Learning When to Plan: Efficiently Allocating Test-Time Compute for LLM Agents

Fuente: arXiv
Saved in:
Bibliographic Details
Main Authors: Paglieri, Davide, Cupiał, Bartłomiej, Cook, Jonathan, Piterbarg, Ulyana, Tuyls, Jens, Grefenstette, Edward, Foerster, Jakob Nicolaus, Parker-Holder, Jack, Rocktäschel, Tim
Format: Preprint
Published: 2025
Subjects:
Online Access:
Tags: Add Tag
No Tags, Be the first to tag this record!
_version_ 1866917277301997568
author Paglieri, Davide
Cupiał, Bartłomiej
Cook, Jonathan
Piterbarg, Ulyana
Tuyls, Jens
Grefenstette, Edward
Foerster, Jakob Nicolaus
Parker-Holder, Jack
Rocktäschel, Tim
author_facet Paglieri, Davide
Cupiał, Bartłomiej
Cook, Jonathan
Piterbarg, Ulyana
Tuyls, Jens
Grefenstette, Edward
Foerster, Jakob Nicolaus
Parker-Holder, Jack
Rocktäschel, Tim
contents Training large language models (LLMs) to reason via reinforcement learning (RL) significantly improves their problem-solving capabilities. In agentic settings, existing methods like ReAct prompt LLMs to explicitly plan before every action; however, we demonstrate that always planning is computationally expensive and degrades performance on long-horizon tasks, while never planning further limits performance. To address this, we introduce a conceptual framework formalizing dynamic planning for LLM agents, enabling them to flexibly decide when to allocate test-time compute for planning. We propose a simple two-stage training pipeline: (1) supervised fine-tuning on diverse synthetic data to prime models for dynamic planning, and (2) RL to refine this capability in long-horizon environments. Experiments on the Crafter environment show that dynamic planning agents trained with this approach are more sample-efficient and consistently achieve more complex objectives. Additionally, we demonstrate that these agents can be effectively steered by human-written plans, surpassing their independent capabilities and highlighting the potential for safer and more collaborative agentic systems.
format Preprint
id arxiv_https___arxiv_org_abs_2509_03581
institution arXiv
publishDate 2025
record_format arxiv
spellingShingle Learning When to Plan: Efficiently Allocating Test-Time Compute for LLM Agents
Paglieri, Davide
Cupiał, Bartłomiej
Cook, Jonathan
Piterbarg, Ulyana
Tuyls, Jens
Grefenstette, Edward
Foerster, Jakob Nicolaus
Parker-Holder, Jack
Rocktäschel, Tim
Artificial Intelligence
Training large language models (LLMs) to reason via reinforcement learning (RL) significantly improves their problem-solving capabilities. In agentic settings, existing methods like ReAct prompt LLMs to explicitly plan before every action; however, we demonstrate that always planning is computationally expensive and degrades performance on long-horizon tasks, while never planning further limits performance. To address this, we introduce a conceptual framework formalizing dynamic planning for LLM agents, enabling them to flexibly decide when to allocate test-time compute for planning. We propose a simple two-stage training pipeline: (1) supervised fine-tuning on diverse synthetic data to prime models for dynamic planning, and (2) RL to refine this capability in long-horizon environments. Experiments on the Crafter environment show that dynamic planning agents trained with this approach are more sample-efficient and consistently achieve more complex objectives. Additionally, we demonstrate that these agents can be effectively steered by human-written plans, surpassing their independent capabilities and highlighting the potential for safer and more collaborative agentic systems.
title Learning When to Plan: Efficiently Allocating Test-Time Compute for LLM Agents
topic Artificial Intelligence
url https://arxiv.org/abs/2509.03581