Fact-Augmented Lookahead Planning for LLM Agents

Samuel Holt, Max Ruiz Luyten, Thomas Pouplin, Mihaela van der Schaar
Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, PMLR 300:3205-3213, 2026.

Abstract

Large Language Models (LLMs) are increasingly capable, but LLM agents still struggle to plan effectively in interactive, partially observable, long-horizon environments when search is unguided or recent history is insufficient. We introduce LWM-Planner, a fact-augmented lookahead planning framework that improves agent behavior purely through in-context learning. After each episode, the agent extracts task-critical atomic facts from its trajectories, validates candidates with a lightweight predictive-consistency filter (and optionally compresses them), and uses the resulting fact set to condition action proposal, single-step latent world-model simulation, and state-value estimation. Planning then proceeds via recursive, depth-limited lookahead over candidate trajectories conditioned on the accumulated facts and recent history, enabling online improvement without parameter updates. We provide abstraction-style motivation—treating facts as reducing state aliasing (proxy $\epsilon_{\mathrm{sim}}$) and fact-conditioned simulation as lowering one-step error (proxy $\delta_{\mathrm{model}}$)—without claiming formal guarantees. Empirically, on text FrozenLake variants, CrafterMini, and ALFWorld, the approach improves cumulative return over ReAct/Reflexion and search-only baselines, suggesting that additional test-time search is most useful when grounded by compact, experience-derived facts.

Cite this Paper

BibTeX
@InProceedings{pmlr-v300-holt26a, title = { Fact-Augmented Lookahead Planning for LLM Agents }, author = {Holt, Samuel and Luyten, Max Ruiz and Pouplin, Thomas and van der Schaar, Mihaela}, booktitle = {Proceedings of The 29th International Conference on Artificial Intelligence and Statistics}, pages = {3205--3213}, year = {2026}, editor = {Khan, Emtiyaz and Li, Yingzhen and Solin, Arno and Ramdas, Aaditya}, volume = {300}, series = {Proceedings of Machine Learning Research}, month = {02--05 May}, publisher = {PMLR}, pdf = {https://raw.githubusercontent.com/mlresearch/v300/main/assets/holt26a/holt26a.pdf}, url = {https://proceedings.mlr.press/v300/holt26a.html}, abstract = { Large Language Models (LLMs) are increasingly capable, but LLM agents still struggle to plan effectively in interactive, partially observable, long-horizon environments when search is unguided or recent history is insufficient. We introduce LWM-Planner, a fact-augmented lookahead planning framework that improves agent behavior purely through in-context learning. After each episode, the agent extracts task-critical atomic facts from its trajectories, validates candidates with a lightweight predictive-consistency filter (and optionally compresses them), and uses the resulting fact set to condition action proposal, single-step latent world-model simulation, and state-value estimation. Planning then proceeds via recursive, depth-limited lookahead over candidate trajectories conditioned on the accumulated facts and recent history, enabling online improvement without parameter updates. We provide abstraction-style motivation—treating facts as reducing state aliasing (proxy $\epsilon_{\mathrm{sim}}$) and fact-conditioned simulation as lowering one-step error (proxy $\delta_{\mathrm{model}}$)—without claiming formal guarantees. Empirically, on text FrozenLake variants, CrafterMini, and ALFWorld, the approach improves cumulative return over ReAct/Reflexion and search-only baselines, suggesting that additional test-time search is most useful when grounded by compact, experience-derived facts. } }
Endnote
%0 Conference Paper %T Fact-Augmented Lookahead Planning for LLM Agents %A Samuel Holt %A Max Ruiz Luyten %A Thomas Pouplin %A Mihaela van der Schaar %B Proceedings of The 29th International Conference on Artificial Intelligence and Statistics %C Proceedings of Machine Learning Research %D 2026 %E Emtiyaz Khan %E Yingzhen Li %E Arno Solin %E Aaditya Ramdas %F pmlr-v300-holt26a %I PMLR %P 3205--3213 %U https://proceedings.mlr.press/v300/holt26a.html %V 300 %X Large Language Models (LLMs) are increasingly capable, but LLM agents still struggle to plan effectively in interactive, partially observable, long-horizon environments when search is unguided or recent history is insufficient. We introduce LWM-Planner, a fact-augmented lookahead planning framework that improves agent behavior purely through in-context learning. After each episode, the agent extracts task-critical atomic facts from its trajectories, validates candidates with a lightweight predictive-consistency filter (and optionally compresses them), and uses the resulting fact set to condition action proposal, single-step latent world-model simulation, and state-value estimation. Planning then proceeds via recursive, depth-limited lookahead over candidate trajectories conditioned on the accumulated facts and recent history, enabling online improvement without parameter updates. We provide abstraction-style motivation—treating facts as reducing state aliasing (proxy $\epsilon_{\mathrm{sim}}$) and fact-conditioned simulation as lowering one-step error (proxy $\delta_{\mathrm{model}}$)—without claiming formal guarantees. Empirically, on text FrozenLake variants, CrafterMini, and ALFWorld, the approach improves cumulative return over ReAct/Reflexion and search-only baselines, suggesting that additional test-time search is most useful when grounded by compact, experience-derived facts.
APA
Holt, S., Luyten, M.R., Pouplin, T. & van der Schaar, M.. (2026). Fact-Augmented Lookahead Planning for LLM Agents . Proceedings of The 29th International Conference on Artificial Intelligence and Statistics, in Proceedings of Machine Learning Research 300:3205-3213 Available from https://proceedings.mlr.press/v300/holt26a.html.

Related Material