acceptodds
Under review as a conference paper at ICLR 2027

Plan-and-Patch: Diffusion Language Models for Agentic Planning

Abstract

Planning is increasingly important for long-horizon agents, where successful execution requires coordinating subgoals, tool use, and intermediate outcomes over many steps. Yet assumptions made during planning may be invalidated by the environment, tools may return unexpected results, or actions may fail. Effective agents must therefore not only generate plans, but also revise them. Such revisions often affect only part of a plan, leaving the preceding and subsequent structure intact. Rather than regenerate the entire plan and risk unnecessary changes, repair can regenerate the affected region conditioned on the preserved prefix and suffix. We introduce Plan-and-Patch, a plan-and-act framework in which a diffusion language model (dLLM) generates a structured, program-like plan through parallel unmasking and repairs it by filling in selected regions while keeping the surrounding steps fixed. We compare DreamReasoner-8B and Qwen3-8B as diffusion and autoregressive (AR) planners. On Natural Plan without task-specific training, diffusion (53.7%) achieves nearly twice the plan repair success rate of AR (27.0%). After task-specific training on agentic benchmarks, ALFWorld and TextCraft, the planners achieve similar observed success in plan generation, while diffusion reduces mean plan-generation latency by 39–46% relative to AR. Our results show that Plan-and-Patch provides a framework for faster plan generation and effective plan repair in long-horizon agents.

Primary area:
generative models
Keywords:
Diffusion language models, agentic planning, plan-and-act, localized plan repair, bidirectional infilling, non-autoregressive generation

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_wwauOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper proposes Plan-and-Patch, an agentic planning framework based on diffusion LLMs. The method uses a dLLM to generate structured plans and performs localized plan repair through infilling when execution fails, modifying selected regions based on environment feedback while preserving the remaining plan. The authors also introduce a joint training approach that enables a single model to perform both plan generation and repair, using training data constructed from real execution failures and synthetic perturbations. Experiments are conducted on ALFWorld, TextCraft, and Natural Plan. The results show that diffusion and AR planners achieve similar plan generation success rates on ALFWorld and TextCraft, while diffusion reduces generation latency by 39% - 46%. On Natural Plan, without task specific training, the diffusion planner achieves substantially higher repair success than the AR planner. However, the AR planner still performs slightly better on plan repair in interactive environments.

Critical strengths
  1. The paper effectively leverages the bidirectional infilling capabilities of diffusion LLMs to unify plan generation and localized repair. The diffusion planner can modify selected regions using both preceding and following context while preserving unaffected parts of the plan. The authors also introduce a joint training approach that supports both tasks within a single model. This design naturally aligns the strengths of diffusion LLMs with the requirements of agentic planning.

  2. The method improved plan generation efficiency with competitive performance. The diffusion planner reduces mean plan generation latency by 39% on ALFWorld and 46% on TextCraft compared with the AR planner, while maintaining similar task success rates. The comparison uses the same task splits, training data, plan representations, and executor. These results demonstrate the practical value of diffusion LLMs for efficient long-horizon planning.

Critical weaknesses
  1. The necessity of dLLM is not sufficiently demonstrated. Traditional AR LLMs can also generate replacement content through prompting, given the prefix and suffix of a plan. Although the authors compare diffusion and AR planners, the AR planner achieves higher success rates in the history based repair experiments on both ALFWorld and TextCraft. Diffusion's clear advantage mainly appears in the non-interactive Natural Plan repair setting. The paper does not sufficiently demonstrate the unique benefits of dLLMs over traditional AR LLMs for practical agentic plan repair.

  2. Plan repair performance in interactive environments remains limited. In the history based repair setting, the diffusion planner achieves repair success rates of only 14.3% on ALFWorld and 24.5% on TextCraft. On the matched failure set, history based region selection also substantially reduces repair success compared with reference assisted selection. In addition, the interactive evaluation restarts repaired plans from the initial environment state rather than continuing from the point of failure. These limitations raise concerns about the reliability and practical efficiency of the proposed repair approach in long-horizon agent tasks.

Questions for authors
  1. Given that AR LLMs can also perform localized infilling, what unique advantages does diffusion provide for plan repair beyond faster plan generation?

  2. Diffusion significantly outperforms AR on Natural Plan repair but performs worse on history based repair in ALFWorld and TextCraft. Can the authors explain this discrepancy and clarify under what conditions diffusion provides an advantage over AR for plan repair?

3 comments

We thank the reviewer for the positive assessment and for recognizing the motivation behind combining parallel generation and bidirectional infilling.

1. Why diffusion? AR models can also perform localized infilling, as demonstrated by our jointly trained AR baseline. The distinction lies in how the two models perform this operation. Masked diffusion can jointly reconstruct selected positions conditioned on both preceding and following context, including multiple noncontiguous regions, whereas our AR baseline generates replacements sequentially. Recent work has demonstrated related capabilities in variable-length code infilling [1] and template-guided generation, including trip planning [2]. PLAN-AND-PATCH combines these capabilities with parallel plan generation and execution-feedback-driven repair in a single jointly trained model. As shown in Table 1 and Figure 5, diffusion achieves comparable generation success with 39–46% lower generation latency. Our objective is to exploit this efficiency advantage while maintaining competitive planning quality and supporting localized revision.

2. Different repair results. Natural Plan evaluates reconstruction of supplied regions in otherwise valid reference plans, with correctness assessed by constraint validators. In contrast, ALFWorld and TextCraft evaluate jointly trained planners repairing failed generated plans using execution feedback, where success additionally depends on identifying the appropriate region and executing the revised plan. Recent work has also identified challenges involving action consistency and structured outputs in diffusion-based agentic workflows [3].

On Natural Plan, diffusion achieves 53.7% versus 27.0% validator acceptance, with a particularly strong advantage on trip planning (93.3% versus 35.7%), where replacements must satisfy dependencies across different parts of the plan (Table A.10). In interactive environments, the jointly trained AR baseline achieves higher history-based repair success (Table 1). Our region-selection analysis further shows that history-based selection reduces repair success on ALFWorld but improves it on TextCraft (Table A.9). These findings demonstrate diffusion's effectiveness in supplied-region reconstruction while identifying the additional challenges of failure localization and interactive execution. Our repair evaluation uses re-execution from the initial state rather than continuation from the failure state.

References

[1] Wu et al. DreamOn: Diffusion Language Models for Code Infilling Beyond Fixed-size Canvas. ICLR 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/7c020dd308d7413b5078686723b2f857-Abstract-Conference.html

[2] Lee et al. Unlocking the Potential of Diffusion Language Models through Template Infilling. ACL 2026. https://aclanthology.org/2026.acl-long.284/

[3] Lu et al. The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check. ACL 2026. https://aclanthology.org/2026.acl-long.2036/

The author's response addressed my concern. I will keep my score

Dear Reviewer,

Thank you for your response and for acknowledging that our clarifications addressed your concerns. We sincerely appreciate your constructive feedback and positive consideration of our work.

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

This paper proposes Plan-and-Patch, a framework for agentic plan generation and localized repair using diffusion language models. The method first generates a structured plan and, upon execution failure, identifies the region requiring modification based on the execution history and environment feedback. It then leverages the bidirectional infilling capability of diffusion language models to regenerate the selected region while preserving the unaffected parts of the plan. The authors also introduce a joint training strategy for plan generation and repair. Experiments on ALFWorld, TextCraft, and Natural Plan show that diffusion models achieve comparable plan-generation success rates to autoregressive models while reducing generation latency and demonstrating better performance in certain controlled plan-repair settings.

Critical strengths
  1. The paper addresses a practically meaningful problem by focusing on localized plan repair after execution failures, providing an approach to avoiding unnecessary full-plan regeneration.
  2. Leveraging the bidirectional infilling capability of diffusion language models for localized plan repair is a natural design that preserves unaffected plan regions and supports variable-length replacements.
  3. The experimental evaluation covers plan generation, repair, and failure-region selection. The diffusion planner reduces mean plan-generation latency and achieves promising results in certain controlled repair settings.
Critical weaknesses
  1. The idea of preserving unaffected parts of a plan while repairing only the erroneous regions is conceptually similar to existing plan-repair and execution-feedback-driven replanning methods. Therefore, the main novelty appears to lie in applying diffusion-based infilling rather than introducing a fundamentally new plan-repair mechanism. The authors should further clarify the technical distinctions from prior work and the unique advantages of diffusion models in this setting.
  2. The most substantial improvements in plan-repair performance are obtained in settings where the repair regions are provided in advance. In the more realistic history-based region-selection setting, the diffusion planner achieves lower repair success rates than the autoregressive baseline, and its end-to-end task success rates show no clear advantage. This suggests that failure localization remains an important factor limiting overall repair performance. However, the proposed framework primarily relies on an existing LLM for failure localization, with limited technical advances in this component. The authors should further investigate whether diffusion models meaningfully improve the overall agentic failure-recovery process rather than merely performing infilling over predefined repair regions.
  3. The evidence for generalization and long-horizon planning capabilities remains limited. The 8B models are fine-tuned separately for ALFWorld and TextCraft, while the Natural Plan experiments use the original pretrained checkpoints, leaving the cross-environment generalization of the jointly trained models unverified. Furthermore, the diffusion planner's structured-plan parser acceptance rate drops substantially on ScienceWorld, which involves longer plans. The authors should provide stronger cross-environment transfer and long-horizon evaluations to substantiate the claimed generality and applicability to long-horizon agentic tasks.
Questions for authors

Please refer to the weaknesses above.

2 comments

We thank the reviewer for the constructive assessment and for recognizing the motivation and potential of our framework.

1. Novelty. We agree that localized plan revision has important precedents, but would like to distinguish our contribution from these methods. AdaPlanner [1] supports in-plan refinement, PIVOT [2] preserves validated prefixes while regenerating affected suffixes, and Atomic Task Graph [3] repairs affected task-graph regions. Our contribution addresses a different question: how to jointly exploit diffusion's parallel generation and masked infilling capabilities within an agentic planning framework.

This distinction is particularly relevant given the emerging literature on diffusion-based agents. Lu et al. [4] identify difficulties with action repetition and structured tool calls in existing diffusion agents. Our work studies a complementary plan-and-act formulation, separating plan generation from execution and enabling the planner to revise selected regions using execution feedback.

Our contribution combines addressable plan regions, repair supervision from execution failures and structured perturbations, and joint training with objective-specific block-size sampling. Joint training is important because sequential repair fine-tuning substantially degraded generation performance in our preliminary experiments (Section 4.3). As shown in Table 1 and Figure 5, the resulting planner achieves comparable generation success to AR with 39–46% lower generation latency. Thus, our contribution goes beyond applying an existing repair method, demonstrating how diffusion can support efficient generation and feedback-driven revision within a single agentic planner.

2. Failure localization. Our framework separates region selection from infilling to evaluate repair under controlled conditions, with both planners receiving identical regions and feedback. We additionally evaluate history-based repair without reference plans (Table 1), using execution history and environment feedback to identify regions for revision. Table A.9 further examines the effect of region selection, showing reduced repair success on ALFWorld but improved success on TextCraft. These experiments distinguish the effects of failure localization from those of the repair models, while identifying factors that limit interactive recovery.

3. Generalization and longer horizons. Our evaluation spans different planning environments to investigate the framework's effectiveness, rather than to establish cross-environment transfer. ALFWorld and TextCraft demonstrate the generation-efficiency result, Natural Plan separately evaluates infilling without task-specific training, and ScienceWorld examines the challenges associated with longer plans. We will clarify the scope of these evaluations.

References

[1] Sun et al. AdaPlanner: Adaptive Planning from Feedback with Language Models. NeurIPS 2023. https://proceedings.neurips.cc/paper_files/paper/2023/hash/b5c8c1c117618267944b2617add0a766-Abstract.html

[2] Zhang et al. PIVOT: Bridging Planning and Execution in LLM Agents via Trajectory Refinement. arXiv:2605.11225, 2026. https://arxiv.org/abs/2605.11225

[3] Zhang et al. Atomic Task Graph: A Unified Framework for Agentic Planning and Execution. arXiv:2607.01942, 2026. https://arxiv.org/abs/2607.01942

[4] Lu et al. The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check. ACL 2026. https://aclanthology.org/2026.acl-long.2036/

Dear Reviewer,

Thank you for your feedback. With the discussion period approaching its end, we hope our rebuttal has addressed your concerns and would be grateful for an opportunity to provide any further clarifications. If your concerns have been resolved, we would appreciate your reconsideration of the score.

Thank you for your time and consideration.

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

This paper introduces PLAN-AND-PATCH, a plan-and-act framework that uses a single masked diffusion language model for complete-plan generation via parallel unmasking and localized plan repair via infilling conditioned on a preserved prefix, suffix, and execution feedback. To support this, the authors design a hierarchical delimited plan representation, build repair data from execution failures and synthetic perturbations, and jointly train the planner with task-specific block-size sampling. Evaluations comparing DreamReasoner and Qwen3 show that on fine-tuned ALFWorld and TextCraft tasks, both planners reach similar task success while diffusion reduces generation latency by 39–46%, whereas on zero-shot Natural Plan, the autoregressive planner leads in generation and the diffusion planner nearly doubles repair success.

Critical strengths
  1. Controlled study of joint diffusion training for plan generation and infilling repair:

The paper presents a well-engineered joint training recipe with task-specific block-size curricula that prevents the format degradation caused by sequential fine-tuning, achieving 100% parser compliance on ALFWorld and TextCraft and reducing initial plan-generation latency by 39–46% relative to the autoregressive baseline at comparable generation success.

  1. Comprehensive diagnostic and sensitivity analyses:

The authors disentangle oracle reference-assisted region selection from execution-history detection and provide thorough ablations across executors, plan representations, failure locations, and task horizons, offering concrete empirical insights into the behavior and bottlenecks of discrete diffusion planners.

Critical weaknesses
  1. Unsubstantiated advantage of diffusion infilling for plan repair and missing replanning baselines:
  • On both fine-tuned interactive benchmarks (ALFWorld and TextCraft), the autoregressive baseline (AR-PLAN) matches or outperforms the diffusion planner (DIFF-PLAN) in initial generation (45.5% vs. 44.0%; 66.0% vs. 63.0%), realistic history-based repair (17.5% vs. 14.3%; 30.6% vs. 24.5%), and end-to-end task success (60.6% vs. 59.9%; 72.2% vs. 68.7%).
  • For repair latency, DIFF-PLAN is slower than AR-PLAN on ALFWorld (9.5s vs. 7.6s) because full-canvas diffusion infilling denoises across the entire sequence without key-value caching, whereas autoregressive fill-in-the-middle decodes only the short replacement span.
  • The headline repair advantage on Natural Plan (53.7% vs. 27.0%) stems from an artificial zero-shot setup that blanks spans of ground-truth reference plans with oracle replacement-length hints, where off-the-shelf Qwen3-8B lacks fill-in-the-middle fine-tuning and fails primarily by repeating suffix lines in trip planning (while AR-PLAN outperforms DIFF-PLAN on meeting scheduling, 19.2% vs. 18.1%).
  • Although the paper motivates bidirectional prefix-and-suffix infilling over global revision, it omits baselines that regenerate the full plan or only the post-failure suffix, despite the fact that repaired plans are re-executed from the initial environment state rather than resumed from the failure point.
  1. Confounded executor evaluation, counterproductive delimited plan format, and long-horizon collapse:
  • Reference plans are verified using Claude Sonnet 5, whereas primary evaluation uses an unfinetuned Qwen3-8B executor. In ALFWorld, 103 of 106 identified failing leaf steps are identical to the verified reference plan, and swapping the executor to Claude Sonnet 5 on the exact same generated plans raises success from 44.0%–45.5% to 94.8%–96.3% on ALFWorld and from 63.0%–66.0% to 94.0% on TextCraft, showing that most observed failures are executor grounding errors rather than planning errors.
  • Under the same Qwen3-8B executor, replacing the proposed delimited format with numbered prose steps boosts single-shot generation to 61.9% (DIFF-PLAN) and 67.9% (AR-PLAN) on ALFWorld, and 69.0% and 78.0% on TextCraft, surpassing the entire end-to-end delimited generation-plus-repair pipeline (59.9%–60.6% and 68.7%–72.2%), while AR-PLAN outperforms DIFF-PLAN by 2.6x–6.0x on prose repair (55.1% vs. 21.3% and 30.0% vs. 5.0%).
  • While the paper targets long-horizon agents, ALFWorld and TextCraft reference plans average only 5.8 and 9.8 steps; on longer ScienceWorld plans (averaging 27.1 steps), DIFF-PLAN parser acceptance drops to 34.3% overall and 7–9% at 28 or more steps due to severe action repetition (93.3% median duplicated actions vs. 40.0% for AR-PLAN).
Questions for authors
  1. Comparison with suffix-only and full-plan regeneration:

Since repaired plans are re-executed from the initial state, how does prefix-and-suffix infilling compare in task success and latency against (a) preserving the validated prefix and regenerating the entire remaining suffix, and (b) regenerating the full plan from scratch given execution feedback?

  1. Delimited vs. prose representation and executor mismatch:

Given that single-shot prose generation (61.9%–67.9% on ALFWorld; 69.0%–78.0% on TextCraft) outperforms end-to-end delimited generation plus repair (59.9%–60.6%; 68.7%–72.2%) under the Qwen3-8B executor, and that 103/106 failing ALFWorld steps match the reference plan, why was the delimited format with an unfinetuned Qwen3-8B executor chosen as the primary pipeline instead of end-to-end prose plans or a fine-tuned executor?

  1. End-to-end episode latency:

Reported latency excludes environment execution and shows that diffusion repair is slower than autoregressive repair on ALFWorld (9.5s vs. 7.6s). What is the total wall-clock latency per episode when including step-by-step executor inference, failure detection, and full re-execution from the initial state after repair?

  1. Fairness of the Natural Plan repair baseline:

On Natural Plan, Qwen3-8B is prompted zero-shot in chat format without fill-in-the-middle training, whereas DreamReasoner-8B uses native masked infilling with oracle length hints. How does an autoregressive baseline perform on Natural Plan repair when using native fill-in-the-middle tokens or task-agnostic fill-in-the-middle fine-tuning?

3 comments

We thank the reviewer for the detailed assessment and for recognizing our joint training approach, efficiency improvements, and diagnostic analyses.

1. Motivation and repair baselines. Our objective is to exploit diffusion's computational advantages while maintaining competitive planning quality, rather than claim uniformly superior accuracy. As shown in Table 1 and Figure 5, diffusion achieves comparable generation success with 39–46% lower generation latency. Localized repair complements this advantage by exploiting diffusion's masked-infilling capabilities. Suffix regeneration and full replanning are meaningful alternatives, but involve different repair scopes. Our evaluation focuses on localized infilling, holding failed plans, selected regions, and feedback fixed across models to compare their repair performance directly. We do not claim that localized infilling necessarily outperforms these alternative strategies, and will clarify this distinction.

2. Representation and executor. Structured, program-like plans and the separation of planning from execution have precedents in prior work [1, 2]. Our delimited representation serves a specific purpose within this framework, providing addressable steps and subgoals that allow selected regions to be repaired while preserving the surrounding plan. We use a common Qwen3-8B executor across planners to enable a controlled comparison. Prior work has also demonstrated that plan representation affects agent performance [3]. Consistent with this, our analyses in Table A.7 show that prose plans improve task success with Qwen3-8B, while switching to Sonnet 5 substantially improves execution success on the same generated plans. These experiments characterize how representation and executor choice affect the overall framework.

3. Latency. Our reported speedup concerns plan generation, not total episode latency. Repair latency varies by environment (Figure 5), and total episode latency additionally depends on executor inference, failure detection, and re-execution. Our repair evaluation restarts execution from the initial state rather than reusing completed execution. We will make these distinctions explicit.

4. Natural Plan. We thank the reviewer for suggesting stronger AR infilling baselines. We first clarify that both models receive the same prefix, suffix, repair feedback, and approximate replacement length, as well as the same five benchmark demonstrations (Section A.3.3). This experiment evaluates the original checkpoints without task-specific fine-tuning and is intended as a training-free evaluation of plan generation and repair performance, while our interactive experiments additionally compare jointly trained diffusion and AR planners, with FIM training for AR. On Natural Plan, diffusion achieves substantially higher repair success (53.7% versus 27.0%), particularly on trip planning (93.3% versus 35.7%), and outperforms AR across all five repair constructions (Table A.10). These results demonstrate diffusion's effectiveness at reconstructing plan regions under surrounding constraints, particularly when repairs involve dependencies across different parts of the plan.

5. Longer horizons. We agree that ScienceWorld identifies challenges with longer plans. We include this analysis to characterize current limitations, rather than claim that reliable generation at substantially longer horizons has been established.

References

[1] Singh et al. ProgPrompt: Generating Situated Robot Task Plans using Large Language Models. ICRA 2023. https://arxiv.org/abs/2209.11302

[2] Erdogan et al. Plan-and-Act: Improving Planning of Agents for Long-Horizon Tasks. ICML 2025. https://proceedings.mlr.press/v267/erdogan25a.html

[3] Zambrano et al. Does The Way You Plan Matter? An Empirical Study of Planning Representations for LLM Web Agents. REALM 2026. https://aclanthology.org/2026.realm-1.96/

Dear Reviewer,

Thank you for your feedback. With the discussion period approaching its end, we hope our rebuttal has addressed your concerns and would be grateful for an opportunity to provide any further clarifications. If your concerns have been resolved, we would appreciate your reconsideration of the score.

Thank you for your time and consideration.

Thanks for the response. However, since the rebuttal mostly narrows the claims rather than providing the requested baselines or latency numbers, my main concerns remain:

  1. Without suffix-only or full-plan regeneration baselines, it is still hard to see the practical benefit of prefix-and-suffix infilling. Since repaired plans are re-executed from the initial state anyway, regenerating the post-failure suffix or the full plan is a natural alternative that avoids the region-selection drop in Table A.9, especially when AR-PLAN already matches or beats DIFF-PLAN end-to-end.

  2. On the plan format and executor setup, the numbered prose format in Section A.3.1 already has addressable steps (and is used for region-selected repair in Table A.8). Yet single-shot prose generation beats the full delimited generation-plus-repair pipeline under Qwen3-8B, and AR-PLAN is much stronger than DIFF-PLAN on prose repair (55.1% vs. 21.3% on ALFWorld; 30.0% vs. 5.0% on TextCraft). Also, given that 103/106 failing ALFWorld steps match the reference plan and Sonnet 5 gets 94-96% single-shot success on the exact same plans, most failures in Table 1 still look like executor grounding issues with the delimited format rather than planning errors.

  3. For latency, clarifying in text that the 39-46% speedup only applies to initial plan generation does not really resolve the question. Around 37-56% of episodes fail on the first try and go through step-by-step execution, failure detection, repair (where DIFF-PLAN is slower on ALFWorld, 9.5s vs. 7.6s), and full re-execution from scratch. Without total episode wall-clock time, it is unclear how much speedup remains in practice.

  4. On Natural Plan, prompting chat-tuned Qwen3-8B without native FIM tokens or FIM adaptation makes it hard to separate diffusion vs. AR capability from zero-shot formatting issues (such as the suffix repetition in Figure A.5). In ALFWorld and TextCraft, where AR-PLAN is trained with FIM, diffusion shows no repair advantage. Together with the drop on longer ScienceWorld plans (7-9% parser acceptance at 28+ steps), I am still not convinced of the practical advantage of diffusion here.

For these reasons, I will keep my current rating.

Comments on the paper

2 comments
AuthorsOpenReview ↗

General Comment

We thank the reviewers for their thoughtful evaluations and constructive feedback. We appreciate their recognition of our joint training framework, generation-efficiency improvements, and experimental analyses.

Our objective is not to establish that diffusion language models universally outperform autoregressive (AR) models in planning quality, but to exploit their computational advantages while maintaining competitive task performance. Recent work has identified important tradeoffs between parallel decoding efficiency and generation quality [1–3], while applying diffusion models to agentic workflows introduces further challenges involving action consistency and structured outputs [4].

PLAN-AND-PATCH investigates a plan-and-act framework combining parallel plan generation with execution-feedback-driven repair through masked infilling. Our main contributions are:

Faster generation at comparable quality. On ALFWorld and TextCraft, diffusion achieves comparable plan-generation success to AR with 39–46% lower generation latency (Table 1, Figure 5).

Joint generation and repair. We develop a training procedure combining execution-derived and synthetic repair examples with objective-specific block-size sampling, supporting both operations in a single model (Sections 4.2–4.3).

Repair capabilities and limitations. Diffusion achieves substantially higher supplied-region reconstruction success on Natural Plan (53.7% versus 27.0%), while our interactive experiments examine the effects of failure localization, plan representation, and executor choice.

While localized plan revision has precedents, our contribution lies in jointly adapting diffusion's parallel generation and masked infilling capabilities to agentic planning. We address the specific concerns below and will clarify the motivation and scope of our findings in the revised manuscript.

References

[1] Feng et al. Theoretical Benefit and Limitation of Diffusion Language Model. NeurIPS 2025. https://proceedings.neurips.cc/paper_files/paper/2025/hash/2318d75a06437eaa257737a5cf3ab83c-Abstract-Conference.html

[2] Kang et al. ParallelBench: Understanding the Trade-offs of Parallel Decoding in Diffusion LLMs. ICLR 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/9156b0f6dfa9bbd18c79cc459ef5d61c-Abstract-Conference.html

[3] Israel et al. Planned Diffusion. ICLR 2026. https://proceedings.iclr.cc/paper_files/paper/2026/hash/9fbae6362539ef7a151303c86cb51c00-Abstract-Conference.html

[4] Lu et al. The Bitter Lesson of Diffusion Language Models for Agentic Workflows: A Comprehensive Reality Check. ACL 2026. https://aclanthology.org/2026.acl-long.2036/

The author-reviewer discussion period will end soon. Please review and acknowledge the authors’ responses. If your concerns have been addressed, please reconsider your score.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.