acceptodds
Under review as a conference paper at ICLR 2027

Under Observation: Mitigating Review-Induced Bias in LLM Reflection

Abstract

Self-reflection serves as a critical capability for large language models to detect errors, revise prior outputs, and ensure reliable agentic operations. To enhance this process, many evaluation frameworks and practical agent systems explicitly inform the model that its output is being reviewed, assuming this evaluative context will encourage more rigorous reassessment. However, we demonstrate that these review cues fail to improve overall performance and instead induce a systematic bias where models persistently maintain their initial answers. Further analysis reveals that this bias stems from a stable review-induced state shift that causally alters model judgments, even when accurate correctness information remains preserved within internal representations. To address this issue, we propose ReSET, a Review-Shift Editing Training framework designed to restore reliable self-reassessment. ReSET combines correctness-verified teacher responses with answer-distribution distillation to train models to re-solve problems before making judgments and effectively neutralize the review-induced shift. Comprehensive evaluations demonstrate that ReSET consistently reduces wrong-answer persistence while preserving correct-answer retention, leading to improved judgment discrimination and strong cross-task generalization. Ultimately, this work reveals the inherent behavioral shift of large language models under external observation, offering valuable insights for the future optimization of reliable LLM self-evaluation.

Primary area:
foundation or frontier models, including LLMs
Keywords:
Large Language Models, Self-reflection, Evaluation awareness, Distillation

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_mctpOpenReview ↗
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
  1. The review-induced bias appears to be a novel and interesting phenomenon. This bias is worth studying further.
Critical weaknesses
  1. 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.

  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.

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
AuthorsOpenReview ↗

Response to Reviewer mctp (1/3)

Thank you for your valuable comments.

Weakness 1

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

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 (%).

MethodMMLU-Pro Acc ↑\uparrowWP ↓\downarrowCP ↑\uparrowCMMLU Acc ↑\uparrowWP ↓\downarrowCP ↑\uparrow
Base54.885.389.278.085.794.1
Vanilla distillation64.982.288.178.990.194.7
SFT (same inputs as ReSET)60.462.986.874.190.296.8
ReSET w/o response text56.2100.099.276.7100.0100.0
ReSET71.435.194.080.265.597.7

(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 (%).

MethodQwen3.8-27B Acc ↑\uparrowQwen3.8-27B WP ↓\downarrowGemma-4-12B Acc ↑\uparrowGemma-4-12B WP ↓\downarrowGemma-4-26B Acc ↑\uparrowGemma-4-26B WP ↓\downarrow
Backbone54.850.553.847.859.065.9
Vanilla distillation48.281.957.487.555.872.1
SFT (same inputs as ReSET)35.869.532.278.640.073.4
ReSET88.61.282.82.690.01.9
AuthorsOpenReview ↗

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.

SeedMMLU-Pro ReSET WP ↓\downarrowMMLU-Pro ReSET CP ↑\uparrowMMLU-Pro w/o WP ↓\downarrowMMLU-Pro w/o CP ↑\uparrowCMMLU ReSET WP ↓\downarrowCMMLU ReSET CP ↑\uparrowCMMLU w/o WP ↓\downarrowCMMLU w/o CP ↑\uparrow
2026082435.194.037.093.065.597.767.095.3
1139.098.044.697.656.598.775.098.7
2235.996.740.098.442.198.860.0100.0
3335.996.741.598.057.197.571.498.7
Mean36.596.440.8*96.855.398.268.4*98.2
AuthorsOpenReview ↗

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.

AuthorsOpenReview ↗

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.

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

This paper studies how evaluative framing affects LLM self-reassessment. The authors find that informing a model that its response is being reviewed increases its tendency to endorse its initial answer, including incorrect answers, while producing relatively small changes in final-answer accuracy. They investigate a review-associated direction in hidden representations and propose ReSET, which combines verified reassessment responses, state supervision, and answer-distribution distillation. The work presents an interesting behavioral phenomenon with potential relevance to multi-agent systems, although its practical significance and the representativeness of the benchmark-based evidence require further clarification.

Critical strengths

A clear and controlled behavioral investigation. The paired design holds the question and initial answer fixed while changing the review framing, providing a useful way to examine how evaluative cues influence self-reassessment.

Potential relevance to multi-agent interactions. The findings could inform systems in which agents reassess their outputs under reviewer or verifier roles. Understanding whether such framing encourages accurate correction or merely changes endorsement behavior is a worthwhile research direction.

Supporting mechanistic and intervention evidence. Bidirectional activation steering and random-direction controls strengthen the analysis beyond behavioral correlations. ReSET also shows substantial improvements in reassessment accuracy and judgment discrimination across several evaluation settings.

Critical weaknesses

The connection between the social-framing phenomenon and practical agent reliability remains underdeveloped. The central finding concerns a model’s response to being told that it is under observation, which primarily resembles a social or behavioral framing effect. This is interesting, but the experiments provide limited evidence about its consequences for substantive reasoning or downstream task success, particularly because review cues cause relatively small changes in final-answer accuracy. The potential application to multi-agent systems is promising, yet remains largely motivational: the evaluated reviewer supplies neither substantive feedback nor an alternative solution. Experiments involving actual reviewer-agent feedback, repeated interactions, and downstream outcomes would help establish the practical importance of the phenomenon.

The combination of relatively advanced models and established benchmarks may limit the representativeness of the results. Much of the evaluation uses mature, static QA benchmarks, including MMLU-Pro, CMMLU, MedQA, and MMMU. For recent models, performance on these datasets may reflect benchmark familiarity or prior exposure alongside task-solving ability. Moreover, the remaining initially incorrect examples may form a relatively small and atypical subset, affecting the interpretation and stability of conditional metrics such as wrong-answer persistence. These possibilities do not establish that the results are biased, but they warrant investigation. Evaluation on fresh, contamination-controlled problems and realistic agent tasks, together with sample counts and uncertainty estimates for the initially incorrect subsets, would strengthen the claims.

Questions for authors

see weakness

8 comments
AuthorsOpenReview ↗

Response to Reviewer ytku (1/5)

Thank you for the detailed feedback.

Weakness 1

The connection between the social-framing phenomenon and practical agent reliability remains underdeveloped. The central finding concerns a model’s response to being told that it is under observation, which primarily resembles a social or behavioral framing effect. This is interesting, but the experiments provide limited evidence about its consequences for substantive reasoning or downstream task success, particularly because review cues cause relatively small changes in final-answer accuracy. The potential application to multi-agent systems is promising, yet remains largely motivational: the evaluated reviewer supplies neither substantive feedback nor an alternative solution. Experiments involving actual reviewer-agent feedback, repeated interactions, and downstream outcomes would help establish the practical importance of the phenomenon.

We appreciate the reviewer’s recognition of this phenomenon and of its potential for multi-agent systems.

We would like to clarify that our study targets the language model itself: how a review cue changes the way a model reassesses its own answer. Recently studies of related behaviors, the ELEPHANT benchmark for social sycophancy [1] and activation steering of evaluation awareness [2], likewise study the model alone.

We agree that extending the phenomenon to agents is valuable. To verify it, we evaluated Qwen3.5-9B and Qwen3.5-4B on four widely used agent benchmarks: τ2\tau^2-bench, AgentDojo, BFCL multi-turn, and AppWorld, and we further ran an executor–reviewer experiment with actual reviewer feedback. The review effect extends to agent tasks, and it changes downstream outcomes once the model’s own judgment gates a later step. Specifically:

AuthorsOpenReview ↗

Response to Reviewer ytku (2/5)

(1) On all four agent benchmarks, the review cue makes Qwen3.5-9B more likely to judge a failed execution as completed. Table R2.1 (added as Table 9) reports how often each model judges its own execution as completed, split by whether the benchmark’s checker confirms it. The review-induced bias appears at both Qwen3.5 sizes, more strongly for 9B: under Reviewed, judgment persistence rises on initially correct and initially wrong executions alike.

Table R2.1. Judgment persistence (%) on agent benchmarks, i.e., the rate of judging the original execution as completed, under Control (C) and Reviewed (R).

BenchmarkInitial AccInitially correct (C) ↑\uparrowInitially correct (R) ↑\uparrowInitially wrong (C) ↓\downarrowInitially wrong (R) ↓\downarrowAll (C)All (R)
Qwen3.5-4B
τ2\tau^2-bench60.599.299.23.83.861.561.5
AgentDojo81.76.017.30.02.34.914.6
BFCL60.010.017.56.28.88.514.0
AppWorld4.285.785.76.26.29.69.6
Qwen3.5-9B
τ2\tau^2-bench75.897.098.022.738.179.083.5
AgentDojo86.528.951.212.626.826.747.9
BFCL53.042.562.312.827.728.546.0
AppWorld12.585.7100.047.652.452.458.3

(2) With actual reviewer feedback over three rounds, the review cue lowers the final accuracy of the pipeline. On MMLU-Pro, a tool-using reviewer agent built on the same model returns a written analysis and a suggested answer to the Qwen3.5-9B executor. The executor’s first judgment decides whether an answer is sent to the reviewer. The Reviewed cue halves the wrong answers that reach the reviewer and lowers the final accuracy by 3.5 points (95% interval excludes zero).

Table R2.2. Executor–reviewer pipeline on MMLU-Pro (Qwen3.5-9B, %).

OutcomeControlReviewed
Wrong answers sent to the reviewer ↑\uparrow26.213.1
Final accuracy of the pipeline ↑\uparrow59.856.3

(3) Small accuracy changes hide a consistent rise in wrong-answer persistence. Across the 92 model–benchmark pairs in Table 8, the cue raises accuracy in 41 and lowers it in 41, yet raises WP in 75 and lowers it in only 5. This cost is invisible to accuracy and appears once the judgment gates a later step, as in (1) and (2).


[1] Cheng, M., Yu, S., Lee, C., Khadpe, P., Ibrahim, L., & Jurafsky, D. (2026). ELEPHANT: Measuring and understanding social sycophancy in LLMs. In International Conference on Learning Representations.

[2] Hua, T. T., Qin, A., Marks, S., & Nanda, N. (2026). Steering evaluation-aware language models to act like they are deployed. In International Conference on Learning Representations.

AuthorsOpenReview ↗

Response to Reviewer ytku (3/5)

Weakness 2

The combination of relatively advanced models and established benchmarks may limit the representativeness of the results. Much of the evaluation uses mature, static QA benchmarks, including MMLU-Pro, CMMLU, MedQA, and MMMU. For recent models, performance on these datasets may reflect benchmark familiarity or prior exposure alongside task-solving ability. Moreover, the remaining initially incorrect examples may form a relatively small and atypical subset, affecting the interpretation and stability of conditional metrics such as wrong-answer persistence. These possibilities do not establish that the results are biased, but they warrant investigation. Evaluation on fresh, contamination-controlled problems and realistic agent tasks, together with sample counts and uncertainty estimates for the initially incorrect subsets, would strengthen the claims.

We appreciate the reviewer’s careful point.

We would like to clarify that exposure to these benchmarks during training is not established, and any exposure is identical in Control and Reviewed, which share the same questions and initial answers.

We agree that fresh problems extend the value of the study. Therefore, we added RealTime QA (only questions released after each model’s release date) and MMLU-CF (a contamination-free rebuild of MMLU) to Table 8. The effect holds on these fresh problems, and the initially incorrect subset is neither small nor atypical. Specifically:

Reviewer_aF1hOpenReview ↗
Rating
3
Confidence
2
Informativeness
3
Presentation
3
Soundness
3
Read review
Paper summary

The paper finds that telling a model its answer is being reviewed makes it more likely to say that its initial answer is correct, even when the initial answer is wrong. It indicates that the model’s self-assessment less reliable for systems that use it to decide whether further checking is needed. To understand this effect mechanically, the authors use activation steering to test whether the difference affects the model’s judgment, and linear probes to test whether answer correctness can still be read from its hidden states. Motivated by this analysis, they propose ReSET, which trains the model to solve the problem again before judging its initial answer.

Critical strengths
  1. This paper observes the discrepancy between self-judgment and final answer accuracy. In the aggregated results across 14 models, review prompts consistently increased the approval rate for incorrect answers (WP), but the overall accuracy of the final answers remained largely unchanged. By using the same initial answers under both conditions, the authors were able to isolate the impact of the review prompts.

  2. The improvements brought by ReSET go beyond merely making the model more prone to saying that original answer is wrong. In experiments where the same initial answers were fixed, the authors also observed an actual improvement in error correction.

  3. The authors test transfer to a different task and answer format. This tests whether a model trained on multiple-choice reassessment can also improve on free-form mathematics.

Critical weaknesses
  1. The intervention shows an effect on judgment, but its connection to answer correction is unclear. In Section 4.1, the authors intervene just before the model outputs its judgment, after it has already generated its reasoning and final answer. The experiment therefore shows that the intervention can change the judgment; it does not by itself show that the same state shift caused the earlier answer or reasoning. I am also unsure how to interpret the changes in final-answer accuracy reported in Table 2 under this intervention.

  2. It is unclear how much state editing contributes to ReSET. On Qwen3.5-9B and MMLU-Pro, removing state alignment gives 70.0% accuracy and 37.0% WP, close to the full method's 71.4% and 35.1% (Table 4). On Qwen3.5-4B, the no-state variant even has slightly lower WP, although its accuracy also drops (Table 13). The SFT and vanilla-distillation baselines are less direct comparisons because they do not use the same verified reassessment responses and judgment labels as ReSET (Appendix A.2). It would help to compare the no-state variant with ReSET on the fixed-initial-answer and transfer experiments, and across training seeds. Also, Equation 3 preserves the difference between the Control and Reviewed target states. What evidence shows that ReSET removes the review-induced shift, rather than improving judgments despite that difference?

Questions for authors

Please check my weakness section

3 comments
AuthorsOpenReview ↗

Response to Reviewer aF1h (1/3)

We appreciate your valuable feedback and support of our work.

Weakness 1

The intervention shows an effect on judgment, but its connection to answer correction is unclear. In Section 4.1, the authors intervene just before the model outputs its judgment, after it has already generated its reasoning and final answer. The experiment therefore shows that the intervention can change the judgment; it does not by itself show that the same state shift caused the earlier answer or reasoning. I am also unsure how to interpret the changes in final-answer accuracy reported in Table 2 under this intervention.

Thank you for this careful feedback, and we apologize for the confusion. The intervention is applied before the reasoning and the final answer are generated, and it changes both the judgment and the answer correction that follows.

(1) The intervention precedes the reasoning and the answer. Section 4.1 uses the order judgment →\rightarrow reason →\rightarrow answer (Figure 2; lines 208–210), with the intervention applied immediately before the judgment (lines 218–220). Only ReSET (Section 4.2) uses reason →\rightarrow answer →\rightarrow judgment. The two orders serve different purposes. Judging first places everything the model generates downstream of a single intervention, so its causal effect on the judgment, the reasoning, and the final answer can be measured in one pass. Re-solving first makes ReSET’s judgment rest on a fresh solution rather than on the initial answer.

(2) The shift controls persistence in both directions (Table 2), and answer correction moves with it. Adding Dˉ\bar{D} to Control, which contains no reviewer cue, raises wrong-answer persistence to the Reviewed level, and subtracting it under Reviewed lowers persistence. The final answers follow: under Reviewed −2Dˉ-2\bar{D}, the share of initially wrong answers that end up corrected rises from 13.6% to 20.7%, and when the judgment does not flip, the reasoning and answer rarely change.

(3) Accuracy stays flat because reopening an answer yields both corrections and new errors. Under Reviewed −2Dˉ-2\bar{D}, the flipped cases include 16 wrong answers corrected and 11 correct answers changed to incorrect ones, so accuracy stays within 60.1–61.0%.

AuthorsOpenReview ↗

Response to Reviewer aF1h (2/3)

Weakness 2

It is unclear how much state editing contributes to ReSET. On Qwen3.5-9B and MMLU-Pro, removing state alignment gives 70.0% accuracy and 37.0% WP, close to the full method's 71.4% and 35.1% (Table 4). On Qwen3.5-4B, the no-state variant even has slightly lower WP, although its accuracy also drops (Table 13). The SFT and vanilla-distillation baselines are less direct comparisons because they do not use the same verified reassessment responses and judgment labels as ReSET (Appendix A.2). It would help to compare the no-state variant with ReSET on the fixed-initial-answer and transfer experiments, and across training seeds. Also, Equation 3 preserves the difference between the Control and Reviewed target states. What evidence shows that ReSET removes the review-induced shift, rather than improving judgments despite that difference?

We thank the reviewer for the constructive suggestions. We would like to clarify that state editing constrains the hidden state at the judgment position, where we extract the review-induced direction and show by steering that it causally changes endorsement (Sections 4.1 and 5.2). Its contribution and the fairness of the comparison are supported from three aspects: (1) across four training seeds, state alignment consistently lowers WP; (2) Eq. 3 removes the review-induced displacement where it is harmful; and (3) with identical inputs and judgment labels, ReSET remains better than SFT. Specifically:

First, both the loss ablations (Table 4) and the mechanistic analyses of the judgment state (Section 5.2, Table 2) support the effectiveness of this component. In addition, we added an ablation across four training seeds (Table 18): removing state alignment raises WP in every seed (Table R1.1), which demonstrates that its contribution is consistent.

We would like to note that state alignment targets the judgment, not accuracy, so a small accuracy change is expected. On Qwen3.5-4B (Table 14), the WP difference on MMLU-Pro is marginal while state alignment clearly lowers WP on CMMLU, so its effectiveness still holds.

Table R1.1. ReSET with and without (w/o) state alignment across four training seeds (Qwen3.5-9B, Reviewed, %; from Table 18). * significant difference from ReSET.

SeedMMLU-Pro ReSET Acc ↑\uparrowMMLU-Pro ReSET WP ↓\downarrowMMLU-Pro w/o Acc ↑\uparrowMMLU-Pro w/o WP ↓\downarrowCMMLU ReSET Acc ↑\uparrowCMMLU ReSET WP ↓\downarrowCMMLU w/o Acc ↑\uparrowCMMLU w/o WP ↓\downarrow
2026082471.435.170.037.080.265.579.967.0
1172.439.069.944.683.056.581.075.0
2271.435.972.040.090.042.187.060.0
3371.235.971.441.585.057.184.071.4
Mean71.636.570.840.8*84.655.383.068.4*
AuthorsOpenReview ↗

Response to Reviewer aF1h (3/3)

Second, Eq. 3 sets the target of an initially wrong answer under Reviewed to exactly its Control state, and after ReSET the WP under Reviewed is no longer higher than under Control (51.7 vs. 53.4 in Table 15).

Third, we added SFT trained on exactly the same inputs and judgment labels as ReSET to the ablation and transfer experiments: ReSET remains clearly better in both accuracy and WP.

Table R1.2. Comparison of ReSET with the SFT and distillation baselines and the response-text ablation on Qwen3.5-9B under Reviewed (%).

MethodMMLU-Pro Acc ↑\uparrowWP ↓\downarrowCP ↑\uparrowCMMLU Acc ↑\uparrowWP ↓\downarrowCP ↑\uparrow
Base54.885.389.278.085.794.1
Vanilla distillation64.982.288.178.990.194.7
SFT (same inputs as ReSET)60.462.986.874.190.296.8
ReSET w/o response text56.2100.099.276.7100.0100.0
ReSET71.435.194.080.265.597.7

Table R1.3. Transfer from MMLU-Pro to MATH-500 under Reviewed (%).

MethodQwen3.8-27B Acc ↑\uparrowQwen3.8-27B WP ↓\downarrowGemma-4-12B Acc ↑\uparrowGemma-4-12B WP ↓\downarrowGemma-4-26B Acc ↑\uparrowGemma-4-26B WP ↓\downarrow
Backbone54.850.553.847.859.065.9
Vanilla distillation48.281.957.487.555.872.1
SFT (same inputs as ReSET)35.869.532.278.640.073.4
ReSET88.61.282.82.690.01.9

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.