- Rating
- 1
- Confidence
- 4
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper aims to develop an early stopping for reasoning models. To this end, the paper first introduces three signals: readiness, hazard, and steadiness. Then, the authors propose training a classifier that gets hidden states at specific parts of reasoning called “anchors" and predicts readiness and hazard signals. At inference time then, the model stops reasoning if readiness and steadiness are larger than their corresponding thresholds and hazard is lower than its threshold. Across two reasoning models and three math datasets, the authors showed that their proposed model outperforms some but a limited number of baselines.
Critical strengths
I think introducing the readiness signal is the main strength of the paper. However, lack of more analysis and sufficient ablation for it makes it hard to understand the effect of these.
Critical weaknesses
-
The comparisons are generally limited in datasets, models, and baselines. The authors could strength their results by adding at least the AIME dataset, another non Qwen model, and baselines such as RCPD[1], CoDE-Stop[2], and On-policy SFT [3]. As of now, the results are very limited.
-
I think this paper is heavily suffering from not having a strong motivation and ablation behind its method. The authors mentioned that they have used the last 3 anchors to get the signals. Why not 2 or 5? Or why should the steadiness signal work? what is the motivation behind it? These are interesting engineering techniques but without much justification.
[1] Hosseini, Parsa, et al. "Early Stopping for Large Reasoning Models via Confidence Dynamics." arXiv preprint arXiv:2604.04930 (2026).
[2] Wang, Xi, et al. "Conformal thinking: Risk control for reasoning on a compute budget." arXiv preprint arXiv:2602.03814 (2026).
[3] Zhao, Anhao, et al. "On-policy supervised fine-tuning for efficient reasoning." arXiv preprint arXiv:2602.13407 (2026).
Questions for authors
-
Please take a look at the weaknesses.
-
The reported Acc of DEER on DeepSeek-R1-Distill-Qwen-7B and GSM8k has been around 90% [1]. However, the reported Acc for DEER on DeepSeek-R1-Distill-Qwen-14B in the paper is around 80%. What is the setup and hyperparam used for DEER?
-
Could the authors add the baselines to Table 3. Right now it doesn't convey much information about the transfer capabilities. Also, I am curious about transfer to AIME because it generally requires more reasoning budgets.
[1] Zhao, Anhao, et al. "On-policy supervised fine-tuning for efficient reasoning." arXiv preprint arXiv:2602.13407 (2026).
4 comments
W2
First, a clarification. The method does not use the last three anchors. In §3.2 and Figure 2, "three early answers per anchor" refers to performing three truncations at each anchor and collecting the corresponding answers, and they are used to construct the Readiness label (an anchor is labelled ready if all three truncated answers are correct).
Most early-stopping methods in this literature rely on the model's internal confidence — how sure it is of the answer it is about to give, or whether it can already produce a correct answer at this point. Readiness implements exactly that signal. It is necessary but not sufficient, for two reasons we observed while building the training data:
A confident stop can destroy an answer that was going to be right. At many anchors, the model can still reach the correct answer if it keeps going. However, if the model answers immediately with only a short prompt-forced finalization, it may produce a wrong answer. Confidence-based gating cannot see this case, because the signal is computed from the partial reasoning rather than from what the forced answer would actually be. Hazard detects precisely this event: it is trained on "the full trajectory is correct but early answering is not".
Readiness jumps. We observed discontinuous jumps in Readiness from one anchor to the next, when the model switches strategy, guesses, or has an "aha moment". Such a jump is sometimes good for solving the problem, but is not a reliable stopping target, because the jump happens when the model finishes a sub-goal of the reasoning, not when the model is able to solve the problem as a whole. A threshold on a single anchor can fire at the jump, and the controller then stops before the remaining reasoning is produced, leading to an incorrect final answer. Steadiness addresses this by requiring the previous anchor's Readiness to also exceed a threshold.
Why one preceding anchor, and not two or five. A lag of one is the minimal confirmation that removes single-anchor spikes, adds no parameter and no extra head, and preserves the earliest safe stopping point. A longer lag delays every stop by at least one anchor and reduces the token saving, without contributing information that the Readiness heads do not already encode.
The ablation supports this reading. Removing one gate condition at a time while keeping the fitted heads and thresholds (Table 2) gives:
| Configuration | GSM8K | MATH-500 | AMC |
|---|---|---|---|
| SteadyStop | 84.84 / 1137 | 80.80 / 3110 | 65.06 / 6671 |
| Without Steadiness | 80.14 / 787 | 74.00 / 2173 | 53.01 / 4211 |
| Δ accuracy | −4.70 pp | −6.80 pp | −12.05 pp |
Removing Steadiness makes the gate more permissive and the token count collapses. The accuracy loss (4.70 to 12.05 pp) is far larger than threshold variation alone would produce.
Takeaway. The three signals are not arbitrary additions: each is motivated by observed failures of the prior design, the lag and the number of early answers are deliberately minimal rather than tuned, and the ablation quantifies the contribution of each signal.
Q2
Short answer: The two figures are produced under different protocols and different basic setups, so they are not directly comparable. Moreover, when we raise the budget substantially, our 14B GSM8K accuracy rises to a normal, competitive level, which confirms that the earlier figure reflects the budget rather than the model.
Experimental infrastructure, decoding parameters, and generation budgets vary across different papers. Experimental data reported for a baseline in one paper are measured under its protocol, and comparing it with data measured on ours mixes two different measurements. This is why we do not compare against numbers reported elsewhere at all, and every entry in Table 1 is produced by us under a single protocol and a single setup.
The 14B model is substantially stronger than the smaller checkpoints used in several baselines, so we think a large generation budget is not needed for it to reach a stable answer. We use 4,096 tokens for GSM8K, 8,192 for MATH-500, and 10,240 for AMC. The fraction of traces that exhaust the budget is correspondingly small, so the differences between methods reflect their stopping behaviour rather than truncation.
Excerpted from Table 1 (GSM8K, 14B, mean of three runs):
| Method | Acc.↑ | Tokens↓ |
|---|---|---|
| Full reasoning | 83.90 | 804.82 |
| TokenSkip | 84.05 | 719.46 |
| SBT-D | 83.24 | 696.83 |
| DEER | 83.47 | 702.84 |
| T-Calibration | 83.67 | 748.27 |
| SteadyStop (ours) | 85.97 | 677.57 |
Our method has both the highest accuracy and the fewest tokens at this setting, which is the comparison the paper actually claims.
With a 16,384-token budget, the 14B model reaches a normal accuracy level:
| GSM8K, DeepSeek-R1-Distill-Qwen-14B @ 16,384 | Acc.↑ | Tokens↓ |
|---|---|---|
| Full reasoning | 92.49 | 868.2 |
| SteadyStop | 91.13 | 706.9 |
Raising the budget moves full reasoning from 83.90 % to 92.49 %, and it is close to the 93.9 % that DEER reports for this checkpoint at 16,384 tokens. Also, at the larger budget, the controller no longer gains accuracy: it removes 18.6 % of the tokens at a cost of 1.36 pp. The positive gap in Table 1 is therefore specific to the smaller budget, where part of full reasoning's output consists of traces that reach the cap without a usable answer.
DEER is run under exactly the same settings as our main experiment — the same prompts, the same 4,096 / 8,192 / 10,240 budgets and the same decoding parameters — so every number in Table 1 comes from a single protocol.
Takeaway. The apparent gap is explained by differences in protocol and basic setup rather than by an unfair comparison, and under a strictly shared budget and decoding configuration our method is favourable against every compared baseline.
Q3
We thank the reviewer for both suggestions.
Due to time and compute constraints, we have not yet finished the relevant experiments, and these runs will be added in the follow-up experiments. For the AIME extension, we refer the reviewer to the AIME section of our responses to W1, where the added benchmark and its generation budget are described.
I appreciate the authors for providing more experimental results, but my main concerns mostly remain.
W1.
I think the paper needs a comparison with methods that already show improvements over DEER. For example, this can be Conformal Thinking if the authors reimplement their method, or CoDE-Stop and On-policy SFT.
W2.
Thank you for the clarification about this. I misinterpreted this part from the paper. However, I still think the supporting evidence and analysis for why these new signals work is limited.
Q2.
Thanks for providing the explanation. This resolves my issue.
Q3.
I appreciate the new experiments. However, again, without having other baselines on AIME, I am not sure how informative the results are.
In general, I still think some baselines stronger than DEER are missing and the method lacks deeper analysis. Therefore, I will maintain my score.
W1
Short answer: We have added a non-Qwen model (DeepSeek-R1-Distill-Llama-8B, four benchmarks) and AIME 2024/2025 (all three models). Of the three suggested baselines, Conformal Thinking releases no runnable code. On-policy SFT trains the target model, whereas our method leaves the target model untouched. CoDE-Stop scores answers with a GPT-4o-mini judge, but our method relies on no external model at any stage, since the labels come from the target model's own rollouts and answers are checked by symbolic equivalence. We consider our method a clear advance over the last two approaches, so we do not present these three as aligned comparisons, and our primary comparison set remains the methods in Table 1.
Added model and data.
See Table R3 in our response to Reviewer APpp, W2 & Q2.
The same trade-off is reproduced on a non-Qwen architecture, which indicates that the signals are not specific to the Qwen family.
See Table R4 in our response to Reviewer APpp, W2 & Q2.
On AIME, the controller removes 14.8–25.0 % of tokens and loses 1.67–3.33 pp. The reduction is smaller than on GSM8K or MATH-500, which is expected: harder problems determine their answer later, so fewer tokens can be removed safely.
Following the references in the review: [1] is CoDE-Stop (arXiv:2604.04930), [2] is Conformal Thinking (arXiv:2602.03814) and [3] is On-policy SFT (arXiv:2602.13407).
CoDE-Stop (arXiv:2604.04930) scores each candidate answer with a GPT-4o-mini judge and generates a trial answer at every reasoning step, so its per-query cost grows with the length of the trace. Our controller needs neither: it reads one hidden state and applies two linear heads with 66 parameters, which costs 0.2–4.5 ms per query.
For Conformal Thinking (arXiv:2602.03814, ICML 2026), the code is not publicly available for this work, so we are unable to reproduce its results.
On-policy SFT (arXiv:2602.13407) requires training a model: it needs a data pipeline, a filtering or reward design, a training run, and the result is a new checkpoint. SteadyStop does not train models: it reads hidden states at inference and applies two linear heads with 66 parameters. This keeps the base checkpoint intact, needs no training infrastructure and applies to frozen models, which is a clear step forward over the training-based recipe.
Both CoDE-Stop (arXiv:2604.04930) and On-policy SFT (arXiv:2602.13407) are preprints without a peer-reviewed version at the time of writing, so we did not adopt them as the primary comparison set; our primary baselines are the peer-reviewed methods with released code in Table 1, all run under identical protocols.
Takeaway. The revised paper now covers a non-Qwen model and a harder benchmark, and all three suggested baselines are addressed with explicit reasons rather than silently omitted. The training-free, judge-free design of our controller is stated as a clear advantage over the alternatives.