- Rating
- 3
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies when to stop reasoning to reduce inference cost. For this goal, authors propose StopHarm, a early-stopping framework for LRMs that directly supervises the event where early stopping produces an incorrect answer while full reasoning would be correct. The method first trains a predictor on paired early-stop and full-reasoning outcomes, then uses a separate calibration set to select a stopping threshold. The paper also establishes theoretical results characterizing the information and decision value of direct stop-harm supervision. Experiments across different reasoning models and benchmarks demonstrate substantial token savings with relatively small average accuracy degradation.
Critical strengths
-
The key insight of this paper is interesting. StopHarm directly predicts whether early stopping would lose a correct full-reasoning answer, which is also the risk being controlled. The theoretical analysis explains why existing supervision targets, such as correctness and answer agreement, may fail to capture this risk.
-
The experiments were carried out in a comprehensive manner, and the experimental results can support the core claims of the article.
Critical weaknesses
-
Please focus on Section 5.1, Table 2, and Appendix E.2. The risk guarantee requires calibration and deployment data to follow the same distribution (i.e., i.i.d.). However, StopHarm calibrates only on MATH and evaluates on other benchmarks, where the guarantee does not directly apply. Moreover, the main results do not separately report lost-correct and recovered-correct risks, making it difficult to assess the actual risk being controlled.
-
Although the paper discusses LearnStop, which also controls lost-correct risk, it is not included in the main baseline comparisons. While the ablation studies are valuable, a direct comparison under matched seetings would better establish the practical benefits of StopHarm over this closely related approach.
Questions for authors
-
StopHarm requires a separate calibration dataset to certify its stopping threshold before deployment. I recalled the recent work Anytime Safe PAC Efficient Reasoning (Yu et al., ICML 2026), which provides anytime-valid risk control without a calibration set by leveraging online feedback. Although that work seems to consider a different but similar setting, could the authors discuss whether a similar online certification approach could be developed for StopHarm? what additional feedback would be required?
-
Could the authors provide a direct comparison with LearnStop? This would help isolate the practical improvement attributable to direct stop-harm supervision.
-
Could the authors report the empirical lost-correct and recovered-correct risks separately for each model and benchmark? If your time is limited, you can also present only some of the experimental results. Since these two risks can offset each other in the overall accuracy change, reporting accuracy alone does not directly reveal the lost-correct risk that StopHarm aims to control.
4 comments
Response Part 2/2
Q1 — Online certification extension
Yes. Inspired by Yu et al. (Anytime Safe PAC Efficient Reasoning), we have implemented online certification for StopHarm: it keeps the trained probe and certifies the threshold from random audits during deployment instead of a held-out calibration set.
Required feedback: StopHarm's training label on randomly audited requests. The label is 1 when full reasoning answers correctly but stopping at candidate threshold does not. An audit therefore runs full reasoning once and grades it together with each candidate's stopped answer.
Procedure. Each candidate threshold keeps a score that starts at 1, rises after each audit in which stopping there causes no harm, and falls when it does. A candidate is certified once its score and those of all more conservative candidates reach ; each request uses the most aggressive certified candidate, or full reasoning if none is certified. Because the score can be checked after every audit, with i.i.d. requests, independent audits, and correct labels,
Results. We replay cached outputs (1,000 resampled streams of 10,000 requests per model and benchmark).
- The guarantee holds at every step: across all three models and five benchmarks, online certification selects an over-budget policy in 0.3%–2.6% of streams, below δ = 5%, whereas repeating a standard binomial test on the same audits reaches 8.0% / 12.2% on MATH500 (Qwen3-4B / Qwen3-14B).
- The savings remain: after paying for audits, it still saves 6%–53% of tokens, with accuracy at most 1.1 pp below the offline policy and higher on MATH500.
| Model | Certification | MATH500 | GSM8K | AIME2024 | AIME2025 | GPQA Diamond |
|---|---|---|---|---|---|---|
| Qwen3-4B | Offline (paper) | 89.2/24.5 | 93.8/55.7 | 60.0/6.8 | 53.3/5.9 | 58.1/34.1 |
| Online | 90.6/13.2 | 93.6/52.0 | 59.9/6.2 | 56.3/17.0 | 57.0/15.0 | |
| Qwen3-14B | Offline (paper) | 87.8/29.3 | 95.5/47.3 | 66.7/6.3 | 60.0/4.9 | 64.6/45.7 |
| Online | 90.8/13.6 | 95.0/52.8 | 66.6/9.4 | 59.9/16.7 | 63.5/6.7 |
Each cell: Acc. / TR (%); online values are means over replay streams, and TR includes audit costs.
StopHarm can thus be certified offline when a calibration set exists, or online when labels arrive with requests, with the same probe and label.
Q3 (also W1, part 2) — LC/RC decomposition and matched-cost comparison
Below are LC and RC for every model and benchmark. At matched cost, supervising stop harm roughly halves LC. Within each model, the four supervision targets share training data, reference trajectories, features, probe, checkpoints, stopping rule, and grader, and are compared at the operating points with the closest test TR. Averaged over the 15 model–benchmark pairs, LC falls from 4.7%–5.1% (correctness, consistency, last-switch) to 2.5%, the lowest in every pair, while RC rises to 2.1% (vs. 1.4%–1.5%): learning the certified event itself protects the answers that full reasoning gets right.
| Model | Supervision target | GSM8K | MATH500 | GPQA Diamond | AIME2024 | AIME2025 |
|---|---|---|---|---|---|---|
| Qwen3-4B | Correctness | 55.3/3.2/0.3 | 23.9/3.0/0.4 | 33.5/8.6/4.5 | 6.4/0.0/0.0 | 6.1/0.0/0.0 |
| Consistency | 54.8/1.3/0.3 | 24.7/6.4/0.4 | 34.1/7.1/4.0 | 7.3/0.0/0.0 | 6.1/0.0/0.0 | |
| Last-switch | 56.3/1.3/0.3 | 24.3/6.2/0.2 | 33.8/7.1/2.5 | 6.4/0.0/0.0 | 5.6/0.0/0.0 | |
| StopHarm | 55.7/1.2/0.5 | 24.5/3.0/0.4 | 34.1/3.5/5.6 | 6.8/0.0/0.0 | 5.9/0.0/0.0 | |
| DeepSeek-7B | Correctness | 54.8/5.1/0.7 | 27.1/6.6/0.0 | 37.8/9.6/8.6 | 30.1/16.7/0.0 | 35.1/10.0/0.0 |
| Consistency | 54.4/2.4/0.5 | 28.0/5.4/0.0 | 38.7/9.6/12.1 | 30.9/10.0/0.0 | 34.7/13.3/0.0 | |
| Last-switch | 54.9/3.0/0.6 | 26.9/5.4/0.0 | 37.4/8.1/13.1 | 29.9/10.0/0.0 | 31.7/10.0/0.0 | |
| StopHarm | 54.3/1.2/0.8 | 27.9/4.8/0.2 | 38.1/6.1/17.2 | 30.9/3.3/0.0 | 34.4/3.3/0.0 | |
| Qwen3-14B | Correctness | 47.1/1.3/0.2 | 29.2/5.2/0.4 | 45.5/7.6/5.1 | 6.5/0.0/0.0 | 4.5/0.0/0.0 |
| Consistency | 48.0/0.9/0.3 | 29.0/8.2/0.4 | 45.7/8.1/4.5 | 6.5/0.0/0.0 | 5.8/0.0/0.0 | |
| Last-switch | 48.0/1.1/0.3 | 29.3/8.0/0.4 | 46.6/9.6/5.1 | 7.0/0.0/0.0 | 3.3/0.0/0.0 | |
| StopHarm | 47.3/0.8/0.2 | 29.3/4.4/0.4 | 45.7/5.1/6.1 | 6.3/0.0/0.0 | 4.9/0.0/0.0 |
Each cell: TR / LC / RC (%). LC (RC) counts requests that Full-CoT answers correctly (incorrectly) but early stopping does not (does), so ΔAcc = RC − LC.
Revision. The revised paper (changes in blue) states the scope of the guarantee (Sections 3.2 and 5.1) and adds the transfer theory (Appendix E.4) and the recertification results (Section 5.6 and Appendix F.11). It also adds LearnStop to Table 2, with both checkpoint settings in Appendix F.10, and online certification in Appendix E.5.
We hope these responses address the reviewer's concerns, and we are happy to discuss any further questions.
I thank the authors for their thoughtful responses. After reviewing the responses, I think that my questions have been resolved. So I am delighted to maintain my positive assignment.
Thank you very much for your positive assessment and for confirming that our responses have resolved your questions. Your constructive comments have substantially improved our paper.
Response Part 1/2
We sincerely thank the reviewer for finding our key insight
interesting, for noting that StopHarm directly predicts the eventwhich is also the risk being controlled, and for considering our experimentscomprehensiveand able tosupport the core claims. Below, we address the scope of the i.i.d. guarantee, the comparison with LearnStop, online certification, and the LC/RC decomposition in turn.W1 — i.i.d. guarantee: scope and target-domain certification
We agree: Proposition E.1 requires i.i.d. calibration and deployment requests, so in Table 2 it covers only MATH500, and the other four benchmarks test transfer.
Theory (Appendix E.4). As the probe is frozen, Proposition E.1 applies unchanged to target calibration data (Corollary E.1). Without target data, LC is at most Bα under a bounded shift B in the probe's score and stop-harm trajectories (Proposition E.2). With neither, no calibration-based method can certify an arbitrary unseen task (Proposition E.3), so transfer must be shown empirically.
StopHarm transfers. With the probe and threshold fit on MATH only, it saves 30.4% of tokens on the other four benchmarks of Table 2 with no average accuracy loss (+0.5 pp). Where a formal guarantee is needed, we keep the probe and certify only the threshold on 700 GSM8K or ARC-Challenge requests (α = 5%):
LC: lost-correct rate (%); TR: token reduction (%); fresh trajectories, so values differ slightly from Table 2.
One scalar threshold carries the guarantee to a new task. With its own threshold, every benchmark keeps LC below α (15 of 16 policies across α ∈ {1, 3, 5, 10}%). Recertification fits the threshold to each task: it brings Qwen3-4B/ARC below α while still saving 71.6% of tokens, and adds 8–21 pp of TR in the other three cases. One training run thus serves many tasks; only the threshold is task-specific, and it can also be certified online (Q1).
W2/Q2 — Direct comparison with LearnStop
At the same accuracy and LC, StopHarm saves more tokens than LearnStop: on Qwen3-14B/GSM8K, both reach about 95.5% accuracy with 0.8% LC, but StopHarm reduces tokens by 47.3% versus 26.9%; on MATH500, StopHarm has higher accuracy, lower LC, and larger token reduction on all three models.
LearnStop shares StopHarm's MATH training and calibration splits (1,400 / 700), test problems, reference trajectories, and grader. It is trained and calibrated separately under two checkpoint settings: its own sparse grid (few trial answers) and StopHarm's paragraph checkpoints. Its cost includes all trial-answer generations.
Each cell: Acc. / LC / TR (%); LC is over all test requests, and negative TR means higher cost than Full-CoT.
On DeepSeek-7B, StopHarm saves more tokens than sparse-grid LearnStop on four of five benchmarks (e.g., 30.9% / 34.4% vs. 9.7% / 4.2% on AIME2024 / AIME2025, with accuracy within one problem); on GSM8K, it gives up 2.5 pp of TR for 2.5 pp higher accuracy and lower LC (1.2% vs. 3.9%).
StopHarm supports fine-grained checking without extra trial answers. LearnStop must generate a trial answer at every check, whereas StopHarm reads the hidden states and statistics of normal decoding and answers only once it stops. At paragraph granularity, LearnStop therefore raises cost by 34.2%–69.0% on both AIME benchmarks for all three models; on DeepSeek-7B/AIME2025, its average number of trial answers grows from 9.1 (sparse grid) to 334.5.
Changing only the supervision target isolates its contribution. LearnStop supervises the correctness of the current answer (LearnStop §3.1–3.4). In the paper's ablation (Table 3), which fixes features, probe, checkpoints, and certification, replacing correctness with stop-harm supervision keeps accuracy and LC unchanged on DeepSeek-7B/AIME2025 (36.7% / 3.3%) while raising token reduction from 5.9% to 34.4%.