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

• URL: http://arxiv.org/abs/2606.30852v2
• Date: Sat, 04 Jul 2026 03:36:21 GMT
• Title: When Does Learning to Stop Help? A Cost-Aware Study of Early Exits in Reasoning Models
• Abstract summary: LearnStop is a hidden-state-free logistic stopper over prefix-observable features.<n>A trajectory decomposition predicts the regime: learning pays where answers oscillate and correctness evidence is spread.<n>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.
• Score: 4.8190992438931035
• License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/
• Abstract: 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.

Related papers

• The Dice Roll Method: A Standardized Protocol for Repeated-Query Auditing of Large Language Model Brand Recommendations [0.0]
  We formalize the Dice Roll Method as a reusable protocol for repeated-query auditing of brand recommendations.<n>Three tiers of iteration guidance emerge from the D-study: exploratory, confirmatory, and rigorous.<n>A pre-registered external validation on three independent corpora reproduces the D-study reliability prediction in 37 of 39 cells with no failures.
  arXiv  Detail & Related papers  (2026-09-03T16:22:06Z)
• Knowing When to Ask for Help: Bayesian Self-Escalation in Hierarchical LLM Agents [0.0]
  Current agent systems decide delegation before reasoning begins.<n>We study a third regime: an agent that recognises, during its own reasoning, that it is unlikely to succeed.
  arXiv  Detail & Related papers  (2026-08-25T05:35:07Z)
• DreamQAS: Learning a Decision-Useful World Model for VQE-Efficient Quantum Architecture Search [9.679776057194056]
  Reinforcement-learning-based quantum architecture search (RL-QAS) repeatedly optimize a variational quantum eigensolver (VQE) after extending a circuit.<n>We introduce DreamQAS, a model-based RL framework that preserves exact circuit dynamics and learns only the expensive post-VQE feedback.
  arXiv  Detail & Related papers  (2026-07-31T14:58:23Z)
• SVR: Self-Verifying Refinement via Joint Verdict-Confidence Reinforcement Learning for Adaptive Test-Time Compute [62.3458279176813]
  Self-Verifying Refinement is an oracle-free multi-turn reinforcement learning framework.<n>It learns to use self-verification as a compute-control policy.<n>It achieves a macro-average accuracy of 0.563 with only 2.99 inference turns on average.
  arXiv  Detail & Related papers  (2026-07-30T16:20:58Z)
• Know Before You Fetch: Calibrated Retrieval-Budget Allocation for Retrieval-Augmented Generation [10.60293484295824]
  We formulate adaptive RAG as calibrated retrieval-budget allocation.<n>We calibrate sequence log-probability and prefix-logit uncertainty signals into probabilities of correctness.<n>For sequence log-probability, ECE drops from 0.275 to 0.062 on TriviaQA, 0.643 to 0.009 on NQ, and 0.711 to 0.031 on MS MARCO.
  arXiv  Detail & Related papers  (2026-06-29T08:36:33Z)
• TASR: Training-Free Adaptive Stopping for Iterative Retrieval [12.177557521540082]
  Iterative retrieval-augmented generation agents commonly overspend by continuing to retrieve after the model has converged on an answer.<n>We propose TASR (Training-Free Adaptive Stopping Rule), a one-line predicate that fires when the model repeats its previous-round normalized answer.
  arXiv  Detail & Related papers  (2026-06-11T18:35:14Z)
• Unveiling the Entropy Dynamics of Chain-of-Thought Reasoning [81.57028614960576]
  We show that the Confidence Region possesses two critical properties: 1) High Reliability -- answers in the confidence region become highly accurate and stable, and 2) High Redundancy -- models generate unnecessary tokens long after reaching the correct answer.<n>These properties unlock more efficient and reliable inference strategies.
  arXiv  Detail & Related papers  (2026-06-01T10:11:14Z)
• Understanding and Mitigating Premature Confidence for Better LLM Reasoning [76.16007941549857]
  Long chains of thought (CoT) from current language models frequently contain logical gaps and unjustified leaps.<n>We find such a signal in how the model's confidence evolves during reasoning.<n>We exploit this in progressive confidence shaping, a reinforcement learning objective that trains models to update their confidence as they reason rather than commit early.
  arXiv  Detail & Related papers  (2026-05-23T04:42:45Z)
• Step-wise Rubric Rewards for LLM Reasoning [72.17879367869503]
  Reinforcement Learning with Verifiable Rewards (RLVR) is widely used to improve reasoning in large language models.<n>We find 18.2% of steps in correct responses are wrong yet positively rewarded.<n>Across six mathematical reasoning benchmarks, SRaR improves average accuracy over RaR by 3.57 points.
  arXiv  Detail & Related papers  (2026-05-17T07:08:14Z)
• Pause and Reflect: Conformal Aggregation for Chain-of-Thought Reasoning [8.024041325202612]
  Chain-of-thought (CoT) reasoning with self-consistency improves performance by aggregating multiple sampled reasoning paths.<n>We introduce a conformal procedure for CoT reasoning that directly addresses aggregation uncertainty.<n>Our approach replaces majority voting with weighted score aggregation over reasoning paths and calibrates an abstention rule using conformal risk control.
  arXiv  Detail & Related papers  (2026-05-13T20:33:59Z)
• Confidence-Aware Alignment Makes Reasoning LLMs More Reliable [65.44962502963378]
  We introduce CASPO, a framework that aligns token-level confidence with step-wise logical correctness without training a separate reward model.<n>During inference, we propose Confidence-aware Thought (CaT), which dynamically prunes uncertain reasoning branches with negligible O(V) latency.<n>Experiments across ten benchmarks and multiple model families show that CASPO consistently improves reasoning reliability and inference efficiency.
  arXiv  Detail & Related papers  (2026-05-08T07:08:25Z)
• Correction and Corruption: A Two-Rate View of Error Flow in LLM Protocols [51.56484100374058]
  We propose a paired-outcome measurement interface for auditing a single protocol step on exact-match tasks.<n>For each instance, the interface records a baseline correctness bit and a post-step correctness bit.<n>These rates predict accuracy changes and define a reusable empirical interface testable across seeds, mixtures, and pipelines.
  arXiv  Detail & Related papers  (2026-04-20T13:25:40Z)
• ZIP-RC: Optimizing Test-Time Compute via Zero-Overhead Joint Reward-Cost Prediction [57.799425838564]
  We present ZIP-RC, an adaptive inference method that equips models with zero-overhead inference-time predictions of reward and cost.<n> ZIP-RC improves accuracy by up to 12% over majority voting at equal or lower average cost.
  arXiv  Detail & Related papers  (2025-12-01T09:44:31Z)
• Uncertainty-Guided Checkpoint Selection for Reinforcement Finetuning of Large Language Models [27.97382399449914]
  Reinforcement learning (RL) finetuning is crucial to aligning large language models (LLMs), but the process is notoriously unstable.<n>In practice, selecting the best checkpoint is challenging: evaluating checkpoints on the validation set during training is computationally expensive and requires a good validation set.<n>We introduce an uncertainty-guided approach for checkpoint selection (UGCS) that avoids these pitfalls.
  arXiv  Detail & Related papers  (2025-11-13T01:46:58Z)
• Sample Smart, Not Hard: Correctness-First Decoding for Better Reasoning in LLMs [72.82403830490084]
  We argue that the decoding rule should be calibrated by correctness, not confidence alone.<n>We propose simple strategies that achieve this goal: Greedy-Threshold makes sampling greedy at very low confidence steps.<n>Together, our findings challenge prevailings about decoding under uncertainty and show gains across math and general reasoning benchmarks.
  arXiv  Detail & Related papers  (2025-10-07T14:46:12Z)
• Reinforcement Learning with Verifiable yet Noisy Rewards under Imperfect Verifiers [90.50039419576807]
  Reinforcement Learning with Verifiable Rewards (RLVR) trains policies against automated verifiers to avoid costly human labeling.<n>To reduce vulnerability to verifier hacking, many RLVR systems collapse rewards to binary $0,1$ during training.<n>This choice carries a cost: it introduces textitfalse negatives (rejecting correct answers, FNs) and textitfalse positives (accepting incorrect ones, FPs)
  arXiv  Detail & Related papers  (2025-10-01T13:56:44Z)
• Entropy After $\langle \texttt{/Think} \rangle$ for reasoning model early exiting [38.93424884988798]
  We show that large reasoning models overthink, continuing to revise answers even after reaching the correct solution.<n>We propose Entropy After /Think> (EAT) for monitoring and deciding whether to exit reasoning early.<n>EAT reduces token usage by 13 - 21% without harming accuracy.
  arXiv  Detail & Related papers  (2025-09-30T16:59:37Z)
• Centaur: Robust End-to-End Autonomous Driving with Test-Time Training [84.78837437133234]
  We propose Centaur, which updates a planner's behavior via test-time training without relying on hand-engineered rules or cost functions.<n>We develop a novel uncertainty measure, called Cluster Entropy, which is simple, interpretable, and compatible with state-of-the-art planning algorithms.
  arXiv  Detail & Related papers  (2025-03-14T17:59:41Z)
• Upper Confidence Primal-Dual Reinforcement Learning for CMDP with Adversarial Loss [145.54544979467872]
  We consider online learning for episodically constrained Markov decision processes (CMDPs) We propose a new emphupper confidence primal-dual algorithm, which only requires the trajectories sampled from the transition model. Our analysis incorporates a new high-probability drift analysis of Lagrange multiplier processes into the celebrated regret analysis of upper confidence reinforcement learning.
  arXiv  Detail & Related papers  (2020-03-02T05:02:23Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.