Reasoning Has Structure; Confidence-Based Stopping Cuts Thinking Tokens by 25%
datawithsuman · x · 2026-10-11
In a Simons Institute talk at UC Berkeley, Soheil Feizi argues that thinking tokens are becoming a new scaling axis for AI — but more thinking doesn't mean smarter thinking. The talk covers four recent papers:
- Reasoning has structure: traces share a common "heartbeat" and recurring loops that can predict correctness (ThinkARM, ACL 2026 Oral).
- Better reasoning without reward-model training: a larger model guiding a smaller one's reasoning matches a trained process reward model (EMNLP 2026 Main).
- Knowing when to stop matters: models can learn to stop based on confidence, using up to 25% fewer thinking tokens at matched accuracy (ConfSFT).
- Thinking harder isn't always better (details truncated in the source post).
The takeaway: the predictable structure of reasoning traces can be exploited for efficiency and reliability rather than simply scaling up thinking.
More from Models
- Last three models on OpenRouter are all labeled as 'decision' models — adonis_singh · 2026-10-11
- Anthropic hands out Claude API credits to users — daluoseo · 2026-10-11
- Dev lists OpenAI product questions that went unanswered for months despite its chatty employees — Angaisb_ · 2026-10-11
- Researcher: Copilot Premium turned one-line data analysis into long-winded prompt begging — lescarr · 2026-10-11
- Odyssey-3: a 14B generative world model enters free public research preview — The Decoder · 2026-10-11
- Researcher questions whether 'release first, improve later' metrics are worth hillclimbing — suchenzang · 2026-10-11