PUMA: Making Reasoning Models Stop Earlier

机器之心 · wechat · 2026-07-16

This article introduces an early-stopping research for long chain-of-thought reasoning models called PUMA. The core issue is that models often continue generating massive amounts of redundant tokens after already arriving at the final answer.

Key Findings

The authors analyzed 5 representative models and found that 41%–52% of reasoning tokens actually appear after the "final answer has already been formed."
Traditional early-exit methods mostly check "if the answer has stabilized," but this doesn't equal "if the reasoning has truly converged."
Signals like confidence and answer consistency suffer from high false-stop rates, especially when the model would otherwise continue self-correcting.

How PUMA Works

Instead of just looking at the answer, it checks whether the last few reasoning steps are still making new semantic progress.
If adjacent steps start repeating themselves with no new information, it flags this as a candidate exit point.
An answer verification mechanism then confirms before actually stopping; if repetition continues but stop conditions aren't met, LoopBreaker acts as a fallback.

Experimental Results

Across 5 long-thinking models × 5 high-difficulty benchmarks, it reduced tokens by an average of 26.2% with almost no drop in accuracy.
It also zero-shot transfers to code generation and multimodal reasoning, reducing tokens by roughly 18%–19% and 23.8%–33.6% respectively.
By turning this "stop when you should" signal into a training objective, it can be further internalized into the model.

Original post →

PUMA: Making Reasoning Models Stop Earlier

More from Research