PaperlayerView exact revision on arXiv

Source-linked AI summary

LeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents

Shen Liu, Ze Zhou, Zhenguo Xu, Shaopu Wang, Yike Gao, Chunlei Wang

arXiv:2610.32412v1cs.AI

TL;DR

Recursive LLM agents can still spend host-model inference translating explicit leaf tasks into executable actions, even after high-level planning has produced them. LeafSkill reuses retrieved task–action experience through state-bound candidate selection and abstention. Across Robotouille and ALFWorld, it reduces host tokens while leaving the host planner in place.

  • Problem

    After recursive agents make a leaf task explicit, they may still invoke the large host model to translate it into an environment-accepted action, a recurring use of host tokens.

    Source: §1 · p. 1

  • Method

    LeafSkill retrieves canonical task–action templates, binds them to current entities and admissible actions, and uses a lightweight controller to select a candidate or abstain to the host.

    Source: §1 · p. 1

  • Results

    14.4–41.3%: LeafSkill reduces Qwen3.5-27B host tokens across Robotouille and ALFWorld on jointly solved tasks.

    Source: §5 · p. 7

  • Takeaways & Limitations

    The results support state-bound action reuse as a practical complement to recursive LLM planning rather than a replacement for it.

    Source: §5 · p. 7

    Evidence covers one planner, one host, two environments, and one selector family; LeafSkill also assumes recurring parameterizable leaves, a discrete admissible-action interface, and an environment adapter.

    Source: §5 · p. 7

Abstract

from arXiv · show

Recursive large language model (LLM) agents solve long-horizon tasks through decomposition, execution, and environmental feedback. Yet after producing an atomic subtask, they often invoke the same large model again to translate a familiar instruction into an executable action. We observe that high-level planning is task-specific, whereas leaf-level action grounding recurs across tasks and can be checked against environment-admitted actions. We introduce LeafSkill, an external reuse layer that retrieves semantically related grounding experience, binds action templates to current objects and admissible actions, and uses a lightweight 1.7B selector to choose a bound action or defer the unchanged subtask to the original agent. Skills remain outside the model parameters, allowing a frozen selector to use post-training additions. Under lexical and syntactic paraphrases, exact matching retrieves none of the reference actions, while semantic retrieval preserves 96.0% and 97.0% candidate coverage; the frozen selector adopts 97.34% of unseen skill types. On an independent confirmation set of 200 Robotouille tasks, LeafSkill completes 137 tasks versus 121 for ReCAP (68.5% vs. 60.5%, exact McNemar p=.0259). On the 106 tasks completed by both, it reduces Qwen3.5-27B tokens by 14.42% and calls by 14.08%. A separate held-out repeat yields a consistent 9.0-point success gain and 18.98% token reduction. Reusable leaf grounding can thus improve task completion while reducing host-model inference without replacing high-level planning.

1 Introduction

LeafSkill targets the repeated action-grounding step that remains after recursive agents decompose a task, leaving high-level planning and recovery with the host. Its constrained reuse layer addresses this recurring cost and reduces host-model token use across two environments.

Motivation

  • 22.9%, 23.7%, and 44.5% of host-model tokens go to strict leaf grounding in two Robotouille splits and ALFWorld, respectively.The paper studies this opportunity with ReCAP while leaving its planning, recovery, and completion mechanisms unchanged.Source: §1 · p. 1
  • Reusable grounding must handle semantic recall, state-dependent instantiation, candidate disambiguation, and safe abstention because similar phrases can encode different intents.Retrieved arguments may be absent from the current state, and the same task can ground differently under different observations.Source: §1 · p. 1

Approach

  • LeafSkill retrieves semantic task–action templates, binds them to current entities and admissible actions, then selects a candidate or abstains to the unchanged host.Its controller cannot generate free-form commands; planning, recovery, and goal completion remain host-owned.Source: §1 · pp. 1–2

Results

  • 23.1% on Robotouille and 41.3% on ALFWorld: LeafSkill reduces host-model tokens on jointly solved tasks with the same ReCAP planner and Qwen3.5-27B host.The reductions are averaged over all nine cross-run pairings of three independent repeats.Source: §1 · p. 2

Contributions

  • Ablations identify complementary roles for retrieval, binding, and learned selection, while post-training library growth can increase savings without selector retraining.The contribution summary describes evidence across Robotouille and ALFWorld and compatible additions to the library.Source: §1 · p. 2

2 Related Work

Related work spans recursive planning, experience reuse, constrained grounding, and cost-aware routing. LeafSkill is positioned as a narrower delegation boundary: it reuses task–action templates and allows only a local choice among admissible actions or abstention.

Recursive agent backbones

  • Recursive-agent approaches primarily address planning, context management, and recovery, whereas LeafSkill operates after an adapter exposes an eligible leaf task and does not take over decomposition or tree-state control.The paper instantiates this boundary using ReCAP as the recursive host.Source: §2 · p. 2

Experience and skill reuse

  • Experience-reuse systems retrieve lessons, demonstrations, trajectories, programs, workflows, or structured skill graphs, while LeafSkill retrieves task–action templates and binds them to the current state.The table frames the comparison around reuse granularity and delegated decisions.Source: §2 · p. 2, Table 1 · p. 2

Constrained grounding and routing

  • Unlike feasibility constraints on plans or constrained decoding, LeafSkill binds retrieved templates to currently admissible actions and delegates only a leaf-level choice or abstention.Cost-aware routing selects models for a complete request; LeafSkill acts after the host determines the leaf task.Source: §2 · p. 2

3 Method

LeafSkill separates offline construction of reusable task–action experience from online grounding at eligible leaf tasks. Its delegation boundary returns either a selected admissible action or control to the host through abstention.

Overview

  • Online, LeafSkill retrieves related templates at adapter-eligible leaf tasks, binds them to the current admissible action set, then selects one action or abstains.Offline, host trajectories are converted into canonical task–action templates and selector supervision.Source: Overview · p. 2, Figure 2 · p. 3

Leaf-Level Delegation Boundary

  • Leaf-level grounding maps an already specified task node to an action in the current admissible set without changing decomposition, a parent plan, or completion state.The eligible node, local observation, and admissible action set are denoted by τt, ot, and At.Source: Leaf-Level Delegation Boundary · p. 3
  • When the selector abstains, no delegated action executes and the same node returns to the host; the adapter determines eligibility and the selector decides whether to execute a bound candidate.Retrieval proposes related experience but cannot authorize execution; binding removes commands outside At.Source: Leaf-Level Delegation Boundary · p. 3

Retrieve–Bind–Select

LeafSkill retrieves related task–action templates, binds them to the current request, then lets a local selector choose an admissible candidate or abstain.

Retrieve–Bind–Select

  • Each library skill pairs a canonical task with an action template, and multiple skills may share a task key to preserve state-dependent alternatives.Skills also retain observed surface forms and provenance.Source: Retrieve–Bind–Select · p. 3
  • Semantic retrieval ranks canonical task keys against the leaf task and passes templates associated with the top-k keys above threshold γ to binding.Retrieval is recall-oriented: retrieved templates cannot execute on their own.Source: Retrieve–Bind–Select · p. 3
  • State-bound binding instantiates typed template slots with current entities and retains only commands exactly matching an action in the admissible set.Duplicate skill–action pairs are merged into the finite candidate set.Source: Retrieve–Bind–Select · p. 3
  • The selector chooses a shuffled request-local candidate identifier or abstains; the chosen action is checked against admissible actions again before execution.An empty candidate set returns control to the host.Source: Retrieve–Bind–Select · p. 3

Environment Adapters and State-Bound Binding

Environment adapters determine when a task node is eligible for grounding, while state-bound binding instantiates templates against the current interface actions.

Environment Adapters and State-Bound Binding

  • Robotouille exposes atomic leaves directly, while ALFWorld admits non-root nodes on the downward execution path and excludes root, return, and revisit phases.For example, “check fridge 1” is eligible during descent but not when revisited during recovery.Source: Environment Adapters and State-Bound… · pp. 3–4, Table 2 · p. 4
  • The binder canonicalizes action strings, requires concrete task objects to appear in the action, and enumerates current interface strings rather than synthesizing slot values.It retains only concrete commands matching an action in the current admissible set.Source: Environment Adapters and State-Bound… · p. 4
  • Canonicalization joins spaced object identifiers, tokenizes alphanumeric fields, and replaces instance suffixes with typed placeholders while leaving paraphrase matching to retrieval.For example, “onion1” becomes “onion#.”Source: Environment Adapters and State-Bound… · p. 4

Selector and Deterministic Guards

The selector sees only request-local information and cannot generate commands; deterministic guards return control to the host whenever a valid bound choice is unavailable.

Selector and Deterministic Guards

  • The selector receives the task, local observation, admissible actions, and localized candidate descriptions, but not the root goal or planning history.Randomized request-local identifiers require it to read candidate content rather than memorize labels.Source: Selector and Deterministic Guards · p. 4
  • The selected identifier is resolved through its request-local mapping and checked against admissible actions again before execution.The selector chooses an identifier or abstains; it cannot generate a free-form command.Source: Selector and Deterministic Guards · p. 4
  • Empty retrieval or binding, explicit abstention, an unknown identifier, or a failed final admissibility check returns control to the host.A skill linked to failed local execution may be masked for the episode, with the mask reset between episodes.Source: Selector and Deterministic Guards · p. 4, Table 3 · p. 4

Library and Supervision Construction

The library is built from host-proposed task nodes and their subsequently executed environment actions, then canonicalized into task keys and action templates.

Library and Supervision Construction

  • Library records pair a host-proposed task node with the environment action subsequently executed, then canonicalize that pair into a task key and action template.This local attribution does not claim the action belongs to a globally optimal trajectory.Source: Library and Supervision Construction · p. 4
  • The Robotouille source library contains 446 skills from 200 group-A trees, and grows to 588 skills using group-B trees without controller updates.Group C is not used for construction, model selection, or adaptation.Source: Library and Supervision Construction · p. 4
  • The ALFWorld library retains 1,091 logged instances and produces 226 task keys and 228 skills from successful valid_seen source episodes.The instances include 1,089 unique executable contexts.Source: Library and Supervision Construction · p. 4

Lightweight Selector Adaptation

Selector adaptation mirrors deployment-time retrieval and binding, teaching the local model both which candidate to choose and when to defer. Training uses staged data splits, while compatible skills can be added later without retraining the selector.

Lightweight Selector Adaptation

  • Adaptation supervises both candidate selection and preservation of the host fallback, with request-local identifiers shuffled so persistent skill IDs cannot become the prediction target.Source: Lightweight Selector Adaptation · p. 4
  • Training reproduces deployment retrieval and binding, using mixed-rank examples with intent-conflicting distractors and reserving 124 of 413 Robotouille keys for a post-training adoption diagnostic.Source: Lightweight Selector Adaptation · p. 4
  • The default ALFWorld route adapts a Robotouille-trained selector on valid_seen; transfer-only and target-only controls omit one of the two adaptation stages.Source: Lightweight Selector Adaptation · p. 4
  • Each adaptation stage uses one epoch of rank-16 LoRA, and the resulting selector is frozen for evaluation; Robotouille Group C and ALFWorld valid_unseen are excluded from optimization.Source: Lightweight Selector Adaptation · p. 5
  • Post-training additions are compatible only within the existing action grammar, typed slots, retriever, admissible-action binding, and request-local description schema.Request-local identifiers do not provide new binders or action semantics.Source: Lightweight Selector Adaptation · p. 5

Implementation Details

Implementation uses a fixed semantic retriever and lightweight selector, with both libraries and selector parameters frozen during evaluation.

Implementation Details

4 Experiments

Across Robotouille and ALFWorld, LeafSkill reduces host-model cost on jointly solved tasks while observed success is higher in the reported comparisons. The experiments also test the roles of retrieval, binding, and selection, as well as whether a frozen selector can use compatible skills added after training.

4 Experiments

  • The evaluation asks how much host inference is saved, what retrieval, binding, and selection contribute, and whether frozen selectors use post-training skills.Robotouille repeats estimate rollout variability on the same 200 tasks; evaluation data are excluded from library construction and controller adaptation.Source: §4 · p. 5, Experimental Setup · p. 5, Table 4 · p. 5

Experimental Setup

  • All systems retain the ReCAP planner and Qwen3.5-27B host; host-cost reductions use jointly solved tasks to avoid conflating grounding cost with failure timing.The default routes use a Robotouille-trained selector, with additional valid_seen adaptation for ALFWorld.Source: Experimental Setup · p. 5

End-to-End Host Efficiency (RQ1)

Post-Training Library Growth (RQ3)

  • After training, adding Group-B traces expands the library from 446 to 588 skills; on 88 shared successes, token reduction over ReCAP rises from 18.35% to 25.64%.The frozen selector adopts 97.3% and retains 98.9% of 124 held-out compatible keys.Source: Table 10 · p. 7, Post-Training Library Growth (RQ3) · p. 7

Component Contributions (RQ2)

  • Semantic Top-5 retrieval retains 96.0% lexical and 97.0% syntactic paraphrase coverage, while state-bound templates with direct Top-1 execution yield 13.08% token reduction.Exact reuse of raw task–action pairs changes paired host tokens by 0.20%; semantic retrieval over raw pairs yields 8.08%.Source: Table 8 · p. 6, Component Contributions (RQ2) · p. 7
  • SFT raises mixed-rank accuracy from 81.9% to 93.3%, while controller-only fallback is 7.5% on the all-negative OOD diagnostic.On a direct R2 comparison, SFT is 9.0 points more successful; adding true planner context yields 90.7% mixed-rank accuracy versus 93.3% with the local interface.Source: Component Contributions (RQ2) · p. 7, Table 9 · p. 7

Controller Transfer and Adaptation

5 Discussion and Conclusion

LeafSkill replaces repeated leaf-grounding calls with bounded local action selection while leaving planning and recovery to the host. Its token savings grow with compatible library additions, but the evidence and interface assumptions limit the demonstrated scope.

5 Discussion and Conclusion

  • Expanding the library from 446 to 588 skills deepens fixed-set host-token savings from 18.3% to 25.6% without selector retraining.Request-local identifiers support compatible additions.Source: §5 · p. 7
  • Retrieval supplies paraphrase coverage, binding enforces admissibility, and selection resolves candidates, while planning and recovery remain with the host.LeafSkill replaces repeated 27B leaf-grounding calls with a finite local choice.Source: §5 · p. 7
  • LeafSkill assumes recurring parameterizable leaves, discrete admissible actions, and an environment adapter; new action grammars need adapters and binders.Open-ended or continuous actions may require richer executable-action interfaces; current evidence covers one planner, one host, two environments, and one selector family.Source: §5 · p. 7
  • Across Robotouille and ALFWorld, LeafSkill reduces Qwen3.5-27B host tokens by 14.4–41.3% while separating recurring leaf grounding from recursive reasoning.The layer complements rather than replaces the host planner.Source: §5 · p. 7

Related papers

Loading 2610.32412v1…