Research map / Large language models / Agents, code and math
Fleet of Agents: Coordinated Problem Solving with Large Language Models
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 Fleet of Agents (FoA), a genetic-type particle filtering runtime that runs a fleet of LLM agents over dynamic tree searches with resampling by heuristic value, improving cost-quality trade-off on Game of 24, Mini Crosswords and WebShop.
Problem
LLM reasoning frameworks trade off cost and quality poorly: single-query methods are cheap but weak, while multi-query tree-search methods are accurate but very expensive.
Approach
Treat LLM agents as particles in a genetic particle filter: each agent mutates its state independently for k steps, then the fleet is resampled with replacement according to a heuristic value function, with backtracking to past visited states and discounting by gamma, giving dynamic branching and predictable cost.
Key results
- On Game of 24 with GPT-4, FoA reaches 76.0% success at $62.93 versus 74.0% for ToT at $75.02.
- On WebShop with GPT-3.5, FoA scores 75.6 average score at $1.68 versus 66.1 for LATS at $232.27.
- On average across tasks and LLMs, FoA gains about 5% quality at about 40% of SOTA cost; Llama3.2-11B + FoA surpasses Llama3.2-90B.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use FoA instead of ToT | Game of 24 with GPT-4 · Game of 24 test | success rate (%): 74.0 -> 76.0 | Helped |
| Use FoA instead of GoT | Mini Crosswords with GPT-4 · Mini Crosswords test | overlap (%): 41.2 -> 46.0 | Helped |
| Use FoA instead of the strongest reported baseline, WebGUM | WebShop · WebShop test | average score: 67.5 -> 75.6 | Helped |
| Add the selection phase (resampling) rather than running agents without selection (ablation) | FoA · Game of 24 | success rate | Helped |
| Use resampling rather than assigning every agent to the highest-value state (ablation) | FoA · Game of 24 | success rate | Helped |
| Add backtracking with discounted past-state resampling rather than setting γ = 0 (ablation) | FoA · Game of 24 | success rate | Helped |
| Use discounted past-state resampling rather than considering all past states without decay (γ = 1) (ablation) | FoA · Game of 24 | success rate | Helped |
| Enable caching rather than disabling the cache (ablation) | FoA · Game of 24 | success rate | Helped |
| Enable batching rather than disabling batching (ablation) | FoA · Game of 24 | quality | No clear effect |
Surprising findings
- Disabling the cache not only increases cost but also lowers success rate, which the authors attribute to inconsistent re-evaluations of the same state.
Limitations
- Fixed fleet size across tasks; adaptive allocation left to future work.
- Homogeneous fleet of identical agents; no hierarchical coordination explored.
- GPT-4 results were run once due to cost, and many baselines were not run with GPT-4 or on every task.
Experimental setup
Datasets: Game of 24 Mini Crosswords WebShop
Benchmarks: Game of 24 Mini Crosswords WebShop
Metrics: success rate overlap average score cost (US$)
Baselines: IO prompting CoT CoT-SC AoT ToT GoT RAFA Act ReAct Reflexion LASER LATS
Builds on: Tree of Thoughts genetic particle filtering LATS Chain-of-Thought
Closest papers by meaning
- Language Agent Tree Search Unifies Reasoning, Acting, and Planning in Language Models (ICML 2024 · similarity 0.92)
- AgentSwift: Efficient LLM Agent Design via Value-Guided Hierarchical Search (AAAI 2026 · similarity 0.91)
- AgentTTS: Large Language Model Agent for Test-time Compute-optimal Scaling Strategy in Complex Tasks (NeurIPS 2025 · similarity 0.90)
- SE-Agent: Self-Evolution Trajectory Optimization in Multi-Step Reasoning with LLM-Based Agents (NeurIPS 2025 · similarity 0.90)
- Chain of Agents: Large Language Models Collaborating on Long-Context Tasks (NeurIPS 2024 · similarity 0.90)
- Distilling LLM Agent into Small Models with Retrieval and Code Tools (NeurIPS 2025 · similarity 0.90)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.