acceptodds
Under review as a conference paper at ICLR 2027

Thinking Longer, Refusing Less: Reasoning Shifts What Models Accept

Abstract

Thinking longer is expected to help reasoning models catch what is wrong with a request, but it may instead talk them into accepting it. To find out which, we cut the natural reasoning of ten open-weight models at several points, so only how much of one trajectory precedes the answer varies, and record whether each answer accepts harmful and benign requests, false and true premises, and wrong and correct proposed answers. Thinking longer shifts what models accept. They accept more harmful requests and false premises, most in the longest trajectories, while they refuse fewer benign requests and accept no more proposed wrong answers. In nearly every trajectory that ends in wrong acceptance, the model writes a reason to accept, usually an assumption of its own, and removing this reason lowers wrong acceptance. Motivated by this observation, we propose Stance-Contrastive On-Policy Self-Distillation (SC-OPSD), which distills into the weights a verification that the same model writes without seeing its reasoning, out of reach of such reasons. On Qwen3.5-4B and OLMo-3-7B-Think, SC-OPSD accepts fewer harmful requests and false premises than the base model on every source, especially where the base model thinks longest. Its harmful acceptance changes little as it thinks longer and at 80% of its thinking stays below the base model's at 20%, demonstrating that thinking longer need not mean refusing fewer harmful requests.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
CoT, Reasoning Models, LLM safety, Interpretability, Alignment

Reviews

Among:
Shape:
1234this paper, Bottom 10%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.67 over 3 reviews, above 10% of the others.
Reviewer_ewpXOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how longer reasoning affects whether reasoning models accept or reject user requests. Across ten open-weight models, the authors find that later reasoning prefixes tend to increase acceptance of harmful requests and false premises, and they attribute this partly to self-generated justifications that shift the model’s stance. They further propose SC-OPSD to reduce this acceptance drift by introducing an independent verification signal and distilling it into the model.

Critical strengths
  1. The paper is well-motivated. It provides clear examples of the motivation issue.
  2. The paper is easy to understand.
  3. The performance of the proposed method is good.
Critical weaknesses
  1. The claim of “thinking longer” is somewhat narrower than the paper’s framing. The model does not independently generate trajectories under different reasoning budgets; instead, the authors truncate a single completed trajectory at different points and resample the final answer. This establishes the effect of exposing the answer policy to a longer prefix of the same reasoning trace, rather than the effect of naturally allocating more inference-time compute.

  2. The generalization to frontier reasoning models remains unclear. The paper only evaluates 10 open-weight models which are relatively small and weak compared wtih frontier models such as GPT-6-Astra. It is therefore unclear whether the same phenomenon persists in stronger frontier models with different safety mechanisms. This limits the scope of the paper.

  3. The detailed results show heterogeneity that is partially obscured by the averaged results. In particular, GPT-OSS-20B and GPT-OSS-120B show the opposite trend on harmful acceptance, and GPT-OSS-120B also shows decreased false-premise acceptance with longer reasoning. There is a lack of more specific discussion on this condition.

  4. The explanation remains largely behavioral rather than mechanistic. While the paper shows that self-generated justifications can causally shift final answers through removal and insertion interventions, it does not identify the internal mechanism that produces or propagates this stance change during reasoning. The current evidence is better viewed as a behavioral account of the phenomenon rather than a mechanistic

  5. The method contribution feels relatively limited compared with the empirical part of the paper. SC-OPSD appears largely as an application of existing OPSD with a different form of privileged teacher context, rather than a novel method. More importantly, the mitigation method does not further explain or validate the core phenomenon of reasoning-induced acceptance drift, so this section feels somewhat disconnected from the paper’s main scientific question.

Questions for authors
  1. Does the observed acceptance shift persist when “thinking longer” is induced by genuinely larger inference-time reasoning budgets, rather than by truncating different prefixes of the same completed reasoning trajectory?

  2. Does the observed relationship between reasoning length and acceptance generalize to substantially stronger frontier reasoning models with different post-training and safety mechanisms?

  3. What explains the substantial cross-model heterogeneity in the effect of longer reasoning, particularly the reversed trend observed in GPT-OSS-20B and GPT-OSS-120B?

  4. What internal mechanism causes longer reasoning to shift the model from withholding to acceptance?

  5. Does SC-OPSD provide a technical contribution beyond applying existing OPSD, and does it reveal anything new about the mechanism underlying acceptance drift?

Reviewer_ppiiOpenReview ↗
Rating
1
Confidence
3
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

This paper first argues that the longer a model reasons, the less it refuses, leading to unsafe outcomes. It then analyzes the issue and finds that wrong acceptance may be caused by assumption-based justifications. Finally, it proposes SC-OPSD, which uses DSV verification and a stance-free summary as two teacher contexts to train the model and improve safety.

Critical strengths
  • The experimental design is good, the experiments are thorough, and the paper covers many models and datasets;
  • Some of the controlled-variable experiments are rigorous.
Critical weaknesses
  • There is a mismatch between title and experimental results. The title says "Think longer, refuse less." However, in the "accepting is wrong" panels of Figure 2, most curves do not continue to increase from 40% to 100%. Does this suggest that increasing reasoning length from 40% to 100% has no effect, and that the most important stage is 0%–40%? Does this further suggest that some behavior in the early stage of reasoning drives wrong acceptance?
  • The causal explanations behind some numbers are not described clearly enough.
    • According to Sec. 3.3, more than 90% have justifications, but the introduction says that removing the justification only lowers it by 6 points. I am not sure whether this value is significant.
    • The decoupled verification also differs by only 2–3%. Does this not suggest that the impact of justification on the model is very limited?
Questions for authors
  1. It would make the paper’s logic smoother if there were experiments verifying that after SC-OPSD, justifications become fewer and this change is highly correlated with performance changes.
  2. It is true that both 20% and 80% are truncated rather than naturally ended, but the reasoning stages at different positions are different. Can they really be compared? For example, at the 20% position, is it possible that the first round of reasoning has not even finished and the justification has not yet been generated?
  3. In the counterfactual experiment in Sec. 3.3, the paper says the model is allowed to continue thinking. How does it ensure that the model will not generate a new justification while continuing to think?
  4. It is suggested to use the same flagship model with different reasoning efforts (low, medium, high, max) for comparison. Otherwise, I think the title does not match the content of the paper.
Reviewer_Jh7hOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
1
Soundness
2
Read review
Paper summary

The paper asks whether longer reasoning helps reasoning models push back on a user's direction, or talks them into accepting it. "Direction" covers three cases: a harmful request, a false premise, or a proposed wrong answer. The authors cut natural reasoning traces from 10 open-weight models at 0–100% and let the model answer from each prefix. From the 20% cut to the 80% cut, acceptance rises on harmful requests (+1.5 to +3.0 points) and false premises (+3.7). It does not rise on wrong proposed answers, and benign requests get more help. An LLM-coded analysis finds that almost every wrongly-accepting trace contains a written "justification", usually resting on an assumption the model made up. Removing that justification lowers wrong acceptance. The authors then propose SC-OPSD. It distills a verification written without the reasoning into the weights, using a per-token reward contrasted against a stance-free summary. It lowers harmful and false-premise acceptance on Qwen3.5-4B and OLMo-3-7B-Think.

I sincerely believe the novelty and content of this work has potential to be a strong conference paper. The writing, however, made my job of understanding this paper significantly more difficult than it should be. Should the paper be written in better English, it deserves a score of 4. But unfortunately given the current draft, I can only give it a 2. I also don’t think this paper can be improved by just rewriting 1 or 2 paragraphs for rebuttal. My biggest advice is to break down some super-long sentences (e.g., line 087-090) into shorter ones and give a proper definition every time you use a new term.

Critical strengths
  1. The main research question matters: LLMs could accept more harmful requests or false premise after thinking. Moreover, separating acceptance from correctness is a real conceptual contribution. Scores on a single axis hide shifts that point opposite ways on harmful and benign requests.

  2. The within-trajectory cut design holds the model, the request and the sampled trace fixed. The controls are thoughtful: benign requests, true premises, correct candidates, a with/without closing-clause check, and a raised answer budget. The judge validation is thorough: a second judge, a blind third-rater audit, and checks for judge drift across cut points.

  3. The removal experiment includes a paraphrase control and a placebo edit, and the insertion experiment suggests the same way.

Critical weaknesses
  1. Very long sentences full of appositives make the paper hard to follow. Examples are lines 087–090, 112–114, the abstract's "so only how much of one trajectory precedes the answer varies", and lists like "the three harmful and the two false-premise sources". Splitting them into shorter sentences would help a lot.

  2. Many terms are used before they are defined, or never defined. Regarding SC-OPSD, what is “the verification” and “a stance-free summary of the same request”. Why do we need the summary rather than the original request? Is the verification written by the base policy? Please give an actual example of these terms in Figure 5.

  3. SC-OPSD is hard to understand from the main text. The paper should show a real H and N for one request in Fig. 5. It should also state in Sec. 4.2 that x appears in both teacher contexts, who writes H and N, and that H is filtered by label (and, for benign requests, by GPT-5.4-mini).

  4. SC-OPSD's gains don't match the mechanism the paper proposes. DSV at inference barely moves false-premise acceptance: +0.2 on Qwen3.5-4B, −2.6 averaged over models. Yet SC-OPSD cuts false-premise acceptance by 33–54 points. Several factors that are not "distilling the verification" could explain this. H is filtered by label. The training pool includes FalseQA train, UMWP and synthetic unanswerable math, which closely match the test sources. And the "another request's N" ablation works on OLMo. The authors should explain where the gain comes from, for example with an ablation that uses unfiltered H.

Questions for authors

See my weakness points above.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.