acceptodds
Withdrawn from ICLR 2027

A Hint from the Past: Guiding Diffusion Language Models under Context Updates

Withdrawn from ICLR 2027, so left out of the paper lists and the map.

Abstract

Reasoning assistants must accommodate changes to a user's requirements while a solution is being generated. In multi-step problem solving, this requires adapting to new constraints or objectives after reasoning has begun. The partial answer provides potentially useful derivations and solution strategies that can guide the response to these changes. We investigate whether selectively using text from the previous partial answer during generation can improve correctness beyond retaining it in context. To this end, we introduce HiPE, a method that guides diffusion language models under updated conditions by injecting selected passages as revisable hints during generation. Implemented in DiffusionGemma, HiPE uses the current generation state to determine intervention timing, attention to extract candidate passages, and predictive distributions to select their destinations. It requires no additional training and constructs interventions from existing decoder forward passes. On AIME 2024 and 2025, HiPE improves accuracy over regeneration by 2.22 and 3.33 percentage points, respectively, with fewer decoder forward passes on average. It achieves higher accuracy than all evaluated initialization-based reuse methods across mathematics and coding tasks.

Primary area:
applications to computer vision, audio, language, and other modalities
Keywords:
DiffusionGemma, Diffusion Language model, context update

Reviews

Among:
Shape:
1234this paper, Bottom 2%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.33 over 3 reviews, above 2% of the others.
Reviewer_Z5EzOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies how diffusion language models can reuse previous reasoning after user requirements change. HiPE is a training-free method that inserts selected passages from a previous partial answer during denoising, allowing the model to revise them in later steps. It uses the current generation state, attention, and token predictions to decide when to intervene, which passages to select, and where to place them. Experiments on mathematical reasoning and coding tasks show modest accuracy gains over RESTART on AIME and LiveCodeBench, with similar accuracy on GSM8K and MATH-500. HiPE also achieves higher accuracy than the evaluated warm-start baselines.

Critical strengths

Training-free intervention without additional decoder calls. HiPE uses the accepted-token fraction, along with attention and predictive distributions from existing forward passes, to decide when, what, and where to inject. The inserted text remains revisable during subsequent denoising, offering a practical way to reuse earlier reasoning without modifying model weights or decoding rules.

Controlled comparisons reveal meaningful differences between reuse strategies. Under matched contexts and generation budgets, HiPE achieves higher average accuracy than RESTART on AIME and coding tasks and surpasses all three warm-start baselines across the evaluated tasks. Although gains over RESTART are relatively modest, these results suggest that initialization-based reuse and selective injection during generation may have different effects.

Critical weaknesses

W1. Missing clean-restart and continuation baselines. RESTART keeps the original question and previous reasoning in context, while the warm-start methods begin a new generation rather than continue from the interrupted answer. The paper does not test continuation under the updated conditions, nor does it include a clean restart that solves the fully updated problem without the earlier question or reasoning. Given the relatively small gains over RESTART, it is unclear how HiPE compares with these two alternatives.

W2. No comparison with simple update guidance. The condition-update tasks come from Wu et al. [1], who show that brief prompt guidance can improve adaptation in their evaluated models. However, the paper does not test this simpler approach with DiffusionGemma under the same experimental settings. It is therefore unclear whether HiPE offers an advantage over simply reminding the model to follow the updated requirements.

[1] Wu et al. Are Large Reasoning Models Interruptible? ICML 2026.

Questions for authors

Q1. Practical efficiency. Could you report wall-clock latency under the same hardware and decoding settings, including the overhead of selecting and placing historical passages? NFE alone does not show whether HiPE provides a practical speed advantage over RESTART.

Q2. Intervention timing. How were the default accepted-fraction bounds of 0.25 and 0.45 and the cutoff of t < 12 selected? Table 10 shows that different windows can lead to different accuracy and NFE results, particularly on AIME 2025. Were these settings fixed before evaluation, selected using held-out data, or tuned on the evaluation tasks?

Minor comment on figure readability. The axis labels and tick labels in Figure 3 are relatively small. In Figure 4, some curves and distributions are difficult to distinguish, especially where they overlap. Larger labels and more distinct colors or line styles would improve readability.

1 comment

hank you for the concrete suggestions concerning baselines, practical efficiency, intervention timing, and figure readability.

Regarding W1 and W2, we conducted follow-up experiments with clean restart and explicit prompt guidance baselines, using the same problems, generation seeds, and output budgets within each dataset. Clean restart solves the fully updated problem without earlier reasoning, while prompt guidance explicitly instructs the model to adapt to the update. These comparisons did not establish a consistent advantage for HiPE over these alternatives. Clean restart achieved higher accuracy point estimates with greater NFE, while prompt guidance showed mixed results. We plan to expand these comparisons to include a strict continuation baseline and evaluate accuracy–computation trade-offs more systematically in a future submission.

Regarding Q2, the original accepted-fraction bounds and the (t<12) cutoff were chosen heuristically and require stronger justification. In our follow-up experiments, we removed this cutoff and selected the margin and accepted-fraction window using a MATH validation subset, then fixed the configuration for evaluation. We intend to further investigate sensitivity to intervention timing and parameter selection.

Regarding Q1, we agree that NFE reductions alone are insufficient to demonstrate wall-clock improvements. A comprehensive, controlled latency evaluation will be important for establishing practical efficiency. We also acknowledge the concerns about figure readability and plan to increase label sizes and make overlapping curves easier to distinguish.

We appreciate these suggestions and will incorporate the lessons from these evaluations into a more carefully controlled study for a future submission.

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

The paper studies the problem of generation under context updates (i.e., "interruptable" generation where the prompt/task is updated midway during the generation). Concretely, the authors study the setting of diffusion LLMs (DiffusionGemma model) and whether there is some benefit to not only providing the existing (outdated) generation in the context for the new generation, but also inserting it in the current canvas being denoised/generated. To this end, they propose an insertion mechanism that selects only relevant parts of the outdated generation (a1) and also determines at which sampling step and where exactly in the current canvas the selected part of the outdated generation will be inserted. They compare their proposed method with a baseline where the outdated generation is provided only in the context (RESTART baseline) on DiffusionGemma model and math/coding tasks (GSM8k/MATH/AIME/LCB)

Critical strengths
  • The setting of generation under context update is interesting and I haven't seem much work on using dLLMs for it (though I have to say that it's my first time reading/reviewing a paper on generation under context update so I'm not familiar with the existing techniques/literature there)
  • The proposed mechanism of selecting parts of the outdated generation makes sense and it doesn't introduce too large of an computational overhead (since it's done only once per canvas; though this overhead is not measured in the paper which is actually a limitation since the comparison with RESTART baseline in terms of NFEs might not be fair)
Critical weaknesses
  • Main experimental results look very weak, see Table 1. There the proposed HiPE methods outperforms RESTART baseline for more than 1% only on AIME datasets where it does so for 2-3%. But making claims on AIME datasets is tricky because these are very small datasets iirc (~30 questions per year)? So gap of 2-3% on AIME probably means a difference of a single question being correct/wrong which I wouldn't say is statistically significant and, if I put on my practitioner hat, it doesn't justify the additional implementation and potentially inference overhead (which as mentioned above is not measured, only NFEs are provided)
  • Also can we see std bars in Table 1? sometimes the gap is so small (e.g., 0.3%) that this is probably within error bars or?
  • While I think the proposed selection mechanism for the parts of outdated generation makes sense (via their attention scoring), I struggle a bit to understand the motivation for why not only provide the outdated generation in the context? Could the authors expand on this a bit? Because the current weak experimental results in Table 1 make me feel like the motivation is not really there?
Questions for authors

See weaknesses above

1 comment

Thank you for emphasizing the small evaluation sets, statistical uncertainty, and practical motivation.

We agree that the AIME results warrant cautious interpretation. Each dataset contains only 30 problems, and the three generation seeds per problem should not be treated as independent problems. We therefore estimated uncertainty by first averaging the paired seed-level differences within each problem and then resampling problems. The resulting confidence intervals included zero, highlighting the need for more comprehensive evaluations. We plan to present uncertainty estimates more prominently in a future submission.

Regarding the motivation, Restart already receives the previous partial answer in its context. Our hypothesis was that making selected passages available as revisable generation states could influence adaptation differently from providing those passages only as context. The practical benefit of this distinction, however, requires further empirical validation.

We also agree that avoiding additional decoder calls does not necessarily imply lower computational cost or latency. Future work will examine practical overhead more systematically and investigate when interventions help, have little effect, or disrupt generation.

We appreciate your constructive feedback, which will help us strengthen both the empirical evaluation and the practical motivation for the method.

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

In this paper, the authors aim to tackle the scenario where a reasoning trace is interrupted by users and a new context is provided. Inspired by the intuition that the partial answer provides potentially useful derivations and solution strategies that can guide the response to these changes, the authors proposed HiPE, a method that guides diffusion language models under updated conditions by injecting selected passages as revisable hints during generation. The authors provide a detailed heuristic protocol for which past tokens should be placed where. When implemented in DiffusionGemma, HiPE is claimed to achieve superior performance on reasoning benchmarks like AIME 2024 and AIME 2025 while requiring no additional training.

Critical strengths
  • The problem setting, while kind of trivial, is practically useful.
  • The algorithm is simple and does not require extensive computation overhead.
  • The authors are honest in terms of presenting their experimental results.
Critical weaknesses
  • In Table 8, the authors provide error bars for their results for both the benchmark performance improvement and NFE decrease. However, all the confidence intervals include 0, meaning that all their results are statistically insignificant. Yet, the authors claim to have consistent and sufficient improvement across math and code reasoning benchmarks.
  • The efficiency gain is dubious. The NFE decreases across benchmarks are minimal, with a maximum value of around six. Nevertheless, HiPE introduces computation overhead for the intervened decoding steps. Therefore, sampling wall-clock time instead of NFE should be a better proxy for sampling efficiency, and NFE is biased in favor of this HiPE algorithm.
  • Despite the minimal improvement in benchmark performance and efficiency, the authors use no valid set and directly tuned hyperparamters on the test sets. This makes me doubt whether this HiPE algorithm is introducing any real benefit or the so-called gains shown in the paper are just some random noise.
Questions for authors
  • Can the authors explain why their method's improvement always lies in the confidence interval?
  • Can the authors demonstrate how HiPE will perform when a separate validation set is used to tune the hyperparameters?
  • Can the authors test HiPE's performance on at least one more diffusion language model backbone?
1 comment

Thank you for identifying these concerns. Your feedback prompted us to revisit both the evaluation protocol and the strength of our conclusions.

We agree that confidence intervals that include zero do not establish a reliable improvement and that our original claims should be more carefully qualified.

To address the parameter-selection concern, we conducted follow-up experiments using an explicit MATH-500 validation split. We selected the margin and intervention window from 12 configurations using only the validation subset, then fixed the configuration for subsequent evaluations. The resulting accuracy differences relative to Restart were modest, and the problem-level bootstrap 95% confidence intervals included zero on both AIME datasets. We plan to build on this validation-based protocol with more comprehensive evaluations in a future submission.

We also agree that NFE alone cannot establish practical efficiency. A comprehensive wall-clock evaluation is needed to assess computational benefits reliably. Similarly, evaluations on additional backbones are necessary to establish generalizability beyond DiffusionGemma.

We sincerely appreciate your feedback, which has helped us identify the evidence needed for a more rigorous evaluation and more appropriately qualified conclusions.

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

Related papers

Discussion (0)

Sign in to comment.