acceptodds
Under review as a conference paper at ICLR 2027

Guiding Discrete Diffusion with Distilled Process Rewards

Abstract

Masked diffusion language models such as LLaDA generate text by iteratively committing tokens from a fully masked canvas. An erroneous early commitment can influence subsequent denoising steps and lead to an incorrect final answer. Process reward models (PRMs) can localize these errors, but running a separate 7B-parameter PRM at inference is expensive. We present a method for distilling process rewards into a frozen diffusion language model. A value head summarizes the quality of a partially masked state and supports post-hoc repair; a step-issue head predicts token-level issue scores and guides commitment order during the original denoising loop. We evaluate a single 256-token canvas to isolate adaptive commitment ordering from fixed block boundaries: lower-risk positions can be committed in parallel, while higher-risk positions wait for additional context. In one reported GSM8K run with LLaDA-8B-Instruct, step-issue guidance changes accuracy from 44.7% to 53.9% (+9.2 pp) with a 1.1M-parameter MLP. Matched controls in that run do not reproduce this change with random or shuffled issue scores. On MATH-500, the corresponding GSM8K-to-MATH-500 transfer run changes accuracy from 24.0% to 25.8%. Direct PRM repair reaches a higher point estimate but requires the full PRM at inference and an additional regeneration pass. In a matched 100-example GSM8K timing run, guidance was only 0.4% slower than baseline, versus 23.6% for distilled repair and 35.1% for direct PRM repair.

Primary area:
generative models
Keywords:
Discrete diffusion language models, Masked diffusion (LLaDA), Process reward models (PRMs), Value-head distillation, Test-time repair / suffix regeneration, Error localization

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_4dhgOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a method to leverage the scores given by process reward models (PRM) to guide the decoding process of a discrete diffusion language model. The authors propose three different methods. “PRM repair” scores complete generations and regenerates from the first step whose score is below a threshold. “PRM-distilled repair” follows the same logic but uses a distilled value head to resample mid-generation if scores fall below a threshold. Finally, “step-issue guidance” uses a step-issue head to define a decoding policy that favors decoding positions that would lead to higher PRM scores. Both the value head and the step-issue head are trained to predict PRM scores, but the value predicts scores for the entire sequence (average of four rollouts) and step-issue head predicts scores per token position. The methods were tests on two reasoning datasets GSM8K and MATH500, improving over the baseline LLaDA model with no guidance.

Critical strengths
  • All three proposed methods improve over the baseline in all cases, which suggests the PRM scores are used effectively and the distilled head is capable of capturing at least some of the behavior of the original PRM.
  • The simplicity of the methods is a strength, as it facilitates their application by practitioners. Moreover, each of the methods represent different trade-offs between computational cost and accuracy, which might be interesting for downstream applications.
Critical weaknesses
  • The proposed guidance interventions, albeit interesting, are not yet well motivated. They are based on thresholding predicted scores or on additive changes to the decoding priority, neither of which is sufficiently justified in the text. This is in contrast to much of the test-time guidance literature, which similarly estimates value functions but derives their guidance from KL-regularized reinforcement learning instead (see references below). The guidance interventions proposed in this paper might have merit and be a valid addition to the literature, but it is hard to gauge that since these prior works are not discussed or evaluated in the paper.
  • Following on the previous point, the related work does not discuss prior work on aforementioned score-guided inference methods for autoregressive [1, 2, 3] and diffusion [4, 5, 6] models, as well as learned unmasking policies [7].
  • The experiments are limited.
    • The only baseline seems to be standard LLaDA and no other guidance method is considered. For instance, SVDD [5] is a relevant baseline to PRM repair, while step-issue guidance is essentially an unmasking policy and [7] could be a relevant baseline.
    • The only use case considered is math reasoning with a single model and PRM scorer. While this might be sufficient to demonstrate the proposed methods, I do not feel confident the gains would be easily translated to other domains because of the guidance interventions are not well motivated and rely on heuristics like resampling from the first digit in the generation.
    • Unless I missed it in the text, experiments cover a single seed. For instance, the gains of PRM-distilled repair over the baseline on MATH500 might not be statistically significant (Table 1), and random scores seem to outperform the baseline on GSM8K (Table 2).
  • Some parts of the text are quite hard to follow. The inference paragraphs for both PRM-distilled repair and step-issue guidance are particularly confusing.

Minor issues

  • At the top of page 2, there is a box that feels out of place. What is the purpose of this example? If it is useful, it should be better integrated into the text, with a caption.
  • The sentence “This inconsistency motivates guidance; the final output alone does not identify which commitment decisions caused it.” also feels out of place. What inconsistency is the text referring to here?

References

[1] Yang, Kevin, and Dan Klein. "FUDGE: Controlled text generation with future discriminators." Proceedings of the 2021 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies. 2021.

[2] Han, Seungwook, et al. "Value augmented sampling for language model alignment and personalization." arXiv preprint arXiv:2405.06639 (2024).

[3] Zhao, Stephen, et al. "Probabilistic inference in language models via twisted sequential monte carlo." arXiv preprint arXiv:2404.17546 (2024).

[4] Uehara, Masatoshi, et al. "Inference-time alignment in diffusion models with reward-guided generation: Tutorial and review." arXiv preprint arXiv:2501.09685 (2025).

[5] Li, Xiner, et al. "Derivative-free guidance in continuous and discrete diffusion models with soft value-based decoding." arXiv preprint arXiv:2408.08252 (2024).

[6] Singhal, Raghav, et al. "A general framework for inference-time scaling and steering of diffusion models." arXiv preprint arXiv:2501.06848 (2025).

[7] Jazbec, Metod, et al. "Learning unmasking policies for diffusion language models." arXiv preprint arXiv:2512.09106 (2025).

Questions for authors
  1. The distilled value head tries to map incomplete generation to the scores the PRM assigns to complete generations. This essentially learns a proper value function in the reinforcement learning sense [8]. Thus, one possible principled guidance objective is to maximize expected PRM reward while penalizing KL divergence from a specified reference completion distribution. This has a closed form solution as a tilted distribution that favors higher-scoring completions [4, 5]. Why not derive guidance from this objective? What advantages or practical constraints motivate that choice?
  2. The value function is trained with samples from two specific policies, one greedy and another stochastic. However, these policies do not account for the guidance intervention. That is, after the first intervention, the value function might not be a faithful representation of the expected score, since the test time sampling policy is different from the one used at training time. Have the authors considered that issue?
  3. There is a large gap in performance between PRM repair and the two other methods that use distilled heads. Could the authors comment on what is the primary cause for this? Is it the difference in performance between distilled head and original PRM or is it the different decoding policy induced by value and step-issue heads?

References

[8] Mudgal, Sidharth, et al. "Controlled decoding from language models." arXiv preprint arXiv:2310.17022 (2023).

3 comments
AuthorsOpenReview ↗

Response to Reviews

Why use the proposed intervention instead of KL-derived guidance?

The intervention is chosen to act on the decoder decision that determines which positions receive further context before commitment: using a lightweight predictor from frozen hidden states to decide which masked positions to commit later. It reuses the current backbone pass and leaves its token probabilities at that state intact. This directly targets commitment order while avoiding PRM evaluation and regeneration for the guidance method. The recorded computation figures below support the feasibility of that design.

We agree that the additive priority rule and repair thresholds are heuristic. We do not derive them as the optimizer of a KL-regularized reward objective. The scalar head predicts a PRM-derived reward proxy under the synthetic-fill procedure used for supervision; this is not automatically the expected return of the actual guided continuation policy. A KL-based objective is a principled alternative, but showing that our particular intervention implements its optimal transitions would require a derivation we do not provide. Our contribution is the supervision mapping and practical commitment-order mechanism, not an optimality result.

Relevant prior work and missing baselines

The reviewer is right that the contribution should be situated more clearly within future-aware prediction, value-guided diffusion inference, and learned unmasking policies. FUDGE uses partial-sequence attribute predictions to adjust autoregressive generation probabilities. SVDD uses soft values for reward-guided diffusion inference. Learning Unmasking Policies for Diffusion Language Models learns unmasking decisions through reinforcement learning. Our method instead learns from completed-step PRM scores projected onto masked canvas tokens and uses those predictions to change commitment priority with a frozen backbone.

The contribution lies in the PRM-to-masked-canvas supervision and the resulting inexpensive commitment policy. It does not require claiming that value-based guidance or learned unmasking itself is new. These distinctions position the method, while empirical superiority over those alternatives remains unestablished. We have not evaluated SVDD or the cited learned unmasking policy and therefore cannot claim an advantage over them. The existing experiments compare the proposed mechanism with unguided decoding, direct PRM repair, distilled repair, and score controls. We acknowledge that a broader comparative evaluation remains a limitation.

Distribution shift after guidance

The concern correctly identifies a limit of the scalar value proxy, but it is useful to distinguish that repair baseline from the primary token-level guidance method. The issue head supplies a learned ordering score; its operation does not require that a scalar head equal the expected return under the deployed policy. Intervention nevertheless changes the future state distribution for both approaches, so the concern remains relevant to prediction reliability. The training labels come from completed fills of logged states, and the resulting value proxy need not remain calibrated after guidance or candidate-masking interventions. We do not claim policy-consistent value estimation or an on-policy improvement guarantee. The end-task results evaluate the complete deployed procedures in the reported settings, but they do not isolate or rule out this distribution-shift problem.

Why does direct PRM repair outperform the distilled methods?

The gap reflects methods with materially different information and computation budgets, so matching the full PRM is not the claim required for the lightweight guidance contribution. The current comparison does not isolate one cause. Direct PRM repair scores completed reasoning and regenerates from the first low-scoring step. Distilled repair uses a scalar prediction at a partially masked state and candidate-mask probes to choose a digit-bearing boundary. Guidance uses token-level predictions to change the original decoding trajectory without regeneration. Scorer quality, available information, boundary selection, intervention timing, and computation therefore differ together.

On GSM8K, direct PRM repair, distilled repair, and guidance reach 63.53%, 48.07%, and 53.90%, respectively; on MATH-500 they reach 29.8%, 24.8%, and 25.8%. These observations support an empirical quality/cost trade-off within the tested methods. They do not establish whether teacher approximation error or the different intervention policy is primarily responsible for the gap. The digit-based boundary heuristic belongs to the distilled-repair baseline; it is not required by the primary step-issue guidance method.

Random controls and useful learned scores

Random scores do improve GSM8K from 44.73% to 46.70% in the reported run, so arbitrary changes to commitment order can affect accuracy favorably. Learned scores reach 53.90%, exceeding the random control by 7.20 percentage points; shuffled learned scores reach 44.66%. On MATH-500, random and shuffled scores reach 21.2% and 20.8%, versus 25.8% for learned scores and 24.0% for baseline. The learned-score advantage over both random and shuffled controls is evidence that the trained assignments contain useful information beyond the particular arbitrary perturbations tested. The controls do not establish that every gain is attributable solely to precise error localization.

The controls have limits. Shuffling occurs over the full prompt-plus-canvas sequence, and the implementation's control labeled “confidence” uses a hidden-norm proxy rather than token confidence. We identify it as a hidden-norm control in the manuscript. These details limit the mechanistic interpretation, and we do not claim that the controls establish significance across independent seeds.

Computational motivation and the scope of timing evidence

The primary guidance method uses an approximately 1.1M-parameter head, updates its predictions every 16 steps, and requires no PRM inference or separate regeneration. In the recorded 100-example GSM8K timing run, baseline and guidance each use 6,400 backbone calls. Guidance adds 400 issue-head calls; total times are 564.56 and 566.85 seconds, a 2.29-second increase, approximately 0.41%. Distilled repair takes 697.73 seconds and direct PRM repair 762.41 seconds in that timing run.

This is direct evidence for small observed guidance overhead under that configuration. It is a single timing run on a subset, and we do not claim that 0.4% is a statistically resolved or hardware-independent overhead. The 100-example runtime experiment is distinct from the full-benchmark accuracy evaluations.

Thank you for responding to my comments and questions.

Overall, the ideas of the paper have merit, but at the current stage, the proposed methods are still not presented in a convincing manner. As I see it, two main problems remain. One pertains to evaluation. While it is not necessarily a problem to rely on heuristics, the lack of a rigorous theoretical motivation would have to be compensated by an extensive empirical evaluation, which the current version of the paper does not provide. A larger number of datasets, models, baselines and random seeds would be required to support the main claims of the paper. The other issue is the presentation. The text is unnecessarily hard to follow, and the methods are not well motivated in the context of the broader guidance literature.

Reviewer_BMWrOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
1
Soundness
1
Read review
Paper summary

This work aims to improve generation of masked diffusion models (MDMs) by distilling the scores produced by process reward models, which provide scores for each step in a reasoning trace, into lightweight classifier heads that guide the decoding order of an MDM and determine low quality generated sequences that should be remasked and regenerated.

Critical strengths

Determining token ordering in MDM generation is an important problem.

Critical weaknesses

I unfortunately found the current manuscript difficult to follow. Below I detail a (non-exhaustive) list of examples that highlight the issues with this work.

  • I found the abstract to not be clearly written.
  • Lines 23-24: What is “random or shuffled issue scores”? This is not a standard term in my opinion
  • Line 53: What are “scaled issue score”?
  • The example in the intro (Lines 59-63) is not labeled as a figure nor is it referenced in the text.
  • The introduction seems to abruptly change from describing the method in line 57 to discussing some previously described “inconsistency” in line 64?
  • The paragraph on PRMs (Lines 94-97) is very difficult to follow. I am not sure what the last sentence is supposed to be saying. More details on how PRMs work is needed.
  • Section 3.1 does not seem comprehensive. The authors introduce notation to describe decoding in MDMs, but the provided information would not be sufficient to describe the generation algorithm from these models.
  • Notation is overloaded: pp is used for probabilistic model and as the vocab index; ff is used to denote the NN model and for the denoising fraction.
  • Lines 134-135: what does “We sample states at denoising fractions” mean? Are the authors keeping x0.25,x0.5,x0.75\mathbf{x}^{0.25}, \mathbf{x}^{0.5}, \mathbf{x}^{0.75} only? If so, why just those 3 sampled times? Also why not be consistent in the notation?
  • Lines 139-140: “We decode only the generated suffix (the prompt is retained separately) and segment it at blank-line boundaries, discarding empty or whitespace-only segments” what is a “blank line boundary?”
  • Lines 143-144: “A fill is considered malformed if the number of PRM scores differs from the number of nonempty segments; such a fill is rejected rather than assigned an ambiguous label” what does a non-empty segment refer to? A diagram of the methodology would be very helpful.
  • Lines 145-155 are hard to follow
  • The term “Issue” scores is referenced frequently without every being properly introduced or defined
  • Lines 165-166 describe methodology that is not introduced. What is a “suffix boundary” what is an “advantage test”?
  • Line 205, τval\tau_{val} and τadv\tau_{adv} are never defined, neither is “midpoint masked state”
Questions for authors

Please see the weaknesses section above.

3 comments

Plain explanation of the method and issue scores

A PRM scores completed reasoning segments. To obtain supervision for a partially masked diffusion canvas, we fill its masked positions offline, score the resulting completed segments, and map each segment score back to its token span. A token belonging to a segment with PRM score s receives the soft issue target 1-s. The issue head predicts this target from frozen backbone hidden states and timestep features. It is trained to reproduce projected PRM feedback, rather than independently annotated token errors.

During decoding, a masked position's commitment priority is its current top-token probability minus lambda times its predicted issue score. In the main setting, lambda is 0.75. High issue scores therefore lower priority and defer commitment while other positions are revealed. Scores are updated every 16 denoising steps and cached between updates. The head does not directly alter logits at a fixed canvas, but changed commitment decisions can alter subsequent canvases and hence later predictions.

“Random” and “shuffled” issue scores refer to experimental controls. The shuffled control reassigns learned scores to positions; the random control draws new scores at update steps. The constant control subtracts the same score at every position, preserving commitment ranking under fixed-count selection. These are control definitions rather than assumed standard terminology. We agree they need to be introduced before use.

Sampling intermediate states, PRM segmentation, and token alignment

The training corpus retains three partially masked states per trajectory, at denoising fractions 0.25, 0.50, and 0.75. These are intermediate states in a trajectory, rather than three separate complete generations. The corpus contains 9,000 state rows. This sampling design is heuristic; we do not claim that three fractions are optimal.

For each state, the greedy fill assigns the logged argmax token to each mask. Three additional fills sample masked positions from stored normalized top-5 distributions. The PRM receives the original question and a completed generated response. The response is split at blank-line boundaries, and empty or whitespace-only segments are discarded. Each retained segment receives a PRM score. A fill is rejected if its number of scores differs from its number of retained segments, avoiding ambiguous score-to-text alignment.

Segment boundaries are mapped back to token spans by decoding prefixes of the exact fill sequence and adding the prompt-token offset. Tokens in a scored span inherit one minus that segment's score; targets are averaged over aligned fills, and positions without a valid target are excluded from the issue loss. For example, a segment scored 0.2 supplies issue target 0.8 to its aligned tokens. This example illustrates the target definition, not a new experimental result.

The state-level value head uses the mean fill-level PRM score instead of token-level targets. Its separate diagnostic uses nine fractions from 0.1 to 0.9, which should not be confused with the three fractions used to construct the training corpus.

Repair terminology and thresholds

A suffix boundary is the generated-token position from which the output is masked again for repair. For distilled repair, the value head scores the midpoint masked state. The current thresholds are tau_val = 0.5 and tau_adv = 0. If the predicted value falls below 0.5, revealed generated positions whose decoded token contains a digit are considered as boundary candidates. Each candidate is temporarily masked and rescored. The policy chooses the largest predicted value increase and performs repair only if that increase is positive. It then masks from that position through EOS and performs one deterministic 64-step infill pass.

This is the counterfactual advantage test: it compares the value prediction after a candidate masking intervention with the original value prediction. It does not measure an experimentally established causal advantage. Candidate restriction to digit-bearing positions is a math-specific heuristic.

Direct PRM repair uses a different rule: score the completed reasoning, find the first step with PRM score below 0.5, and regenerate the suffix starting at that step. The primary step-issue guidance method does neither of these repair procedures; it adjusts commitment priorities during the original decoding loop.

Presentation criticism and the motivating example

We accept that the overloaded notation and insufficiently introduced terms make the method harder to follow. To state the mechanism clearly, let i index a position, c_i denote token confidence, and u_i the issue prediction: the priority is c_i - lambda u_i. Separating position indices, vocabulary symbols, backbone functions, and completion fractions removes the ambiguities identified by the reviewer. The pipeline figure and repair pseudocode in the current manuscript provide procedural detail, which will be provided in the additional page given

The motivating popsicle example illustrates a final arithmetic expression inconsistent with its own preceding numbers: 90-75 should be 15. It does not establish which commitment caused that error or demonstrate a measured correction by guidance. We agree that the example needs a caption and an explicit textual reference, and that “this inconsistency” should identify the arithmetic discrepancy directly. These are presentation shortcomings rather than additional empirical claims.

Thank you for providing your response.

Unfortunately, I believe the current manuscript still requires significant revision, which renders it not fit for publication in its current form.

Moreover, the empirical evidence currently provided is insufficient, in my opinion. The authors "Common Response" above focuses on the improvement over the LLaDA instruct model results on the single GSM8K benchmark under a very specific setting reported in the paper which yields a baseline result of ~45% accuracy; this is significantly lower than the performance of this model reported elsewhere in the literature.

Reviewer_mDkpOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper proposes to distill step-level rewards from Qwen2.5-Math-PRM into two heads on top of a frozen LLaDA-8B. In order to do so, it unmask the tokens from partially masked states using the process reward model (PRM), and aligning the step scores back to token spans. The proposed method i.e., a token-level step-issue head, subtracts a scaled issue score from each position’s commitment priority during denoising, so suspicious positions are committed later without touching the logits or needing a regeneration pass. A state-level value head serves as a distilled post-hoc suffix-repair baseline. On GSM8K, guidance raises accuracy from 44.7% to 53.9% (random, shuffled, and constant score controls do not reproduce this) with only 0.4% time overhead. Gains are much smaller on MATH-500 (24.0% to 25.8%), and direct PRM repair stays more accurate (63.5% / 29.8%).

Critical strengths

None.

Critical weaknesses
  • "All accuracy comparisons use one evaluation run" (Lines 451-452) is a critical weakness. Since the claimed transfers from GSM8K training to MATH-500 is 24.0% to 25.8%, 1.8% difference is very small and requires multiple runs to confirm the difference is significant. (Lines 464-465)
  • The experiments only used one diffusion LM i.e., LLaDA 8B and only one process reward model i.e., Qwen2.5-Math-PRM. We therefore cannot distinguish does the proposed method generalizes towards other diffusion language models.
Questions for authors
  • Are there experiments using other diffusion language model and process reward models? Otherwise the experimental results could be model-specific.
2 comments
AuthorsOpenReview ↗

Response to review

Single-run evidence and MATH-500

The main empirical claim rests on the GSM8K improvement and its score controls, rather than statistical significance of the smaller MATH-500 transfer result. The 24.0% to 25.8% MATH-500 change should be read as a point estimate. It represents 120 versus 129 correct answers out of 500. The 24.8% distilled-repair result is an even smaller change. Our supported claim is that the GSM8K-trained heads were evaluated on MATH-500 with the reported point estimates; we do not claim that these differences establish a reliable improvement across seeds.

The principal empirical result is the GSM8K change from 44.73% to 53.90%, a 9.17 percentage-point increase in the recorded run. Constant and shuffled-score controls give 44.73% and 44.66%, while random scores give 46.70%. These comparisons provide evidence for the utility of the learned score assignments in this setting, although they do not establish repeated-training robustness. We acknowledge that limitation explicitly rather than using the MATH-500 result as the main evidence of effectiveness.

Model and teacher scope

The paper establishes an implementation and evaluation for the concrete LLaDA-8B-Instruct/Qwen2.5-Math-PRM-7B combination. Its mechanism addresses a specific representational mismatch: a completed-text PRM cannot directly label masked positions, while the proposed fill–score–align pipeline supplies supervision for those positions. That construction and its measured utility can be assessed within this combination; generality to another backbone or teacher remains untested. MATH-500 is a dataset-transfer evaluation, and the supplementary ASDiv-CoT-LLaMA evaluation also uses LLaDA; “CoT-LLaMA” names the evaluation prompt variant, not a second backbone. The ASDiv point estimates are 74.19% for baseline and 75.36% for guidance over 2,305 examples, with 128-token blocks. These observations extend dataset coverage within the same model/teacher combination. We agree that they do not demonstrate generality across model families or teachers and restrict our claim accordingly.

Thank you for providing your response.

After reading it, I believe the current manuscript still requires significant revision.

The paper establishes an implementation and evaluation for the concrete LLaDA-8B-Instruct/Qwen2.5-Math-PRM-7B combination.

So the insights gained from the current manuscript has no evidence beyond LLaDA-8B-Instruct/Qwen2.5-Math-PRM-7B combination, which is significantly limited and the title "Guiding Discrete Diffusion with Distilled Process Rewards" is significantly misleading the readers due to the scope being too broad. Furthermore, the weakness I addressed in the comments such as using other masked diffusion language models like SDAR [1] and/or Dream [2] is not addressed. Therefore, I stick to my original rating.

[1] https://huggingface.co/JetLM/SDAR-8B-Chat

[2] https://huggingface.co/Dream-org/Dream-v0-Instruct-7B

Comments on the paper

1 comment
AuthorsOpenReview ↗

Common Response

We thank the reviewers for their feedback. The central contribution is a supervision pipeline that makes completed-step PRM feedback usable during partially masked diffusion decoding: complete intermediate canvases offline, score reasoning segments, project those scores onto token spans, and train a small head to guide commitment order using frozen backbone states. The resulting guidance requires neither PRM inference nor a regeneration pass. In the reported GSM8K evaluation, it improves accuracy from 44.73% to 53.90%, exceeding shuffled scores at 44.66% and random scores at 46.70%. In the separate 100-example timing run, guidance uses the same number of backbone evaluations as baseline and adds approximately 0.41% observed time.

This combination of an explicit supervision mapping, a lightweight inference intervention, and controlled empirical results is the contribution we ask reviewers to assess. The smaller MATH-500 and ASDiv changes are supporting dataset-transfer observations, with significance and broader model/teacher generality remaining open. Below we clarify the mechanism, explain the design choices, and distinguish the primary guidance method from the repair baselines.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.