LATS (Language Agent Tree Search)

Unified Reasoning-Action-Planning Framework for AI Agents L APIs & Messaging

Basic Information

  • Type: Unified Reasoning-Action-Planning Framework for AI Agents
  • Paper: "Language Agent Tree Search Unifies Reasoning Acting and Planning in Language Models" (2024)
  • Authors: Andy Zhou et al. (University of Illinois)
  • Publication: ICML 2024
  • GitHub: https://github.com/lapisrocks/LanguageAgentTreeSearch

Paradigm Description

LATS is the first general framework that unifies the reasoning, action, and planning capabilities of LLMs. It integrates Monte Carlo Tree Search (MCTS) into language agents, combining LLM-driven value functions and self-reflection mechanisms to achieve efficient exploration and enhanced decision-making. The core innovation of LATS lies in introducing external environmental feedback, providing a more prudent and adaptive problem-solving mechanism than existing technologies.

Core Mechanisms

  • Monte Carlo Tree Search: Introduces MCTS into language agents for systematic search
  • LLM Triple Roles:
  • Action Generator: Samples reasonable actions at each tree node
  • Value Function: Estimates future expected rewards
  • Reflection Mechanism: Generates self-criticism for suboptimal trajectories
  • External Environment Feedback: Uses environmental feedback to guide search direction
  • Backtracking Capability: Can revert to previous states for re-exploration

Performance

  • HumanEval Programming Task: GPT-4 achieves 92.7% pass@1 accuracy (state-of-the-art at the time)
  • WebShop Web Navigation: GPT-3.5 achieves an average score of 75.9, comparable to gradient fine-tuning methods
  • Achieves near-fine-tuning performance without gradient updates

Technical Evolution (2025)

  • Genetic Particle Filtering: Search variant replacing MCTS
  • Bayesian Tree Optimization: Acquisition function with uncertainty guidance
  • Hierarchical Agent Design: Layered agent search architecture
  • Stepwise Q-Guided Search: Uses Q-values to guide search direction
  • Autonomous Agent Orchestration: Scalable multi-agent reasoning

Relationship with Other Paradigms

  • vs ReAct: LATS adds systematic exploration through tree search on top of ReAct
  • vs ToT: LATS combines ToT's tree search with ReAct's environmental interaction
  • vs Reflexion: Reflexion uses cross-trial reflection, while LATS uses intra-trial reflection to guide search

Relationship with OpenClaw Ecosystem

LATS provides the most powerful unified reasoning-action-planning framework for OpenClaw agents. When agents face complex tasks requiring systematic exploration (e.g., multi-solution comparison, optimal path search), LATS can help agents efficiently search for optimal solutions in the action space. Due to high computational costs, it is recommended that OpenClaw enable LATS only for high-complexity tasks.