Research map / Multi-agent systems and games / LLM agents
AlphaAgentEvo: Evolution-Oriented Alpha Mining via Self-Evolving Agentic Reinforcement Learning
Ziyi Tang, Xuexiong Yin, Weixing Chen, Zechuan Chen, Yongsen Zheng, Wenxuan Ye, Keze Wang, Liang Lin
Read the paperCheck your idea against 137,217 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 AlphaAgentEvo, a self-evolving agentic reinforcement learning approach for alpha factor mining that aims to evolve alphas using natural-language instructions and insights from failed attempts.
Problem
Alpha mining for predictive financial alpha factors from a vast, noisy search space; existing methods such as genetic programming cannot interpret natural-language instructions and fail to learn from unsuccessful attempts, hindering systematic alpha evolution.
Approach
Frames alpha mining as a self-evolving agentic reinforcement learning process so that the agent learns from unsuccessful attempts and interprets natural-language instructions to evolve alphas, in contrast to genetic programming.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| train Qwen3-4B-thinking with self-evolving agentic RL | Qwen3-4B-thinking · AlphaEvo500 HS300 | pass@3: 0.36 -> 0.97 | Improved |
| train Qwen3-1.7B with agentic RL | Qwen3-1.7B · AlphaEvo500 HS300 | pass@3: 0.08 -> 0.77 | Improved |
| use hierarchical reward ARL instead of ToolRL | Qwen3-4B · AlphaEvo500 HS300 | pass@5: 0.81 -> 0.97 | Improved |
| use AlphaAgentEvo-4B instead of GEPA with GPT-5-mini | evolution framework · AlphaEvo500 CSI500 | pass@3: 0.86 -> 0.93 | Improved |
| add exploration reward (ablation) | AlphaAgentEvo-4B · AlphaEvo500 | pass@3: 0.54 -> 0.65 | Improved |
| add consistency (direction-aware) reward (ablation) | AlphaAgentEvo-4B · AlphaEvo500 | pass@3: 0.51 -> 0.65 | Improved |
| add exploration reward (ablation) | AlphaAgentEvo-4B · Alpha158 | pass@3: 0.513 -> 0.581 | Improved |
| add consistency reward (ablation) | AlphaAgentEvo-4B · Alpha158 | pass@3: 0.510 -> 0.581 | Improved |
| use ARL-trained agent instead of genetic programming | alpha evolution · Alpha158 (2024-2025) | pass@5: 0.132 -> 0.994 | Improved |
| increase GP offspring per generation | GP · Alpha158 | pass@3: 4 offspring 0.003 -> 20 offspring 0.125 -> 50 offspring 0.094 (2024-2025) | Mixed |
| ARL training compared with untrained base model on diversity | AlphaAgentEvo-4B · top-20 alpha similarity | maximum similarity: 0.600 -> 0.263 (vs Qwen3-4B) | Improved |
Experimental setup
Baselines: genetic programming
Builds on: genetic programming
Closest papers by meaning
- AlphaPareto: Formulaic Alpha Discovery with LLM-Guided Multi-Objective Reinforcement Learning (NeurIPS 2026 · similarity 0.90)
- AlphaPROBE: Alpha Mining via principled retriveval and on-graph biased evolution (NeurIPS 2026 · similarity 0.87)
- AlphaQCM: Alpha Discovery in Finance with Distributional Reinforcement Learning (ICML 2025 · similarity 0.86)
- AlphaForge: A Framework to Mine and Dynamically Combine Formulaic Alpha Factors (AAAI 2025 · similarity 0.86)
- Self-evolving LLM agents with in-distribution Optimization (ICML 2026 · similarity 0.85)
- HiveMind: Contribution-Guided Online Prompt Optimization of LLM Multi-Agent Systems (AAAI 2026 · similarity 0.85)
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.
