acceptodds
Under review as a conference paper at ICLR 2027

Beyond Spurious Reflection: Strategic Self-Correction as State Replacement in LLMs

Abstract

Large reasoning models excel by scaling inference-time Chain-of-Thought, yet their performance is brittle: an early mistake corrupts all subsequent reasoning. Reinforcement-learning-trained reasoners often produce reflective phrases like "wait, let me reconsider," but growing evidence shows these reflections are largely spurious–they neither alter reasoning paths nor final answers. Existing self-reflection methods keep the faulty chain in context, forcing the model to overcome its own errors. We propose REFLECT+, an end-to-end framework that directly optimizes reflection-triggered self-correction. Building on the Clean-Slate Reflection paradigm–where the model emits a structured <reflection> block summarizing verified sub-conclusions, diagnosed mistakes, and corrected directions, then continues from the problem and reflection alone–REFLECT+ adopts a two-stage training scheme combining supervised cold-start and role-aware reinforcement learning. This enables strategic decisions on when to reflect, what to retain, and how to resume. Experiments on Qwen2.5-Math-1.5B / Qwen3-0.6B-Base / Qwen3-4B-Base across MATH500, AMC23, GSM8K, LogiQA, GPQA, MMLU-Pro, BBH, CLUTRR and DROP evaluate both self-correction quality and reasoning accuracy, and outperforms GRPO on accuracy with selective reflection.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
Large Language Models, Self-Correction, Reflection, Reinforcement Learning, Chain-of-Thought Reasoning, Autonomous Reasoning

Reviews

Among:
Shape:
1234this paper, Top 15%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.00 over 3 reviews, above 85% of the others.
Reviewer_TCceOpenReview ↗
Rating
3(+1)
Confidence
4(+1)
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

REFLECT+ trains a model to interrupt its reasoning, produce a structured reflection, and resume from the problem and reflection while discarding the previous reasoning prefix. It combines supervised initialization with modified GRPO rewards and advantages, and reports improvements across several reasoning benchmarks and model sizes.

Critical strengths

The idea is practical, and the gains on the 1.5B model are encouraging. Figure 4 provides a useful comparison between retaining and discarding the faulty prefix, while the reflection-source and fixed-reflection continuation experiments go beyond reporting aggregate accuracy.

Critical weaknesses
  1. The paper has not separated self-correction from a fresh start. Figure 4 supports removing the faulty prefix, but does not compare against restarting from the problem alone at the same trigger point and with the same remaining budget. The reflection may preserve useful progress, or the model may simply benefit from abandoning a failed attempt. These are different explanations, and the current experiments do not tell them apart. This leaves the central claim of corrective state replacement unresolved.

  2. The RL objective encourages reflection, but its corrective value is not established. A failed reflective rollout receives 0 while a direct failure receives -0.25; when all rollouts fail and both types are present, the failed reflection still gets positive advantage. This is understandable as exploration shaping, but it does not establish suppression of spurious reflection. The paper also changes several components together without isolating the proposed baseline and advantage floor. I cannot tell whether the gains require this credit-assignment scheme or mainly come from the initialization, reward shaping, and context reset.

  3. The main baselines do not establish an advance over existing self-correction methods. Base, SFT, and vanilla GRPO are useful training comparisons, but they are not a substitute for competing self-correction approaches. The paper discusses methods such as Self-Refine and ReActR without comparing against them. A matched-backbone, matched-budget feedback-and-retry baseline and a closely related training-based method would make the contribution much clearer. As it stands, outperforming vanilla GRPO does not show that REFLECT+ is a better way to learn self-correction.

  4. The general diagnosis goes beyond the models tested. The complete method is evaluated on 0.6B, 1.5B, and 4B Qwen-family backbones. This does not distinguish a general limitation of reasoning models from small models struggling to recover in the presence of misleading context. Larger models used as external continuers provide useful evidence, but do not test the complete learned policy. The paper acknowledges capacity-dependent gains, yet its broader claims about existing self-correction still need stronger evidence or a narrower scope.

Questions for authors
  1. Does the reflection actually help beyond a fresh restart from the problem alone, using the same model and remaining token budget?
  2. With the same supervised initialization, clean-slate rollout, and reward, what do the proposed baseline and advantage floor add over ordinary GRPO? What evidence shows they reward useful correction rather than merely encourage reflection?
  3. How does REFLECT+ compare with an existing feedback-and-retry method and a closely related self-correction training method under matched conditions?
  4. Does clean-slate reflection remain necessary for stronger, well-trained reasoners, or are the gains mainly compensating for limited correction abilities in small models?
5 comments

Thank you for identifying four issues: the need for a question-only restart control, the attribution of gains to individual training components, comparisons with related methods, and the scope of our conclusions. We address your four weaknesses and four questions separately below.

Weakness 1. Does retaining the reflection help beyond restarting from the question alone?

To distinguish the effect of retaining the reflection from that of restarting, we conducted a paired comparison using the same REFLECT+ checkpoint. Both conditions discard the prior reasoning at the same reflection trigger point and use the same continuation budget. Clean-Slate restarts from the original question and the reflection, while Reset-only restarts from the original question alone.

DatasetQuestion + Reflection Mean@8Question-Only Restart Mean@8Gain (pp)
MATH50031.40%27.13%+4.26
GSM8K43.68%37.27%+6.41
AMC2334.69%34.38%+0.31
BBH26.24%24.88%+1.36
DROP34.55%31.45%+3.10

Under the evaluation settings shown in the table, Clean-Slate achieved higher Mean@8 point estimates than Reset-only on all five datasets. The gains were 4.26 and 6.41 percentage points on MATH500 and GSM8K, respectively, and 0.31, 1.36, and 3.10 percentage points on AMC23, BBH, and DROP. Since both conditions discard the prior reasoning, these results show that retaining the reflection can improve continuation accuracy beyond restarting alone.

Weakness 2. The corrective value of reward shaping and the contributions of the advantage baseline and floor

We agree with your analysis. In a group where all rollouts fail and both direct and reflective rollouts are present, failed direct rollouts receive −0.25 and failed reflective rollouts receive 0. A failed reflective rollout therefore receives a raw advantage of +0.25. This design gives the model an incentive to explore reflection. In an early training run where both kinds of failure received 0, the reflection trigger rate fell from 32.2% to 0.8%, and the model increasingly answered directly.

To assess corrective ability, Table 2 holds the original question and the reflection fixed. Average continuation accuracy rose from 34.6% to 38.1% after the full RL stage, indicating that the trained model makes better use of a given reflection. The paired comparison in Weakness 1 also shows that retaining the reflection can yield higher accuracy than restarting from the question alone under the reported settings.

In the revision, we will describe the positive advantage as an exploration incentive and limit our claims about corrective ability to the continuation gains observed in these experiments. The individual contributions of the direct-response baseline and the advantage floor still need to be tested while holding the SFT initialization, restart policy, and reward fixed.

Weakness 3. How does REFLECT+ compare with existing self-correction methods?

Thank you for pointing out that comparisons with Base, SFT, and vanilla GRPO leave open how REFLECT+ performs relative to existing self-correction methods. We added three inference baselines on five datasets: Self-Refine, feedback-only restart, and ReActR-style. All methods use the same questions, four samples per question, and the same per-question cap on total generated tokens, up to 3,072. Self-Refine and feedback-only restart use the same base model and share the initial answer and feedback. They differ in whether the initial answer is retained in the context for revision.

DatasetBase + Self-RefineBase + Feedback-Only RestartBase + ReActR-styleREFLECT+
MATH50026.30028.15032.00064.350
GSM8K41.54746.32345.79280.402
AMC238.75011.87514.37536.250
GPQA22.34822.85421.97026.263
LogiQA27.38527.43325.74726.272
Average25.26627.32727.97746.707

The table reports Mean@4 (%), with an unweighted average across the five datasets. REFLECT+ achieved an average accuracy of 46.707%, 18.730 percentage points above ReActR-style, the highest-scoring inference baseline on average. The gains were larger on the mathematics tasks. On LogiQA, both Self-Refine and feedback-only restart scored slightly higher than REFLECT+, showing that the relative performance varies by task.

ReActR-style is a training-free inference baseline. It uses the base model to review its initial answer, decide whether revision is needed, and answer again when appropriate. The original ReActR training pipeline also uses a teacher model to generate and filter multi-turn correction trajectories, followed by SFT and GRPO. The authors report processing over 100 million tokens for data synthesis alone, so reproducing the training pipeline from scratch requires substantial additional data construction and training. The table compares feedback and retry strategies at inference time. ReActR, Appendix A We plan to reproduce the full ReActR training pipeline, compare it with REFLECT+ under matched training and evaluation conditions, and report the results in the final version of the paper.

Weakness 4. Do the conclusions extend to stronger reasoning models?

We agree that the current results do not establish whether stronger reasoning models still need a Clean-Slate restart. We evaluated the full REFLECT+ policy on Qwen backbones at 0.6B, 1.5B, and 4B. On Qwen3-4B, REFLECT+ increased average Pass@32 across six benchmarks from 88.96% to 92.21% relative to vanilla GRPO. This shows that the improvement in solution coverage also extends to the 4B backbone we tested.

Figure 4 provides complementary evidence about the role of context. With the reflection generated by REFLECT+ held fixed and Qwen2.5-7B-Instruct used for continuation in both conditions, removing the prior reasoning prefix increased average Mean@8 across three mathematics benchmarks from 46.1% to 55.1%. This suggests that the old prefix can still affect subsequent reasoning at these fixed reflection points, even with a larger continuation model. Because the experiment fixes both the trigger point and the reflection content, it tests how the larger model uses a given reflection.

In the revision, we will limit our conclusions to the models and protocols tested. In future work, we will evaluate the full policy on stronger reasoning models and compare three conditions using the same model, trigger point, and remaining budget: retaining the old prefix, restarting from the question alone, and restarting from the question and the reflection.

Question 1. Does reflection help beyond restarting from the question alone with the same model and remaining budget?

Under the settings reported in Weakness 1, retaining the reflection yielded higher Mean@8 point estimates on all five datasets, with gains of 4.26 and 6.41 percentage points on MATH500 and GSM8K, respectively.

Question 2. What do the advantage baseline and floor add over vanilla GRPO, and what evidence shows that training improves correction?

With the original question and reflection held fixed, the full RL stage improved average continuation accuracy from 34.6% to 38.1%, indicating that the trained model makes better use of a given reflection. The positive advantage for failed reflective rollouts encourages exploration. Matched ablations are needed to quantify each component’s contribution to correction.

Question 3. How does REFLECT+ compare with existing feedback-and-retry methods and related training-based methods?

Across the five datasets in the inference comparison, REFLECT+ achieved an unweighted average Mean@4 of 46.707%, exceeding the strongest inference baseline by 18.730 percentage points. REFLECT+ achieved the highest accuracy on four of the five datasets, with the largest gains on the mathematics tasks.

Question 4. Do stronger reasoning models still need a Clean-Slate restart?

The current experiments do not establish whether it is necessary. The 4B results show that the full method can improve solution coverage, and the 7B comparison shows that removing the old prefix can help a larger model use a given reflection. The full policy still needs to be evaluated directly on stronger reasoning models.

Thank you for taking the time to explain what evidence would help you assess our contribution. Your feedback guided the additional comparisons and helped us state more precisely what our results support. If these experiments and clarifications address your concerns, we would be grateful if you would reconsider your rating in light of the new evidence.

Reviewer_p2UbOpenReview ↗
Rating
3
Confidence
2(−1)
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes REFLECT+, a self-correction framework that replaces an erroneous reasoning prefix with a structured reflection before continuing generation. It combines supervised cold-start training with role-aware reinforcement learning to learn when to reflect, what information to retain, and how to resume reasoning. Experiments on three Qwen models and multiple reasoning benchmarks show higher average performance than vanilla GRPO and selective reflection behavior; on Qwen2.5-Math-1.5B, REFLECT+ also produces shorter responses on average.

Critical strengths

REFLECT+ introduces a clear clean-slate self-correction mechanism that discards the faulty reasoning prefix and resumes from the original problem and a structured reflection.

Critical weaknesses
  1. The evaluation lacks direct comparisons with closely related reflection-aware methods. The main results compare REFLECT+ with Base, SFT, and vanilla GRPO, but not with methods such as REA-RL, or ReActR. This makes it difficult to determine whether the proposed framework improves upon existing self-correction approaches.

  2. The paper provides functional evidence that reflection improves downstream continuation, but it does not directly evaluate whether the verified states, error diagnosis, and corrected direction are individually accurate and faithful to the preceding reasoning. Since the reward is primarily based on final-answer correctness and the format gate only enforces structural requirements, an inaccurate reflection could still receive positive credit if the continuation reaches the correct answer. Field-level semantic evaluation or controlled perturbation experiments would better support the claim that REFLECT+ avoids semantically spurious reflection.

Questions for authors
  1. How does REFLECT+ compare with existing reflection-aware methods such as REA-RL or ReActR under the same training and evaluation settings?

  2. Could the authors provide field-level annotations or controlled perturbation experiments to verify that the stated error diagnosis and corrected direction are semantically accurate, rather than merely useful for producing a correct final answer?

  3. The cold-start stage relies on 7,632 reflection instances generated by a commercial API teacher, which may contribute substantially to the model’s reflection and correction abilities. How much of the reported improvement depends on this teacher, and does REFLECT+ remain effective when the cold-start data are generated by an open-source model or by the base model itself?

4 comments

Thank you for recognizing the clean-slate mechanism and for requesting stronger comparisons and semantic validation. We provide the new empirical evidence and its scope below.

Weakness 1. Comparisons with reflection-aware methods

Thank you for pointing out that comparisons with Base, SFT, and vanilla GRPO leave open how REFLECT+ performs relative to existing reflection-aware methods. We added three inference baselines on five datasets: Self-Refine, feedback-only restart, and ReActR-style. All methods use the same questions, four samples per question, and the same per-question cap on total generated tokens, up to 3,072. Self-Refine and feedback-only restart use the same base model and share the initial answer and feedback. They differ in whether the initial answer is retained in the context for revision.

DatasetBase + Self-RefineBase + Feedback-Only RestartBase + ReActR-styleREFLECT+
MATH50026.30028.15032.00064.350
GSM8K41.54746.32345.79280.402
AMC238.75011.87514.37536.250
GPQA22.34822.85421.97026.263
LogiQA27.38527.43325.74726.272
Average25.26627.32727.97746.707

The table reports Mean@4 (%), with an unweighted average across the five datasets. REFLECT+ achieved an average accuracy of 46.707%, 18.730 percentage points above ReActR-style, the highest-scoring inference baseline on average. The gains were larger on the mathematics tasks. On LogiQA, both Self-Refine and feedback-only restart scored slightly higher than REFLECT+, showing that the relative performance varies by task.

ReActR-style is a training-free inference baseline. It uses the base model to review its initial answer, decide whether revision is needed, and answer again when appropriate. The original ReActR training pipeline also uses a teacher model to generate and filter multi-turn correction trajectories, followed by SFT and GRPO. The authors report processing over 100 million tokens for data synthesis alone, so reproducing the training pipeline from scratch requires substantial additional data construction and training. The table compares feedback and retry strategies at inference time. ReActR, Appendix A We plan to reproduce the full ReActR training pipeline, compare it with REFLECT+ under matched training and evaluation conditions, and report the results in the final version of the paper.

Weakness 2. Accuracy and faithfulness of reflection fields

We agree that final-answer correctness alone cannot establish whether a reflection's stated intermediate conclusions and diagnosis are accurate. The format gate checks the structure of the reflection and cannot determine whether it preserves all the necessary information. In our automatic audit, 365 of 3,397 reflections (10.74%) were flagged as omitting some of the prior progress needed for the next reasoning step.

To test how continuation depends on the retained content, we replaced one field at a time with the corresponding field from a different problem in the same dataset, keeping the original problem and the other two fields fixed. The replacement field was 0.5–2 times the length of the original field and was selected without using gold answers or outcome labels. All arms discarded the previous reasoning prefix, used the same REFLECT+ checkpoint, and sampled eight continuations per arm under a shared continuation budget capped at 1,024 tokens.

The table reports Mean@8 with the original reflection minus Mean@8 after field replacement, in percentage points. Positive values indicate that replacement reduced accuracy.

DatasetReplace SReplace DReplace U
GSM8K+10.23−0.49+11.56
MATH500+8.34+1.82+6.92
AMC23+10.26+1.92+8.01

Replacing S or U lowered the Mean@8 point estimate on all three datasets, suggesting that the model relies to some extent on the state summary in S and the direction for further reasoning in U. Replacing D produced smaller changes that varied in direction, suggesting that the model may rely less on the explanation of the earlier error.

Question 1. Matched comparison with REA-RL or ReActR

W1 provided new evaluation-budget-aligned inference baselines. We have expanded the evaluation with Self-Refine, feedback-only restart, a ReActR-style inference adaptation, and a Reflexion implementation.

Question 2. Field annotations or controlled perturbations

W2 provides paired single-field replacements. And we also sampled 3,400 reflections across nine datasets and obtained 3,397 valid API-based annotations. The table reports the field-level results across all 3,397 annotated reflections. We will describe these mixed findings and the partial automatic-audit status alongside the functional results.

Reflection fieldSemantically correctSemantically incorrectUncertain
S: State summary69.53%30.47%0.00%
D: Diagnosis63.41%36.50%0.09%
U: Update direction72.59%27.38%0.03%

Question 3. Dependence on the commercial cold-start teacher

The reflection-source experiment in Figure 3 replaces the inference-time reflection with one generated by open-source Qwen2.5-3B or 7B models while keeping the continuer fixed. It provides evidence that the continuer can use reflections from these sources. The cold-start checkpoint in that experiment was still trained on the original teacher-generated data. We will state the training-data provenance explicitly and keep this limitation separate from inference-time reflection-source transfer.

Thank you for raising these questions. Your feedback prompted us to broaden our comparisons with self-correction methods, directly evaluate the semantic quality and functional role of the reflection fields, and clarify the distinction between inference-time reflection-source transfer and dependence on the cold-start teacher. If these additional experiments and clarifications address some of your concerns, we would be grateful if you would reconsider your rating in light of the new evidence.

Thank you for the detailed response and additional experiments. The new inference baselines, field-level audit, and perturbation results strengthen the paper and clarify the scope of the claims. The training-level comparison with related methods and the dependence on the commercial teacher remain open, but I appreciate that the authors have acknowledged these limitations clearly. I will maintain my current rating.

Thank you for carefully considering our rebuttal and recognizing the value of the additional evidence. Your feedback has helped us strengthen the empirical evidence and define the scope of our claims more precisely. We will incorporate these results into the revision as soon as possible, together with the explicit limitations concerning training-level comparisons and dependence on the commercial teacher.

We hope that the expanded experiments and clearer scope provide a firmer basis in the findings. Thank you again for your time and constructive feedback of our work.

Reviewer_Ksn1OpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper targets spurious reflection, a known failure where RL‑trained LLMs produce empty reflective text without fixing reasoning mistakes. Instead of appending reflection to corrupted reasoning context, the authors propose REFLECT+, a two‑stage framework based on Clean‑Slate Reflection. When detecting errors, the model outputs structured reflection summarizing verified conclusions, error diagnosis and repair directions. A dedicated rollout engine discards the faulty reasoning prefix and restarts generation only from the original question plus reflection.

The pipeline includes supervised cold‑start SFT to learn reflection format, followed by a modified GRPO that optimizes when to reflect, what to reflect on, and how to resume reasoning. New RL components contain format gate, reflection efficiency penalty and difficulty‑gated advantage floor to avoid training collapse. Evaluated on Qwen‑series models across nine reasoning benchmarks, REFLECT+ outperforms vanilla GRPO overall and learns selective reflection triggered mainly on hard inputs, with rich ablations and case studies supporting core claims.

Critical strengths
  1. The studied problem is interesting and solution is somehow reasonable.

  2. The paper compares adaptive reflection triggering against forced/random interruption; evaluates different continuation context configurations; reports reflection trigger rate and hard‑subset recovery metrics across multiple model sizes.

  3. Hyper‑parameters, templates, dataset statistics and evaluation protocols are documented in appendices. Code will be released post‑acceptance. It builds on verl with LoRA and does not require vocabulary modification.

Critical weaknesses
  1. REFLECT+ cannot run under naive vanilla generation. Without external prompt‑rebuilding logic after `` emission, reflection is merely appended to faulty traces and performance degrades. It is unclear how large the performance drop would be for offline approximate simulation via post‑processing outputs. This critical deployment constraint deserves more discussion.

  2. Throwing away the entire prior trace eliminates error contamination but introduces a new failure mode: logical discontinuity if the reflection block fails to capture all valid intermediate conclusions. The current design relies on template constraints and indirect reward signals, with no explicit check for reflection completeness. Potential mitigations could include dependency‑guided partial context pruning (e.g., provenance tracking), pre‑resumption reflection validation, or multi‑choice policy over full‑continue / partial‑prune / full‑restart. These failure modes should be explicitly analyzed.

Questions for authors
  1. What fraction of regex‑valid reflections are still semantically spurious during RL training?

  2. For environments without custom rollout engine: what performance degradation can we expect for offline post‑processing simulation of clean‑slate restart?

1 comment

Thank you for recognizing the motivation, the targeted controls, and the implementation details. We address the two weaknesses and two questions separately below.

Weakness 1. Deployment with a standard generation API

REFLECT+ requires the serving system to rebuild the prompt at the end of a reflection block. If the entire trajectory is generated in one uninterrupted call, both the previous reasoning and the reflection remain in context when the model continues.

A lightweight wrapper around a standard generation API can implement the same restart as our custom rollout engine. The wrapper stops the current generation when the model emits a complete reflection block, rebuilds the prompt using the original question and the reflection, and starts a new request with the remaining token budget. The previous reasoning prefix is thus removed from the context before the continuation is generated.

The same procedure can be used offline. When the model weights, prompt tokens, stopping rules, decoding settings, and remaining token budget are matched, the wrapper and the custom engine implement the same inference policy. We therefore do not expect a systematic drop in accuracy from using a standard generation API. If offline post-processing only deletes the previous reasoning prefix from a completed output, the answer was still generated with that prefix in context. We have not measured the accuracy gap between this edit-only approximation and the actual restart. We will compare the standard API wrapper and the custom engine under matched conditions and report the results in the final version of the paper.

Weakness 2. Information loss after discarding the previous reasoning

We agree that a reflection can omit intermediate conclusions needed for the next step. The format gate checks the structure of the reflection and cannot determine whether it preserves all the necessary information. In our automatic audit, 365 of 3,397 reflections (10.74%) were flagged as omitting some of the prior progress needed for the next reasoning step.

To test how continuation depends on the retained content, we replaced one field at a time with the corresponding field from a different problem in the same dataset, keeping the original problem and the other two fields fixed. The replacement field was 0.5–2 times the length of the original field and was selected without using gold answers or outcome labels. All arms discarded the previous reasoning prefix, used the same REFLECT+ checkpoint, and sampled eight continuations per arm under a shared continuation budget capped at 1,024 tokens.

The table reports Mean@8 with the original reflection minus Mean@8 after field replacement, in percentage points. Positive values indicate that replacement reduced accuracy.

DatasetReplace SReplace DReplace U
GSM8K+10.23−0.49+11.56
MATH500+8.34+1.82+6.92
AMC23+10.26+1.92+8.01

Replacing S or U lowered the Mean@8 point estimate on all three datasets, suggesting that the model relies to some extent on the state summary in S and the direction for further reasoning in U. Replacing D produced smaller changes that varied in direction, suggesting that the model may rely less on the explanation of the earlier error.

Question 1. How many reflections that meet the format requirements still have semantic issues?

We sampled 3,400 reflections across nine datasets and obtained 3,397 valid API-based annotations. The table reports the field-level results across all 3,397 annotated reflections.

Reflection fieldSemantically correctSemantically incorrectUncertain
S: State summary69.53%30.47%0.00%
D: Diagnosis63.41%36.50%0.09%
U: Update direction72.59%27.38%0.03%

Question 2. How much accuracy loss should we expect when reproducing the restart offline?

If an offline script rebuilds the prompt and regenerates the continuation under matched decoding conditions, we do not expect a systematic drop in accuracy. Editing a completed output leaves the original continuation intact. We have not measured the accuracy gap between this approximation and the actual restart.

Thank you for raising these questions. Your feedback prompted us to assess the quality of the reflection fields directly and clarify how the restart can be reproduced in practice. If the additional experiments and clarifications address some of your concerns, we would be grateful if you would reconsider your rating in light of the new evidence.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.