acceptodds
Under review as a conference paper at ICLR 2027

AlphaS: Semantic–Symbolic Co-Evolution for Alpha Mining with Coding Agents

Abstract

Quantitative alpha mining aims to discover predictive signals that generalize under noisy and non-stationary markets. Large language model (LLM) agents provide a promising direction by combining financial reasoning with executable factor generation. However, existing methods primarily evolve complete alpha expressions within a fixed representation or search space. As a result, knowledge discovered through expensive LLM exploration is largely reused only at the level of individual factors, rather than being distilled into reusable structures that can reshape subsequent search. Our key insight is that effective alpha discovery should co-evolve both the semantic hypotheses being explored and the symbolic language used to express and search them. We introduce \tool, a semantic-symbolic co-evolution framework for alpha mining with coding agents. A semantic LLM agent generates and evolves executable alphas from financial hypotheses, while an operator-construction agent distills recurring relations from successful factors into typed symbolic operators. These operators expand the grammar of a genetic-programming miner, enabling efficient recombination into new factors. Selected symbolic descendants are then returned to the LLM archive, closing a bidirectional co-evolution loop: semantic exploration evolves the symbolic search space, while symbolic exploration broadens future semantic discovery. Across CSI300, CSI500, and CSI1000, \tool achieves strong predictive and economic performance, with ICIRs of 0.3549, 0.2092, and 0.517 and annualized excess returns of 9.08%, 13.22%, and 16.58%, respectively. These results demonstrate that co-evolving semantic knowledge and symbolic representations provides an effective mechanism for scalable, reusable, and adaptive alpha discovery.

Primary area:
other topics in machine learning (i.e., none of the above)
Keywords:
LLM Agent, Self-Evolving System

Reviews

Among:
Shape:
1234this paper, Bottom 25%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.75 over 4 reviews, above 25% of the others.
Reviewer_fnBQOpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

This paper proposes AlphaS², a semantic-symbolic co-evolution framework for automated alpha mining. The main idea is to go beyond reusing previously discovered factors by extracting reusable computational structures from them and incorporating these structures into subsequent search.

The framework combines three components. A semantic LLM agent generates and refines executable Python factors based on financial hypotheses and evaluation feedback. An operator-construction agent extracts reusable patterns from these factors and turns them into typed operators. A genetic programming (GP) miner then searches expressions using the expanded operator library. Selected GP factors are fed back to the LLM as parents or references for further evolution.

The method is evaluated on CSI500 and CSI1000. The authors report competitive predictive performance and transaction-cost-adjusted portfolio returns, with ICIRs of 0.2092 and 0.517 and annualized excess returns of 13.22% and 16.58%, respectively. Ablation experiments on CSI1000 and recorded program reuse are used to support the proposed bidirectional transfer mechanism.

Critical strengths

1. A meaningful approach to reusing knowledge from previous alpha discoveries.

The paper addresses an interesting limitation of existing alpha-mining systems: useful structures discovered during factor search are often reused only as parts of complete factors. Extracting them into reusable GP operators is a reasonable way to make previous discoveries more accessible to subsequent symbolic search. The FlowDecaySkew example in Figure 2 and Appendix A illustrates how a learned operator can be applied to new inputs and reused within more complex expressions.

2. A concrete implementation of bidirectional transfer, supported by program-level evidence.

The framework connects LLM-based factor generation with GP search through both operator construction and program feedback. Unlike a purely conceptual proposal, the paper provides operator definitions, validation procedures, prompt templates, and recorded examples of GP-to-LLM reuse. Table 5 gives useful evidence that symbolic discoveries actually entered subsequent LLM generations. The full system also achieves the highest portfolio IR among the reported CSI1000 ablation configurations.

Critical weaknesses

1. Insufficient evidence for the effectiveness of the proposed mechanisms.

The paper proposes two key mechanisms: extracting reusable operators from LLM-generated factors to improve GP search, and feeding GP discoveries back to the LLM to improve subsequent factor generation. However, the experiments provide limited direct evidence for either claim.

First, the benefit of learned operators remains unclear. In Table 3, removing the operator-construction agent yields a higher portfolio IR than the archive-only configuration with learned operators (1.742 vs. 1.563). More direct evidence is needed to show that these operators improve GP search efficiency or factor quality under comparable search budgets.

Second, while the full system outperforms archive-only feedback in portfolio IR (1.922 vs. 1.563), and Table 5 confirms that GP discoveries are reused, neither result directly shows whether GP feedback improves the quality of subsequent LLM-generated hypotheses or factors.

Since these mechanisms are central to the paper's contribution, their effectiveness should be evaluated more directly, rather than mainly through downstream portfolio performance.

2. Insufficient clarity and rigor in the search and evaluation protocols.

The semantic evolution procedure is not sufficiently specified. Section 4.3 reports 500 initial and 500 continuation LLM proposals, but it remains unclear how proposals use previous evaluation feedback, how the search is scheduled, and how many complete co-evolution cycles are performed. This makes the proposed iterative process difficult to reproduce.

The experimental comparisons are also limited. Table 2 relies on published baseline results rather than rerunning methods under a unified evaluation protocol, and the CSI1000 comparison with AlphaMaster is not fully aligned. Moreover, several closely related LLM-based alpha mining methods, such as HARLA, CogAlpha, and Navigating the Alpha Jungle[1], are discussed or relevant but not included in the main performance comparisons. Comparisons with representative methods under consistent settings would better establish the proposed framework's advantages.

[1]Yu Shi, Yitong Duan, Jian Li. Navigating the Alpha Jungle: An LLM-Powered MCTS Framework for Formulaic Factor Mining

Finally, the factor-selection procedure is less clearly specified for CSI500 than for CSI1000. These issues weaken the reproducibility and reliability of the empirical conclusions.

Questions for authors

Q1. Could the authors clarify the actual execution process of the semantic evolution agent and the overall co-evolution loop?

Section 4.3 reports 500 initial and 500 continuation LLM proposals, but I find it difficult to reconstruct how these proposals are generated and how information flows between them.

Could the authors describe the actual execution process of the initial 500 proposals, particularly how previously generated factors and their evaluation feedback are incorporated into subsequent LLM generations? I would also like to understand how the continuation stage builds on the initial semantic search and the discoveries made by GP.

More broadly, could the authors explain the complete execution timeline of the semantic-symbolic co-evolution process, including how many full co-evolution cycles are performed and how information is carried forward across these cycles?

A concrete description or pseudocode of the actual search procedure would help clarify how the proposed iterative knowledge reuse mechanism is implemented.

Q2. What is the actual benefit of learning new GP operators?

The paper motivates operator construction as a way to improve search under limited computational budgets. However, Table 3 shows that removing the operator-construction agent yields a higher IR than the archive-only configuration with learned operators (1.742 vs. 1.563).

Could the authors clarify the exact configuration differences between these two variants? Is there evidence that the learned operators improve the quality or efficiency of GP search under matched evaluation budgets?

Q3. Can the benefit of direct GP-to-LLM feedback be isolated more clearly?

The full system outperforms archive-only feedback in portfolio IR, but Section 4.3 notes that the search settings differ. Could the authors specify these differences and clarify whether the improvement remains when the initial archive, GP outputs, continuation budget, and downstream selection procedure are held fixed?

It would also help to understand why Table 5 reports 308 continuations with a direct GP parent on CSI500 but only 18 on CSI1000, and whether direct GP-parent usage is associated with improved subsequent factor quality.

Q4. Could the authors clarify the evaluation and factor-selection protocols?

Table 2 uses external baseline results without rerunning them, and the authors acknowledge that the CSI1000 comparison with AlphaMaster is not fully protocol-matched.

In addition, could they provide the final factor-selection rules for CSI500 at the same level of detail as CSI1000, along with the number of selected factors and uncertainty estimates for each ablation configuration?

Reviewer_6goQOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper treats alpha discovery as a loop between two representations. A language-model agent writes executable factors from financial hypotheses, an operator agent turns recurring structure in those factors into typed operators, and a genetic program searches the expanded grammar. Selected symbolic programs return to the language model as parents and references. On CSI500 and CSI1000 the authors report information coefficients, information ratios, and cost-inclusive excess returns against numbers taken from earlier papers, plus ablations of the return path and of operator construction, and a record that later language-model edits really do start from genetic-programming parents.

Critical strengths
  1. The research question is sharper than another agent loop. Reusing whole factors, or storing verbal lessons, does not change the grammar a symbolic miner can search. Compiling a relation into a typed operator with a fixed panel signature, then letting crossover bind it to new inputs, is a real proposal about search bias under a limited budget. The paper is also honest that this need not increase theoretical expressiveness.

  2. The return path is documented as program reuse, not only asserted. Later language-model continuations are described as conditioned on genetic-programming parents and references, and the case study shows an edit that rewrites a compiled operator rather than merely avoiding its correlation. Recording that lineage is the right kind of evidence for a co-evolution claim, and it is more informative than a leaderboard delta alone.

  3. The economic report is more careful than a frictionless equity curve. Both markets deduct stated costs, returns are excess over the index, and on CSI500 the authors separate discovered factors from the public base library. That separation shows the headline portfolio is a combination, which is the disclosure a reader needs.

Critical weaknesses
  1. The two markets are difficult to read as one comparison, because the protocols differ. For CSI500, the baselines come from a published benchmark and are not rerun in this setup; the correlation gain is modest, and the excess-return drawdown is larger. For CSI1000, the numbers also come from another paper, with a ten-day target and a different portfolio, and the predictive difference is small.

  2. The look-ahead rule is stated in the prompt, but the execution setup does not appear to enforce it. The semantic program receives the full panel, including the evaluation years, in one call, so a full-sample normalization could affect the training score, and the future-access check is described for registered operators rather than for the free-form programs. Releasing more of the accepted formulas would help readers follow the economic logic and see how future information is avoided in the expressions themselves.

  3. The ablations are useful, but they also change the search settings, so the role of bidirectional transfer is hard to separate. Removing operator construction still raises the information coefficient and keeps most of the cost-adjusted ratio, and the semantic agent alone gives the most stable correlation. It would help to hold the language-model stage fixed and compare a budget-matched genetic program on the original primitives with the expanded grammar.

  4. The reported result comes from a single run of an external API. The paper notes that another run can produce different factors, and a second run, a frozen model snapshot, or a seed distribution is not included. Public base factors also contribute substantially to the CSI500 information ratio, so the share attributable to co-evolution is still difficult to judge.

Questions for authors
  1. If the baselines were rerun on the same panel, horizon, mask, cost, and predictor across several seeds, would the gains in correlation and excess return remain?

  2. When the panel is masked at the evaluation date, what fraction of the accepted free-form programs change their training score or their selection?

  3. With the language-model and genetic-programming budgets held fixed, how does test performance differ between the original grammar and the grammar after the new operators are added?

Reviewer_MWzxOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper resolves utilizing Large language model (LLM) agents for quantitative alpha mining problems which aim to discover predictive signals that generalize under noisy and nonstationary markets. Existing methods primarily search complete alpha expressions within a fixed representation or search space where expensive LLM exploration is largely reused only at the level of individual factors, rather than being distilled into reusable structures that can reshape subsequent search. This work solves this problem by proposing a semanticsymbolic coevolution framework for alpha mining with coding agents. The proposed framework includes a semantic LLM agent that generates and evolves executable alphas from financial hypotheses, while an operatorconstruction agent distills recurring relations from successful factors into typed symbolic

Critical strengths

This paper solves a key problem of quantitative alpha mining in financial area. They make the following contributions:1)A perspective on alpha discovery which frame alpha mining as the joint evolution of financial hypotheses and symbolic search primitives, making discovered factors a source of reusable structure for future exploration. 2)A domain-specific co-evolution framework which distills financial relations from LLM-generated factors into typed operators for genetic programming and returns selected symbolic descendants as parents and references for semantic exploration. 3)An empirical evaluation demonstrate predictive and economic performance, with ICIRs of up to 0.517 and cost-inclusive annualized excess returns of up to 16.58%, and examine the transfer mechanisms through component comparisons and recorded search lineages.

Critical weaknesses
  1. As an academic paper, this work has significant shortcomings in scientific presentation. For instance, several abbreviations are used directly without clear definition or reference, such as GP on Page 3, Line 109. There are also issues like incorrect figure citations.
  2. The mathematical formulations in the paper are poorly defined. Specifically: (1) The notations t and T appear to be mixed‑up. (2) Some symbols are left undefined (e.g., certain terms in Equation (2)), which renders the definitions and method descriptions ambiguous and unscientific.
  3. On Page 3, Line 152, the notation alphaT,i\text{alpha}_{T,i} is inconsistent with the prior definition of αT\alpha_T given in the equations.
  4. The symbol g in Figure 1 is not defined and neither used in the following sections.
  5. The ablation study fails to sufficiently validate the key contributions claimed in the paper, which include: 1) search‑space reduction, 2) enrichment of the primitive vocabulary, and 3) representation forms for reusable structures that reshape subsequent search.
  6. The LLM plays a limited role in the overall pipeline; most innovations lie in tool‑oriented techniques for finance, which makes the paper misaligned with the scope of the ICLR conference.
Questions for authors
  1. What does GP stand for on Page 3, Line 109?
  2. Could you provide a detailed example illustrating how the operators expand the grammar of a genetic‑programming miner, as well as an analysis of their practical implications?
  3. Could you present a detailed ablation study to verify the mechanism that distills recurring relations from successful factors into typed symbolic operators? Additionally, please analyze these typed symbolic operators and explain how they facilitate alpha‑mining discovery.
  4. On Page 4, Line 168, is there a misreference to Figure 2? I believe this should instead refer to Figure 1.
  5. On Page 5, Line 230, how are "high‑quality" factors measured?
  6. On Page 5, Line 238, does L denote dates or timestamps? What is the underlying relationship?
Reviewer_LeKWOpenReview ↗
Rating
1
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces AlphaS^2, a framework combining LLM agents with Genetic Programming (GP) for formulaic alpha mining. The proposed approach attempts to establish a bidirectional co-evolution loop: an LLM semantic evolution agent generates factor programs from market observations; an operator-construction agent distills recurring patterns into typed symbolic operators that expand the GP grammar; and symbolic evolution explores combinations of these operators, with selected descendants fed back into the LLM archive as parents or prompt references. The authors evaluate the framework on CSI500 (1-day horizon) and CSI1000 (10-day horizon) universes and report improvements in predictive (IC, ICIR) and economic metrics (AER, IR).

Critical strengths

Conceptual Framing of Grammar Expansion: The high-level motivation of abstracting reusable sub-structures/motifs discovered by an LLM into typed symbolic operators for genetic programming, rather than merely treating the LLM as a raw expression generator or prompt-level mutator, is an intuitive direction for domain-specific neuro-symbolic search.

Critical weaknesses
  1. Incomparable Baseline Comparisons (Section 4.1 & Table 2): The baseline results are borrowed from AlphaAgent and AlphaMaster without re-running them under identical backtesting protocols (Lines 314–316, 327–329). On CSI1000, the authors acknowledge that portfolio settings are not fully specified, making the comparison only approximate. Differences in transaction costs, rebalancing, and portfolio construction could explain the reported gains. More importantly, Table 4 shows that the mined factors alone achieve an IR of only 0.945 on CSI500, below AlphaAgent's reported 1.488. The apparent improvement largely comes from combining mined factors with existing public factors.

  2. Contradictory Ablation Results (Section 4.3 & Table 3): The ablations do not support the necessity of the full architecture. On CSI1000, removing the Operator Construction Agent actually improves IC (0.07249 vs. 0.07012), while the Semantic Evolution Agent alone achieves the highest ICIR (0.538 vs. 0.517). These results cast doubt on whether the proposed operator construction mechanism provides meaningful benefits.

  3. Limited Novelty and Unclear Added Value: LLM-guided evolutionary alpha mining is already extensively explored (e.g., AlphaGen, AlphaAgent, AlphaForge, AlphaAgentEvo, QuantAlpha, AlphaPROBE). The proposed approach mainly combines LLM-generated operators with standard GP crossover and mutation. This appears incremental, especially when the ablations fail to show consistent benefits from the central operator construction component.

  4. Unconstrained Code Generation vs. Standard Alpha DSLs: Allowing arbitrary Pandas/PyTorch implementations introduces several concerns. First, the expanded search space makes comparisons with grammar-constrained baselines potentially unfair. Second, prompt-based restrictions and heuristic checks are insufficient to guarantee the absence of look-ahead leakage in generated code. Third, arbitrary Python operators raise questions about execution efficiency and practical deployment. These issues need stronger validation.

  5. Overfitting Risk and Missing Compute Budget: The framework evaluates a large number of LLM-generated candidates and GP individuals on historical data, raising concerns about data snooping and selection bias. The evaluation lacks sufficient independent out-of-sample tests to address this risk. Moreover, token usage, runtime, and computational costs are not reported, making it difficult to assess whether the gains justify the additional search budget.

Questions for authors
  1. Why were all external baselines not re-implemented and evaluated under the exact same data pipeline, rebalancing rules, and transaction cost assumptions? Can you provide an apples-to-apples re-run of key baselines (e.g., AlphaGen, AlphaAgent) under your exact evaluation script?
  2. In Table 3, "w/o Operator Construction Agent" achieves higher IC (0.07249) than "Full" (0.07012), and "Semantic Evolution Agent Only" achieves the highest ICIR (0.538). How do the authors reconcile these findings with the claim that distilling semantic relations into typed symbolic operators provides a superior search space?
  3. In Table 4, removing public base factors drops the IR from 1.614 to 0.945 on CSI500. Did the baseline AlphaAgent in Table 2 include the exact same public base factor suite in its reported IR of 1.488?
  4. Given the abundance of existing evolutionary and agentic alpha mining pipelines (e.g., AlphaGen, AlphaAgent, AlphaAgentEvo), what fundamental conceptual advance does AlphaS^2 offer beyond an engineering wrapper that combines prompt-guided factor generation with GP mutation? In light of your ablation results where simpler variants without operator abstraction perform comparably or better, what decisive evidence proves that this co-evolution design is practically necessary?
  5. Why did the authors choose free-form PyTorch/Pandas code generation for operators rather than evaluating within a standardized, sandboxed formulaic DSL (e.g., Qlib's expression engine)? How do you ensure that the performance gains are not simply artifacts of giving the framework a far larger, unconstrained functional space that baseline formulaic miners (e.g., AlphaGen) were strictly forbidden from accessing?
  6. Beyond prompt instructions and basic smoke checks, how does the system formally guarantee that arbitrary generated code does not introduce subtle forward-looking leakage (look-ahead bias) across rolling and cross-sectional axes?
  7. What is the total computational cost (API tokens, GPU hours, and wall-clock time) required to complete one full mining run of AlphaS^2 compared to a standard GP run or a pure LLM agent?

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.