Paper: arXiv 2609.33524

Authors: Siyuan Li, Jiangfeng Zhang, Rui Yao, Weihua Qiu, Mingyang Xu, Zixuan Yuan

Abstract

Self-evolving agents aim to turn research feedback into reusable skills, tools, and research rules. Whether these accumulated capabilities continue to improve later research requires controlled evaluation. Long-horizon alpha discovery provides a state-dependent setting: once a new factor enters the portfolio, the predictive information already covered changes, so the value of the same candidate or experience may change over time. We introduce EverMine, an empirical framework for studying self-evolving research capabilities in long-horizon alpha discovery. EverMine decomposes the research state into history (Hist), the current factor portfolio (Frontier), and reusable capabilities (Cap). Under matched resource limits, we compare complete runs with fixed or evolving Cap, and replace Cap while holding Hist and Frontier fixed to estimate the conditional value of accumulated capabilities. We also combine full trajectories with historical-state replay to examine how experience-based decisions affect candidate selection and portfolio outcomes. Across 18 long-horizon trajectories, end-to-end comparisons show no consistent gain from Cap evolution. Across 48 continuation branches from shared Hist and Frontier states, accumulated Cap also does not consistently outperform the initial Cap. Parameter tuning of existing factor structures can still improve the portfolio. In an exploratory replay of two screening batches from one Evolving trajectory, some screened-out candidates have positive marginal value at the original state, yet submitting all screened-out candidates sequentially slightly lowers final portfolio IC in both batches. These results show that candidate value depends on the evolving portfolio and submission order, and motivate evaluating self-evolving research capabilities through end-to-end outcomes, conditional capability value, and the consequences of experience-based decisions.

Complexity vs Empirical Score

  • Math Complexity: 6.0/10
  • Empirical Rigor: 7.0/10
  • Quadrant: Holy Grail — high math complexity, high empirical rigor

Why this score: This paper introduces a novel framework, EverMine, to dissect the self-evolution of research capabilities in long-horizon alpha discovery. It combines rigorous empirical evaluation with a clear conceptual model, providing valuable insights into the effectiveness of evolving research agents. The findings, while showing no consistent gain from Cap evolution, are significant for future research in this area.

Research Flowchart

  flowchart TD
    A["Research Goal: Evaluate Self-Evolving Research Capabilities in Alpha Discovery"] --> B{"Key Methodology: EverMine Framework"};
    B --> C["Inputs: History (Hist), Factor Portfolio (Frontier), Reusable Capabilities (Cap)"];
    C --> D{"Computational Process: Comparative Runs"};
    D --> E["Outcomes: End-to-end Comparisons, Conditional Cap Value, Experience Replay"];
    E --> F{"Key Finding 1: No Consistent Gain from Cap Evolution"};
    E --> G{"Key Finding 2: Candidate Value Depends on Portfolio & Submission Order"};
    F & G --> H["Conclusion: Evaluate Self-Evolving Capabilities via End-to-End Outcomes, Conditional Value, & Experience Effects"];

Evidence Card

What the authors claim, on what data, and who has checked it. Read from the paper's own text by a language model: every quote is verbatim, nothing was re-run. How to read this card

Proposes

The paper proposes EverMine, an empirical framework that separates research history, current factor portfolio, and reusable capabilities to study self‑evolving research in long‑horizon alpha discovery.

Source sentence
We introduce EverMine, an empirical framework for studying self-evolving research capabilities in long-horizon alpha discovery.

Assumes 4

  • Research uses a point‑in‑time universe of 60 USDT‑denominated spot trading pairs selected by turnover, with no future information used. market
    Source sentence
    as the only raw market‑data source. The universe contains USDT‑denominated spot trading pairs, with all timestamps in UTC
  • All features, labels, and asset selection at time t use only information visible within the corresponding evaluation boundary. data
    Source sentence
    All processing follows one rule: asset selection, features, and labels at time t use only information visible within the corresponding evaluation boundary.
  • The factor pool has a fixed capacity of 30 and follows a fixed update rule during the development period. model
    Source sentence
    The evaluator maintains a factor pool with capacity 30 under a fixed update rule and measures predictive
  • Research runs are resource‑matched across conditions, ensuring comparable computational and evaluation budgets. other
    Source sentence
    Under matched resource limits, we compare complete research runs with either fixed or continuously evolving Cap.

Data

  • Binance Spot dynamic‑U60 dataset · Binance Data Vision archive market-prices public
    2025-01-01 to 2025-12-31 (development), · Top 60 USDT spot trading pairs by turnover each month
    Source sentence
    as the only raw market‑data source… The universe contains USDT‑denominated spot trading pairs

Recognised sources: Binance

Costs

Not applicable The paper reports no trading or execution result.

Finds 3

  • Across 18 long‑horizon trajectories and 48 continuation branches, there is no consistent improvement from explicit capability evolution. negative
    Source sentence
    Across 18 long‑horizon trajectories, end‑to‑end comparisons show no consistent gain from Cap evolution
  • Parameter tuning of existing factor structures yields a positive net improvement in 16 of 18 trajectories. performance
    Source sentence
    Parameter tuning of existing factor structures contributes a positive net improvement in 16 of 18 trajectories, showing that familiar structures can remain useful targets for further research.
  • Submitting all screened‑out candidates sequentially slightly lowers the final portfolio IC in both replayed batches. negative
    Source sentence
    After all screened‑out factors are sequentially submitted, the final portfolio IC decreases slightly in both batches.

Checked by others

No other paper in the archive has been found to support or contradict these findings (compared with 9 related papers). Coverage is partial, so this is neither replication nor refutation.

Code

No repository by the authors found in the paper.

Extracted 2026-11-14 · same card as JSON · A language model chose these items: check the source sentence before relying on one.

Cited References (47)

  1. AEL: Evolving Agent Harness in Open-Ended Environments (2026) — link
  2. AlphaAgent: LLM-Driven Alpha Mining with Regularized Exploration to Counteract Alpha Decay (2025) — link
  3. AlphaAgentEvo: Evolution-Oriented Alpha Mining via Self-Evolving Agentic Reinforcement Learning (2026) — link
  4. AlphaMemo: Structured Search-Process Memory for Self-Evolving Alpha Mining Agents (2026) — link
  5. The Deflated Sharpe Ratio: Correcting for Selection Bias, Backtest Overfitting, and Non-Normality (2014) — link
  6. AlphaSAGE: Structure-Aware Alpha Mining via GFlowNets for Robust Exploration — → 2509.25055
  7. Cognitive Alpha Mining via LLM -Driven Code-Based Evolution (2026) — link
  8. Approximate Submodularity and its Applications: Subset Selection, Sparse Approximation and Dictionary Selection (2018) — link
and 39 more
  1. DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression (2026) — link
  2. DeepSeek Harness: Everything is a Plugin (2026)
  3. Large Language Model Agent in Financial Trading: A Survey — → 2408.06361
  4. FactorMiner: A Self-Evolving Agent with Skills and Experience Memory for Financial Alpha Discovery — → 2602.14670
  5. Taming the Factor Zoo: A Test of New Factors (2020) — link
  6. FinAcumen: Financial Multimodal Reasoning via Self-Evolving Experience Memory Harness (2026) — link
  7. FinEvolveBench: A Benchmark for Self-Evolving Agents on Low-Repetition Tasks with Implicit Rewards (2026) — link
  8. FinMem: A Performance-Enhanced LLM Trading Agent With Layered Memory and Character Design — → 2311.13743
  9. The Characteristics that Provide Independent Information about Average U.S. Monthly Stock Returns (2017) — link
  10. AQuA: Recursively Self-Improving Quantitative Trading Research Agents (2026) — link
  11. … and the Cross-Section of Expected Returns (2015) — link
  12. Evaluating Memory in LLM Agents via Incremental Multi-Turn Interactions (2026) — link
  13. Reasoning on Time-Series for Financial Technical Analysis — → 2511.08616
  14. From Knowing to Doing: A Memory-Controlled Benchmark for LLM Trading Agents on Stock Markets (2026) — link
  15. LLMs Get Lost In Multi-Turn Conversation (2026) — link
  16. Time Travel is Cheating: Going Live with DeepFund for Real-Time Fund Investment Benchmarking (2025) — link
  17. AutoScientist-Quant: Self-Evolving Coding Agents for Automatic Research in Quantitative Investment (2026) — link
  18. AI Trading’s Alpha Singularity: Emergent Market Reasoning through Agent-to-Agent Self-Evolution (2026) — link
  19. AlphaBench: Benchmarking Large Language Models in Formulaic Alpha Factor Mining (2026) — link
  20. Harness-Aware Self-Evolving: Co-Evolving Model Weights, Harness, and Task Solutions (2026) — link
  21. Does Academic Research Destroy Stock Return Predictability? (2016) — link
  22. Agentic Empirical Asset Pricing: Methodological Foundations (2026) — link
  23. QuantaAlpha: An Evolutionary Framework for LLM-Driven Alpha Mining — → 2602.07085
  24. R&D-Agent-Quant: A Multi-Agent Framework for Data-Centric Factors and Model Joint Optimization — → 2505.15155
  25. A Survey of Self-Evolving Agents: What, When, How, and Where to Evolve on the Path to Artificial Super Intelligence (2026) — link
  26. Your Agent May Misevolve: Emergent Risks in Self-evolving LLM Agents (2026) — link
  27. A Programming Paradigm for Spatiotemporal Composability (2026) — link
  28. Reflexion: Language Agents with Verbal Reinforcement Learning (2023) — link
  29. The Illusion of Diminishing Returns: Measuring Long Horizon Execution in LLMs (2026) — link
  30. Trade in Minutes! Rationality-Driven Agentic System for Quantitative Financial Trading (2026) — link
  31. Qwen3.8-Max : A New Bar for Coding and Cowork (2026) — link
  32. Tree-of-Experience: Hierarchical Experience Management for Self-Evolving Agents (2026) — link
  33. BacktestBench: Benchmarking Large Language Models for Automated Quantitative Strategy Backtesting (2026) — link
  34. Qlib: An AI-oriented Quantitative Investment Platform (2020) — link
  35. Generating Synergistic Formulaic Alpha Collections via Reinforcement Learning — → 2306.12964
  36. A survey on large language model-based alpha mining (2025) — link
  37. Agentic Context Engineering: Evolving Contexts for Self-Improving Language Models (2026) — link
  38. ExpeL: LLM Agents Are Experiential Learners (2024) — link
  39. A lpha QCM : Alpha Discovery in Finance with Distributional Reinforcement Learning (2025) — link

8 of 47 cited works are in the corpus; resolved via arXiv source bundles, Semantic Scholar and OpenAlex. Corpus-wide: 2596 linked papers.