---
title: CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining
url: https://www.ml-quant.com/papers/arxiv/2611.04555/
site: ML-Quant (https://www.ml-quant.com)
updated: 2026-11-06
license: Summaries CC BY 4.0; links go to the original sources
index: https://www.ml-quant.com/llms.txt
identifier: arXiv:2611.04555
source_url: https://arxiv.org/abs/2611.04555
featured: 2026-11-06
citations: unknown
topic: ML & AI Methods
---


# CosmosAlpha: Multi-Relational Graph Search with Pareto-Guided Exploration for LLM-Driven Alpha Mining

A graph-structured framework discovers alpha factors using an LLM, connecting factors by generation, similarity, and correlation while applying Pareto-guided search to balance quality, fertility and diversity.

- Source: https://arxiv.org/abs/2611.04555
- Identifier: arXiv:2611.04555
- Released: 2026-11-04
- First featured: Quant Letter No. 138 (2026-11-06): https://www.ml-quant.com/issues/2026-11-06/
- Citations (Semantic Scholar): not tracked
- Published in: not yet
- Topic: ML & AI Methods
- Authors: Zhan-He Wang, Zi-Yang Lang, Yi-Yang Li, Qing-Yu Lan, Yun-Song Piao

## Abstract (arXiv, CC0)

Discovering predictive alpha factors from noisy financial data is a fundamental yet challenging problem in quantitative finance. LLM-driven formulaic alpha mining offers an attractive paradigm by combining interpretable symbolic expressions with iterative generation and refinement guided by quantitative feedback. However, existing generation-centric and tree-based search strategies primarily organize factors through generation lineage, leaving structural and statistical relationships across search branches underutilized. Consequently, information remains localized along parent--child trajectories, which can confine exploration to redundant regions. Moreover, effective factor discovery must account for multiple distinct objectives that cannot be adequately captured by a single search criterion. We introduce CosmosAlpha, a graph-based LLM factor-mining framework that represents discovered factors as a multi-relational graph encoding generation lineage, expression-structure similarity, and cross-sectional exposure correlation. A Graph Flow mechanism propagates search utilities across the graph, allowing each factor to be assessed using relational evidence from nearby factors rather than in isolation. Pareto-guided selection jointly considers quality, fertility and diversity, explicitly modeling their trade-offs within a multi-objective search framework. Across CSI 300 and CSI 500, CosmosAlpha-Graph achieves the best mean rank over seven predictive and portfolio metrics and outperforms its controlled CosmosAlpha-MCTS counterpart on every reported metric, showing a consistent advantage for relational graph search in LLM-driven factor mining. The source code is available at https://github.com/wzhmtldswz/CosmosAlpha

## Related

- [AlphaPilot: Multi-Turn Tool-Using LLM Agents for Adaptive Alpha Mining with Reinforcement Learning](https://www.ml-quant.com/papers/arxiv/2610.15488/): A reinforcement-learning framework trains multi-turn tool-using LLM agents to discover factor expressions by conditioning on market descriptions and querying real-data feedback.
- [Jumping Ahead: Improving Reconstruction Fidelity with JumpReLU Sparse Autoencoders](https://www.ml-quant.com/papers/arxiv/2407.14435/): The paper presents JumpReLU Sparse Autoencoders (SAEs), which provide high-quality reconstruction of language model activations at a specific sparsity level, while maintaining interpretability.
- [AgentClinic: a multimodal agent benchmark to evaluate AI in simulated clinical environments](https://www.ml-quant.com/papers/arxiv/2405.07960/): AI Evaluation in Clinical Environments: The paper introduces AgentClinic, a benchmark for assessing large language models in simulated clinical environments, highlighting the significant impact of biases on diagnostic accuracy and patient interactions.
- [Learning Performance-Improving Code Edits](https://www.ml-quant.com/papers/arxiv/2302.07867/): The research presents a framework for optimizing programs using large language models, achieving a mean speedup of 6.86, outperforming average individual programmers.
- [Generative agent-based modeling with actions grounded in physical, social, or digital space using Concordia](https://www.ml-quant.com/papers/arxiv/2312.03664/): Concordia is a library designed to help build and operate Generative Agent-Based Models (GABMs), using Large Language Models (LLMs) to simulate physical or digital environments.
- [Neuro-Symbolic AI in 2024: A Systematic Review](https://www.ml-quant.com/papers/arxiv/2501.05435/): Neuro-Symbolic AI has grown since 2020, focusing on learning and inference, but still lacks in areas like explainability, trustworthiness, and Meta-Cognition.