acceptodds
Under review as a conference paper at ICLR 2027

Post-Reasoning: Improving the Performance of Non-Thinking Models at No Cost

Abstract

As the widespread adoption of Large Language Models (LLMs) accelerates, token consumption from intermediate reasoning traces increasingly contributes to infer- ence latency and operational cost. Recent studies suggest that many real-world tasks require little to no explicit reasoning, with additional reasoning sometimes even degrading performance. In this work, we propose Post-Reasoning, a simple yet effective approach that improves instruction-tuned models by conditioning them to justify their answers after generating the final response. By design, it enables the final answer to be obtained without additional latency or token cost, while still improving performance through simple instruction augmentation. We evaluate Post-Reasoning across 117 model–benchmark settings spanning 13 open and proprietary models, 7 model families, and 9 diverse reasoning and knowledge- intensive benchmarks, including AMC, HMMT, GSM8K, GPQA, MMLU-Pro, and BIG-Bench Hard. Post-Reasoning improves performance in over 87.18% of evaluated settings, achieving a mean relative improvement of 16.30%. Furthermore, we propose supervised post-reason tuning, which further improves performance in over 91.11% of evaluated settings, and exceeds the prompt-based post-reasoning baseline by an average of 8.01%, demonstrating that post-reasoning can be effec- tively internalized through training. Ultimately, Post-Reasoning establishes a new performance ceiling for direct-answer capabilities.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Post-Reasoning, Large Language Models, Answer-First Generation, Efficient Inference, Reasoning Efficiency, Instruction Prompting, Supervised Fine-Tuning

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_gS6MOpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper studies whether asking a model to answer first and justify afterward improves its initial answer. It also introduces fine-tuning on post-answer explanations while masking the answer loss. Experiments cover 13 models and nine benchmarks, with generation optionally stopped before the explanation. The broad evaluation is useful, but the contribution needs clearer positioning against recent work, and several repeated results require clarification.

Critical strengths
  1. The study addresses a practical question: improving answer accuracy without generating a reasoning chain before the answer. The coverage across model families and tasks is a strength.

  2. The masking ablation and measurements of latency when stopping after the answer provide useful evidence for the proposed training and deployment choices.

Critical weaknesses
  1. The novelty is overstated. Think or Not Think [1] already introduces Think-After-Answer, explicitly requesting an answer followed by reasoning, and reports improvements over reasoning-first training in some settings. Aux-Think [2] also studies reasoning before versus after the action and uses reasoning supervision with direct prediction at inference. These are multimodal settings, so they do not establish the same prompting gains reported here. Nevertheless, the central output ordering is already established, and neither paper is discussed.

  2. The training comparison does not sufficiently establish the remaining contribution. TwT [3] already transfers reasoning benefits into direct-answer inference through staged distillation. Table 16 usefully tests answer masking, but comparisons with answer-only fine-tuning and staged reasoning distillation on the same data would clarify what this particular objective adds.

  3. Some repeated results disagree without a documented change in setting. For Qwen3.5-9B on AMC 8, prompt-only Post-Reasoning scores 21.12 in Table 1 but 44.03 as the baseline in Table 9. For Qwen3.5-27B, the corresponding values are 30.67 and 20.67. Most other repeated entries match exactly. These discrepancies affect the claimed incremental training gains and need resolution.

[1] Li M, Zhong J, Zhao S, et al. Think or not think: A study of explicit thinking in rule-based visual reinforcement fine-tuning[J]. arXiv preprint arXiv:2503.16188, 2025.

[2] Wang S, Wang Y, Li W, et al. Aux-think: Exploring reasoning strategies for data-efficient vision-language navigation[J]. Advances in Neural Information Processing Systems, 2026, 38: 30892-30916.

[3] Xu J, Zhou M, Liu W, et al. TwT: Thinking without Tokens by Habitual Reasoning Distillation with Multi-Teachers' Guidance[J]. arXiv preprint arXiv:2503.24198, 2025.

Questions for authors
  1. What distinguishes the proposed approach from Think-After-Answer beyond the model and task coverage?

  2. Does explanation-only training outperform answer-only fine-tuning and staged reasoning distillation under a common data and evaluation setup?

  3. Which AMC 8 baseline values are correct, and how do the corrections affect the reported training improvements?

2 comments

===== COMMENT 1/2 =====

We thank you for the close reading, down to the cross-check of repeated table entries, and for recognizing the practical question and the coverage across model families. Your three points need three kinds of evidence: that the finding is new relative to earlier answer-first work, that the training objective adds something specific, and that the reported numbers hold.


[Clarification:1] Distinction from Think-After-Answer and Aux-Think

We will cite and discuss both works in Related Work. They differ from our work in task, in method and, most importantly, in the direction of the result:

Think-After-Answer [R1]Aux-Think, Post-Think [R2]Post-Reasoning (ours)
TaskVision: classification, spatial perception, puzzles, visual mathematicsVision-language navigationText: mathematics, science, knowledge, multi-step reasoning
ModelsQwen2-VL 2B and 7B (one family)NVILA-lite 8B (one model)13 models, 7 families, 4B–70B, open and proprietary
How the order is inducedRL fine-tuning (GRPO-style)SFT, with loss on the action and the reasoningOne instruction (no training), and SFT with the answer masked
Stop after the answer at inferenceNot reportedNo (28.97 s per step, against 1.25 s for No-Think)Yes, measured (Appendix E)
Answer-first result against no reasoningLower than No-Thinking RFT (PuzzleVQA, 2B: 64.70 against 70.85)Lower than No-Think (success rate 29.0 against 35.1)Higher than Direct in 99 of 117 settings

The last row is the key difference. The authors of [R1] state that Think-After-Answer still underperforms No-Thinking RFT. The authors of [R2] state that Pre-Think and Post-Think both fall short of No-Think. In both works, reasoning after the answer gives lower accuracy than no reasoning. Our paper shows the opposite for instruction-tuned text LLMs: a request for a post-answer justification gives higher accuracy than direct answering, in 99 of 117 settings, with no tokens before the answer. Thus, the shared output order does not make the finding known. The prior results predict the opposite of what we observe.

Two more differences follow from the table:

Prompting without training : [R1] and [R2] get the answer-first order only through fine-tuning. Our main result needs no training. One instruction is sufficient, and it also applies to proprietary models whose weights are not available (Tables 4–6, 20–22).

Deployment cost : Neither work stops generation at the answer or measures the cost to the answer. The latency of Aux-Think shows that Post-Think generates the full reasoning. Our paper measures the stop at the answer: 10.9 tokens to the answer, within 0.02–0.07 s of Direct (Appendix E).

Earlier text-only evidence is also limited. In the original CoT paper, a "chain of thought after answer" prompt gave about the same accuracy as the standard prompt [R3]. Thus, before our work, the evidence did not show that an answer-first instruction increases the accuracy of the answer.


[Clarification:2] What the training objective adds, compared with answer-only training and staged distillation

TwT solves a different problem : TwT [R4] removes reasoning from the output. It does not place the answer before a justification. Its pipeline also needs resources that ours does not need:

TwT [R4]Post-reason tuning (ours)
TeachersSeveral large LLMs (GPT-4, GPT-4o-mini, Mistral-Large) give pseudo-labels and rationalesNone. The model justifies a given answer in its own words (self-distillation)
StagesThree: full reasoning, then compressed reasoning, then no reasoningOne
Loss on the answerYesNo (masked)
Output at inferenceAnswer onlyAnswer first. The justification is optional
Evaluation2 students (Mistral-7B, Phi-3.5-mini), 3 benchmarks10 models, 9 benchmarks

Teacher rationales are weaker on the same data : Table 8 compares the sources of rationales under a common setup: the same questions, the same LoRA configuration and the same evaluation. Only the source of the justification changes (Mistral-Small-24B):

Rationale sourceMATHGPQAMMLU-ProBBH
Expert (stronger teacher model)26.5036.4054.0059.53
Rephrased26.1838.4354.0359.50
Self (ours)27.2145.1754.5359.94

(Continued in the next comment.)

===== COMMENT 2/2 =====

(Continued from the previous comment.)

Self-distillation gives the highest accuracy on 4 of 4 benchmarks, with +8.77 points on GPQA over teacher distillation. Teacher rationales are the central input of TwT. In our setting, they give lower accuracy than the model's own justifications.

More answer supervision decreases accuracy : The unmasked objective in Table 16 uses the same questions, answers, LoRA configuration and seeds, and adds a direct loss on the answer tokens. This added supervision decreases accuracy in 9 of 9 settings, all significant at p < 0.05. If supervision on the answers explained the gain, we expect the opposite result.

New result: answer-only fine-tuning : As you requested, we trained the three models of Table 16 on the same 3,500 questions with the loss on the answer tokens only. The data, LoRA configuration and seeds are identical. The loss target is the only difference:

ModelAMC 10 Answer-onlyAMC 10 MaskedΔGPQA Answer-onlyGPQA MaskedΔMMLU-Pro Answer-onlyMMLU-Pro MaskedΔ
Qwen-3.5 (4B)9.2510.86+1.61*39.8442.87+3.03*44.5045.70+1.20*
Ministral-3 (14B)13.1716.01+2.84*36.1439.68+3.54*51.8854.91+3.03*
Llama-3.3 (70B)29.5833.15+3.57*47.6949.36+1.67*52.1153.31+1.20*

Mean accuracy (%) over three seeds. Δ is Masked minus Answer-only. * p < 0.05, one-sided Welch t-test, n = 3 seeds per arm.

Explanation-only training exceeds answer-only fine-tuning in 9 of 9 settings, all significant, by +1.20 to +3.57 points (mean +2.41). Mean accuracy orders as answer-only (36.02) < unmasked (36.53, Table 16) < masked (38.43). Accuracy increases as the loss moves from the answer to the justification. Distilling Step-by-Step [R5] puts a direct loss on the label. Our objective puts no loss on the answer, and these results test exactly this design choice.

The staged pipeline, component by component : The two parts of staged distillation are already tested under a common setup. Its final stage trains on the question and the answer only, with no reasoning. This is the answer-only arm above, which is lower in 9 of 9 settings. Its earlier stages depend on teacher rationales, which are lower on 4 of 4 benchmarks (Table 8). Thus, each part of staged distillation gives lower accuracy than our objective on the same data.


[Clarification:3] AMC 8 values in Table 9

Table 1 gives the correct prompt-based values: 21.12 for Qwen-3.5-9B and 30.67 for Qwen-3.5-27B. Both columns of these two cells in Table 9 have a transcription error. The other 88 repeated entries in Tables 9 and 10 match Tables 1–3 are accurate. The correct AMC 8 values are:

ModelPrompt-based PostPost-Reason SFTΔ
Qwen-3.5-9B21.1220.94−0.18
Qwen-3.5-27B30.6732.35+1.68

The correction does not change the count. Post-Reason SFT exceeds prompt-based post-reasoning in 66 of 90 settings in Tables 9–10, before and after the correction. Against Direct, Post-Reason SFT improves 82 of 90 settings (91.11%), which is the value that the paper reports. The revision corrects the four cells.


All of the authors sincerely appreciate your questions and we are eager to address any further questions. We look forward to a constructive discussion.

References:

[R1] Li, Ming, et al. "To think or not to think: A study of thinking in rule-based visual reinforcement fine-tuning." Advances in Neural Information Processing Systems 38 (2025): 167356-167407.

[R2] Wang, Shuo, et al. "Aux-Think: Exploring reasoning strategies for data-efficient vision-language navigation." Advances in Neural Information Processing Systems 38 (2025): 30892-30916.

[R3] Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems 35 (2022): 24824-24837.

[R4] Xu, Jingxian, et al. "TwT: Thinking without tokens by habitual reasoning distillation with multi-teachers' guidance." Findings of the Association for Computational Linguistics: EMNLP 2025 (2025): 16475-16489.

[R5] Hsieh, Cheng-Yu, et al. "Distilling step-by-step! Outperforming larger language models with less training data and smaller model sizes." Findings of the Association for Computational Linguistics: ACL 2023 (2023): 8003-8017.


Reviewer_xBRdOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
3
Soundness
2
Read review
Paper summary

The paper proposes a prompting strategy called post-reasoning to improve LLM reasoning. Similar to CoT, post-reasoning appends a prompt suffix: "State the final answer immediately and justify your answer." It intends to get the model to respond directly and subsequently justify its answer. Post-reasoning improves performance against the base model and can be used for tuning the model behaviour to improve reasoning.

Critical strengths

The authors show that appending the prompt-suffix improves reasoning across a broad spectrum of models and can be used for tuning model behaviour.

Critical weaknesses

The originality and novelty of the method are limited.

I believe the comparison to CoT, other prompting strategies, and reasoning models should be in the main body of the paper. I do not see a reason to adopt post-reasoning over CoT or instead of using reasoning models. With respect to in-context learning, it is more intuitive to base the answer on prior reasoning rather than answering directly with subsequent justification.

The statement "enables the final answer to be obtained without additional latency or token cost" is not supported. Similar to CoT, the generated output tokens before the EOS token are expected to be longer than those without the instructed prompting behaviour. Additionally, the instruction itself already introduces latency via the additional input tokens.

Questions for authors

What happens if the first answer is wrong? Does the model realise this in the justification and revise the answer? Which of these answers is taken for the final answers of the evaluations?

How is post-reasoning deployed for the evaluations? Do you generate the full sequence until EOS or just the first answer and stop after? What are the token costs of the base model vs post-reasoning vs CoT?

If you want the "final answer to be obtained without additional latency or token cost", how do you stop generation after the answer? How do you know the answer is fully generated? And why would you instruct the model for justification in the first place if you stop after the answer anyway?

How do you extract the answer? Some evals just check whether the solution is contained in the output string or whether the first tokens align with the solution. What happens if you instruct the model just to answer directly?

6 comments

===== COMMENT 1/2 =====

We thank you for the review and for noting that the paper is clear and that the gain holds across a broad range of models. Your points come back to one question: why a system that can use CoT or a reasoning model needs post-reasoning, and whether the cost claim holds at deployment. The answer to the first part changes how the other points read. Post-reasoning is not an alternative to CoT or to reasoning models. It improves the no-thinking to low-thinking effort path, where CoT is not used in low-latency deployment where speed is critical.


[Clarification:1] Post-reasoning adds to reasoning models. It does not compete with them.

The paper evaluates post-reasoning with reasoning enabled (Appendix D). Gemini-2.5-Flash, Claude-Haiku-4.5 and GPT-5.4-Mini run with thinking set to low. Only the suffix changes, so both arms reason before the answer. The gain of post-reasoning over this reasoning baseline is:

BenchmarkClaude-Haiku-4.5GPT-5.4-MiniGemini-2.5-Flash
AMC 8+6.34+2.52+1.49
AMC 10+6.97+4.09+2.30
AMC 12+9.23+2.21+2.37
HMMT Feb+15.86+3.37+1.68
HMMT Nov+3.00+3.55+1.36
GSM8K+2.70+4.00+2.72
GPQA+1.35+3.22+11.01
MMLU-Pro+3.83+3.74+1.37
BBH+8.42+3.37+2.89

Thinking Post minus Thinking Direct in accuracy points (Tables 20–22).

Post-reasoning improves 27 of 27 settings, with a mean gain of +4.26 points. Here, the Direct arm is itself a reasoning baseline. With full thinking on seven open models, post-reasoning improves 48 of 63 settings (Tables 17–19). Thus, the choice is not "post-reasoning or CoT." A system that uses CoT adds the instruction and gets a more accurate answer at the same reasoning cost.


[Clarification:2] Why the no-thinking and low-thinking paths matter

CoT is not the most accurate choice for every query, and the literature documents this:

Where CoT helps : CoT gives large gains mainly on mathematics and symbolic reasoning. On other task categories, the gain is less than 1 point (56.8 against 56.1) [R1].

Overthinking : No model thinks optimally, and thinking models spend hundreds of tokens on simple queries [R2, R3]. On some task types, CoT decreases accuracy by up to 36.3 points [R4].

Routing in practice : For these reasons, a large body of work routes each query to a thinking or a no-thinking path [R5–R8] (Appendix A).

In each of these systems, a large share of queries goes to the no-thinking or low-thinking path. Post-reasoning makes that path more accurate, with no tokens before the answer. In contrast, CoT needs a median of 115 times more tokens before the answer (range 5 to 596 times). The measured time to the answer is:

StrategyLlama-3.1-8BLlama-3.3-70BQwen-3.5-9B
Direct0.02–0.05 s0.07–0.17 s0.04–0.08 s
Post (stopped at the answer)0.05 s0.15–0.24 s0.06–0.09 s
CoT0.6–18 s4–17 s1.5–26 s

Single request, H200 GPU, range of per-benchmark medians over 9 benchmarks (Table 23). The CoT row uses the same setup.


[Clarification:3] The "no additional latency or token cost" claim is measured

The claim refers to the cost to get the final answer. The justification is optional (Section 2, footnote 1). Appendix E measures both sides:

Output tokens : Generation stops at the end of the answer field. Median tokens to the answer: 10.5 for Direct, 10.9 for Post. The larger limit in Table 13 applies only when the justification is kept.

Input tokens : The suffix adds four words. Prefill processes all input tokens in parallel, and decode takes most of the end-to-end time [R9, R10]. On BLOOM-176B, the prefill of a 1,500-token prompt takes the same time as 6 output tokens [R9]. Our measured time to first token is 26–77 ms (Table 24).

Wall-clock time : Post stays within 0.02–0.07 s of Direct, from 8B to 70B (Table 23).


[Clarification:4] Answers to the four questions

Q1. What if the first answer is wrong? : We grade only the first answer, which comes before the justification. A change in the justification does not count, so a self-correction cannot increase the score.

Q2. How is it deployed, and what are the token costs? : The main tables grade the answer field only, with greedy decoding (temperature 0.0, Table 13). With greedy decoding, the tokens before the stop are identical with or without truncation. Thus, Tables 1–6 give exactly the accuracy of the truncated deployment. Tokens to the answer: 10.5 (Direct), 10.9 (Post), a median of 115 times more for CoT.

(Continued in the next comment.)

===== COMMENT 2/2 =====

(Continued from the previous comment.)

Q3. How do you stop, and why request a justification? : The answer is in a delimited answer field, and generation stops at the end-of-answer marker (Section 2). It parsed correctly on at least 99.6% of examples in 90 settings (Appendix E). We request the justification because the request itself causes the gain (Equation 2). With the same answer-first format, a justification request beats a confidence or summary request on 3 of 3 benchmarks (Table 11).

Q4. How do you extract the answer, and what if you instruct a direct answer? : A regular expression reads the answer field, not a substring of the full output (Appendix B.2). The direct-answer instruction is the Direct baseline: "State the final answer immediately." (Appendix B.6). Every Δ in the paper compares with it.


[Clarification:5] Novelty, and the intuition about in-context learning

Reasoning before the answer helps in CoT. This does not exclude a second effect: the instruction itself conditions the answer. Earlier results on answer-first order are null or negative, and ours is positive:

WorkAnswer-first result
CoT after the answer, few-shot prompting [R11]About equal to the standard prompt
Explanations after answers in few-shot examples [R12]Helps only large models
Think-After-Answer, RL on 2 vision models [R13]Lower than no-thinking
Aux-Think Post-Think, navigation [R14]Lower than No-Think (29.0 against 35.1)
Ours, one instruction, 13 text LLMsHigher than Direct in 99 of 117 settings

Our response to Reviewer gS6M (Clarification 1) gives the full comparison.


All of the authors sincerely appreciate your questions and we are eager to address any further questions. We look forward to a constructive discussion.

References:

[R1] Sprague, Zayne, et al. "To CoT or not to CoT? Chain-of-thought helps mainly on math and symbolic reasoning." The Thirteenth International Conference on Learning Representations (2025).

[R2] Aggarwal, Pranjal, et al. "Optimalthinkingbench: Evaluating over and underthinking in llms." arXiv preprint arXiv:2508.13141 (2025).

[R3] Chen, Xingyu, et al. "Do NOT think that much for 2+3=? On the overthinking of long reasoning models." Proceedings of the 42nd International Conference on Machine Learning, PMLR 267 (2025): 9487-9499.

[R4] Liu, Ryan, et al. "Mind your step (by step): Chain-of-thought can reduce performance on tasks where thinking makes humans worse." Proceedings of the 42nd International Conference on Machine Learning, PMLR 267 (2025): 38489-38517.

[R5] Fang, Gongfan, Xinyin Ma, and Xinchao Wang. "Thinkless: Llm learns when to think." Advances in Neural Information Processing Systems 38 (2025): 151268-151295.

[R6] Jiang, Lingjie, et al. "Think only when you need with large hybrid-reasoning models." Advances in Neural Information Processing Systems 38 (2025): 159355-159382.

[R7] Lou, Chenwei, et al. "Adacot: Pareto-optimal adaptive chain-of-thought triggering via reinforcement learning." arXiv preprint arXiv:2505.11896 (2025).

[R8] Zhang, Jiajie, et al. "Adaptthink: Reasoning models can learn when to think." Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing (2025): 3716-3730.

[R9] Patel, Pratyush, et al. "Splitwise: Efficient generative llm inference using phase splitting." 2024 ACM/IEEE 51st Annual International Symposium on Computer Architecture (ISCA) (2024): 118-132.

[R10] Agrawal, Amey, et al. "Taming throughput-latency tradeoff in LLM inference with Sarathi-Serve." 18th USENIX Symposium on Operating Systems Design and Implementation (OSDI 24) (2024): 117-134.

[R11] Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems 35 (2022): 24824-24837.

[R12] Lampinen, Andrew K., et al. "Can language models learn from explanations in context?" Findings of the Association for Computational Linguistics: EMNLP 2022 (2022): 537-563.

[R13] Li, Ming, et al. "To think or not to think: A study of thinking in rule-based visual reinforcement fine-tuning." Advances in Neural Information Processing Systems 38 (2025): 167356-167407.

[R14] Wang, Shuo, et al. "Aux-Think: Exploring reasoning strategies for data-efficient vision-language navigation." Advances in Neural Information Processing Systems 38 (2025): 30892-30916.


Thank you for the clarifications. They answer my questions regarding grading, answer extraction, and stopping after the answer. I believe this important aspect of the method should be made more prominent and clear throughout the paper.

My main concern still remains, and I do not see a theory for why the request itself should cause the gain. Equation 2 only states that the answer is conditioned on the instruction. It does not explain why a few additional input tokens, which provide no meaningful information about the task itself, improve accuracy. I also agree with Reviewer gS6M that the claim of a “latent reasoning trajectory” is not supported by the experiments.

I also agree with the other reviewers (Kt8D, gS6M) that there might be some inconsistencies in the reported numbers. The share of improved settings is 87.18% in the abstract, 88.19% in the introduction, and 84.62% in the rebuttal. Also, I agree that the claim of "mean relative improvement of 16.30%" is misleading, and the absolute delta (+2.11 points) is more meaningful here. Otherwise, small absolute gains appear as large relative gains. For example, Llama-3.1-8B on HMMT Feb increases from 1.44 to 2.88 (Table 2). This is only +1.44 points, but a relative gain of 100%.

Sec 2.1 only shows the prompt without any end-of-answer token defined. For the sake of generality and deployability, it would be worth adding it to the prompt. Otherwise, the method remains only deployable on these specific answer parsing benchmarks. Further, I am wondering whether all the deployed evals actually contain these benchmark-specific end-of-answer tokens, or did you add them yourselves? Also, since the Post-Reason prompt (“State the final answer immediately and justify your answer.”) asks for a justification as part of the task, I would assume some of these models might set the end-of-answer token after the justification. Does this happen?

Additionally, it might be interesting to investigate how token usage changes for the reasoning models. I would expect the model to think longer when it also has to justify its answer. Appendix D reports no thinking-token counts. This would also help to support the claim of “the same reasoning cost”. Some Thinking Direct baselines are surprisingly low (Qwen-3-8B on AMC 8: 23.51 vs. 91.79 with Thinking Post). What happened here? Is there a format problem?

The statement “no additional latency or token cost” is still not supported and has not been removed from the manuscript. Even when stopping after the answer, the cost increases for both prefill and decoding. The authors should speak about negligible costs instead, also in the title.

The rebuttal addresses some of my concerns, and I will consider raising the score once the authors have updated their manuscript and elaborated on the points above.

Reviewer_Kt8DOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper investigates prompting instruction-tuned language models to answer immediately and then justify their response. Generation can stop at the answer boundary, avoiding the requested justification at deployment. It also proposes supervised post-reason tuning, where the model learns to generate justifications conditioned on a supplied answer, while answer-token losses are masked. Evaluation covers 13 models and 9 benchmark settings, with additional training experiments on ten models.

This addresses a relevant pragmatic question, namely, can the quality of a short answer be improved without paying for a long reasoning trace before receiving it? The approach is easy to implement and could be a useful default baseline for systems operating under tight answer-latency constraints. Its simplicity is a strength. The broad positive pattern is consistent across model families.

Critical strengths
  1. Broad evidence for a useful and inexpensive intervention. Evaluation across 13 models and 9 benchmark settings suggests that the effect is not restricted to one architecture or task. The paired prompt change is simple and reproducible in principle.

  2. The study includes meaningful controls and a plausible training extension. Alternative summary and confidence continuations test whether any post-answer instruction would be sufficient. The answer-masking ablation holds data, configuration, and seeds fixed, and reports higher observed accuracy for the masked objective in nine settings. Training across ten models and transfer to non-mathematical benchmarks add a generalization argument beyond prompting.

Critical weaknesses
  1. The scientific interpretation and novelty claims are overstated. The experiments establish an instruction-conditioning effect instead of directly identifying an improved latent reasoning trajectory. Likewise, masking answer-token loss does not prevent learning answer associations, since the correct answer remains in the training context and parameters are shared. Closer prior work also narrows the novelty. These limitations affect the claimed explanation and incremental training contribution.

  2. The magnitude and reliability of the aggregate improvements need clearer support. Strong positive pattern but inconsistent with the headline percentages. Averaging relative gains can amplify changes from low baselines. These issues are important for assessing effect size and generality.

Questions for authors
  • How much of the training benefit is specific to answer-masked post-answer justification, beyond ordinary supervised adaptation on the same questions? The existing masked-versus-unmasked comparison is informative but a matched answer-only baseline would clarify the incremental contribution. If this unavailable, please narrow the training-mechanism claim accordingly.

  • How should the contribution be distinguished from prior answer-first prompting and answer-before-explanation training?

  • What is the central scientific contribution of post-reasoning? Given precedents for answer-first generation and explanation-based training, is the main advance the breadth of the empirical finding, the masked training objective, or a new understanding of how instructions influence direct-answer performance?

  • Under what conditions should post-reasoning be expected to help? Can the variation across models and tasks support a general account of its benefits and limitations, beyond the observation that it improves most evaluated settings?

  • What does supervised post-reason tuning add beyond the prompting effect? How clearly does the evidence distinguish benefits specific to the proposed objective from those of additional task training or rationale supervision?

3 comments

===== COMMENT 1/2 =====

We thank you for the careful reading and for recognizing the breadth of the evidence and the value of post-reasoning as a default baseline. Your points ask for one thing: proof that the explanation in the paper matches what the experiments isolate, and that each part of the method adds something of its own.


[Clarification:1] What supervised post-reason tuning adds beyond prompting

New result: matched answer-only baseline : We trained the three models of Table 16 on the same 3,500 questions with the loss on the answer tokens only. The data, LoRA configuration and seeds are identical. The loss target is the only difference.

ModelAMC 10 Answer-onlyAMC 10 MaskedΔGPQA Answer-onlyGPQA MaskedΔMMLU-Pro Answer-onlyMMLU-Pro MaskedΔ
Qwen-3.5 (4B)9.2510.86+1.61*39.8442.87+3.03*44.5045.70+1.20*
Ministral-3 (14B)13.1716.01+2.84*36.1439.68+3.54*51.8854.91+3.03*
Llama-3.3 (70B)29.5833.15+3.57*47.6949.36+1.67*52.1153.31+1.20*

Mean accuracy (%) over three seeds. Δ is Masked minus Answer-only. * p < 0.05, one-sided Welch t-test.

The masked objective wins in 9 of 9 settings, all significant (mean +2.41 points). Mean accuracy orders as answer-only (36.02) < unmasked (36.53, Table 16) < masked (38.43). Accuracy increases as the loss moves from the answer to the justification.

Other controls on the same questions : Over prompt-based post-reasoning, tuning is higher in 66 of 90 settings (Tables 9–10). Self-distillation beats teacher rationales on 4 of 4 benchmarks (Table 8). A loss added on the answer lowers accuracy in 9 of 9 settings (Table 16).

Thus, neither more task training nor rationale supervision explains the gain. Unlike Distilling Step-by-Step [R1], our objective puts no loss on the answer.


[Clarification:2] Distinction from earlier answer-first prompting and training

Earlier answer-first work reports null or negative results:

Answer-first prompting : A chain of thought after the answer gives about the same accuracy as the standard prompt [R2]. Explanations after the answers of few-shot examples help only large models [R3].

Answer-first training : Think-After-Answer still underperforms No-Thinking RFT on vision tasks [R4]. Post-Think falls short of No-Think on navigation (success rate 29.0 against 35.1) [R5].

Distillation : TwT removes reasoning from the output and does not use answer-first order [R6].

No earlier work isolates the instruction or shows a gain over direct answering at this scale. Our response to Reviewer gS6M (Clarification 1) gives the full comparison.


[Clarification:3] Central scientific contribution

Our contribution: a request to justify the answer changes the computation that produces the answer itself. With zero tokens before the answer, the model uses reasoning that direct answering leaves unused.

The instruction induces this effect. The effect itself is reasoning:

PropertyResult
General99 of 117 settings improve (Tables 1–6)
Reasoning-specificJustification > confidence > summary, same answer-first format (Table 11)
Separate from explicit reasoning+4.26 points on top of low-effort thinking, 27 of 27 settings (Tables 20–22)
Trained from reasoning aloneLoss only on the justification improves the answer over Direct in 81 of 90 settings (Tables 9–10)

The fourth row is the most direct evidence of a latent reasoning trajectory. In masked training, the gradient comes only from the justification tokens, which come after the answer. The answer still improves, more than with a loss on the answer itself (Clarification 1). The gradient reaches the answer only through the shared weights that compute it. Thus, training the model to reason about its answer changes how it computes the answer.

The third row shows that this is not the CoT effect, which earlier work credits to the reasoning tokens before the answer [R2]. A model that already reasons still gains +4.26 points in every setting.


[Clarification:4] When post-reasoning helps

Post-reasoning helps when the answer path does not use the full capability of the model. The gain decreases when the model already spends that capability before the answer. The three levels of reasoning effort show this:

Reasoning effortSettings improvedMean Δ (points)
Off (Tables 4–6)24 / 27+2.26
Low (Tables 20–22)27 / 27+4.26
Full (Tables 17–19)48 / 63—

With full thinking, Qwen-3.5-4B and 9B (Thinking-Direct mean 64–65%) improve on 9 of 9 benchmarks. Qwen-3.5-27B (79.5%) improves on 1 of 9. Per task, gains are largest on GPQA (+4.29) and smallest on GSM8K (+0.17), where direct answers are already strong (Section 7, [R7, R8]).

(Continued in the next comment.)

===== COMMENT 2/2 =====

(Continued from the previous comment.)


[Clarification:5] The gain comes from reasoning, not from the prompt format

The instruction induces the effect, and three controls exclude explanations other than reasoning:

One suffix is the only difference : Direct: "State the final answer immediately." Post adds "and justify your answer." Prompt format alone can change accuracy by up to 76 points [R9], so we keep all else identical (Appendix B.6).

The requested content matters : With the same answer-first format, a justification request beats a confidence or summary request on 3 of 3 benchmarks (Table 11). Only the request to reason gives the full gain. This is the latent reasoning that Section 6 describes.

Fixed persona, temperature 0.7, 3 seeds : The gain is positive in 12 of 12 cells and significant in 10 of 12:

BenchmarkQwen-3.5-4BGemma-3-12BMinistral-3-14BLlama-3.3-70B
AMC 10+3.75*+3.60*+3.23*+5.20*
GPQA+12.00*+0.93*+1.84*+0.70
MMLU-Pro+4.75*+0.75*+0.83+0.94*

Post minus Direct in points, mean over 3 seeds. * p < 0.05, one-sided Welch t-test.


[Clarification:6] Answer-token masking excludes answer association

The answer is in the context by design (Equation 3), but it is never a training target. Three results exclude answer association:

More answer supervision lowers accuracy : Masked beats unmasked in 9 of 9 settings (Table 16), and answer-only is lower again (Clarification 1).

The gain transfers to unseen answer formats : Training uses integer-answer mathematics only. On GPQA, MMLU-Pro and BBH, the tuned models exceed Direct in 25 of 30 settings (Tables 3 and 10).

In-format data does not help more : MMLU training gives lower MMLU-Pro accuracy than mathematics training in 10 of 10 models (Table 26).


[Clarification:7] Effect size in absolute points

The absolute gains per benchmark, over the 13 models of Tables 1–6, are:

BenchmarkAMC 8AMC 10AMC 12HMMT FebHMMT NovGSM8KGPQAMMLU-ProBBH
Improved13/1313/1311/1310/1312/138/1311/1310/1311/13
Mean Δ+3.99+2.94+2.07+1.04+1.00+0.17+4.29+1.73+1.78

Low baselines do not drive the effect. The correlation between Direct accuracy and absolute gain is r = 0.02 over 117 settings. The mean absolute gain is +2.11 points, and the median relative gain is +6.98%. Without the settings where Direct is below 10%, 72 of 84 settings still improve.

Tables 1–6 give 99 improvements, 2 ties and 16 decreases (84.62%). The revision reports these counts, and absolute deltas in every table. Main results use greedy decoding (Table 13). Clarification 5 gives the 3-seed results.


All of the authors sincerely appreciate your questions and we are eager to address any further questions. We look forward to a constructive discussion.

References:

[R1] Hsieh, Cheng-Yu, et al. "Distilling step-by-step! Outperforming larger language models with less training data and smaller model sizes." Findings of the Association for Computational Linguistics: ACL 2023 (2023): 8003-8017.

[R2] Wei, Jason, et al. "Chain-of-thought prompting elicits reasoning in large language models." Advances in Neural Information Processing Systems 35 (2022): 24824-24837.

[R3] Lampinen, Andrew K., et al. "Can language models learn from explanations in context?" Findings of the Association for Computational Linguistics: EMNLP 2022 (2022): 537-563.

[R4] Li, Ming, et al. "To think or not to think: A study of thinking in rule-based visual reinforcement fine-tuning." Advances in Neural Information Processing Systems 38 (2025): 167356-167407.

[R5] Wang, Shuo, et al. "Aux-Think: Exploring reasoning strategies for data-efficient vision-language navigation." Advances in Neural Information Processing Systems 38 (2025): 30892-30916.

[R6] Xu, Jingxian, et al. "TwT: Thinking without tokens by habitual reasoning distillation with multi-teachers' guidance." Findings of the Association for Computational Linguistics: EMNLP 2025 (2025): 16475-16489.

[R7] Aggarwal, Pranjal, et al. "Optimalthinkingbench: Evaluating over and underthinking in llms." arXiv preprint arXiv:2508.13141 (2025).

[R8] Chen, Xingyu, et al. "Do NOT think that much for 2+3=? On the overthinking of long reasoning models." Proceedings of the 42nd International Conference on Machine Learning, PMLR 267 (2025): 9487-9499.

[R9] Sclar, Melanie, et al. "Quantifying language models' sensitivity to spurious features in prompt design or: How I learned to start worrying about prompt formatting." The Twelfth International Conference on Learning Representations (2024).


AuthorsOpenReview ↗

Final revision uploaded: summary of your questions and the new evidence

Dear Reviewer Kt8D,

Thank you again for your review, and for describing post-reasoning as "a useful default baseline for systems operating under tight answer-latency constraints." We uploaded the final revision, with all changes in red. Below, each of your questions is mapped to the new evidence and its place in the paper.

Your questionEvidence in the revisionLocation
What does post-reason tuning add beyond answer supervision?The matched answer-only baseline you requested: same questions, LoRA configuration and seeds, with only the loss target changed. Masked training is highest in 9 of 9 settings, all significant. Mean accuracy orders answer-only (36.02) < unmasked (36.53) < masked (38.43).App. C.3 (Table 16)
How does this differ from earlier answer-first work?Earlier answer-first results are null or negative, or distil a teacher into ≤2B students with a loss on the answer. None tests the instruction alone; ours shows that it improves 99 of 117 settings, and our training objective uses no teacher and no answer loss.Related Work
What is the central contribution?A request to justify the answer improves the answer itself, with zero tokens before it. Section 6 presents the pre-caching mechanism with two supporting tests: the content control (Table 11) and masked training.§1, §6
When does post-reasoning help?Gains are largest where the direct answer leaves capability unused: +4.29 points on GPQA and +3.00 on AMC, against +0.17 on GSM8K. They also hold on top of low-effort thinking: +4.26 points in 27 of 27 settings.§3, §4, App. F (Table 25)
Size and reliability of the aggregate gainsOne count throughout: 99 of 117 settings, with a mean gain of +2.11 points (r = 0.02 with the Direct baseline). Exact McNemar test: p = 1.0 × 10⁻⁴³. Sampled decoding over 3 seeds: significant in 10 of 12 cells.§4, App. F

Reviewer gS6M raised their rating to Weak Accept after the same answer-only baseline and paired analysis. They wrote that "the corrected reporting, the new paired analysis, and the revised cost claim address my remaining concerns about the empirical presentation."

We understand that this is a busy period. We hope the revision addresses your questions, and we would be grateful if you would take it into account in your final assessment. We are happy to answer any further questions before the discussion closes.

Sincerely,

The Authors

Comments on the paper

2 comments

===== COMMENT 1/2 =====

We thank you for the follow-up, and for confirming that the AMC 8 correction and the answer-only baseline address two of your concerns. Your remaining points concern the relation to answer-first distillation, the full TwT pipeline, the mechanism, and the reporting. We answer each point below. All changes are in red in the revised PDF.


[Clarification:1] Relation to PST and CoT-after-label distillation

We added both works to Related Work. Both train a small student on teacher rationales, with the answer before the rationale. The comparison with our work is:

PST [R1]CoT after label [R2]Post-reasoning (ours)
ModelsGPT2-Large (774M), T5-Large (700M)GPT-2, Phi-1.5, Gemma-2B13 instruction-tuned models, 4B–70B, open and proprietary
Rationale sourcegpt-3.5-turbo (teacher)Mistral-7B-Instruct (teacher)None for prompting. The model itself for training
Loss on the answerYes (weighted)YesNo (masked)
Effect of an instruction, no trainingNot testedNot tested99 of 117 settings improve

Our main result is the prompting result. Neither work tests if an instruction alone changes the answer of a model that is not trained for it. Thus, this result is not an extension of answer-first distillation: the effect is already present in instruction-tuned models, before any training.

For training, the two design choices that separate our objective from both works are tested directly:

Loss on the answer : A loss on the answer lowers accuracy in 9 of 9 settings (Table 16), and answer-only training is lower again in 9 of 9 settings.

Source of the rationale : Teacher rationales give lower accuracy than the model's own justifications on 4 of 4 benchmarks (Table 8).

The two effects also differ in kind. [R2] reports that its gains are robust to permutations of the rationale tokens. In our results, the content of the request matters: justification > confidence > summary, in the same answer-first format (Table 11).

(Continued in the next comment.)

===== COMMENT 2/2 =====

(Continued from the previous comment.)

[Clarification:2] The full TwT pipeline

Our claim about the training objective is specific: under a matched setup, a loss on the answer and teacher rationales each lower accuracy. The answer-only arm, Table 8 and Table 16 test exactly these two choices, with the same questions, LoRA configuration and seeds.

A full TwT run [R3] is not a matched test of this claim:

Different supervision : TwT uses pseudo-labels and rationales from GPT-4, GPT-4o-mini and Mistral-Large, with three training stages for each model. Its result combines the effect of the pipeline with the effect of stronger teachers. Our objective uses no teacher.

Different problem : TwT removes reasoning from the output, and its final model gives the answer only. Post-reasoning places the answer first and keeps the justification available.

Thus, the comparison that tests our claim changes only the loss target or only the rationale source. Both comparisons favor our objective: 9 of 9 settings over answer-only and unmasked training, and 4 of 4 benchmarks over teacher rationales.


[Clarification:3] The mechanism

A language model computes information for future tokens at the current position, and this "pre-caching" increases with model scale [R4]. With post-reasoning, the tokens after the answer must justify the answer. Thus, the computation at the answer position also prepares a justification. Our masked objective trains through exactly this gradient path: the loss is only on the justification, but the answer improves over Direct in 82 of 90 settings (Tables 9–10). Section 6 of the revision states the claim in this measured form. Our response to Reviewer xBRd (Clarification 1) gives the full argument.


[Clarification:4] Reporting

Counts : The revision reports one count everywhere: 99 of 117 settings improve (84.62%). Post-Reason SFT improves 82 of 90 settings over Direct (91.11%) and 66 of 90 over prompting (73.33%). Each Δ in every table is now the difference in accuracy points.

New result: paired significance : We tested Direct against Post on the same items with an exact McNemar test, in the 108 settings that have paired item logs:

Items wrong → rightItems right → wrongSignificant gainsSignificant decreasesPooled p
Direct → Post12,40110,31228 settings1 setting1.0 × 10⁻⁴³

Cost : Post-reasoning generates zero reasoning tokens before the answer, and the answer comes first in 99.97% of the 130,420 Post outputs. The revised title and text state the cost as negligible, not zero.


We hope that these answers, the new results and the revised manuscript address your remaining concerns. If so, we kindly ask you to consider raising your score.

References:

[R1] Chen, Xiaoshu, et al. "Post-semantic-thinking: A robust strategy to distill reasoning capacity from large language models." arXiv preprint arXiv:2404.09170 (2024).

[R2] Wadhwa, Somin, Silvio Amir, and Byron C. Wallace. "Investigating mysteries of CoT-augmented distillation." Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing (2024): 6071-6086.

[R3] Xu, Jingxian, et al. "TwT: Thinking without tokens by habitual reasoning distillation with multi-teachers' guidance." Findings of the Association for Computational Linguistics: EMNLP 2025 (2025): 16475-16489.

[R4] Wu, Wilson, John X. Morris, and Lionel Levine. "Do language models plan ahead for future tokens?" First Conference on Language Modeling (2024).


From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.