acceptodds
Under review as a conference paper at ICLR 2027

Learning to Refine Partial Solutions with Recursive Discrete Denoising

Abstract

Structured reasoning requires predictions whose parts jointly satisfy constraints or an inferred rule. Reusing a small neural network across computation steps offers a parameter-efficient approach, but each intermediate answer can contain both useful assignments and errors. Learning to refine such answers is challenging when supervision provides completed solutions without intermediate reasoning trajectories. We propose the Discrete Reasoning Diffusion Model (DRDM), which trains a looped Transformer to reconstruct solutions from partially correct candidates. Masking and token replacement create completion and correction tasks while preserving the observed problem. A shared core performs several latent updates before predicting solution tokens; progressive decoding adds assignments, allows revision, and carries latent memory between steps. Across three training seeds, the approximately 7M-parameter model achieves 85.2% exact accuracy on Sudoku-Extreme, 87.3% on Maze-Hard, and 50.5% pass@2 on ARC-AGI-1, exceeding self-attention TRM by 9.7, 2.7, and 7.3 percentage points, respectively.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
Structured reasoning, Discrete diffusion, Recursive reasoning, Iterative refinement

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_ZmGNOpenReview ↗
Rating
4(+2)
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

The paper presents DRDM, a small recurrent model trained to reconstruct solutions from candidates containing correct, masked, and incorrect assignments. A shared Transformer performs latent updates before predicting tokens, while progressive decoding fills the candidate and permits revision. Experiments report improvements over self-attention TRM on Sudoku-Extreme, Maze-Hard, and ARC-AGI-1, with comparable performance on ARC-AGI-2.

Critical strengths

I find this an interesting and substantial piece of work. Constructing refinement problems from completed answers is a natural way to train a recurrent solver without annotated reasoning trajectories. The combination is coherent, and the gains over TRM on Sudoku and ARC-AGI-1 give it practical value. The evaluation covers several forms of structured reasoning and includes extensive analysis of training choices and intermediate predictions.

Critical weaknesses

1. The contribution needs a sharper intellectual boundary

TRM already provides shared recursive computation, GIDD combines masking and token noise, and Loopholing preserves latent information across denoising steps [1–3]. A list of these ingredients does not establish a new reasoning principle. For me, the interesting contribution is the choice of which imperfect answers a recurrent solver should learn to refine. The paper should make that insight carry the argument and explain what changes when the training problem is defined this way. The current discussion identifies the predecessors, but does not sharply enough distinguish the conceptual advance from the assembled architecture.

2. Random corruption is an imperfect model of reasoning errors

The method learns from independently masked or replaced entries in a correct solution. Its own mistakes at inference can be correlated, mutually reinforcing, and confidently wrong. These are different refinement problems, and the paper moves too quickly from learning to undo synthetic corruption to learning to repair intermediate reasoning. The final accuracies show that the recipe works; they do not fully explain why this training distribution is appropriate for the errors the solver actually makes. I would like a stronger argument for that modeling choice and a clearer discussion of where it may fail.

3. Consistency does much of the work, but its role is underexplained

Removing consistency regularization drops Sudoku accuracy from 85.2% to 71.4% and Maze accuracy from 87.3% to 78.6%. This is a central part of the method, yet its justification largely stops at encouraging agreement. Candidates at different corruption levels contain different amounts of useful information, so agreement in the target does not by itself explain why their predictive distributions should agree. This deserves a more careful account, particularly because prediction confidence also controls which assignments are revealed. The paper establishes that consistency helps, but leaves an important part of its own success unexplained.

Questions for authors

Given the weaknesses above, I tentatively assign a score of 2. I look forward to the authors’ responses and would be happy to revise my score if these concerns are resolved.

  1. What does training on imperfect candidate answers teach the recurrent solver beyond standard supervision on final answers? I see this as the paper’s most interesting idea, but its contribution needs to be clearer.

  2. Why should learning to repair randomly corrupted answers help the model correct its own reasoning mistakes, which may be structured and confidently wrong?

  3. Consistency regularization contributes substantially to performance. Why should predictions agree across corruption levels when the candidates contain different amounts of useful information?

5 comments
AuthorsOpenReview ↗

What candidate training adds (1/3)

Thank you for identifying the central training question and for your positive assessment of our formulation and its practical value. Your questions about structured errors and consistency prompted us to examine what the solver learns from imperfect candidates and how it uses its predictions during decoding.

Q1. What does training on imperfect candidates add?

A complete solution provides the target, while the candidate determines the update the recurrent solver must learn. In Sudoku, a blank cell requires completion; a plausible but incorrect digit requires correction using the clues and surrounding assignments. Retained correct entries provide context. Each completed solution thus supplies a family of refinement tasks. Our design builds on shared recurrence in TRM, mixed masking and replacement in GIDD, and latent information carried across denoising steps in work such as Loopholing. Our contribution is to specify and evaluate the candidate states on which the recurrent update is trained.

We compared these training choices using the same architecture, data, latent updates, and time conditioning, with 50,000 optimizer updates per seed and λ = 0. We retrained all configurations with cross-entropy averaged over all editable positions, giving them a common supervision rule. This controlled comparison uses an averaged loss, whereas the submitted objective sums over perturbed positions. With the same progressive decoder, all-mask, partial-mask, and partial-mask-plus-replacement training achieve 60.00%, 69.00%, and 72.00% on Sudoku, and 70.00%, 77.00%, and 79.00% on Maze. The corresponding ARC-AGI-1 results are 42.50%, 47.75%, and 48.83%. For ARC, a test input receives a score of 1 if either candidate matches the full output grid. Scores are averaged within each task and then equally across tasks. All values in this candidate-training comparison are means over three seeds. Both retaining useful assignments and including incorrect ones improve performance in this comparison. At comparable inference cost, evaluating the all-mask model with all-mask inputs yields 68.00% on Sudoku, 75.00% on Maze, and 45.00% on ARC-AGI-1. Mixed-candidate training remains ahead on these tasks.

AuthorsOpenReview ↗

Transfer to the solver's own errors (2/3)

Q2. Why should random-corruption training help correct the solver's own errors?

With a known recovery target, random corruption lets us control how much useful information a candidate retains. Correcting an independently placed wrong Sudoku digit still requires reasoning about its row, column, and box. To test transfer to the solver's own errors, we collected candidates after step 8 of 16. For each input, we constructed a random candidate with exactly the same mask positions and number of incorrect assigned tokens. Before the next update, we initialized the latent state for both candidates using the same input-dependent rule. The diagnostic uses seed 42 and includes 1,000 Sudoku puzzles, 1,000 mazes, and all test inputs from the 400 ARC-AGI-1 and 120 ARC-AGI-2 tasks. We score ARC token changes offline by aligning predictions to the reference grid. The denoiser does not receive correctness labels. The table reports the fraction of incorrect assigned tokens corrected in one update.

DatasetSolver-generated candidateMatched random candidate
Sudoku55/302 = 18.212%88/302 = 29.139%
Maze370/1,682 = 21.998%572/1,682 = 34.007%
ARC-AGI-1256/1,970 = 12.995%433/1,970 = 21.980%
ARC-AGI-2210/2,103 = 9.986%358/2,103 = 17.023%

In this paired comparison, DRDM repairs its own errors in all four settings. Those errors are harder to repair than matched random corruptions and more often carry high confidence. We use a threshold of 0.90 for the probability assigned to the current wrong token. The high-confidence fractions are 37.01% versus 18.02% on ARC-AGI-1 and 40.99% versus 22.02% on ARC-AGI-2. These confidence figures describe the errors present in the candidates; the repair rates in the table are measured over all incorrect assigned tokens.

We also measured revisions along the model's native inference trajectories, with the latent history retained. In the 1,000-puzzle Sudoku diagnostic, 680 of 3,100 erroneous assigned-token exposures were corrected, while 260 of 452,940 correct assigned-token exposures became incorrect. These counts accumulate over steps. With revision enabled, three-seed Sudoku accuracy is 85.20%, compared with 79.50% when assigned tokens are frozen. These results show that DRDM corrects self-generated errors during inference and that enabling revision improves final solution accuracy.

AuthorsOpenReview ↗

Why consistency helps (3/3)

Q3. Why encourage consistency across corruption levels?

Adjacent views share the task and clean target, but differences in their information content can justify different confidence levels. The regularizer favors predictions that remain consistent as artificial corruption changes. Equation 5 balances this preference against reconstruction through a finite KL coefficient. The KL term acts on the full clean-token distribution, including confidence. The default λ = 0.1 gives the highest means in the new sweep over 0, 0.01, 0.1, and 0.5. This supports the chosen balance between consistency and reconstruction within the tested range. All configurations use the same optimizer-update count; consistency training requires additional computation.

We compared the models on the same corrupted candidates at t = 4/16 and t = 12/16. At t = 12/16, consistency lowers Sudoku token NLL from 0.460 to 0.390 and 15-bin ECE from 5.70% to 4.10%. The accuracy of newly revealed tokens rises from 92.50% to 94.60%. These diagnostics use seed 42 and 1,000 puzzles. We then compared ordering policies with the reveal counts held fixed.

Training and reveal policySudokuMaze
λ = 0, random order65.20 ± 2.0074.00 ± 1.80
λ = 0, confidence order71.40 ± 2.1078.60 ± 1.90
λ = 0.1, random order76.10 ± 2.0081.20 ± 1.50
λ = 0.1, confidence order85.20 ± 1.7087.30 ± 1.50

For each training seed, we repeated random ordering three times and averaged the results. We then computed the mean and SD across three training seeds. Consistency improves mean accuracy with both ordering policies, and confidence ordering gives a larger gain in the model trained with consistency. Together, the paired-candidate diagnostics and ordering comparison support improvements in token prediction and in the use of confidence to select reveals.

Thank you again for your constructive questions about candidate training and consistency, and for the opportunity to discuss this work. Please let us know if anything remains unclear or if you have further questions; we would be glad to continue the discussion. We sincerely hope you will reconsider our work in light of these explanations and the new evidence. Your thoughtful review and engagement are a great encouragement to us, and we are deeply grateful for your time and consideration.

Reviewer_J3CJOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper develops the Discrete Reasoning Diffusion Model (DRDM), a recursive denoiser trained to complete and correct these candidates. By using reconstruction problems to train a compact shared-core denoiser, DRDM improves accuracy on Sudoku-Extreme and Maze-Hard and pass@2on ARC-AGI-1 over self-attention TRM.

Critical strengths
  1. This paper is well organized and well described.
  2. Experimental results on three reasoning tasks show better accuracy than several baselines.
  3. The results of different training configurations are reported.
Critical weaknesses
  1. What is the exact meaning of partial solutions in this paper? Is it partially correct candidates for any task?
  2. In Figure 2, the candidates to refine include useful, missing, and wrong. How can the candidate type be predicted automatically?
  3. In Equation 5, what is the effect of lambda?
  4. The accuracy of the Transformer is 0 for all tasks. Is it correct? If yes, why is it so poor?
Questions for authors

See weaknesses.

3 comments
AuthorsOpenReview ↗

Partial solutions, consistency weight, and Transformer diagnostics

Thank you for your constructive questions and positive assessment of the paper's organization, results, and training comparisons. Your questions led us to clarify the candidate states and examine λ and the direct Transformer baseline in more detail.

Q1. What does "partial solution" mean?

A partial solution is the solver's current candidate answer, which can contain correct assignments, masks, and incorrect assignments simultaneously. In Sudoku, clues remain fixed; editable cells may be blank or contain a proposed digit. This definition also covers Maze path tokens and ARC output colors.

Q2. How are the candidate types in Figure 2 identified?

We construct training candidates from complete solutions, so the retained, masked, and replaced entries are known. The colors in Figure 2 mark these entries. At inference, the solver receives the puzzle and current candidate. Masks indicate blanks; filled cells contain proposed values. The correct/incorrect labels in the figure are not available to the solver. It predicts clean-token distributions directly to guide reveal and revision, without first assigning entries to the three categories. We will clarify in the caption that the labels are illustrative.

Q3. What is the effect of λ?

In Equation 5, λ weights the KL penalty between clean-token predictions at adjacent corruption levels. We set λ = 0.1 and use the time weight exp(−t/T), with β_c = 1.0. We evaluated weights on either side of this default, with 50,000 optimizer updates for each seed. Results are reported as mean ± SD across three seeds, in percent.

λSudokuMazeARC-AGI-1ARC-AGI-2
071.4 ± 2.178.6 ± 1.949.0 ± 2.45.7 ± 0.8
0.0180.00 ± 1.8084.50 ± 1.6049.75 ± 2.756.67 ± 0.83
0.185.2 ± 1.787.3 ± 1.550.5 ± 2.67.4 ± 0.9
0.583.00 ± 1.8086.00 ± 1.5050.00 ± 2.506.94 ± 0.48

For ARC, each test input scores 1 if either candidate matches the complete output grid. We average scores within each task, then equally across tasks. Among the tested settings, the default weight gives the highest mean on each benchmark. Increasing λ to 0.5 lowers the means, consistent with a trade-off between agreement and reconstruction. We used the same number of optimizer updates throughout the sweep; training with consistency required more computation than training with λ = 0.

Q4. Is the Transformer's zero accuracy correct, and why?

We examined the evaluation scores and token predictions for seed 42. This run produced zero exact solutions on Sudoku and Maze and zero cumulative task scores on both ARC datasets.

DatasetExact-solution count or cumulative ARC task scoreToken accuracyOutput-size error rate, first candidate
Sudoku0/422,786approximately 74%0.00%
Maze0/1,00093.00%0.00%
ARC-AGI-10/40076.00%130/419 = 31.03%
ARC-AGI-20/12071.00%65/167 = 38.92%

We compute token accuracy over editable positions in the reference grid, counting missing positions as errors. Sudoku and Maze require the entire solution to be correct. In ARC, a test input contributes to its task score only if the whole output grid is correct. Thus, a model can predict many individual tokens correctly and still receive zero scores. The ARC diagnostics also show frequent output-size errors in the first candidate.

On a small Sudoku training subset, the baseline solves 32/32 puzzles exactly. We checked labels, output sizes, and padding on 128 inputs sampled across the four datasets; all checks passed. Evaluation results on these inputs also matched those from an independent implementation. The model can therefore fit the small training set and correctly predict many individual tokens. These diagnostics use seed 42; submitted Table 1 reports rounded summaries over three runs.

Thank you again for your careful review and encouraging assessment. If you find these clarifications and new results convincing, we would be grateful if you could reconsider the paper in light of them. Your recognition would be a great encouragement as we pursue careful, substantive research amid rapid advances in AI. Your support would also help us share this work with the community, where we hope others will use and build on it. We would be honored to contribute in this way and sincerely appreciate your time and consideration.

AuthorsOpenReview ↗

Revised manuscript and a guide to your questions

Thank you for taking the time to review our work. We recognize that reviewing is a voluntary service to the research community and adds to your already demanding workload. We deeply appreciate the generosity and dedication behind this commitment.

Given the space constraints of the rebuttal format, we have tried to keep our responses focused rather than include every detail. We have made every effort to address your questions clearly, accurately, and concisely.

A more detailed account of our work may also be helpful in informing your evaluation or reconsideration. Your questions have been valuable in helping us improve the paper. We have therefore revised the PDF, including both the main text and the appendix, to provide more detailed explanations and additional results. We hope these additions address your questions more fully and provide the level of detail needed for your assessment.

We have also corrected formatting issues and typos. We are pleased to share this revised version with you and sincerely appreciate the time and care you have devoted to reviewing our work.

For convenience, the passages relevant to your questions are listed below, using the same question numbers as in our earlier responses. Page and line numbers refer to the revised PDF, and all changes are marked in red:

  • Q1. Definition of a partial solution: Sec. 3, p. 4, lines 203–210; the task representations are described in detail in Appendix A.1, p. 15, lines 760–781.
  • Q2. Meaning of the candidate labels in Figure 2: Figure 2 caption, p. 3, lines 126–130, and Sec. 3, p. 4, lines 206–208. These passages clarify that correctness labels are not provided at inference time.
  • Q3. Role and choice of λ: Sec. 3.3, p. 5, lines 262–269, Eq. (5), and Sec. 4.3, p. 8, lines 378–386; Appendix F.4, pp. 27–28, lines 1423–1499, Tables 21–23.
  • Q4. Verification and interpretation of the Transformer's zero exact-solution scores: Sec. 4.2, p. 7, lines 339–342; Appendix F.7, p. 30, lines 1577–1601, Table 27.

We welcome any further thoughts or questions you may have. If our clarifications and additional results have addressed your concerns, we would be grateful if you would consider revising your evaluation accordingly.

Thank you again for your time and thoughtful feedback.

Thanks for the reply. I will keep the original rating.

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

This paper proposes the DRDM, which trains a recurrent Transformer to refine partial solutions through discrete denoising. By masking or replacing tokens in completed solutions, the method constructs completion and correction tasks without requiring intermediate reasoning trajectories. DRDM further incorporates a consistency regularization objective, recurrent latent-state updates, and progressive decoding with token revision. Experiments on Sudoku-Extreme, Maze-Hard, and ARC-AGI benchmarks show improvements over the self-attention TRM baseline on several tasks. However, the current experiments do not fully isolate the contributions of the proposed components or establish whether the gains persist under matched inference-compute budgets.

Critical strengths
  • The paper uses corrupted solution candidates to train models to complete missing assignments and correct erroneous ones, without requiring intermediate-step annotations. This is a natural and potentially useful approach to learning iterative refinement.

  • The proposed method achieves substantial improvements over the self-attention TRM baseline on Sudoku-Extreme and ARC-AGI-1, and smaller improvements on Maze-Hard.

Critical weaknesses
  • The individual contributions of the proposed components are not sufficiently isolated. DRDM combines denoising-based supervision, consistency regularization, recurrent latent-state updates, and progressive decoding with token revision. Although the paper includes several ablations, some comparisons change multiple design choices simultaneously. For example, the non-progressive decoding variant modifies the overall decoding procedure, making it difficult to attribute the performance drop to a specific mechanism.

  • The experimental results suggest that the benefits may depend on the task rather than reflect a consistent improvement in structured reasoning. The paper would benefit from a deeper analysis of these differences and additional evidence supporting its broader claims about general-purpose reasoning.

Questions for authors
  • Can the authors provide controlled experiments that isolate the contributions of denoising supervision, consistency regularization, latent-state carryover, and token revision, particularly under the same architecture and training budget?

  • Why does DRDM fail to improve over TRM on ARC-AGI-2 despite its gains on other benchmarks? Are the reported improvements statistically significant at the task level, and what characteristics of the evaluated tasks explain the variation in performance?

4 comments
AuthorsOpenReview ↗

Candidate training and component controls (1/2)

Thank you for your careful review and for recognizing the value of learning completion and correction from completed solutions. Your comments on component attribution and inference cost helped us refine the experimental controls. We have completed these comparisons and a paired analysis of ARC task scores.

Q1. What does candidate training contribute?

We trained three versions of the same backbone with the same latent updates, time conditioning, and data. We trained each version for 50,000 optimizer updates per seed. All three versions used λ = 0 and the same progressive decoder, and had the same reported training FLOPs. We retrained each configuration with cross-entropy averaged over all editable positions in each example, keeping the supervision scope and normalization the same across conditions. The submitted paper sums the loss over perturbed positions. Results below are means ± SD over three seeds, expressed as percentages. For ARC, a test input scores 1 if either candidate matches the complete reference output, and 0 otherwise. We average the input scores within each task, then take the equally weighted mean across tasks.

Training candidatesSudokuMazeARC-AGI-1ARC-AGI-2
All mask60.00 ± 2.5070.00 ± 2.4042.50 ± 2.503.33 ± 0.83
Partial mask69.00 ± 2.3077.00 ± 2.0047.75 ± 2.505.00 ± 0.83
Partial mask + replacement72.00 ± 2.1079.00 ± 1.9048.83 ± 2.385.83 ± 0.83

Retaining correct assignments improves performance on each benchmark, and adding incorrect assignments improves it further. We also evaluated the all-mask model at a comparable computation budget, using all-mask inputs throughout inference. Its scores are 68.00%, 75.00%, 45.00%, and 4.17%, respectively. Mixed-candidate training remains ahead even when the all-mask model is evaluated with inputs that match its training.

Q2. What are the roles of latent carryover, token revision, and consistency?

We disabled carryover and revision separately while holding the checkpoints, inference horizon, core-call count, and reveal budget fixed. The table also reports results from separate models trained without carryover or consistency. ARC follows the task-averaged pass@2 protocol defined above. Results are means ± SD over three seeds, in percent.

ConfigurationSudokuMazeARC-AGI-1ARC-AGI-2
Full DRDM85.2 ± 1.787.3 ± 1.550.5 ± 2.67.4 ± 0.9
Reset latent state at each step78.00 ± 2.1082.00 ± 1.8044.00 ± 2.505.00 ± 0.83
Freeze assigned tokens79.50 ± 1.9084.10 ± 1.6046.25 ± 2.505.83 ± 0.83
Train and evaluate without carryover81.00 ± 1.8085.00 ± 1.3047.00 ± 2.006.67 ± 0.83
Without consistency71.4 ± 2.178.6 ± 1.949.0 ± 2.45.7 ± 0.8

Resetting the latent state measures how much the trained model depends on carryover. Training a separate model without carryover accounts for the mismatch introduced by resetting states only at inference. Freezing assigned tokens preserves progressive reveal and directly tests revision within the existing decoder.

For the consistency comparison, we kept the architecture, the paper's summed reconstruction loss, and 50,000 optimizer updates per seed. The default weight is λ = 0.1. Removing consistency yields the last row above, which corresponds to the ablation in submitted Table 2. Both models receive the same number of updates; consistency adds computation, as shown in the reported training costs.

AuthorsOpenReview ↗

Inference computation and ARC task-level evidence (2/2)

Q3. Do the gains persist at comparable inference cost?

We compared DRDM at R = 16 with TRM at R = 4. These settings use 80 and 84 core calls per candidate, respectively. Including both ARC candidates brings the cost to 160 and 168 calls per test input. Under the reported FLOP accounting, DRDM uses approximately 4.8% less computation at these settings. Results are means ± SD over three seeds, in percent.

DatasetDRDM, R = 16TRM, R = 4Difference, percentage points
Sudoku85.20 ± 1.7068.64 ± 1.70+16.56
Maze87.30 ± 1.5084.57 ± 0.95+2.73
ARC-AGI-150.5 ± 2.636.33 ± 2.38+14.17
ARC-AGI-27.4 ± 0.95.56 ± 1.27+1.81

We calculated ARC differences from the unrounded seed scores. DRDM has higher mean accuracy at these closely matched core-call budgets.

Q4. Are the ARC gains statistically significant, and how do they vary across tasks?

This analysis compares the models at their original R = 16 horizons, separately from the matched core-call comparison above. For each fixed checkpoint pair, we paired the scores by task for both analyses. We resampled paired task scores 200,000 times to obtain bootstrap 95% confidence intervals for the mean difference. We also applied a two-sided exact paired permutation test to that difference.

DatasetSeedDRDM − TRM, percentage points95% CI, percentage pointsPermutation p
ARC-AGI-142+7.00[+3.00, +11.00]0.0008
ARC-AGI-12026+7.00[+3.00, +11.00]0.0013
ARC-AGI-1100+8.00[+3.75, +12.25]0.0004
ARC-AGI-242+0.83[−1.67, +4.17]1.0000
ARC-AGI-22026−0.14[−3.75, +3.33]1.0000
ARC-AGI-2100−0.69[−4.86, +3.33]0.8438

On ARC-AGI-1, DRDM gains 7.00, 7.00, and 8.00 percentage points over TRM across the three checkpoint pairs. All three 95% confidence intervals are above zero. In the seed-42 exploratory analysis, both output-size groups show gains, with a larger gain for the smaller-output group. The gain is 9.83 points for the 173 tasks whose maximum reference-output area is at most 100 cells, and 4.85 points for the other 227 tasks.

Q5. How should the ARC-AGI-2 results be interpreted?

On ARC-AGI-2, both methods report a mean of 7.4% at their original R = 16 settings. The two models make different numbers of core calls at the same horizon. Under closely matched core-call budgets, DRDM at R = 16 reaches 7.4% and TRM at R = 4 reaches 5.56%. The difference is 1.81 percentage points, calculated from unrounded seed scores. Including both candidates, these settings use 160 and 168 core calls per test input, respectively. Across all four evaluated benchmarks, the controlled comparisons show higher mean accuracy for DRDM at comparable inference cost. These results support the value of learned solution refinement.

Thank you again for your constructive questions and comments, and for discussing the experimental controls and evaluation with us. Please let us know if any part of our response is unclear; we welcome further questions and discussion of any aspects you find interesting. We sincerely hope you will reconsider our work in light of the clarifications and new results. Your careful review and willingness to engage are a great encouragement to us. We deeply appreciate your time and consideration.

AuthorsOpenReview ↗

Revised manuscript and a guide to your questions

Thank you for taking the time to review our work. We recognize that reviewing is a voluntary service to the research community and adds to your already demanding workload. We sincerely appreciate the time, effort, and care you have devoted to evaluating our work.

Given the space constraints of the rebuttal format, we have tried to keep our responses focused rather than give every detail equal space. We have made every effort to address your questions clearly, accurately, and concisely.

We also recognize that a more detailed account of our work may be helpful in informing your evaluation or reconsideration. Your questions have been valuable in helping us improve the paper. We have therefore revised the PDF, including both the main text and the appendix, to provide more detailed explanations and additional results. We hope these additions address your questions more fully and provide the details needed for your assessment.

We have also corrected formatting issues and typos throughout the paper. We are pleased to share this revised version and sincerely appreciate the time and care you have devoted to reviewing our work.

For convenience, the passages relevant to your questions are listed below, using the same question numbers as in our earlier responses. Page and line numbers refer to the revised PDF, and all changes are marked in red:

  • Q1. Contribution of candidate training: Sec. 4.3, p. 7, lines 351–368, Table 2; Appendix F.1, p. 25, lines 1308–1330, Table 16.
  • Q2. Roles of carryover, token revision, and consistency: Sec. 4.3, p. 7, lines 370–377, and p. 8, lines 378–386; Appendix F.2, pp. 25–26, lines 1335–1370, Tables 17–18, and Appendix F.4, p. 27, lines 1423–1437, Table 21.
  • Q3. Comparison at matched inference budgets: Sec. 4.4, p. 8, lines 395–407, Table 3; Appendix F.3, pp. 26–27, lines 1374–1419, Tables 19–20.
  • Q4. ARC task-level statistics and differences across tasks: Sec. 4.4, p. 8, lines 409–415; Appendix F.6, pp. 29–30, lines 1534–1574, Tables 25–26.
  • Q5. Interpretation of the ARC-AGI-2 results: Sec. 4.4, p. 8, lines 395–415, Table 3; Appendix F.6, p. 29, lines 1556–1559, with the full budget comparison in Appendix F.3, p. 26, Table 19.

We welcome any further thoughts or questions you may have. If our clarifications and additional results have addressed your concerns, we would be grateful if you would consider revising your evaluation accordingly.

Thank you again for your time and thoughtful feedback.

AuthorsOpenReview ↗

Follow-up on our experimental updates

Dear Reviewer ToXZ,

Thank you again for your insightful review. We wanted to kindly check in to see if our new controlled ablations, compute-matched evaluations, and task-level statistical analyses (detailed in our previous comments and highlighted in red in the revised PDF) have adequately addressed your concerns regarding component attribution and benchmark variation.

Please let us know if you have any further questions before the discussion window closes—we would be happy to discuss them!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.