Research map / Multi-agent systems and games / LLM agents

AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search

AAAI 2026 · Multi-agent systems and games › LLM agents · approach: LLM multi-agent decision making · 2 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes AgentSwift, which searches a hierarchical space of agentic workflows plus memory, tool-use and planning components, using a distilled value model and uncertainty-guided hierarchical MCTS to cut evaluation cost.

Problem

Automated LLM agent design is limited to narrow search spaces (mostly workflows), and evaluating each candidate agent on benchmarks is expensive, while search strategies explore large design spaces inefficiently.

Approach

Jointly search workflow and functional components (memory, tool use, planning) in one hierarchical space; train a value model on a covering-array plus balanced Bayesian sampled dataset to predict agent performance; run MCTS whose selection favors high-predicted and high-uncertainty candidates, with recombination, mutation and refinement expansions.

Key results

• AgentSwift reaches 0.806 on ALFWorld vs 0.701 for the best baseline (AgentSquare) with GPT-4o-mini, and 0.634 on M3ToolEval vs 0.561.
• Across seven benchmarks AgentSwift achieves the top score on every task in the GPT-4o-mini table; the paper reports an average gain of 8.34% over baselines.
• The value model (Qwen2.5-7B backbone) reaches MSE 0.0054 and Spearman 0.899 on held-out agent designs, far above GPT-4o few-shot (MSE 0.0162, Spearman 0.765).

Surprising findings

• Several hand-crafted agents score near zero on some tasks (e.g., Self-Refine and Step-Back on Travel), showing high sensitivity of fixed workflows to task.
• Full evaluation of every candidate progresses slowest because budget is spent on low-performing agents, so selective prediction-based evaluation beats it.

Limitations

• Mutation and refinement rely on heuristic, manually crafted LLM prompts rather than learned generation.
• The value model still requires labeled agent-performance data, which can be expensive in interactive environments.
• The search is static and does not adapt agents during deployment.

Experimental setup

Datasets: ALFWorld ScienceWorld MATH WebShop M3ToolEval TravelPlanner PDDL

Benchmarks: ALFWorld ScienceWorld MATH WebShop M3ToolEval TravelPlanner PDDL

Metrics: success rate MSE MAE R^2 Spearman correlation

Baselines: CoT CoTSC Tree-of-Thought Fleet-of-Agents Thought Propagation Self-Refine DILU Voyager DEPS Step-Back AgentSquare AFlow

Builds on: AFlow AgentSquare ADAS MaAS

Closest papers by meaning

• AgentSquare: Automatic LLM Agent Search in Modular Design Space (ICLR 2025 · similarity 0.95)
• Automated Design of Agentic Systems (ICLR 2025 · similarity 0.91)
• Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.91)
• Multi-Agent Design: Optimizing Agents with Better Prompts and Topologies (ICLR 2026 · similarity 0.90)
• AOrchestra: Automating Sub-Agent Creation for Agentic Orchestration (ICML 2026 · similarity 0.90)
• AgentBench: Evaluating LLMs as Agents (ICLR 2024 · similarity 0.89)