# When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models

arXiv: https://arxiv.org/abs/2606.30852 · HTML: https://mlsanity.com/paper/2606.30852/

## Key facts

- Cited by 2 papers in the map (data through 14 Oct 2026).
- Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning.
- Builds on 15 of its 18 references that are in the map.
- First posted on arXiv 29 Jun 2026.

## Placed in

- LLM reasoning / [Test-time compute and efficient thinking](https://mlsanity.com/lr/test-time-compute-and-efficient-thinking/)
- Efficient LLM inference and serving / [Efficient reasoning and agent inference](https://mlsanity.com/inference/efficient-reasoning-and-agent-inference/)

## Builds on (most cited first, 15 of 15)

- [Training Verifiers to Solve Math Word Problems](https://mlsanity.com/paper/2110.14168/) (2021)
- [Measuring Mathematical Problem Solving With the MATH Dataset](https://mlsanity.com/paper/2103.03874/) (2021)
- [Let's Verify Step by Step](https://mlsanity.com/paper/2305.20050/) (2023)
- [GPQA: A Graduate-Level Google-Proof Q&A Benchmark](https://mlsanity.com/paper/2311.12022/) (2023)
- [MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmark](https://mlsanity.com/paper/2406.01574/) (2024)
- [s1: Simple test-time scaling](https://mlsanity.com/paper/2501.19393/) (2025)
- [Conformal Risk Control](https://mlsanity.com/paper/2208.02814/) (2022)
- [Token-Budget-Aware LLM Reasoning](https://mlsanity.com/paper/2412.18547/) (2024)
- [Dynamic Early Exit in Reasoning Models](https://mlsanity.com/paper/2504.15895/) (2025)
- [Conformal Language Modeling](https://mlsanity.com/paper/2306.10193/) (2023)
- [Entropy After </Think> for reasoning model early exiting](https://mlsanity.com/paper/2509.26522/) (2025)
- [Thought calibration: Efficient and confident test-time scaling](https://mlsanity.com/paper/2505.18404/) (2025)
- [Conformal Thinking: Risk Control for Reasoning on a Compute Budget](https://mlsanity.com/paper/2602.03814/) (2026)
- [Stop When Reasoning Converges: Semantic-Preserving Early Exit for Reasoning Models](https://mlsanity.com/paper/2605.17672/) (2026)
- [TERMINATOR: Learning Optimal Exit Points for Early Stopping in Chain-of-Thought Reasoning](https://mlsanity.com/paper/2603.12529/) (2026)

## Built on by (most cited first, 2 of 2)

- [Stable Answers, Unfinished Reasoning: Why Self-Consensus Is Not a Safe Early-Exit Signal](https://mlsanity.com/paper/2609.09989/) (2026)
- [Selective Reasoning for Typed Decision Models: When to Start, Read, and Stop](https://mlsanity.com/paper/2610.19644/) (2026)

## Abstract (arXiv)

Reasoning models spend test-time compute unevenly across instances, and a growing family of early-exit rules -- confidence thresholds, entropy monitors, answer-stability checks, and learned stoppers -- promises to reclaim the waste. These rules, however, are evaluated under heterogeneous protocols that leave the deployment question unanswered: at a fixed tolerance for losing correct answers, which policy saves more compute, and does the saving survive probe overhead? We answer this question with a controlled study across 18 task-model settings spanning GSM8K, MATH-500, MMLU-Pro, AIME-90, and GPQA on Qwen3 and DeepSeek-R1-distilled models, using LearnStop, a hidden-state-free logistic stopper over prefix-observable features, as the learned policy instrument. Under matched lost-correct risk at α = 0.15, with the scalar competitor selected on calibration data from confidence, entropy, confidence-leap, and run-stability exits, the answer forms three regimes. Learned stopping wins on all four primary Qwen3 free-form math settings (+3.2 to +21.2 pp additional total-token saving); calibrated scalar exits win on multiple-choice MMLU-Pro; and small hard benchmarks (AIME-90, GPQA) admit no certifiable aggressive policy at all. A trajectory decomposition predicts the regime: learning pays where answers oscillate and correctness evidence is spread across complementary signals, while a single confidence threshold suffices where most instances are already solved at the first checkpoint. Cost accounting sharpens the picture further -- the same policy that saves 32% of tokens under KV-cache forking costs 121% extra under black-box repeated prefilling. Together, these results replace the single-method race with a decision procedure for choosing a stopping rule from the trajectory structure and serving regime of the target workload.