HaReCAP: Habitual-action Grounding for Recursive Large Language Model Agents

Shen Liu; Zhenguo Xu; Shaopu Wang; Yike Gao; Chunlei Wang

cs.AIAug 17, 2026

HaReCAP: Habitual-action Grounding for Recursive Large Language Model Agents

Authors: Shen Liu, Zhenguo Xu, Shaopu Wang, Yike Gao, Chunlei Wang

Organizations: North China Institute of Computer System Engineering, Beijing, China · University of Science and Technology of China, Hefei, China · China Information Security Research Institute Co., Ltd., Beijing, China

Abstract

Long-horizon embodied tasks require LLM agents to iteratively decompose high-level goals, revise plans in response to environmental feedback, and ground leaf-level subgoals into valid executable actions. Recursive context-management methods such as ReCAP improve planning stability through multi-level task decomposition and parent-node refinement, but still repeatedly invoke the LLM at leaf nodes to ground atomic subtasks into exact valid actions. We refer to this final grounding step as last-mile grounding redundancy, which accumulates into substantial LLM-call and token overhead during long-horizon execution. To mitigate this issue, we propose HaReCAP (Habitual-action Grounded ReCAP), a low-intrusion leaf grounding extension for ReCAP. HaReCAP extracts frequent leaf decisions from successful trajectories and compiles them offline into auditable and abstainable one-step leaf-reflex rules. At runtime, it skips the leaf LLM call only when a rule can uniquely determine a legal action in the current valid-action set; otherwise, it falls back to the original ReCAP. This design avoids repeatedly carrying the full recursive context into the LLM for routine leaf action grounding, while preserving the original recursive control flow. We evaluate HaReCAP on Robotouille and ALFWorld with Qwen3.5-27B as the main model. On tasks solved by both ReCAP and HaReCAP, HaReCAP reduces token consumption by 14.67%, 17.93%, and 20.08% on Robotouille synchronous, Robotouille asynchronous, and ALFWorld, respectively. The results show that HaReCAP can serve as a low-intrusion extension to ReCAP-style recursive context-management frameworks, reducing last-mile grounding redundancy across environments and models on commonly successful trajectories.

Figures & tables

Benchmark	ReCAP call composition	Grounding
Robotouille Sync		23.67%
Robotouille Async		29.56%
ALFWorld		45.73%

Figure 1: Last-mile action-grounding redundancy in successful ReCAP trajectories. Orange denotes strict leaf/action grounding calls, while blue denotes the remaining recursive reasoning, decomposition, and refinement calls. Statistics are collected from the Qwen3.5-27B ReCAP baseline.

Figure 2: Leaf-level difference between original ReCAP and HaReCAP. HaReCAP preserves recursive planning, backtracking, and refinement, and only inserts a leaf-reflex library at leaf-level action grounding.

Figure 3: System framework of HaReCAP. Offline construction builds a leaf-reflex library from successful ReCAP trajectories. Online execution retrieves only before leaf-level action grounding and falls back to the original ReCAP when no rule hits or when ambiguity is detected.

Method	Synchronous	Asynchronous
CoT	6%	0%
ReAct	22%	6%
ReCAP	75%	46%
HaReCAP s=1	76%	50%
HaReCAP s=2	76%	48%
HaReCAP s=4	74%	50%

Table 1: End-to-end success rates on Robotouille.

Split	Support	Method	Both Succ.	Avg. Tok.	Avg. Calls
Sync	$s=1$	ReCAP	64	753,455	31.48
	$s=1$	HaReCAP	64	703,732 ( $\downarrow$ 6.60%)	29.19 ( $\downarrow$ 7.30%)
	$s=2$	ReCAP	64	772,053	32.22
	$s=2$	HaReCAP	64	658,777 ( $\downarrow$ 14.67%)	27.66 ( $\downarrow$ 14.15%)
	$s=4$	ReCAP	63	747,628	31.29
	$s=4$	HaReCAP	63	674,415 ( $\downarrow$ 9.79%)	28.25 ( $\downarrow$ 9.72%)

Table 2: Reasoning cost on Robotouille both-success tasks.

Model	Method	Acc.	Avg. Tok.	Avg. Calls
Qwen3.5-27B	ReCAP	94.03%	308,624	29.90
Qwen3.5-27B	HaReCAP s=2	93.28%	244,697 ( $\downarrow$ 20.7%)	23.82 ( $\downarrow$ 20.3%)
Qwen3.5-9B	ReCAP	70.15%	348,457	31.17
Qwen3.5-9B	HaReCAP s=2	71.64%	282,035 ( $\downarrow$ 19.1%)	25.24 ( $\downarrow$ 19.0%)
Gemma4-26B	ReCAP	66.42%	501,186	43.84
Gemma4-26B	HaReCAP s=2	75.37%	443,885 ( $\downarrow$ 11.4%)	36.82 ( $\downarrow$ 16.0%)

Table 3: Cross-environment and cross-model results on ALFWorld.

Model	Method	Both Succ.	Avg. Tok.	Avg. Calls
Qwen3.5-27B	ReCAP	123	314,589	30.25
Qwen3.5-27B	HaReCAP s=2	123	251,411 ( $\downarrow$ 20.08%)	24.22 ( $\downarrow$ 19.94%)
Qwen3.5-9B	ReCAP	83	259,001	27.43
Qwen3.5-9B	HaReCAP s=2	83	222,207 ( $\downarrow$ 14.21%)	22.65 ( $\downarrow$ 17.43%)
Gemma4-26B	ReCAP	82	282,698	30.15
Gemma4-26B	HaReCAP s=2	82	223,069 ( $\downarrow$ 21.09%)	24.09 ( $\downarrow$ 20.11%)

Table 4: Reasoning cost on ALFWorld both-success tasks across models.

Model	Method	Synchronous	Asynchronous
Qwen3.5-9B	ReCAP	32%	5%
Qwen3.5-9B	HaReCAP s=2	44%	10%
Gemma4-26B	ReCAP	53%	26%
Gemma4-26B	HaReCAP s=2	55%	24%

Table 5: Cross-model end-to-end success rates on Robotouille.

Split	Method	Both Succ.	Avg. Tok.	Avg. Calls
Sync	ReCAP	44	909,927	38.66
Sync	HaReCAP s=2	44	611,890 ( $\downarrow$ 32.75%)	26.14 ( $\downarrow$ 32.39%)
Async	ReCAP	10	2,440,140	92.60
Async	HaReCAP s=2	10	2,000,589 ( $\downarrow$ 18.01%)	76.20 ( $\downarrow$ 17.71%)

Table 6: Reasoning cost of Gemma4-26B on Robotouille both-success tasks.

Long-Horizon LLM Agents

Long-horizon large language model agents plan, use tools, and preserve task state across extended, multi-step interactions.

See all 448 papers

LeafSkill: Efficient Leaf-Level Action Grounding for Recursive LLM Agents
Oct 23, 2026Shen Liu, Ze Zhou, Zhenguo Xu +3LLM Agent Skill Retrieval LLM Grounding
North China Institute of Computer System Engineering, Beijing, China · Beijing University of Aeronautics and Astronautics, Beijing, China · University of Science and Technology of China, Hefei, China +1
Recursive Agent Harnesses
Jun 11, 2026Elias Lumer, Sahil Sen, Kevin Paul +1Tool-Augmented Language Model Agents LLM Agent Orchestration
PricewaterhouseCoopers, U.S.
Remembering What Matters for Long-Horizon Robotic World Models
Oct 10, 2026Hao Wu, Tianyu Shao, Shengju Qian +10Video World Models Video Prediction
The Hong Kong University of Science and Technology (Guangzhou) · North China Electric Power University · LIGHTSPEED +6

Oct 23, 2026J/K move · Enter open · S save

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

Shen Liu, Ze Zhou, Zhenguo Xu, Shaopu Wang+2

North China Institute of Computer System Engineering, Beijing, China · Beijing University of Aeronautics and Astronautics, Beijing, China · University of Science and Technology of China, Hefei, China+1

LLM Agent Skill Retrieval LLM Grounding Long-Horizon LLM Agents

Open paper AskPDF arXiv

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.

Abstract

Long-Horizon LLM Agents

Explore similar work

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

Recursive Agent Harnesses

Remembering What Matters for Long-Horizon Robotic World Models