- 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.
-
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.
-
Why should learning to repair randomly corrupted answers help the model correct its own reasoning mistakes, which may be structured and confidently wrong?
-
Consistency regularization contributes substantially to performance. Why should predictions agree across corruption levels when the candidates contain different amounts of useful information?
5 comments
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.
| Dataset | Solver-generated candidate | Matched random candidate |
|---|---|---|
| Sudoku | 55/302 = 18.212% | 88/302 = 29.139% |
| Maze | 370/1,682 = 21.998% | 572/1,682 = 34.007% |
| ARC-AGI-1 | 256/1,970 = 12.995% | 433/1,970 = 21.980% |
| ARC-AGI-2 | 210/2,103 = 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.
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 policy | Sudoku | Maze |
|---|---|---|
| λ = 0, random order | 65.20 ± 2.00 | 74.00 ± 1.80 |
| λ = 0, confidence order | 71.40 ± 2.10 | 78.60 ± 1.90 |
| λ = 0.1, random order | 76.10 ± 2.00 | 81.20 ± 1.50 |
| λ = 0.1, confidence order | 85.20 ± 1.70 | 87.30 ± 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.
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.