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

Syamantak Kumar; Jiang Guo; Hassan Hamad; Hideo Kobayashi; Yi Xiang; Yezhou Yang; Yanjun Qi; Daniele Bonadiman

cs.AIOct 7, 2026

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

Authors: Syamantak Kumar, Jiang Guo, Hassan Hamad, Hideo Kobayashi, Yi Xiang, Yezhou Yang, Yanjun Qi, Daniele Bonadiman, +1 more

Organizations: University of Texas at Austin · Amazon

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.

Figures & tables

Panel of Figure 1 — Figure 1 : The Plan-and-Patch framework. The same planner generates plans and infills selected regions. After an execution failure, the executor acts as the failure detector, selecting a region using execution history and environment feedback. The planner infills it conditioned on repair feedback, with the prefix and suffix fixed. Figure A.2 gives a worked example.

Panel of Figure 2 — Figure 2 : A TextCraft plan and its ordered hierarchy. All five executable steps are shown; reasoning is elided. Arrows show containment; execution visits leaves in depth-first, left-to-right order.

Panel of Figure 3 — Figure 3 : Reference-plan construction. ALFWorld groups navigation–manipulation pairs; TextCraft groups ingredient preparation before the final craft. Numbers map actions to subgoals; leaf boxes show serialized fields. Both domains share the bottom pipeline. Actions and observations are abbreviated.

Panel of Figure 4 — Figure 4 : Joint training with task-specific block sizes. The initial planner supplies failures for data construction; the final planner starts from the original checkpoint and learns generation and repair jointly. We adapt DreamReasoner’s block-size curriculum ( Wu et al., 2026c ) with separate sampling distributions for generation and repair. The lower panel shows weights normalized within each objective; dashes denote unsampled sizes. Inference uses the separate settings in Table A.11 .

Environment	Model	Plan generation	Reference-assisted repair	History-based repair	End-to-end $\dagger$
ALFWorld	ReAct	$38.1\%$	$-$	$-$	$-$
	Diff-Plan	$44.0\%$	$28.4\%$	$14.3\%$	$59.9\%$
	AR-Plan	$45.5\%$	$27.7\%$	$17.5\%$	$60.6\%$
TextCraft	ReAct	$48.0\%$	$-$	$-$	$-$
	Diff-Plan	$63.0\%$	$15.5\%$	$24.5\%$	$68.7\%$
	AR-Plan	$66.0\%$	$18.3\%$	$30.6\%$	$72.2\%$

Table 1 : Task success under the Qwen3-8B executor (%). Generation:

134/100

tasks; reference-assisted/history-based repair:

148/71

and

126/49

failures (ALFWorld/TextCraft), with one attempt and restart.

\dagger

End-to-end provides generation-plus-repair success including recovery on each planner’s own failures using the history-based repair. Dashes denote no repair stage.

Panel of Figure 5 — Figure 5 : Diffusion reduces mean plan-generation latency by $39$ – $46\%$ . Values are synchronized, unbatched calls on one A100 in bf16; prompt construction, parsing, and environment execution are excluded. Repair latency varies by environment.

Task	Model	Validator	GPT-5	Sonnet 5
Generation	Diff-Plan	$33.0\%$	$20.2\%$	$31.8\%$
Generation	AR-Plan	$48.6\%$	$34.0\%$	$46.1\%$
Repair	Diff-Plan	$53.7\%$	$72.9\%$	$81.6\%$
Repair	AR-Plan	$27.0\%$	$42.1\%$	$53.4\%$

Table 2 : Natural Plan without task-specific training. Acceptance rates (%) under the benchmark validator and two judges. Generation uses

1{,}000

tasks; repair infills reference-plan regions in

2{,}661

examples from

696

tasks.

Environment (tasks)	Both succeed	Both fail	Only diffusion fails	Only AR fails
ALFWorld ( $134$ )	$36$	$50$	$25$	$23$
TextCraft ( $100$ )	$50$	$21$	$16$	$13$

Table 3 : Pairwise error analysis of generated plans. Counts partition the test tasks. In each “only” column, the other planner succeeds.

Appendix figures & tables17 assets

Supplementary material from the paper’s appendix.

Appendix

	Task split				Repair-training examples
Environment	train-gen	train-rep	dev	test	real	synthetic	total
ALFWorld	$2{,}665$	$888$	$140$	$134$	$699$	$763$	$1{,}462$
TextCraft	$3{,}824$	$1{,}268$	$50$	$100$	$267$	$1{,}661$	$1{,}928$

Appendix

Table A.1 : Task splits and repair-training examples. Task counts precede reference-plan execution filtering. Repair examples combine failed planner executions and structured perturbations of verified plans.

Environment	Split	Candidates	Success	Mean	Median	Max
ALFWorld	train	$3{,}553$	$91.0\%$	$5.8$	$6$	$8$
	dev	$140$	$90.7\%$	$5.7$	$6$	$8$
	test	$134$	$91.0\%$	$5.6$	$6$	$8$
	all	$3{,}827$	$91.0\%$	$5.8$	$6$	$8$
TextCraft	train	$5{,}092$	$99.8\%$	$9.8$	$7.0$	$66$
	dev	$50$	$100\%$	$9.9$	$7.0$	$91$

Appendix

Table A.2 : Reference-plan candidates. Train is the union of train-gen and train-rep . Counts precede execution filtering; verified train-gen candidates supply generation targets, while verified train-rep references provide repair targets for failed plans. Mean, median, and maximum refer to the number of leaf steps; success is measured by execution.

Panel of Figure 11 — Figure A.1 : Observed success in the two interactive benchmarks. (a) Task success after plan generation, on $134$ ALFWorld and $100$ TextCraft tasks. (b) Task success after reference-assisted repair, on $148$ and $71$ pooled failed plans. Diffusion–AR differences are at most three percentage points; higher is better.

	Plan generation				Reference-assisted repair
Environment	both	Diff-Plan only	AR-Plan only	neither	both	Diff-Plan only	AR-Plan only	neither
ALFWorld	$36$	$23$	$25$	$50$	$24$	$18$	$17$	$89$
TextCraft	$50$	$13$	$16$	$21$	$6$	$5$	$7$	$53$

Appendix

Table A.3 : Paired outcomes. Generation compares the planners on the same test tasks. Reference-assisted repair compares them on pooled failed plans, which may include two plans from one task. Columns indicate which planners succeed.

Task family	Tasks	Both fail	Only diffusion fails	Only AR fails
Light inspection	$18$	$12$	$3$	$0$
Simple placement	$24$	$3$	$2$	$5$
Clean then place	$31$	$12$	$3$	$7$
Cool then place	$21$	$6$	$9$	$2$
Heat then place	$23$	$7$	$4$	$6$
Two-object placement	$17$	$10$	$4$	$3$

Appendix

Table A.4 : Paired ALFWorld generation failures by task family. Each task contributes to at most one failure column; tasks where both planners succeed are omitted from those columns but included in the total.

Recorded stopping reason	Diffusion	AR	Diffusion	AR
	ALFWorld		TextCraft
Failure at a leaf step	$10$	$13$	$10$	$7$
No-progress limit	$2$	$2$	$2$	$3$
Plan completed without terminal success	$4$	$0$	$3$	$1$
Invalid-action limit	$9$	$8$	$1$	$2$
Total	$25$	$23$	$16$	$13$

Appendix

Table A.5 : Stopping reasons on tasks failed by only one planner. Columns identify the failing planner; the other planner succeeds on each counted task. Counts cover every model-exclusive failure.

Action at the identified failing leaf step
Diagnostic	ALFWorld	TextCraft
All failed executions	$148$	$71$
Identified failing leaf step	$106$	$14$
Unidentified failing leaf step	$42$	$57$
take	$89$	$0$
fetch	$0$	$12$

Appendix

Table A.6 : Where failed executions stop. Action and reference-comparison counts cover failures with an identified leaf step and describe the observed failure location.

Panel of Figure 16 — Figure A.2 : Plan repair in ALFWorld. In this training example, a reference-assisted judge selects one leaf step, the reasoning block, and three intent fields for regeneration. Infilling changes countertop to toilet and updates the expected observation. The other seven leaf steps remain fixed. The repaired plan succeeds.

Panel of Figure 17 — Figure A.3 : Interaction before failure and preservation after repair. We analyze $104$ ALFWorld and $38$ TextCraft failed executions with at least one completed step. Plans come from both planners and use the Qwen3-8B executor. (a) Pooled fraction of executor tokens consumed before the recorded failure point. (b) Fraction of plans after reference-assisted repair that parse and preserve every completed step, shown separately for each planner. Preservation compares completed-step text before and after repair; repaired plans are executed from the initial state.

		Qwen3-8B executor		Sonnet 5 executor
Environment	Planner	Delimited	Prose	Delimited	Prose
ALFWorld	Diff-Plan	$44.0\%$	$61.9\%$	$96.3\%$	$95.5\%$
ALFWorld	AR-Plan	$45.5\%$	$67.9\%$	$94.8\%$	$94.8\%$
TextCraft	Diff-Plan	$63.0\%$	$69.0\%$	$94.0\%$	$94.0\%$
TextCraft	AR-Plan	$66.0\%$	$78.0\%$	$94.0\%$	$96.0\%$

Appendix

Table A.7 : Generation under two representations and executors. Entries are task success rates. Within each representation, both executors receive identical stored plan text.

Environment	Executor	Failures	Eligible	Diff-Plan repair	AR-Plan repair
ALFWorld	Qwen3-8B	$51+43=94$	$46+43=89$	$19/89\;(21.3\%)$	$49/89\;(55.1\%)$
TextCraft	Qwen3-8B	$31+22=53$	$14+6=20$	$1/20\;(5.0\%)$	$6/20\;(30.0\%)$
ALFWorld	Sonnet 5	$6+7=13$	$13$	$4/13\;(30.8\%)$	$6/13\;(46.2\%)$
TextCraft	Sonnet 5	$6+4=10$	$5$	$2/5\;(40.0\%)$	$1/5\;(20.0\%)$

Appendix

Table A.8 : Reference-assisted repair of prose plans. “Failures” gives counts from Diff-Plan and AR-Plan , in that order; “Eligible” counts plans with a selected repair region. Both planners receive the same failed plans and regions. Repair entries report task success after repair and restart.

	Diff-Plan		AR-Plan
Environment	Reference-assisted	History-based	Reference-assisted	History-based
ALFWorld ( $n=126$ )	$31.7\%$	$14.3\%$	$32.5\%$	$17.5\%$
TextCraft ( $n=49$ )	$22.4\%$	$24.5\%$	$22.4\%$	$30.6\%$
Pooled ( $n=175$ )	$29.1\%$	$17.1\%$	$29.7\%$	$21.1\%$

Appendix

Table A.9 : Repair under reference-assisted and history-based region selection. Entries are task success rates after repair. Paired columns use the same failed plans and repair feedback; abstentions remain in the history-based denominator.

Panel of Figure 21 — Figure A.4 : Plan generation and repair on Natural Plan. Both models are evaluated without task-specific fine-tuning. Generation uses $1{,}000$ tasks; repair infills supplied regions in $2{,}661$ examples from $696$ tasks. Bars show acceptance rates under the benchmark validator and two language-model judges. Section A.3.3 gives the filtering protocol and complete results.

Domain	$n$	Diff-Plan	AR-Plan
Trip planning	$1{,}261$	$93.3\%$	$35.7\%$
Meeting scheduling	$1{,}400$	$18.1\%$	$19.2\%$
Calendar scheduling	$0$	–	–

Appendix

Table A.10 : Natural Plan repair by domain and construction. Entries are validator acceptance rates on the

2{,}661

retained examples. Calendar references contain only a few words specifying a meeting day and time; no calendar candidates pass the repair filters, so this domain is evaluated only for generation.

Panel of Figure 23 — Figure A.5 : An example of plan repair on Natural Plan. Both planners receive the same task, repair feedback, preserved prefix and suffix, and approximate replacement length. The blank is the prefix of a line, so its replacement must join the preserved suffix. The diffusion output restores the reference phrase; the AR output repeats the following line. This example illustrates the within-task comparison in Figure A.4 .

Panel of Figure 24 — Figure A.6 : ScienceWorld parser acceptance by reference-plan length. Rates are computed over $825$ held-out generations per planner. Acceptance for Diff-Plan falls from $74.7\%$ in the shortest bin to $7$ – $9\%$ at $28$ or more reference steps, with pooled parser acceptance of $34.3\%$ for Diff-Plan and $73.8\%$ for AR-Plan .

Setting	Plan generation	Plan repair
Training endpoint	Convergence of development-set performance
Learning rate	$10^{-5}$	$10^{-5}$
Batch $\times$ gradient accumulation	$1\times 16$	$1\times 16$
Learning-rate warmup	$5\%$ of optimization steps
Maximum sequence length	$4{,}096$ tokens at training and inference
Diff-Plan block sizes	$4,8,16,32,64,$ $128,256,512,1024$	$32,64,128,256,2048,4096$

Appendix

Table A.11 : ALFWorld and TextCraft training and decoding settings. Block weights follow the listed block-size order.

Explore similar work

Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents
Jun 3, 2026Haoyu Sun, Wenxuan Wang, Mingyang Song +5Tool-Augmented Language Model Agents LLM Agent Evaluation
Tongji University · Shanghai AI Laboratory · Harbin Institute of Technology +5
Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution
Sep 29, 2026Subba Reddy Oota, Francisco Herrera, Jordi Cabot Sagrera +2LLM Agents Long-Horizon LLM Agents
ADIA Lab, Abu Dhabi, United Arab Emirates · University of Granada, Granada, Spain · Luxembourg Institute of Science and Technology, Luxembourg +2
GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI
Sep 24, 2026Arunabh Srivastava, Mohammad A., Khojastepour +2LLM Planning
Amir · University of Maryland, College Park, MD · NEC Laboratories America, Inc.

Jun 3, 2026, cs.CLJ/K move · Enter open · S save

Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents

Haoyu Sun, Wenxuan Wang, Mingyang Song, Jujie He+4

Tongji University · Shanghai AI Laboratory · Harbin Institute of Technology+5

Tool-Augmented Language Model Agents LLM Agent Evaluation AI Agent Benchmarks LLM Agent Planning LLM Agent Reliability

Open paper AskPDF arXiv

Planning is central to LLM agents: before acting, an agent must decompose goals, select tools, reason over constraints, and decide when a task is infeasible. Yet existing agent evaluations often report only end-to-end success, making it difficult to determine whether failures stem from planning or execution. We introduce Agent Planning Benchmark (APB), a planning-specific diagnostic benchmark with 4,209 multimodal cases across 22 domains and five settings, covering holistic planning, feedback-conditioned step-wise planning, and robustness under extraneous tools, broken tools, and unsolvable tasks. Across 12 MLLMs, APB reveals systematic weaknesses in long-horizon planning, tool-noise robustness, calibrated refusal, and inference-time refinement. We further validate APB on 200 ToolSandbox tasks and 200

τ^2

-bench tasks, where APB-guided refinement consistently improves plan correctness, plan grade, and downstream execution metrics across three representative models. APB thus serves as an upstream diagnostic complement to execution benchmarks. The APB benchmark and code are available in this URL.

Abstract

Explore similar work

Agent Planning Benchmark: A Diagnostic Framework for Planning Capabilities in LLM Agents

Do LLM Agents Execute the Plans They Declare? From Planning-Mode Declaration to Pattern-Specific Execution

GRASP: Generating, Revising, and Assessing for Strategic Planning with Agentic AI