PaperlayerView exact revision on arXiv

Source-linked AI summary

EverMine: Dissecting the Self-Evolution of Research Capabilities in Long-Horizon Alpha Research

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

arXiv:2609.33524v1cs.AIq-fin.ST

TL;DR

Long-horizon alpha research raises the question of whether accumulated capabilities continue to improve research as portfolio coverage changes candidate value. EverMine compares evolving and fixed capabilities, tests accumulated capabilities from matched states, and replays experience-based screening decisions; its experiments show no consistent advantage for capability evolution or accumulated capabilities.

  • Problem

    Because portfolio changes can alter a candidate’s marginal value, it remains an open question whether accumulated research capabilities improve subsequent work on an evolving frontier.

    Source: §1 · p. 2

  • Method

    EverMine compares fixed and evolving capabilities under matched resource limits, replaces capabilities at shared history and portfolio states, and uses trajectory statistics with historical-state replay.

    Source: §1 · p. 2

  • Results

    Experiments show no consistent end-to-end gains from Cap evolution or consistent advantage for accumulated Cap at matched research states.

    Source: §1 · p. 2

  • Takeaways & Limitations

    Candidate value depends on the evolving portfolio, and experience-based rules should be assessed through the choices they change and the resulting outcomes.

    Source: §1 · p. 2

    The 2026 holdout does not establish that the period was unseen during model training, limiting claims of post-training temporal generalization.

    Source: §6 · p. 9

Abstract

from arXiv · show

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.

1 INTRODUCTION

EverMine studies whether reusable research capabilities improve long-horizon alpha discovery, where candidate value changes as the portfolio evolves. It separates overall system performance from the conditional value of accumulated capabilities and traces how experience-based decisions affect candidate selection and portfolio outcomes.

1 INTRODUCTION

  • Candidate value depends on both its predictive power and the information already covered by the evolving portfolio, so the same candidate can have different marginal value across states.Earlier discoveries can change the remaining search space, and many new factors add limited independent information after controlling for existing factors.Source: §1 · p. 2
  • Prior work studies agents that accumulate experience and research resources, but endpoint comparisons mix changes in history, portfolio, and capabilities.Evaluating conditional capability value therefore requires comparing from the same research state.Source: §1 · p. 2
  • EverMine combines matched-resource Fixed-versus-Evolving runs with Cap replacement at fixed Hist and Frontier to distinguish overall system effects from conditional capability value.It also combines full trajectories with historical-state replay to examine how experience changes candidate selection and portfolio outcomes.Source: §1 · p. 2, Figure 1 · p. 3
  • Experiments show no consistent end-to-end gains from Cap evolution or consistent advantage for accumulated Cap at matched research states.Parameter tuning of existing factor structures contributes positive net improvement in 16 of 18 trajectories; sequentially submitting screened-out candidates slightly lowers final portfolio IC in both replay batches.Source: §1 · p. 2

2 RELATED WORK

Prior work addresses alpha search around existing factor sets, reusable experience in research agents, and evaluation across changing research states. EverMine focuses on capability reuse within one evolving long-horizon trajectory, combining matched-resource comparisons, matched-state replacement, and historical-state replay.

Alpha Discovery on an Evolving Predictive Frontier

  • Alpha discovery research has expanded from proposing individual factors to joint search around existing factor sets and changing factor pools.AlphaGen uses portfolio complementarity (Yu et al., 2023), while AlphaQCM and AlphaSAGE address changing pools and structure-aware generation (Chen et al., 2026).Source: §2 · p. 3

Reusable Experience and Self-Evolving Research Agents

Evaluating Adaptation across Changing Research States

  • Prior evaluation work separates research products from processes or compares frozen skills, whereas EverMine studies the evolving frontier within one long-horizon trajectory.It combines matched-resource end-to-end comparisons and same-Hist-and-Frontier capability replacement with replay of experience-based decisions.Source: §2 · pp. 3–4

3 PROBLEM SETUP

The setup treats alpha research as portfolio-conditioned: candidate value depends on the current portfolio and its covered information. It represents the evolving run through history, portfolio, and reusable capabilities.

3.1 PORTFOLIO-CONDITIONED ALPHA DISCOVERY

  • Long-horizon research consists of many rounds within one task, sharing history and an updated portfolio while data range and evaluation rules stay fixed.The agent proposes hypotheses, constructs factor expressions, and submits candidates to an evaluator.Source: §3.1 · p. 4
  • A candidate’s marginal portfolio gain is evaluated against the current Frontier, whose covered information can change the candidate’s value.Research progress is portfolio predictive performance; in experiments, Q is average per-bar cross-sectional Pearson correlation with the next 10-minute-bar return, and the portfolio is capped at 30 factors.Source: §3.1 · p. 4

3.2 RESEARCH STATE WITH CAPABILITIES

  • The research state is St = (Histt, Frontiert, Capt): history and its research map, the current portfolio, and reusable capabilities.Capabilities are skills, tools, and procedural documents that support later research and can be saved, retrieved, and revised.Source: §3.2 · p. 4
  • A capability can encode a conditional diagnostic rule, such as checking overlap with existing signals before choosing to tune a formula or explore another direction.Later evidence can support the rule or narrow its scope and add checks such as candidate stability.Source: §3.2 · p. 4

4 EXPERIMENTAL DESIGN

EverMine compares fixed and evolving capability states under matched resource limits, then evaluates accumulated capabilities from shared history and portfolio states. It also analyzes research choices and uses long-horizon experiments with specified market data and resource budgets.

4.1 FIXED AND EVOLVING RESEARCH AGENTS

  • Fixed keeps Cap at its initial state, while Evolving can save, retrieve, revise, or disable skills, tools, and procedures.Both conditions can adapt actions using research history, the current portfolio, shared skills, and platform feedback.Source: §4.1 · p. 5

4.2 END-TO-END EVALUATION

  • The end-to-end evaluation compares final portfolio quality under matched resource limits and also records trajectories, post-capacity improvements, and resource use.The comparison includes the costs of creating and using capabilities.Source: §4.2 · p. 5

4.3 CHECKPOINT-BASED CAPABILITY-STATE REPLACEMENT

  • Capability-state replacement branches from shared Hist and Frontier with either checkpoint Cap_t or initial Cap_0, then compares subsequent portfolio-quality gains.Branches are created at 25% and 75% joint-budget progress, with two repetitions per condition; results are averaged equally across parent trajectories.Source: §4.3 · p. 5

4.4 RESEARCH BEHAVIOR ANALYSIS

  • Research behavior analysis classifies post-capacity submissions and replays screened-out candidates independently at the original state and sequentially in their original order.Parameter tuning changes numerical parameters while retaining data fields and operator structure; the sequential replay updates the portfolio after each submission.Source: §4.4 · pp. 5–6

4.5 EXPERIMENTAL SETUP

  • Experiments use 10-minute Binance Spot USDT data, a 30-factor pool, and development and out-of-sample periods; 18 long trajectories and 48 continuation branches are evaluated.Runs use matched multi-resource limits; Figure 2 normalizes submission progress from 0 to 1000 within each trajectory.Source: §4.5 · p. 6, Figure 2 · p. 7, Table 1 · p. 7

5 RESULTS AND ANALYSIS

Across end-to-end runs, capability evolution shows no consistent gain, and accumulated capabilities do not consistently outperform initial capabilities in matched-state continuations. Parameter tuning can still improve portfolios, while screening replays reveal both missed candidate value and risks from sequential additions.

5.1 END-TO-END PERFORMANCE IN LONG-HORIZON RESEARCH

  • All four end-to-end Evolving-minus-Fixed difference intervals include zero, and the comparison shows no consistent gain from Cap evolution.Development-period means are 0.060902/0.061643 for Qwen Fixed / Evolving and 0.068166/0.068071 for DeepSeek Fixed / Evolving.Source: §5.1 · pp. 7–8

5.2 CONDITIONAL VALUE OF ACCUMULATED CAPABILITIES

  • At the 25% and 75% checkpoints, accumulated Cap differences are −0.000405 and +0.000042 in development, with all development and OOS intervals including zero.Both capability states improve the portfolio at both checkpoints; both OOS checkpoint differences favor initial Cap.Source: Table 2 · p. 8, §5.2 · p. 8

5.3 RESEARCH BEHAVIOR ANALYSIS

  • Parameter tuning contributes 12.8%–40.0% of group net improvement and has positive net contribution in 16 of 18 trajectories.Its mean net contribution remains positive for every group after removing any single trajectory.Source: §5.3 · p. 8
  • In two Evolving screening batches, four screened-out candidates had positive original-state gains, all above 10^-5, while the accepted Batch A candidate gained 0.00044635.The best screened-out Batch A candidate gained 0.00014081; the table describes independent original-state evaluation and fixed-order additions.Source: §5.3 · p. 8, Table 3 · p. 9

B. Screened-out factors: original-state gains and sequential-submission outcomes

The replay separates a candidate’s independent value at the original portfolio from the outcome of adding screened-out candidates sequentially. In both batches, the sequential additions lower final IC.

B. Screened-out factors: original-state gains and sequential-submission outcomes

  • Three of four screened-out factors with positive original-state gains become negative when submitted sequentially after earlier portfolio changes.Table 3B compares two complete screening batches from the same Qwen Evolving trajectory.Source: B. Screened-out factors:… · p. 9
  • Submitting all screened-out factors lowers final IC by 0.00001749 in Batch A and 0.00000666 in Batch B versus the original-decision endpoints.The sequential replay submits candidates in original order and updates the portfolio after each addition.Source: B. Screened-out factors:… · p. 9
  • Independent original-state gains and realized sequential-submission outcomes answer different questions about candidate value.The replay includes screened-out candidates with negative or zero original gains as well as those with positive gains.Source: B. Screened-out factors:… · p. 9

6 LIMITATIONS

The study’s conclusions are bounded by its capability-replacement design, limited repetitions, OOS uncertainty, and experimental setting.

6 LIMITATIONS

  • Effect intervals include zero, and the limited number of independent trajectories or parent trajectories reduces the precision of effect estimates.DeepSeek contributes three trajectories per condition, so its bootstrap intervals provide limited support for nominal 95% coverage.Source: §6 · p. 9
  • The OOS split holds 2026 data and labels out of the agent’s research loop but cannot establish that they were unseen during model pretraining.The behavioral analyses are exploratory, and broader controlled evaluation is needed to estimate screening rates across trajectories and long-run policy value.Source: §6 · p. 9
  • Experiments cover structured Cap, a fixed-capacity factor pool, and 10-minute cryptocurrency data; other representations, update rules, asset classes, and frequencies remain untested.Separate effects of skills, tools, and procedural documents require additional interventions.Source: §6 · pp. 9–10

7 CONCLUSION

EverMine finds no consistent improvement from accumulated capabilities, while portfolio gains can come from parameter tuning and research choices have mixed consequences. The authors recommend connecting recorded experience more directly to its use and resulting portfolio feedback.

7 CONCLUSION

  • Across 18 long trajectories and 48 continuation branches, EverMine observes no consistent improvement from capability evolution or replacement.Effect intervals in development and OOS evaluation include zero.Source: §7 · p. 10
  • Parameter tuning contributes a positive net improvement in 16 of 18 trajectories, while replaying every screened-out candidate sequentially slightly lowers final portfolio endpoints in both audited batches.Screening can retain the strongest candidate while missing other candidates with positive marginal value.Source: §7 · p. 10
  • Future capability design and evaluation should connect recorded experience more directly to its use and portfolio feedback.Evaluation should track when judgments are applied, which factors are continued or abandoned, and how new evidence leads to revision.Source: §7 · p. 10

ETHICS AND REPRODUCIBILITY STATEMENT

This material states the work’s ethics and reproducibility commitments, then surveys portfolio-conditioned alpha research and reusable experience in evolving agents.

ETHICS AND REPRODUCIBILITY STATEMENT

  • The authors describe the work as academic research, disclaim investment-advice status, and state that public data and reproducibility details are provided, with code to be released upon acceptance.The authors also state that AI-assisted organization and language editing were manually reviewed.Source: ETHICS AND REPRODUCIBILITY STATEMENT · p. 10, AI USE STATEMENT · p. 10

A.1 PORTFOLIO-CONDITIONED ALPHA DISCOVERY AND RESEARCH PROCESSES

  • Portfolio-conditioned alpha methods address factor complementarity, search-space coverage, and continued search as the factor pool changes.Prior financial studies find many candidates are absorbed by known factors, while formula-based methods target joint factor-set performance or broaden structural exploration.Source: §A.1 · pp. 14–15
  • Alpha-agent research has expanded from one-shot factor generation toward continuing processes that form hypotheses, test expressions, and revise research.The surveyed systems also organize exploration through constraints, tool use, planning, reflection, and feedback.Source: §A.1 · p. 15

A.2 REUSABLE EXPERIENCE AND SELF-EVOLVING RESEARCH AGENTS

  • Self-evolving agents turn interaction histories into reusable resources that can guide later research directions, candidate modifications, and prioritization.The surveyed approaches range from reflections and experience rules to skill libraries, process memories, and evolving research workflows; self-evolution can also reinforce undesirable patterns.Source: §A.2 · p. 15

A.3 EVALUATION OF ADAPTATION UNDER EVOLVING RESEARCH STATES

The reviewed work motivates evaluating adaptation through both full-process outcomes and controlled tests of accumulated experience. The material also describes shared protocol conditions and how Fixed and Evolving capabilities differ.

A.3 EVALUATION OF ADAPTATION UNDER EVOLVING RESEARCH STATES

  • Continual adaptive research calls for assessing both full-process outcomes and the effect of accumulated experience or capabilities on later actions.Prior work uses component comparisons, repeated discovery over time, and controlled variation of histories or interactions to study these questions.Source: §A.3 · pp. 15–16

A.4 EVALUATING ALPHA RESEARCH AND FINANCIAL EVIDENCE

  • Alpha research measures answer different questions at candidate, factor-set, and investment-outcome levels, while continual discovery also requires evaluation across many rounds.EverMine reports development-period search separately from OOS evaluation of the frozen final portfolio.Source: §A.4 · p. 16

B.1.1 SHARED RESEARCH PROTOCOL

  • All conditions share the research task, materials, submission interface, platform feedback, resource accounting, and a pre-specified status mechanism.The status reminder is triggered by the query alone, though query timing and subsequent use remain part of agent behavior.Source: §B.1.1 · p. 17

B.1.2 FIXED CONDITION

  • Fixed keeps the initial Cap read-only while allowing the agent to update history, respond to portfolio changes, and use permitted research tools.This preserves adaptation from historical information while fixing explicit capability evolution.Source: §B.1.2 · p. 17

B.1.3 EVOLVING CONDITION

  • Evolving adds writable, versioned capabilities that store reusable research judgments and execution methods, while specific factors and one-off facts mainly remain in history.Resources can be skills, procedural documents, or tools; updates are tied to research events, retrieval, action alignment, and evidence-based revision.Source: §B.1.3 · pp. 17–18

B.2 DATA

The appendix specifies the point-in-time data and portfolio-evaluation setup, then standardizes resource accounting and trajectory comparisons for reporting outcomes.

B.2 DATA

  • Evaluation uses monthly Binance Spot USDT pairs selected by prior-30-day turnover, ten-minute bars, next-bar return targets, and separate 2025 development and 2026 OOS periods.The point-in-time universe applies eligibility and data-coverage filters; Table 5 summarizes valid assets per bar as minimum, median, and maximum.Source: §B.2 · pp. 18–19, Table 5 · p. 19

B.3 FACTOR EXPRESSIONS AND FACTOR-POOL EVALUATION

  • Factor expressions map market fields to cross-sectional signals, and portfolio IC is the time average of per-bar cross-sectional Pearson correlations.The evaluator standardizes factors and labels within the current universe; candidate admission depends on the joint factor-pool update.Source: §B.3 · pp. 19–20
  • After adding a candidate, the factor pool refits member weights; it rejects candidates exceeding the signed-correlation threshold and removes the smallest-magnitude weight if capacity is exceeded.Capacity is fixed at 30, and a candidate removed by the update causes that update to fail.Source: §B.3 · p. 20

B.4 COMPUTE ENVIRONMENT AND COMMON CONDITIONS

  • Runs use standardized compute resources, while a frozen reference-price ledger counts model usage, including retries and failed requests, with price cards fixed across conditions.The compute setup specifies sandbox CPU, memory, workspace, and GPU allocations.Source: §B.4 · p. 20, §B.5.2 · p. 21

B.5.1 JOINT-BUDGET PROGRESS AND CHECKPOINTS

  • Joint-budget progress tracks six accumulated resources against fixed limits, saves checkpoints at 25%, 50%, and 75%, and retains resource compositions rather than treating resources as interchangeable.The progress measure does not assign exchange rates between resources.Source: §B.5.1 · pp. 20–21

B.6 TRAJECTORY ALIGNMENT UNDER HETEROGENEOUS STOPPING CONSTRAINTS

  • Trajectory comparisons use only observed run ranges, carry portfolio quality forward between completed evaluations, and leave missing checkpoints unestimated rather than interpolating or extrapolating.Wall-clock limits are treated as censoring, not as progress toward the joint budget; Figure 3 shows parent-trajectory effects and means for capability-state replacement.Source: §B.6 · pp. 21–22, Figure 3 · p. 22

C.1 ENDPOINT AND OOS REPORTING

  • Endpoint tables report final deliverable portfolio IC for development and OOS, while capability replacement compares subsequent gains from two capability states at shared starting states.OOS endpoint results cover all 18 end-to-end trajectories and all 24 paired replacement blocks.Source: §C.1 · p. 22
  • At each checkpoint, two parent effects were positive and four negative, while repetitions had opposite signs in five of 12 parent-trajectory–checkpoint pairs.Endpoint reporting uses final deliverable pools; capability-replacement intervals average repetitions within parent trajectories before weighting parents equally.Source: §C.1 · p. 22

C.2 TRAJECTORY AND RESOURCE ROBUSTNESS

Matched-resource views show early portfolio gains and later differences driven mainly by a few long-tail trajectories, while capability-state diagnostics reveal no simple separation between initial and accumulated Cap. Across submission, replay, and geometric analyses, research choices and portfolio state shape realized outcomes.

C.2 TRAJECTORY AND RESOURCE ROBUSTNESS

  • Matched-resource comparisons preserve early portfolio gains and long-tail differences; checkpoint means vary by model, with Qwen Evolving slightly higher at 80% and DeepSeek Fixed higher at all three checkpoints.Figure 4 aligns trajectories by resource consumption; Table 6 reports group means at joint-budget checkpoints, while the main figure reports medians over relative submission progress.Source: §C.2 · pp. 22–23, Figure 4 · p. 23, Table 6 · p. 23

C.3 CAPABILITY-STATE BEHAVIORAL DIAGNOSTICS

  • At both checkpoints, behavioral differences between accumulated and initial Cap are uncertain, while final outputs remain highly correlated with their own starting portfolios under both conditions.All Figure 5 difference intervals include zero; output correlations are 0.9700/0.9755 under initial/accumulated Cap early and 0.9951/0.9924 late.Source: §C.3 · p. 24

C.4 RESEARCH BEHAVIOR EVIDENCE

  • The research-behavior audit verifies submission gains and screening decisions using records, trajectory text, replay results, and restored historical platform states.The audit covers parameter tuning, screening, and reevaluation.Source: §C.4 · p. 24

C.4.1 SUBMISSION CATEGORIES AND CONTRIBUTIONS

  • Parameter tuning contributes positive net gains in 16 of 18 trajectories, while none of 818 known exact-resubmission outcomes exceeds the 10−4 gain threshold.Across groups, mean parameter-tuning contributions remain positive after deleting any one trajectory.Source: §C.4.1 · pp. 25–26

C.4.2 SCREENING DECISIONS AND REPLAY

  • In the audited Evolving trajectory, four screened-out candidates had positive gains at the original state, but sequentially submitting all screened-out candidates lowered both batch endpoints.Across the two batches, 11 candidates were screened out; sequential endpoint differences were −0.00001749396 and −0.00000666454.Source: §C.4.2 · pp. 26–27

C.4.3 REEVALUATION AND RULE REVISION

  • Resubmitted factors can change marginal-gain sign as portfolios change, and a separate delayed evaluation led to withdrawal of a hard screening threshold after portfolio IC rose.Of 311 positive first gains, 79 became negative; the rule-revision case is qualified because a protocol reminder also preceded submission.Source: §C.4.3 · p. 27

C.5 FACTOR-POOL SIMILARITY AND DIVERSITY

  • The factor-pool analysis evaluates final signals on fixed market segments without return labels and interprets geometric measures alongside portfolio quality and submission behavior.The analysis uses standardized factor outputs and distinguishes pool-level comparisons from descriptive views of individual members and projections.Source: §C.5.1 · p. 28

C.5.1 GEOMETRIC MEASURES AND COMPARISON

  • Complete factor pools are the independent units for within-model Fixed-versus-Evolving tests, using exact permutation of pool labels.The interpretation accounts for final pools reflecting both candidate generation and fixed-capacity selection.Source: §C.5.1 · p. 28

C.5.2 OVERALL AND FAMILY-LEVEL RESULTS

Across final factor pools, overall geometric differences between Fixed and Evolving are not statistically significant, although exploratory family coordinates vary by model. Resource use also differs across models and conditions.

C.5.2 OVERALL AND FAMILY-LEVEL RESULTS

  • None of the overall within-model geometric differences is statistically significant, despite slight outward-distance increases for Qwen and decreases for DeepSeek.The reported dimensions and distance measures differ by model, but all listed exact-permutation p-values exceed 0.05.Source: §C.5.2 · p. 28
  • Exploratory family coordinates show lower Qwen affinity for short-horizon price changes and price-volume coupling, while DeepSeek affinity increases slightly across all five families.Qwen decreases are 0.1021 (p = 0.0043) and 0.0559 (p = 0.0455); DeepSeek's largest increase is 0.0668 for volatility and range (p = 0.100).Source: §C.5.2 · p. 28

C.6 REALIZED RESOURCE USE

  • DeepSeek Evolving runs used more CPU core-hours and model cost but completed fewer evaluations than Fixed runs; Qwen resource averages were similar across conditions.DeepSeek Fixed/Evolving averages were 5.8021/16.0003 CPU core-hours, USD 2.3403/4.2973 cost, and 1000/642 evaluations; Qwen averages were 7.2205/7.2329 CPU core-hours, USD 9.4538/9.3240, and 346.0/295.3 evaluations.Source: §C.6 · p. 28
Loading 2609.33524v1…