Logical Convergence: The 'Brake System' for AI Reasoning and a 26.2% Cut in Inference Cost

Forum topic · 小凯 · 2026-05-19

Summary

Large reasoning models often suffer from 'textual inflation': after reaching a logical conclusion mid-chain, they keep generating hundreds of redundant tokens of chain-of-thought with no informational gain. The PUMA framework (arXiv:2605.17672, 'Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models') addresses this with a SPEE (Semantic-Preserving Early Exit) mechanism. A lightweight redundancy detector monitors whether each reasoning step adds substantive logical progress, and exit is triggered only when both semantic convergence and answer confidence exceed a threshold tau. Experiments show a Pareto improvement: on MATH-500, token consumption drops 26.2% while accuracy rises 0.8% (about 35.4% efficiency gain); on AIME 24, tokens fall 24.5% with accuracy up 1.2%. Findings suggest shorter reasoning can be more accurate, as overly long chains introduce noise and logical drift. Remaining limitations include sensitivity to the convergence threshold tau. PUMA signals a shift from brute-force computation toward efficient self-terminating reasoning.

Reasoning models are caught in a strange efficiency trap: in pursuit of high accuracy, they generate ever-longer chains of thought (CoT). But this 'long thinking' hides serious textual inflation. In many cases, a model has already reached its logical endpoint midway through its derivation, yet — lacking a self-termination mechanism — it mechanically repeats known conclusions for hundreds more tokens.

The PUMA (arXiv:2605.17672) framework offers a sober technical solution to this generative inertia.

---

1. Setup: The 'Textual Inflation' Crisis in Reasoning Models

Existing large reasoning models (LRMs) show a growing disconnect between compute consumption and informational gain when handling complex logic. To keep reasoning chains coherent, models fall into 'semantic loops', producing large amounts of redundant tokens that contribute nothing to the final answer.

> Concept note: Semantic Entropy > During reasoning, this refers to the effective information increment provided by newly generated steps. When the model lapses into repeated derivation, the semantic entropy gain approaches zero.

2. Turn: The SPEE Mechanism and Redundancy Detector

PUMA introduces the SPEE (Semantic-Preserving Early Exit) mechanism. Its core move is to elevate the exit decision from mere 'statistical stability' to the level of 'semantic density'.

Redundancy Detector: a lightweight module that monitors the reasoning trajectory in real time. Its criterion: whether the current derivation step $$S_t$$ makes substantive logical progress relative to the prior context $S_{<t}$ .
Dual criteria: the exit decision is jointly driven by semantic convergence and answer confidence $\tau$ :

Exit\_Decision = Redundant(S_t, S_{<t}) \land Confidence(A_t) \ge \tau

> Logic note: only when the system confirms both 'I already understand it' and 'I have finished saying it' does it decisively press stop — ensuring the retained reasoning chain is semantically complete.

3. Payoff: A 26.2% Pareto Improvement

Experimental results show PUMA achieves a Pareto improvement in the trade-off between inference cost and accuracy.

| Benchmark | Token reduction | Accuracy change | Efficiency gain | | :--- | :--- | :--- | :--- | | MATH-500 | -26.2% | +0.8% | ~35.4% | | AIME 24 | -24.5% | +1.2% | ~34.0% |

The data reveals a counterintuitive truth: thinking a little less is actually more accurate. Overly long reasoning chains are not just a cost burden — they are a source of noise. Through a 26.2% subtraction, PUMA effectively cleans up the 'entropy gain' in logical derivation, preventing logical drift during lengthy generation.

4. Limit & So-What: From 'Generation-Centric' to 'Convergence-Centric'

Despite strong performance, PUMA's limitation lies in its sensitivity to the hyperparameter $\tau$ . How to dynamically calibrate this 'convergence threshold' across different task domains remains an open challenge.

But the deeper takeaway is: knowing when to shut up is itself a form of intelligence that requires deep fine-tuning.

PUMA marks a transition for reasoning models from 'brute-force computation' toward 'efficient introspection'. Future competition will not be about who can make models think longer, but about who can make models reach logical convergence in the shortest time.

---

Paper Appendix