acceptodds
Under review as a conference paper at ICLR 2027

Fine-Tuning Diffusion Language Models with Context Selection and Target Weighting

Abstract

Supervised fine-tuning of discrete diffusion language models masks some response tokens and trains the model to recover their original values from the visible context. The masking pattern therefore determines both the context available to the model and the tokens it learns to predict. Uniform random masking does not explicitly account for the interaction between these choices. We introduce GoldiMask, which selects tokens to reveal as context by approximately maximizing a submodular objective. This objective uses model signals to balance the benefit of revealing tokens against their value as prediction targets. GoldiMask then weights the remaining targets according to how they benefit from the selected context and their remaining learning potential. Across three backbones and three training datasets, GoldiMask achieves the highest average accuracy in most evaluated settings, demonstrating gains on both reasoning and code generation. Component ablations show that both context selection and target weighting contribute to the gains. GoldiMask also reduces decoding iterations on GSM8K and MATH-500 under confidence-threshold parallel decoding, while maintaining comparable accuracy at higher confidence thresholds.

Primary area:
unsupervised, self-supervised, semi-supervised, and supervised representation learning
Keywords:
diffusion language models, masked diffusion models, discrete diffusion, supervised fine-tuning, submodular optimization, non-monotone submodular maximization, context selection, loss weighting, learnability

Reviews

Among:
Shape:
1234this paper, Top 4%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 3.33 over 3 reviews, above 96% of the others.
Reviewer_bkkVOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes GoldiMask, a supervised fine-tuning method for Masked Diffusion Language Models (MDLM), which jointly optimizes context selection and target weighting. The method first over-masks the response, then uses model-predicted probabilities and attention to select tokens to be revealed through submodular optimization. Subsequently, weights are assigned to the remaining training targets based on the changes in predicted probabilities before and after revealing and the token-level supervision priority. The authors evaluate GoldiMask on LLaDA and Dream series models using s1K, LIFT-SFT-12K, and KodCode, and compare it with vanilla SFT and various adaptive SFT methods. Experimental results show that GoldiMask achieves good average accuracy across multiple settings and reduces the number of decoding iterations in some parallel decoding settings. The paper also provides component ablation and computational cost analysis.

Critical strengths
  1. GoldiMask jointly considers context selection and target weighting, proposing a reasonable optimization method to address the problem in diffusion SFT where the mask simultaneously determines both the available context and the supervised target. Compared to existing adaptive masking methods, this joint design is innovative.

  2. Experiments cover three backbones and multiple training datasets, and baselines are retrained under uniform settings. Stable average gain across models, component ablation, and multi-seed results provide substantial support for the effectiveness of the method.

Critical weaknesses
  1. GoldiMask's core design lacks sufficient justification. Appendix H.5 shows that the predictive power of attention-based contextual support is weak (R^2 = 0.041 – 0.098), and it's unclear whether it truly captures token dependencies. Furthermore, in Table 4, random reveal + water-filling achieves 39.2, while the complete method achieves 41.1; other priority functions and weighting rules also perform similarly. Therefore, the necessity of complex submodular optimization and specific weighting design remains unclear.
  2. The evaluation protocol has ambiguities. Tables 1–2 do not specify whether the generation length is selected through the validation set, and the performance advantage is significantly reduced in some settings after fixing the length (Table 9). Additionally, some of the average accuracies of HumanEval and MBPP in Table 3 cannot be directly obtained from the standard test set sizes, a single greedy decoding, and three seeds; further explanation of the evaluation and statistical methods is needed.
Questions for authors
  1. How were the generation lengths chosen in Tables 1–2? If the length is chosen using a validation set, does the main performance advantage still hold?
  2. Could you elaborate on the evaluation protocols for HumanEval and MBPP, including the number of test items, decoding strategy, number of samples per item, and random seeds, to explain the average accuracy in Table 3?
  3. Appendix H shows that the prediction ability of the attention-based support proxy is relatively weak. Could you compare it with a simpler strategy that selects reveal tokens solely based on token probability, within the same two-pass pipeline?
  4. All component ablations are conducted on LLaDA-8B-Instruct with s1K, where random reveals with water-filling (39.2) already approach the strongest baseline LIFT (39.5). Could you report the same ablation on LLaDA-1.5 and Dream-7B to verify whether submodular reveal selection still clearly outperforms random reveals?
3 comments

We thank the reviewer for the thoughtful evaluation of our method and experimental protocol. Below, we address the questions and describe the supporting analyses and revisions.

Q1. How were generation lengths chosen?

Following LIFT and d1, we evaluate each method at generation lengths of 128, 256, and 512 tokens and report the highest accuracy separately for each benchmark. Checkpoints are selected using validation loss; generation lengths are selected using benchmark accuracy, rather than the validation split. The same procedure is applied to all methods. We now clarify this in Appendix F in the updated manuscript.

Q2. HumanEval and MBPP evaluation protocol

We thank the reviewer for pointing this out. The three-seed protocol described for Tables 1-2 applied to the reasoning experiments, whereas the previous code-generation results in Table 3 averaged two training seeds. We have now added a third seed for every method and updated Table 3 and its caption to report means and standard deviations across three seeds.

We evaluate all 164 HumanEval problems and the 257 problems in the sanitized MBPP test split, using greedy diffusion decoding with generation length 256 and one sample per problem. A problem is counted as solved if the generated code passes all its unit tests.

The leading methods remain unchanged.

Q3+Q4. Additional ablation on LLaDA-1.5 and Dream-7B, with an additional probability-based selection method

We thank the reviewer for both suggestions. We address Q3 and Q4 together through additional component ablations on LLaDA-1.5 and Dream-7B, including the suggested probability-only selector. We ran the full component ablation on LLaDA-1.5 and Dream-7B (s1K, three seeds) and added the probability-only selector suggested by the reviewer. It uses the same two-pass pipeline and water-filling weights as GoldiMask-D, but selects reveal tokens by token probability alone, without the attention-based support. Entries are mean ± standard deviation over three seeds.

ModelReveal selection + weightingGSM8KMATHCD.SudokuAvg.
LLaDA-1.5Vanilla79.2 ± 0.532.6 ± 1.622.0 ± 1.014.4 ± 1.637.1 ± 0.9
LLaDA-1.5Random + WF81.2 ± 0.337.7 ± 1.026.2 ± 2.616.6 ± 1.140.4 ± 0.7
LLaDA-1.5Probability-only + WF81.3 ± 0.537.5 ± 1.426.8 ± 1.819.0 ± 1.041.2 ± 0.7
LLaDA-1.5GoldiMask-D + uniform81.0 ± 0.237.2 ± 1.324.7 ± 1.415.7 ± 0.839.7 ± 0.6
LLaDA-1.5GoldiMask-D (D + WF)81.9 ± 0.837.6 ± 0.226.6 ± 1.621.6 ± 1.841.9 ± 0.8
Dream-7bVanilla76.1 ± 0.430.6 ± 0.325.0 ± 0.914.8 ± 2.136.6 ± 0.5
Dream-7BRandom + WF79.0 ± 0.541.6 ± 0.936.3 ± 4.319.4 ± 1.244.1 ± 1.1
Dream-7BProbability-only + WF79.2 ± 0.842.5 ± 1.834.9 ± 3.121.0 ± 1.944.4 ± 0.4
Dream-7BGoldiMask-D + uniform78.3 ± 0.141.5 ± 0.836.5 ± 5.016.3 ± 0.443.1 ± 1.2
Dream-7BGoldiMask-D (D + WF)78.8 ± 0.642.8 ± 0.437.0 ± 4.022.8 ± 2.145.3 ± 1.6

These results support the contribution of both context selection and target weighting. Random reveals with water-filling already improve over Vanilla SFT, while submodular selection provides further gains of 1.5 points on LLaDA-1.5 and 1.2 points on Dream-7B. GoldiMask-D also improves over probability-only selection by 0.7 and 0.9 points, respectively. We have included these ablations to Appendix I.3 of the revised manuscript. We also added Studies 4-5 in Appendix H.6-H.7. The selected reveal set outperforms all 16 random alternatives in 80.5% of tested states with t≤0.75t \leq 0.75, and mean prediction improvement increases across estimated-support bins. These analyses complement the downstream ablations by directly examining the benefit of revealed context.

Thanks for the detailed response and the additional experiments. The clarifications on generation length and the code evaluation, together with the new ablations, address my questions. I will keep my positive score.

We thank the reviewer for their comments, which have helped us improve our paper. We are open to answer any other questions in the reminder of the discussion period.

Reviewer_EXY3OpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes GoldiMask, a masking strategy for supervised fine-tuning of masked diffusion language models. The main idea is that masking determines both which tokens provide context and which tokens are used for supervision, a trade-off that uniform random masking does not explicitly consider. GoldiMask uses model confidence and attention to select informative tokens to reveal, then adjusts the weights of the remaining prediction targets based on the selected context. Experiments across several backbones and datasets show promising improvements on reasoning and code tasks, along with more efficient parallel decoding in some settings.

Critical strengths
  • The motivation is clear and intuitive. I like that the paper considers the dual role of masking in providing context and determining supervision targets. This provides a natural motivation for moving beyond uniform random masking.

  • The evaluation covers a good range of settings. The authors test multiple backbones and training datasets, including reasoning and code tasks. Retraining baselines under the same setup and reporting results across multiple seeds are also appreciated.

  • The ablations help explain the contributions of the main components. Both the token selection strategy and the target weighting scheme appear to contribute to the improvements, making the overall design more convincing.

Critical weaknesses
  • The choice of attention features could be better motivated. The method relies on attention from the final layer to estimate contextual support, but it is not clear why this is the most appropriate choice. Since this signal directly influences which tokens are revealed, I would be interested in seeing whether the results are sensitive to the choice of layers or attention heads.

  • It is unclear how well attention reflects the actual benefit of revealing a token. The method uses attention measured at masked positions to estimate how useful those positions would be once revealed. While this seems like a reasonable heuristic, the connection is not directly validated. Some analysis of how well the estimated support aligns with the actual improvement in predictions after revealing tokens would help strengthen the motivation.

  • The additional training cost is not fully accounted for in the comparisons. GoldiMask requires an extra forward pass and has a somewhat higher training cost than standard SFT, while the comparisons use the same number of training steps. It would be useful to see whether the improvements hold under a matched compute budget, where baselines are allowed to train longer.

Questions for authors
  • GoldiMask uses two forward passes and one backward pass per step, yet the training time is only about 1.2× that of vanilla SFT. Could you provide a breakdown of the additional cost.
4 comments

We thank the reviewer for the constructive feedback and helpful suggestions. We respond to each point below and present the additional experiments.

W1. Choice of attention layer and heads

Thank you for raising the importance of the layer choice. We agree that it is interesting to see if this choice affects the results and have therefore performed an ablation, where we take the attention weights from different blocks.

Attention used for selectionGSM8KMATH500CountdownSudokuAvg.
Final Block (ours)80.4 ± 1.239.5 ± 0.827.7 ± 2.716.9 ± 0.541.1 ± 0.7
Block 16 (middle)80.3 ± 0.839.6 ± 0.626.6 ± 4.317.2 ± 1.640.9 ± 1.1
Block 8 (early)80.6 ± 1.237.9 ± 1.125.4 ± 2.416.5 ± 1.440.1 ± 0.8
Mean of last 4 blocks80.7 ± 1.338.1 ± 1.224.9 ± 2.217.4 ± 1.340.3 ± 0.3

The final block achieves the highest average accuracy, although the middle block performs similarly, and averaging the last four blocks gives a higher mean on Sudoku. We therefore retain the final block as a simple default that avoids combining attention across layers. These results suggest that performance is not uniquely dependent on the final block, while earlier-layer attention yields a lower average. We have added this ablation to the revised manuscript in Appendix I.1.

W2. Does attention reflect the actual benefit of revealing a token?

We thank the reviewer for this suggestion. We tested the connection directly on LLaDA-8B (s1K). For 200 masked training states, we revealed 8 random token sets each and measured, for every target that remained masked, its estimated support ϕ(S)\phi(S) (from attention computed before the reveal) and its actual change in correct-token probability after the reveal.

Estimated support φ(S)0–0.20.2–0.40.4–0.60.6–0.80.8–0.90.9–1.0
Mean actual Δp of the target+0.02+0.16+0.30+0.43+0.47+0.54

The actual improvement rises monotonically with the estimated support, so attention measured at masked positions reflects the benefit of revealing those positions. We have included this analysis in Appendix H.7 in the revised manuscript.

W3. Matched compute budget

We thank the reviewer for this suggestion. We trained Vanilla SFT and CART, the two single-forward-pass baselines, for 1.5×1.5\times as many steps, keeping all other settings unchanged. Results are accuracy percentages reported as mean ±\pm standard deviation over three seeds. The 1×1\times schedule denotes the original number of optimizer steps; Δ\Delta rows show the change from 1×1\times to 1.5×1.5\times training in percentage points.

s1K (LLaDA-8B)

MethodStepsGSM8KMATH500CountdownSudokuAvg.
Vanilla SFT1×79.2 ± 0.236.2 ± 0.917.2 ± 2.114.5 ± 1.336.8 ± 0.3
Vanilla SFT1.5×78.9 ± 0.833.3 ± 0.422.1 ± 6.514.9 ± 1.537.3 ± 1.4
Vanilla SFTΔ−0.3−2.9+4.9+0.4+0.5
CART1×80.4 ± 1.035.9 ± 1.624.0 ± 1.414.4 ± 1.638.7 ± 0.6
CART1.5×79.7 ± 0.235.9 ± 0.925.9 ± 1.615.7 ± 1.339.3 ± 0.3
CARTΔ−0.70.0+1.9+1.3+0.6
GoldiMask-D1×80.4 ± 1.239.5 ± 0.827.7 ± 2.716.9 ± 0.541.1 ± 0.7

KodCode

ModelMethodStepsHumanEvalMBPPAvg.
LLaDA-8BVanilla SFT1×31.5 ± 1.540.5 ± 2.336.0 ± 1.1
LLaDA-8BVanilla SFT1.5×33.1 ± 2.139.9 ± 2.436.5 ± 0.9
LLaDA-8BVanilla SFTΔ+1.6−0.6+0.5
LLaDA-8BCART1×33.7 ± 1.340.6 ± 1.637.2 ± 0.8
LLaDA-8BCART1.5×34.3 ± 1.941.2 ± 1.437.8 ± 0.5
LLaDA-8BCARTΔ+0.6+0.6+0.6
LLaDA-8BGoldiMask-D1×39.0 ± 1.244.2 ± 2.641.6 ± 1.9
Dream-7BVanilla SFT1×45.1 ± 4.041.5 ± 2.043.3 ± 2.8
Dream-7BVanilla SFT1.5×46.5 ± 2.544.7 ± 3.145.6 ± 2.8
Dream-7BVanilla SFTΔ+1.4+3.2+2.3
Dream-7BCART1×46.7 ± 3.451.2 ± 1.549.0 ± 1.6
Dream-7BCART1.5×46.1 ± 2.751.2 ± 1.848.7 ± 2.3
Dream-7BCARTΔ−0.60.0−0.3
Dream-7BGoldiMask-R1×47.0 ± 4.352.4 ± 1.249.7 ± 1.6
Dream-7BGoldiMask-D1×45.9 ± 4.149.3 ± 1.147.6 ± 1.5

Longer training improves average accuracy in most configurations. GoldiMask-D remains ahead of both extended-training baselines on LLaDA-8B for s1K and KodCode, while GoldiMask-R achieves the highest average on Dream-7B/KodCode, using fewer optimizer steps. We have added these comparisons to Appendix I.2 of the revised manuscript.

Q1. Breakdown of the additional cost

We thank the reviewer for raising this. The wall-clock measurements reported in Table 11 include model evaluation and checkpoint saving and logging. To provide a comprehensive picture of the computational costs, we present a detailed breakdown of the training time below. These details are also reflected in Section F of the updated manuscript.

ComponentVanilla SFTGoldiMask-DGoldiMask-R
Pass 1: gradient-free forward—275 s275 s
Reveal selection—14 s48 s
Pass 2: forward with gradients265 s265 s265 s
Backward pass434 s438 s438 s
Loss computation and bookkeeping14 s38 s45 s
Gradient sync and optimizer update47 s50 s50 s
Model loading210 s210 s210 s
Periodic validation49 s49 s49 s
Checkpoint saving and logging512 s512 s512 s
Total1,532 s1,850 s1,892 s
× Vanilla (end to end)1.00×1.21×1.23×

Thank you for the detailed rebuttal and additional experiments, which address my main questions. Some minor concerns remain: GoldiMask-D falls below CART on Dream-7B/KodCode, and the per-step overhead appears closer to 1.4× once fixed costs are excluded, which should be stated in the paper. I remain positive and will keep my score of weak acceptance.

We thank the reviewer for their feedback and positive assessment.

GoldiMask-D falls below CART on Dream-7B/KodCode, but has a higher average in the other five settings in Tables 1-3. GoldiMask-R achieves the highest average on Dream-7B/KodCode.

We will state the training overhead separately from the end-to-end runtime in the revised manuscript.

Please also note that we have included additional experiments where we have given the baselines more compute than GoldiMask. Appendix I.2 shows that training Vanilla SFT and CART for 1.5× as many steps slightly improves their results, but does not close the gap to GoldiMask-D on LLaDA-8B/s1K and KodCode, or to GoldiMask-R on Dream-7B/KodCode.

We thank the reviewer again for their comments, which helped us improve our paper. We are, of course, open to further comments and questions!

Reviewer_AsE6OpenReview ↗
Rating
4
Confidence
3
Informativeness
2
Presentation
3
Soundness
3
Read review
Paper summary

The paper proposes GoldiMask, a supervised fine-tuning method for masked diffusion language models that coordinates which response tokens become visible context and how strongly the remaining masked tokens are supervised. It first over-masks the response and uses a forward pass to estimate ground-truth token probabilities and attention scores. It then selects tokens to reveal by approximately maximizing an objective that balances support provided to other targets against supervision forfeited by revealing a token. A second forward pass measures the remaining targets’ responses to the selected context. Their positive log-probability gains and updated confidence determine capped, normalized loss weights. Experiments cover three backbones, two reasoning training corpora, and a code-generation corpus, alongside component ablations and confidence-threshold decoding evaluations. The main findings are improved average reasoning accuracy, setting-dependent code-generation gains, and fewer decoding iterations on GSM8K and MATH-500. My main concern is that the paper shows the mathematical properties of its objective more convincingly than it shows the objective actually helps learning. Token selection relies on attention computed while candidate positions are still masked, assuming this attention predicts how much revealing the true tokens will help. Appendix H.5 gives some support, but the held-out R² is only 0.041 and 0.098, so most of the variation goes unexplained and it’s unclear whether the objective reliably picks useful tokens to reveal. The accuracy gains also need some caveats. On LLaDA-8B/s1K, GoldiMask-D’s 1.6-point lead over LIFT shrinks to 0.6 points over the strongest baseline at fixed generation length. Other settings hold up better, but some headline gains depend quite a bit on generation length, and Appendix F doesn’t say how it was chosen (an ambiguity, not evidence of test leakage). Code-generation results are mixed too: GoldiMask-R is essentially tied with GIFT on LLaDA, and GoldiMask-D trails CART on Dream. AIME24 also shows no consistent gains on harder math.

Critical strengths

The strongest contribution is the explicit formulation of the trade-off between supplying context and retaining useful prediction targets. Equation 5 captures an interaction that independent token rankings can miss: a token’s reveal value depends on both the context already selected and the targets that remain supervised. The submodularity proof appears correct under the stated fixed-attention assumptions, and the water-filling analysis provides a sound characterization of the proposed allocation rule. The authors appropriately avoid claiming that these results guarantee optimal downstream learning. Empirically, the evaluation has meaningful breadth: both principal variants improve average reasoning accuracy over the strongest reported baseline in all four backbone–corpus settings in Tables 1–2. Retrained baselines, three-seed reporting, validation-selected checkpoints, and the fixed-length comparisons in Table 9 are valuable strengths. Table 4 also provides useful evidence for both components: iterative selection with uniform weights reaches 39.9 average accuracy, while adding adaptive weighting raises this to 41.1. The reported training overhead is reasonable in the measured setting—approximately 1.21–1.23 times vanilla SFT, comparable to LIFT’s 1.24 times.

Critical weaknesses

My principal concern is that the mathematical properties of the constructed objective are established more convincingly than its relationship to actual learning benefit. Selection freezes attention computed while candidate positions still contain mask tokens and assumes that attention predicts the support obtained by revealing their true contents. Appendix H.5 offers some evidence, but the saturating predictor achieves held-out within-target R2R^2 values of only 0.041 and 0.098, compared with 0.033 and 0.082 for a linear predictor. These results indicate a usable signal while leaving most measured variation unexplained; they do not establish that the objective reliably ranks useful reveal sets. The accuracy evidence also needs qualification. On LLaDA-8B/s1K, GoldiMask-D’s main-table advantage is 1.6 percentage points over LIFT, whereas its fixed-length advantage over the strongest baseline is only 0.6 points—39.5 versus 38.9 in Table 9. Other settings retain stronger improvements, so this does not invalidate the overall result, but it shows that some headline gains depend materially on generation-length choices. Appendix F clearly specifies validation-based checkpoint selection but does not clearly identify the split or criterion used to select generation length; this is an unresolved evaluation ambiguity, rather than evidence of test leakage. Code-generation gains are likewise uneven: GoldiMask-R scores 38.2 ± 2.4 versus GIFT’s 38.0 ± 0.5 on LLaDA, while GoldiMask-D underperforms CART on Dream, 47.5 versus 49.0. AIME24 results also do not establish consistent improvements on harder mathematics.

Questions for authors

Several implementation details should be more explicit, including the main validation split’s construction and size, the KodCode training subset and schedule, and sampling settings for AIME24 pass@16. The authors should also specify how stochastic layers are handled between the two forward passes: if their randomness differs, the measured probability change can reflect more than the reveal intervention. Tables 4–6 should include uncertainty, particularly because several priority-function and weight-cap differences are small. The evidence supports prioritizing intermediate-confidence targets more strongly than it supports the precise choice p(1−p)3p(1-p)^3; the authors generally acknowledge this appropriately. Comparisons among selectors should also emphasize that D/R enforce an exact reveal budget while T/P can stop early, changing the number of supervised targets.

3 comments

We thank the reviewer for the careful reading and positive assessment of our work. We address the questions and concerns below.

W1. Link between the objective and actual learning benefit

We thank the reviewer for highlighting this. The R2R^2 in Appendix H.5 reflects noisy single-token responses, whereas GoldiMask selects whole reveal sets. We therefore now also tested the selected set directly (Study 4, Appendix H.6). Across 200 masked training states with t≤0.75t \leq 0.75, where the visible context is still informative, GoldiMask’s set outperforms all 16 random sets of the same size in 80.5% of states and yields 1.55× their benefit. Above t = 0.75, where inputs are almost fully masked, it offers no advantage.

W2. Fixed-length gains are smaller

We agree with the reviewer. The margins narrow at fixed generation lengths, although GoldiMask-D remains ahead of every baseline in all settings (Table 9). We now state this in lines 325-326.

Q1 and W3. Implementation details and generation-length selection

We thank the reviewer for raising this. We provide the requested information below, which has also been added to Appendix F.

Generation length. Generation length is not selected on the validation split. Following d1 (Zhao et al., 2025) and LIFT (Parashar et al., 2026), we evaluate each method at generation lengths {128, 256, 512} and report the best length separately for each benchmark, applying the same procedure to all methods. Table 9 additionally reports results with lengths fixed in advance and shared across methods.

Validation split. Following d1 and LIFT, we hold out 1% of the training set after a seeded shuffle. It is used only for checkpoint selection, by lowest validation loss.

KodCode. We train on KodCode-Light-RL-10K for 5 epochs (6,190 steps) with learning rate 5×10−65\times10^{-6} and global batch size 8, identically for all methods.

AIME24 pass@16. Following LIFT, we generate 16 samples per problem with generation length 512 and 256 denoising steps, counting a problem as solved if any sample is correct. We use temperature 0.7 following [1,2][1,2].

Q2. Stochastic layers between the two forward passes

We thank the reviewer for raising this point. The backbones contain no active dropout; dropout is confined to the LoRA attention adapters (p=0.05) and is active in both passes. Using a trained GoldiMask adapter on 200 masked training states, we compared two passes on identical inputs with independent dropout masks. The mean absolute difference in target log-probability was 0.048, compared with a mean increase of 1.33 after applying GoldiMask’s reveal set. The dropout-only difference is therefore approximately 3.6% of the measured reveal-induced increase in log-probability. Under independent dropout draws, the selected reveal sets overlap by 93%, and, for a fixed reveal set, utility weights have a Pearson correlation of 0.985. These measurements suggest a limited effect on reveal selection and utility weighting in the tested setting.

Q3. Uncertainty in Tables 4–6

We have updated Tables 4–6 to report standard deviation across three training seeds.

R1. Exact budget (D/R) versus early stopping (T/P)

Thank you for this suggestion. We have clarified this distinction in Appendix C.1 (lines 884–885) of the updated manuscript.

[1]. Yang, Zhaohui, et al. "Unearthing Gems from Stones: Policy Optimization with Negative Sample Augmentation for LLM Reasoning." EMNLP (Findings). 2025.

[2]. Kim, Jeonghye, et al. "Understanding reasoning in llms through strategic information allocation under uncertainty." arXiv preprint arXiv:2603.15500 (2026).

Reviewer_AsE6OpenReview ↗

comment on responses

I thank the authors for clarifying all of these points, I believe this will help to improve the quality of the submission. I have no further questions at this time, and I will keep my positive score.

We thank the reviewer for their constructive comments, helpful suggestions, and positive assessment. Their feedback has helped us improve the paper.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.