- Rating
- 2
- Confidence
- 3
- Informativeness
- 2
- Presentation
- 2
- Soundness
- 1
Read review
Paper summary
This paper shows that certain review prompts induce a systematic bias where models persistently maintain their initial answers. The authors provide further analysis to demonstrate that the bias stems from a stable review-induced state shift that causally alters model judgments. They then propose ReSET, a Review-Shift Editing Training framework designed to restore reliable self-reassessment
Critical strengths
- The review-induced bias appears to be a novel and interesting phenomenon. This bias is worth studying further.
Critical weaknesses
-
The comparison against SFT and vanilla distillation does not appear to be fair, which makes the reasons of ReSET’s gains unclear. In Appendix A.2, SFT takes the original question and learns the ground-truth answer, while vanilla distillation also takes the original question and learns the teacher’s answer distribution. However, ReSET receives a significantly richer training task such as conversations containing an initial answer and a reassessment instruction, and both Control and Reviewed contexts. This matters because the paper interprets large reductions in wrong persistence (WP) as evidence that ReSET repairs the proposed internal mechanism. I think the soundness of the soundness of this paper needs improvement.
-
The ablations show that most of the improvement survives removal of the state-editing component. Without state supervision, WP already decreases by 48.3 percentage points from the base, compared with 50.2 points for the full method. Thus, most of (or nearly all of) the observed WP reduction remains without state editing. So the claim that state alignment consistently improves judgment selectivity is not well supported.
Questions for authors
Can you please provide more details on how you obtained the results in Table 1? I have concerns on the fairness/soundness across different models.
4 comments
Response to Reviewer mctp (2/3)
Weakness 2
The ablations show that most of the improvement survives removal of the state-editing component. Without state supervision, WP already decreases by 48.3 percentage points from the base, compared with 50.2 points for the full method. Thus, most of (or nearly all of) the observed WP reduction remains without state editing. So the claim that state alignment consistently improves judgment selectivity is not well supported.
Thank you for this careful observation.
We agree that the change brought by state alignment is not very large in the comparison the reviewer cites. However, we would like to clarify that state alignment clearly helps reduce WP. We also believe that a change that is not very large does not necessarily mean that state alignment fails to improve judgment selectivity consistently.
Specifically, we extended the ablation to four training seeds (Table 18). Removing state alignment raises WP in every seed on both benchmarks, the increase in mean WP is significant, and CP stays at the same level. The contribution is also not marginal: on CMMLU, the WP reduction falls from 30.4 to 17.3 points without state alignment, so over 40% of it is lost.
Table R3.3. WP and CP of ReSET with and without (w/o) state alignment across four training seeds (Qwen3.5-9B, Reviewed, %; from Table 18). * significant difference from ReSET.
| Seed | MMLU-Pro ReSET WP | MMLU-Pro ReSET CP | MMLU-Pro w/o WP | MMLU-Pro w/o CP | CMMLU ReSET WP | CMMLU ReSET CP | CMMLU w/o WP | CMMLU w/o CP |
|---|---|---|---|---|---|---|---|---|
| 20260824 | 35.1 | 94.0 | 37.0 | 93.0 | 65.5 | 97.7 | 67.0 | 95.3 |
| 11 | 39.0 | 98.0 | 44.6 | 97.6 | 56.5 | 98.7 | 75.0 | 98.7 |
| 22 | 35.9 | 96.7 | 40.0 | 98.4 | 42.1 | 98.8 | 60.0 | 100.0 |
| 33 | 35.9 | 96.7 | 41.5 | 98.0 | 57.1 | 97.5 | 71.4 | 98.7 |
| Mean | 36.5 | 96.4 | 40.8* | 96.8 | 55.3 | 98.2 | 68.4* | 98.2 |
Response to Reviewer mctp (3/3)
Question
Can you please provide more details on how you obtained the results in Table 1? I have concerns on the fairness/soundness across different models.
Thank you for raising this question. Table 1 is obtained by a paired comparison within each model: each model first answers a fixed set of questions (a 20% stratified sample for benchmarks above 1,000 questions, the full set otherwise), and then reassesses these same questions and its own initial answers under both Control and Reviewed. The two conditions are therefore matched in data quantity, data type, prompts, and decoding settings and differ only in the review framing, which makes the comparison fair and sound for every model.
The details can be found in the paper:
- Evaluation protocol and metric definitions: Section 3 (lines 155–161 and 184–189).
- Control and Reviewed prompts: Appendix B (lines 763–779).
- Decoding settings: Appendix A.2 (lines 733–734).
- Results for each model and each benchmark: Table 1 and Table 8.
We hope this addresses your concerns, and we welcome further discussion if any questions remain.
Looking forward to your feedback
Dear Reviewers,
Thank you again for your detailed feedback. In response, we have added experiments and detailed explanations, such as a fairer SFT baseline, a four-seed ablation of state alignment, and details of the Table 1 protocol, hoping to address your concerns.
As the discussion period draws to a close, we would greatly appreciate any further feedback you may have. We are committed to ensuring that our rebuttal aligns well with your suggestions, as your feedback is incorporated into the enhancement of our work's quality. Please let us know if any concerns remain or if additional clarification would be helpful. We would be more than happy to discuss them further.
Once again, we sincerely appreciate your time and valuable feedback.
Thanks and regards.
Response to Reviewer mctp (1/3)
Thank you for your valuable comments.
Weakness 1
We appreciate the reviewer’s careful reading of the baseline settings.
We would like to clarify that the original SFT and vanilla-distillation baselines were designed to examine whether supervision on correct answers alone mitigates the bias, and it does not.
We agree that a comparison with identical inputs is fairer. Therefore, we retrained SFT on exactly the same training inputs as ReSET and added it to the ablation (Table 4) and the transfer experiment (Table 5). The conclusion is unchanged: ReSET remains clearly better than SFT in both. Specifically:
(1) The ablation shows that same-input SFT remains 11.0 and 6.1 accuracy points below ReSET, with higher WP on both benchmarks. Removing ReSET’s response text, the matched counterpart of vanilla distillation, raises WP to 100.
Table R3.1. Comparison of ReSET with the SFT and distillation baselines and the response-text ablation on Qwen3.5-9B under Reviewed (%).
(2) The cross-task transfer experiment demonstrates the generalizability of ReSET, which same-input SFT lacks. Trained on MMLU-Pro and evaluated on MATH-500, same-input SFT falls below the backbone in accuracy with higher WP on all three models, whereas ReSET improves both.
Table R3.2. Transfer from MMLU-Pro to MATH-500 under Reviewed (%).