Research map / Large language models / Agents, code and math
AFlow: Automating Agentic Workflow Generation
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 AFlow, which automates generation and optimization of agentic workflows by searching over code-represented workflows with MCTS guided by execution feedback.
Problem
Constructing agentic workflows for LLMs requires substantial human effort, limiting scalability and generalizability; the paper targets automating workflow generation and optimization.
Approach
Represents workflows as code-based node graphs and searches over them with Monte Carlo tree search, where an LLM-based optimizer rewrites workflow nodes and edges using execution feedback on validation tasks.
Key results
- Reported to outperform manually designed and prior automated workflow baselines across six benchmarks (per the abstract and tables; exact numbers not in the provided excerpt).
- Enables smaller open-source models to match GPT-4o performance on some tasks at lower inference cost (as claimed in the paper summary).
Limitations
- The provided text is almost entirely the ICLR style template; the method and results sections are not available, so details are inferred from the abstract only.
Experimental setup
Datasets: HumanEval MBPP GSM8K MATH HotpotQA DROP
Benchmarks: HumanEval MBPP GSM8K MATH HotpotQA DROP
Metrics: accuracy pass@1 F1 cost
Baselines: CoT ComplexCoT MultiPersona Self-Consistency LLM-Debate Self-Refine ADAS AutoAgents
Builds on: chain-of-thought ReAct MaAS AutoGPT
Closest papers by meaning
- WorkflowLLM: Enhancing Workflow Orchestration Capability of Large Language Models (ICLR 2025 · similarity 0.89)
- Benchmarking Agentic Workflow Generation (ICLR 2025 · similarity 0.89)
- Optimize Once, Execute Fast: Latency-Aware Multi-Agent Workflow Learning for Recurrent Queries (NeurIPS 2026 · similarity 0.88)
- Steering Large Language Models between Code Execution and Textual Reasoning (ICLR 2025 · similarity 0.88)
- DyFlow: Dynamic Workflow Framework for Agentic Reasoning (NeurIPS 2025 · similarity 0.88)
- Fleet of Agents: Coordinated Problem Solving with Large Language Models (ICML 2025 · similarity 0.88)
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.