acceptodds
Under review as a conference paper at ICLR 2027

MetaDecomp: Adaptive Task Decomposition via Meta-Learned Procedural Memory

11 views

Abstract

Reusing past experience in long-horizon tasks requires more than storing successful procedures: Language agents must distinguish which procedures apply and what work remains as execution progresses. We introduce MetaDecomp, an in-context meta-learning framework for adaptive task decomposition that jointly learns a task-relative state space and reusable procedures. Its outer loop builds procedural memory from successful training executions, representing task progress as abstract states and procedures between states as directed acyclic graphs of operations and dependencies. As new progress distinctions are learned, MetaDecomp revisits historical execution evidence to reassess state assignments and reorganize the procedures they support. Its inner loop uses this memory to refine an initial decomposition and revise it using environmental feedback, combining applicable procedure fragments with the work required by the current task. Evaluated on ALFWorld and ScienceWorld over three seeds, MetaDecomp achieves the highest mean performance among the compared methods in four of five settings, for example, an 86.3% success rate on ALFWorld with Qwen3-4B. These results are achieved with competitive inference costs across all five settings. Ablation results show that learned cross-task procedural knowledge improves mean performance over within-task revision alone in every setting. Code will be released.

Primary area:
transfer learning, meta learning, and lifelong learning
Keywords:
Agent, Memory, Planning, Decomposition

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

MetaDecomp learns a task-relevant state schema and organizes successful training traces into procedural memory. As the schema evolves, historical observations are reprojected and procedures are regrouped into a dependency graph. At inference time, the agent uses the learned state abstraction and procedures to revise task decompositions, while the schema and memory remain frozen. The distinguishing idea is to learn how experience should be represented and organized together.

Critical strengths
  1. Jointly evolving the state schema and procedural organization is a plausible advance over workflow retrieval methods such as Agent Workflow Memory. Reprojecting old experience when the abstraction changes addresses a real problem: a memory can become poorly indexed as the agent learns which state distinctions matter.
  2. The training/test boundary is clearly described. The stronger outer-loop teacher is not used during test execution, and the learned memory is frozen. This makes the reported inference gains easier to interpret.
Critical weaknesses
  1. The main ablation compares the full system with an inner loop that has empty procedural memory. This shows that learned memory helps, but does not establish the contribution of adaptive schemas, historical reprojection, or DAG dependencies individually. These are the parts that carry the novelty claim.
  2. The cost discussion primarily concerns execution. Outer-loop learning and initial plan refinement are excluded from the execution budget, while ALFWorld can use more execution tokens than MPO. It is difficult to judge the total cost advantage without an amortized accounting.
Questions for authors

Please refer to critical weakness.

5 comments
AuthorsOpenReview ↗

Response to Reviewer ZaCt (1/2)

We sincerely appreciate your thoughtful review and your recognition of the value of jointly evolving state representations and procedural knowledge. Your suggestions have helped us strengthen the evidence for individual learning mechanisms and provide a more complete comparison of training and inference costs. We address both concerns point by point below. Manuscript section, appendix, figure, and table numbers refer to the revised paper; changes are highlighted in blue. Please let us know if you have any remaining concerns or would like further clarification; we would be happy to discuss them.

(W1) The main ablation compares the full system with an inner loop that has empty procedural memory. This shows that learned memory helps, but does not establish the contribution of adaptive schemas, historical reprojection, or DAG dependencies individually. These are the parts that carry the novelty claim.

Thank you for recommending these targeted controls; beyond the original comparison with empty procedural memory, Table R3.1 evaluates the contribution of adaptive schemas, historical reprojection, or DAG dependencies individually and confirms that all three mechanisms carrying the novelty claim contribute substantially across both benchmarks.

  • (Adaptive schemas) Allowing adaptive schemas to evolve during training substantially improves performance over a fixed state space.

    • Disabling adaptive schemas by retraining under Fixed initial schema holds the initial state-space definition constant while keeping initial refinement and online revision, reducing performance on both ALFWorld and ScienceWorld in Table R3.1.
  • (Historical reprojection) Performing historical reprojection when the schema updates is essential for keeping stored procedures aligned with evolved states.

    • Disabling historical reprojection by retraining under No historical reprojection permits schema expansion and retains initial refinement, but omits reassessing earlier successful executions under updated schemas, degrading performance on both benchmarks in Table R3.1.
  • (DAG dependencies and joint state–procedure use) State-indexed DAG dependencies and the joint use of states, procedures, and initial refinement in learned memory each contribute directly at inference time.

    • Removing state–procedure links while keeping each transition graph's internal DAG dependencies intact, or restricting learned memory at inference to State only, Procedure only, or No initial refinement, consistently lowers performance relative to Full MetaDecomp in Table R3.1.

    Table R3.1. Component ablations and relative gains of Full MetaDecomp, with Full − Ablation in parentheses. Each row reports one run: ALFWorld success (%) with Qwen3-4B and ScienceWorld score on a 0–100 scale with Qwen3.5-9B.

    ConfigurationInterventionALFWorld success (%)ScienceWorld score
    Full MetaDecompReference88.8172.20
    Fixed initial schemaRetraining79.10 (+9.71)65.02 (+7.18)
    No historical reprojectionRetraining82.84 (+5.97)62.92 (+9.28)
    No state–procedure linksInference81.34 (+7.47)66.29 (+5.91)
    State onlyInference80.60 (+8.21)53.40 (+18.80)
    Procedure onlyInference80.60 (+8.21)62.18 (+10.02)
    No initial refinementInference79.10 (+9.71)56.92 (+15.28)
AuthorsOpenReview ↗

Response to Reviewer ZaCt (2/2)

(W2) The cost discussion primarily concerns execution. Outer-loop learning and initial plan refinement are excluded from the execution budget, while ALFWorld can use more execution tokens than MPO. It is difficult to judge the total cost advantage without an amortized accounting.

Thank you for emphasizing complete cost accounting; this helped us make the training and inference cost comparison fully transparent. Regarding your concern that the cost discussion primarily concerns execution and that Outer-loop learning and initial plan refinement are excluded from the execution budget, we would like to respectfully clarify that the execution budget in original Appendix A.1 (Lines 772–773) referred only to the 1M-token per-episode termination cap during rollout, whereas initial plan refinement was already included in all reported inference token totals, as noted in Line 426 of the original submitted manuscript. Below, Table R3.2 and Table R3.3 report per-task inference costs and one-time training costs separately to support an amortized accounting of the total cost advantage.

  • (Inference cost including initial plan refinement) With initial plan refinement fully included, MetaDecomp achieves higher performance with fewer inference tokens than ReAct and AutoRefine in all five settings, and uses fewer inference tokens than MPO on ScienceWorld.

    • While ALFWorld can use more execution tokens than MPO in exchange for higher task success, on the longer-horizon ScienceWorld benchmark MetaDecomp simultaneously achieves higher task scores and lower per-task inference tokens, including initial plan refinement, than MPO [1], ReAct [2], AutoRefine [4], and SkillOpt [5], as shown in Table R3.2 and manuscript Table 1 and Table 2.

    Table R3.2. Mean inference cost per task in K input-plus-output tokens, including initial planning, initial plan refinement, online control, and actor rollouts across successful and failed episodes. The 4B/8B actors use Qwen3; 9B/27B use Qwen3.5.

    MethodALFWorld 4BALFWorld 8BScienceWorld 8BScienceWorld 9BScienceWorld 27B
    MetaDecomp53.8042.6094.90103.3074.30
    ReAct [2]57.2249.40140.58149.8289.67
    AdaPlanner [3]32.3028.4040.0742.8830.81
    MPO48.3440.27130.28130.3176.05
    AutoRefine [4]64.7551.04142.41146.7092.82
    SkillOpt [5]101.5086.71213.01218.7786.04
  • (Outer-loop learning cost and amortized total cost advantage) Outer-loop learning is a one-time training cost that is smaller than SkillOpt across all five settings and is amortized across tasks after deployment.

    • Across all five settings in Table R3.3, MetaDecomp uses substantially fewer one-time training tokens for outer-loop learning and rollouts than SkillOpt and more than AutoRefine, while ReAct, AdaPlanner, and MPO have no training run in our experiments.
    • Under an amortized accounting of the total cost advantage, because the learned procedural memory is frozen at deployment, no further outer-loop learning cost is incurred at test time, so MetaDecomp's per-task inference savings over ReAct, AutoRefine, SkillOpt, and on ScienceWorld over MPO accumulate directly as test volume grows.

    Table R3.3. Full training cost in M input-plus-output tokens, including training rollouts, initial planning/refinement, and outer-loop learning updates; MPO uses an official checkpoint with unreported upstream training cost. The 4B/8B actors use Qwen3; 9B/27B use Qwen3.5.

    MethodALFWorld 4BALFWorld 8BScienceWorld 8BScienceWorld 9BScienceWorld 27B
    MetaDecomp23.0220.3465.9037.0835.42
    AutoRefine5.294.0811.3514.816.89
    SkillOpt43.6231.4495.00100.5152.97

[1] MPO: Boosting LLM Agents with Meta Plan Optimization. Findings of EMNLP 2025.

[2] ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.

[3] AdaPlanner: Adaptive Planning from Feedback with Language Models. NeurIPS 2023.

[4] AutoRefine: Compiling Trajectories into Validated Typed Agent Artifacts. arXiv preprint 2026.

[5] SkillOpt: Executive Strategy for Self-Evolving Agent Skills. arXiv preprint 2026.

Thank you again for helping us strengthen the evaluation and its presentation. We would be happy to clarify any remaining questions.

AuthorsOpenReview ↗

We would like to hear from Reviewer ZaCt

Hi Reviewer ZaCt,

We sincerely appreciate your dedicated effort in reviewing our manuscript. We would like to follow up to see if our response addresses your concerns or if you have any further questions. We would really appreciate the opportunity to discuss this further if our response has not already addressed your concerns.

Thank you again!

Submission #3603 Authors

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

This paper proposes MetaDecomp, a two-loop in-context meta-learning framework for long-horizon language agents that jointly learns a task-relative abstract state space and reusable transition procedures represented as DAGs. The outer loop builds and reorganizes procedural memory from successful training executions, including reprojecting historical evidence when the state schema evolves. The inner loop uses this memory to refine initial task decompositions and revise them online based on execution feedback. Experiments on ALFWorld and ScienceWorld show competitive performance and inference efficiency.

Critical strengths
  1. Jointly learning task states and reusable procedures while adapting procedural memory to evolving state representations is interesting and well-motivated.
  2. The paper is generally well written and easy to follow.
Critical weaknesses
  1. The method relies on MPO's cached decompositions, and its effectiveness with other initial planners or lower-quality plans remains unclear.

  2. The current ablation conflates learned procedural memory with initial refinement, making it difficult to isolate the contributions of joint state-procedure learning and historical reprojection.

  3. Experiments primarily evaluate unseen instances within predefined task types, leaving transfer to novel task types or out-of-distribution settings unexplored.

Questions for authors

See Weaknesses.

8 comments
AuthorsOpenReview ↗

Response to Reviewer hEyE (1/4)

We sincerely thank you for your careful review and positive assessment of the motivation and clarity of our work. Your suggestions have helped us strengthen the evaluation of alternative initial planners, the contributions of learned memory and initial refinement, and transfer beyond known task types. We respond to each concern point by point below. Manuscript section, appendix, figure, and table numbers refer to the revised paper; changes are highlighted in blue. Please let us know if any concerns remain or if further clarification would be helpful; we would be glad to discuss them.

(W1) The method relies on MPO's cached decompositions, and its effectiveness with other initial planners or lower-quality plans remains unclear.

Thank you for encouraging us to test alternative initial planners; this helped strengthen the evaluation of its effectiveness with other initial planners or lower-quality plans. Regarding your concern that the method relies on MPO's cached decompositions, we would like to respectfully clarify that MPO [1] is only a plug-in initializer rather than a required component of MetaDecomp.

  • (Learning and operating with other initial planners or lower-quality plans) MetaDecomp remains highly effective when trained and tested with other initial planners or lower-quality plans, without MPO's cached decompositions.

    • Under Retrain in Table R2.1, learning and testing with other initial planners such as ReWOO [2], ADaPT [3], or a Task-only initializer that plans directly from public task instructions without MPO's cached decompositions achieves strong performance on both benchmarks, outperforming all non-MPO baselines in manuscript Table 1 and Table 2. Even starting from lower-quality plans, outer-loop memory learning and inner-loop refinement recover nearly the full performance of the MPO-initialized system.
  • (Zero-shot swap of the test-time initializer) A frozen memory learned with MPO also transfers directly to other initial planners at test time, while matched retraining further improves alignment.

    • Under Test swap in Table R2.1, applying the frozen MPO-trained memory directly to other initial planners at test time maintains solid task completion without retraining, while retraining with the matched initializer consistently improves over Test swap across both benchmarks.

    Table R2.1. Performance with alternative initial planners using the actor model for initial planning, refinement, and online control. ALFWorld: success (%) with Qwen3-4B; ScienceWorld: score with Qwen3.5-9B.

    InitializerALFWorld: Test swapScienceWorld: Test swapALFWorld: RetrainScienceWorld: Retrain
    MPO86.3069.586.3069.5
    Task-only77.6057.685.0765.7
    ReWOO76.1061.985.0767.2
    ADaPT74.6062.182.0965.7

[1] MPO: Boosting LLM Agents with Meta Plan Optimization. Findings of EMNLP 2025.

[2] ReWOO: Decoupling Reasoning from Observations for Efficient Augmented Language Models. arXiv preprint 2023.

[3] ADaPT: As-Needed Decomposition and Planning with Language Models. Findings of NAACL 2024.

AuthorsOpenReview ↗

Response to Reviewer hEyE (2/4)

(W2) The current ablation conflates learned procedural memory with initial refinement, making it difficult to isolate the contributions of joint state-procedure learning and historical reprojection.

Thank you for suggesting a more precise assessment of the learning mechanisms; we agree that the original empty-memory comparison conflates learned procedural memory with initial refinement, and we added targeted ablations in Table R2.2 to isolate the contributions of joint state-procedure learning and historical reprojection.

  • (Isolating joint state–procedure learning) With initial refinement retained across all variants, joint state-procedure learning provides substantial gains through adaptive schemas, state–procedure links, and the combination of states and procedures.

    • To isolate adaptive state abstraction in joint state-procedure learning, retraining under Fixed initial schema fixes the initial state space while keeping initial refinement and online revision, which significantly reduces performance on both benchmarks in Table R2.2.
    • At inference time with initial refinement retained, removing state–procedure links or restricting learned procedural memory access to State only or Procedure only consistently lowers performance across both benchmarks in Table R2.2, confirming that states, procedures, and their links in joint state-procedure learning are all needed together.
  • (Isolating historical reprojection) With initial refinement and schema expansion both retained, historical reprojection is critical for keeping earlier procedures aligned with updated state abstractions.

    • Retraining under No historical reprojection permits schema expansion and retains initial refinement, but omits historical reprojection when the schema changes, causing clear performance drops on both benchmarks in Table R2.2.
  • (Isolating initial refinement from online revision) Retaining learned procedural memory for online revision while removing initial refinement isolates the complementary benefit of refining the starting plan.

    • Under No initial refinement, removing initial refinement while keeping learned procedural memory for online revision also lowers performance in Table R2.2, showing that initial refinement and online revision are complementary.

    Table R2.2. Component ablations and relative gains of Full MetaDecomp, with Full − Ablation in parentheses. Each row reports one run: ALFWorld success (%) with Qwen3-4B and ScienceWorld score on a 0–100 scale with Qwen3.5-9B.

    ConfigurationInterventionALFWorld success (%)ScienceWorld score
    Full MetaDecompReference88.8172.20
    Fixed initial schemaRetraining79.10 (+9.71)65.02 (+7.18)
    No historical reprojectionRetraining82.84 (+5.97)62.92 (+9.28)
    No state–procedure linksInference81.34 (+7.47)66.29 (+5.91)
    State onlyInference80.60 (+8.21)53.40 (+18.80)
    Procedure onlyInference80.60 (+8.21)62.18 (+10.02)
    No initial refinementInference79.10 (+9.71)56.92 (+15.28)
AuthorsOpenReview ↗

Response to Reviewer hEyE (3/4)

(W3) Experiments primarily evaluate unseen instances within predefined task types, leaving transfer to novel task types or out-of-distribution settings unexplored.

We appreciate your suggestion to evaluate transfer beyond unseen instances within predefined task types, which helped us investigate when learned procedures help or hurt under transfer to novel task types or out-of-distribution settings in Table R2.3 and Table R2.4.

  • (Transfer to novel task types) Evaluating frozen memories across novel task types in ALFWorld reveals mixed transfer with both gains and losses in Table R2.3, pointing to the possibility that transfer to novel task types depends on how completely source procedures cover the target task's required operations and preconditions.

    • Freezing a memory learned on one ALFWorld task type and evaluating it zero-shot on the other five novel task types in Table R2.3 yields clear gains on some source–target pairs and drops on others relative to Empty memory, showing that sharing an environment alone does not guarantee beneficial transfer.
    • Intuitively, comparing specific pairs suggests a plausible explanation: Clean, Cool, Heat, and Pick-Two memories already encompass the complete search-pickup-placement workflow needed for Pick and boost Pick above Empty, whereas Pick memory covers searching, picking up, and placing an object directly without intermediate conditioning, steering Cool or Clean tasks to skip cooling or cleaning and place the object prematurely, which suggests that reuse helps when source procedures cover the full target workflow but can mislead planning when they cover only part of the required steps.

    Table R2.3. Transfer between ALFWorld task types with Qwen3-4B, with GLM-5.3-Flash used only for source-memory training. Rows are test types; columns are frozen source memories. Success rates (%) average three evaluations. Empty has no procedural memory.

    Test type / source memoryLookPickCleanCoolHeatPick-TwoEmpty
    Look100.0072.2281.4892.5988.8983.3387.04
    Pick88.89100.0095.8391.67100.00100.0087.50
    Clean74.1969.8977.4273.1277.4280.6572.04
    Cool71.4350.7963.4980.9577.7876.1971.43
    Heat84.0666.6769.5778.2688.4185.5175.36
    Pick-Two74.5166.6760.7874.5158.8274.5172.55

(The response to (W3) continues in Response to Reviewer hEyE (4/4).)

Reviewer_1nY6OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The authors propose a framework that learns a procedural world model from task execution experiences. It builds procedural memory by jointly learning descriptions of state changes in accordance with task progress and reusable procedural fragments. When new state distinctions are learned, it revisits earlier executions and reorganizes the stored procedures. For a new task, the agent uses this memory to refine an initial plan and revise it based on execution feedback. Experiments on ALFWorld and ScienceWorld with several Qwen models show improved task performance in most evaluated settings and a favorable inference-token trade-off, while the learned memory remains fixed during testing.

Critical strengths

The authors proposed an agentic task planning method to learn task-progress representations together with reusable procedures and reorganize earlier procedural knowledge when new state distinctions are observed. This provides a structured way to adapt past experience to the work that remains in a new task. For ALFWorld and ScienceWorld using several Qwen models, the proposed method generally outperforms previous methods including ReAct and AutoRefine with fewer inference tokens in every evaluated configuration. The method also consistently improves over within-task revision alone, supporting the value of procedural memory shared across related task instances.

Critical weaknesses

The current contribution list in section I merely states that the authors developed and evaluated MetaDecomp. It should identify the specific advances beyond prior work and connect each claim to supporting evidence.

The roles, state fields, and transition procedures resemble symbolic planning representations and hierarchical task decomposition. Although the paper cites related studies, it does not clearly explain how MetaDecomp differs from existing planning-model learning and LLM–planning approaches.

The paper does not report the exact training splits, task counts and ordering, successful trajectory counts, or number of memory updates. These details are needed for reproducibility and for assessing whether the methods use comparable training resources.

Successful episodes can contain redundant or unnecessary exploratory actions. Excluding invalid attempts does not establish that the remaining actions are useful for reuse. The paper does not measure how effectively such actions are removed from learned procedures.

Performance varies substantially across ScienceWorld tasks, which needs deeper analysis. The results do not distinguish limitations of the base LLM from errors introduced by state abstraction, retrieval, or procedure reuse. Which factors might explain the proposed method's lower performance on ScienceWorld than on ALFWorld?

Evaluation is limited to two text-based simulated environments and mainly tests new instances within known task types. This leaves generalization to unfamiliar procedures and less structured observations uncertain.

Questions for authors
  • How are valid but unnecessary actions identified and how often do they remain in the stored DAGs?

  • How many different DAGs and state changes are depicted per one task type e.g. Look task in AlfWorld? Detailed analysis on per-task-type memory sizes, DAG structures, procedural redundancy, and reuse rates would help better understand the efficacy of the proposed method.

  • How does the proposed method's performance change with the length of the task horizon?

  • Would it be possible to use the proposed method across multi-domains without manual engineering? When applying the proposed method to unseen task types or new domains, which parts must be manually specified or changed? Can the agent reuse previously learned state descriptions and procedures, or must it build a new memory from scratch? If reuse is possible, does it reduce the number of successful executions needed to learn the new tasks?

  • Could the proposed method support continual or lifelong learning by updating its procedural memory as new tasks are encountered after deployment? How would it incorporate new experience while preserving useful procedures and preventing errors from accumulating in memory?

13 comments
AuthorsOpenReview ↗

Response to Reviewer 1nY6 (1/10)

We sincerely appreciate your thorough review and constructive suggestions, as well as your recognition of jointly learning task-progress representations and reusable procedures. Your feedback has helped us clarify the methodological contributions and strengthen the evidence on procedural reuse, performance variation, and generalization. We respond to your comments and questions point by point below. Manuscript section, appendix, figure, and table numbers refer to the revised paper; changes are highlighted in blue. Please let us know if you have any remaining concerns or would like further clarification; we would be happy to discuss them.

(W1) The current contribution list in section I merely states that the authors developed and evaluated MetaDecomp. It should identify the specific advances beyond prior work and connect each claim to supporting evidence.

Thank you for this valuable suggestion, which helped us present the methodological advances more clearly. To identify the specific advances beyond prior work, Section 1 now states three contributions and their supporting evidence.

  • (Meta-learning) MetaDecomp learns how to decompose across tasks rather than starting from scratch on each one (Section 3). In the outer loop, it distills a shared procedural prior from executions on related tasks. In the inner loop, it uses that prior to build and revise a decomposition for the current task based on environmental feedback.

  • (Joint state–procedure learning) Reusable procedures are anchored to explicit task progress (Sections 3.2–3.3 and Table 4). We learn progress abstractions and transition DAGs together, so each procedure links what has already been accomplished to how the remaining work can be done. As the abstractions are refined, we reinterpret past executions so that these links stay consistent.

  • (Performance, cost, and transfer) MetaDecomp is both stronger and cheaper in most settings, and we identify when transfer stops helping (Section 4).

    • It achieves the highest mean performance in four of the five ALFWorld/ScienceWorld settings while using fewer inference tokens than ReAct and AutoRefine in all five.
    • Across eight additional benchmarks, reuse helps when procedures are highly applicable but leads to negative transfer when they only partially apply. Table R1.9 below reports this pattern using operation and precondition coverage.
AuthorsOpenReview ↗

Response to Reviewer 1nY6 (2/10)

(W2) The roles, state fields, and transition procedures resemble symbolic planning representations and hierarchical task decomposition. Although the paper cites related studies, it does not clearly explain how MetaDecomp differs from existing planning-model learning and LLM–planning approaches.

We sincerely appreciate this suggestion, which helped us better situate MetaDecomp within the planning literature. Section 2 now clarifies how MetaDecomp differs from existing planning-model learning and LLM–planning approaches.

  • (Planning-model learning) Different from existing planning-model learning approaches [1, 2], MetaDecomp jointly learns roles, state fields, and transition procedures so new tasks can reuse experience according to their current progress.
    • Whereas Konidaris et al. [1] learn symbolic planning representations for a given set of skills and Bonet and Geffner [2] infer first-order symbolic planning models, MetaDecomp learns task-relative roles and state fields together with the transition procedures connecting observed states, evolving both as experience accumulates.
    • In the cleaning-and-placement example in Appendix B.2.1, possession alone cannot distinguish whether a held object is clean, whereas adding cleanliness separates the remaining transition procedures: clean then place when it is not yet clean, and place directly when it already is.
  • (LLM–planning) Different from existing LLM–planning approaches for hierarchical task decomposition and workflow reuse [3, 4], MetaDecomp grounds and composes learned procedures without an executable dynamics model and reprojects earlier executions when abstractions change.
    • Unlike TheoryCoder-2 [3], which learns symbolic abstractions and an executable dynamics model for hierarchical planning, MetaDecomp uses the state graph to propose candidate continuations and each transition DAG to record operations and prerequisites, while the LLM selects applicable operations, grounds them in the current task, and fills in missing steps from observations.
    • Unlike Agent Workflow Memory [4], which induces reusable workflow subroutines under a static representation, MetaDecomp additionally reprojects the original execution evidence under the updated schema and reassesses which transition procedures are supported between each pair of states.

[1] From Skills to Symbols: Learning Symbolic Representations for Abstract High-Level Planning. JAIR 2018.

[2] Learning First-Order Symbolic Representations for Planning from the Structure of the State Space. ECAI 2020.

[3] Learning Abstractions for Hierarchical Planning in Program-Synthesis Agents. arXiv preprint 2026.

[4] Agent Workflow Memory. ICML 2025.

AuthorsOpenReview ↗

Response to Reviewer 1nY6 (3/10)

(W3) The paper does not report the exact training splits, task counts and ordering, successful trajectory counts, or number of memory updates. These details are needed for reproducibility and for assessing whether the methods use comparable training resources.

Thank you for highlighting the details needed for reproducibility and for assessing whether the methods use comparable training resources. Table R1.1 and Table R1.2 report the exact training splits, task counts and ordering, successful trajectory counts, or number of memory updates, with the full protocol in Appendices A.1 and B.1.

  • (Training splits and ordering) Training tasks are sampled randomly in batches from fixed candidate splits, without a handcrafted task-type curriculum.

    • Regarding the exact training splits, ALFWorld [1] uses 1,200 candidate training instances across six task types and 134 unseen test tasks, while ScienceWorld [2] uses the official train/validation/test splits of 3,592/1,796/1,819 instances and 211 test tasks.
    • Regarding task counts and ordering, the attempted task counts are reported in Table R1.1, and their ordering is determined by random batch sampling from the training pool rather than a handcrafted difficulty or task-type curriculum.
  • (Successes and memory updates) Only fully successful training episodes update memory, and validation selects the memory frozen for testing.

    • Each learning space updates memory every two successful episodes, following a one-success warmup on ScienceWorld 8B/9B, and freezes the checkpoint with the highest validation performance for testing; Table R1.1 reports the resulting task counts, successful trajectory counts, and number of memory updates.

    Table R1.1. Training resources for the selected main-experiment checkpoints. Tasks count distinct attempted training instances, Episodes include retries, Successes count episodes admitted to memory, and Updates include initialization.

    SettingSpacesTasksEpisodesSuccessesUpdates
    ALFWorld, Qwen3-4B61111119246
    ALFWorld, Qwen3-8B61101109447
    ScienceWorld, Qwen3-8B1731333011357
    ScienceWorld, Qwen3.5-9B1721922811960
    ScienceWorld, Qwen3.5-27B1624224218492
  • (Comparable training resources) All trained methods share the same distillation model, and MetaDecomp uses fewer total training tokens than SkillOpt across all five settings.

    • To ensure comparable training resources, MetaDecomp, AutoRefine [3], and SkillOpt [4] all use GLM-5.3-Flash for outer-loop learning or experience distillation, while MetaDecomp uses only the Qwen actor for initial refinement, online control, and rollouts.
    • Across all five settings in Table R1.2, MetaDecomp uses fewer total training tokens than SkillOpt and more than AutoRefine, whereas ReAct [5], AdaPlanner [6], and MPO [7] have no training run in our experiments.

    Table R1.2. Full training cost in M input-plus-output tokens, including rollouts, initial-plan refinement, and learning updates; MPO uses an official checkpoint with unreported upstream training cost. The 4B/8B actors use Qwen3; 9B/27B use Qwen3.5.

    MethodALFWorld 4BALFWorld 8BScienceWorld 8BScienceWorld 9BScienceWorld 27B
    MetaDecomp23.0220.3465.9037.0835.42
    AutoRefine5.294.0811.3514.816.89
    SkillOpt43.6231.4495.00100.5152.97

[1] ALFWorld: Aligning Text and Embodied Environments for Interactive Learning. ICLR 2021.

[2] ScienceWorld: Is your Agent Smarter than a 5th Grader? EMNLP 2022.

[3] AutoRefine: Compiling Trajectories into Validated Typed Agent Artifacts. arXiv preprint 2026.

[4] SkillOpt: Executive Strategy for Self-Evolving Agent Skills. arXiv preprint 2026.

[5] ReAct: Synergizing Reasoning and Acting in Language Models. ICLR 2023.

[6] AdaPlanner: Adaptive Planning from Feedback with Language Models. NeurIPS 2023.

[7] MPO: Boosting LLM Agents with Meta Plan Optimization. Findings of EMNLP 2025.

Comments on the paper

1 comment
AuthorsOpenReview ↗

Revision Note

We sincerely thank all three reviewers (Reviewer 1nY6, Reviewer hEyE, and Reviewer ZaCt) for their thoughtful feedback and constructive suggestions. We have uploaded the revised manuscript PDF with all updates highlighted in blue. Below is a brief summary of the main revisions:

  • Methodological positioning and schema-reprojection case study (Sections 1–2; Appendix B.2.1): Sharpened the contribution list in Section 1, expanded Section 2 to clarify how MetaDecomp's joint state–procedure learning and historical boundary reprojection differ from symbolic planning-model learning and existing LLM–planning approaches, and added a concrete case study on schema expansion and procedure reassignment in Appendix B.2.1.
  • Component and initial-planner ablations (Section 4.3, Tables 4–5; Appendix B.3): Added six controlled retraining and inference-time ablations isolating adaptive schemas, historical reprojection, state–procedure links, state/procedure representations, and initial refinement (Table 4), alongside evaluations with three alternative initial planners (Task-only, ReWOO, and ADaPT) under both zero-shot test swap and matched retraining (Table 5).
  • Cross-type transfer, 8-benchmark OOD evaluation, and continual learning (Section 4.4, Tables 6–7; Appendices B.4–B.5, Tables 11–12): Added zero-shot cross-type transfer across all six ALFWorld task types (Table 6), a 480-task zero-shot OOD evaluation across eight unseen benchmarks grouped by operation and precondition coverage (Table 7; Appendix B.4), and a six-stage continual-learning evaluation across ALFWorld task types (Appendix B.5, Tables 11–12).
  • Task-horizon and category-variance analyses (Section 4.2, Figure 4; Appendix B.6, Table 13): Added a 10-bin task-horizon analysis on ALFWorld and ScienceWorld (Figure 4) and category/subtype variance breakdowns on ScienceWorld separating base-LLM limitations from state abstraction and procedure reuse (Appendix B.6, Table 13).
  • Training resources, procedural reuse statistics, and full token cost accounting (Section 4.1; Section 4.2, Table 3 and Figure 5; Appendix A.2, Figure 6; Appendices B.1, B.2, and B.7, Tables 9, 10, 14, and 15): Reported exact training splits, task/episode/success/update counts (Table 9), per-setting and per-task-type memory sizes and test-time reuse rates (Table 10), and complete one-time training and per-task inference token breakdowns across all five settings (Table 3; Tables 14–15).

All new analyses were conducted using the unmodified MetaDecomp framework and leave all main-table conclusions unchanged. We thank the reviewers again for helping us make the paper clearer and more self-contained.

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.