- Rating
- 2
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper studies how to train an LLM agent for alpha factor discovery. It formulates factor mining as a multi-turn decision process with an explicit information boundary: the agent sees metrics only on a search window, a later verification window is hidden and enters only the reward, and the test period is evaluated once. The method has three components. First, gated hierarchical process rewards score each turn on executability, predictive power, novelty, and stability on the hidden window. Second, turn-level advantages credit each revision only for the rewards that follow it. Third, an experience bank stores both successful and failed attempts, with failure types derived from the reward hierarchy, and retrieves them to guide later exploration. On CSI 300, CSI 500, and S&P 500, EvoMind outperforms the strongest baseline by about 10% in test RankIC and also improves annualized excess return, with all baselines re-implemented on the same Qwen3-4B backbone.
Critical strengths
- The paper builds a complete infrastructure for training an alpha-mining agent, covering the agent loop, the information boundary between the search, verification, and test windows, and turn-level reward assignment. To my understanding, as I'm not an expert in this area, there is limited prior work on this, so the infrastructure is valuable to the community.
- Rather than relying only on prompting and pipeline design with a frozen LLM, the paper trains the policy itself with RL, which is a promising direction. The ablations show that the gains come from the process rewards and turn-level credit assignment, not from RL alone.
Critical weaknesses
- All experiments use a single small backbone, Qwen3-4B. I understand the time and compute constraints. Still, with results from only one model, it is hard to tell whether the carefully designed infrastructure and reward assignment are generally effective methods or just a configuration that happens to work in this setting. Several of the baselines were originally run with much larger or frontier models: AlphaAgent with GPT-3.5-turbo, QuantaAlpha with DeepSeek-V3.2 and GPT-5.2, and AutoScientist-Quant with GPT-OSS-120B and GLM-5. Here they are all re-implemented on the frozen 4B backbone, which may handicap them. If training larger models is not feasible, it would be very helpful to show that, at least in some setting, the trained 4B agent can match these methods with their original backbones, or match directly calling a frontier model API.
- As noted in the strengths, the main contribution is a specific pipeline and its training, while the underlying learning machinery is largely known: GRPO-style objectives, return-to-go advantages, and process rewards. I would like the authors to argue the contribution more clearly. What makes this domain specifically hard for agent training, and how does each design choice address that difficulty? Or is the work mainly one practical implementation that works well?
Questions for authors
Please see the Weaknesses section.
5 comments
Response to W1-2 (Part 2/3): A Second Backbone and Other Rewards, Windows, and Markets
We agree that results on one model alone cannot establish generality. Below we summarize experiments varying four dimensions: backbone family, reward specification, verification windows, and target market (the first three retrain under the full-system protocol, Part 1; the fourth transfers weights with no retraining).
1. Second backbone family: Llama-2-7B-Chat.
We repeated training and reran AutoScientist-Quant and Frozen-EvoMind on an older, weaker model from another family under identical rewards, budgets, windows, and downstream evaluation (shortening response length for its 4k context; five runs each on CSI 300 / S&P 500):
| Test RankIC difference | Qwen3-4B (paper) | Llama-2-7B-Chat |
|---|---|---|
| Gain from training (EvoMind vs. Frozen-EvoMind) | +0.0067 / +0.0058 | +0.0050 / +0.0034 () |
| Lead over AutoScientist-Quant | +0.0049 / +0.0036 | +0.0038 / +0.0019 (, borderline on S&P 500) |
Absolute RankIC drops by 20%–23% for all three methods on Llama-2 and margins narrow (especially on S&P 500), while both differences stay positive and method ordering is unchanged.
2. Summary across reward settings, verification windows, and unseen markets.
| Dimension varied | Perturbation tested | EvoMind Test RankIC / Margin | Reference / Baseline |
|---|---|---|---|
| Reward settings (CSI 300, 9 settings × 3 runs) | Shift 3 thresholds down/up, halve/increase , swap L3 & L4 | 0.0508–0.0533 (all > baseline; min lead +0.0029, ) | AutoScientist-Quant: 0.0479 (default EvoMind: 0.0528) |
| Verification windows (3 markets × 5 runs) | Rotate training windows & exclude 2022 altogether | Lead +0.0037 / +0.0042 / +0.0027 () | Paper fixed-2022 lead: +0.0049 / +0.0052 / +0.0036 |
| Target market (policy-only, 5 runs, no retraining) | Transfer S&P 500 CSI 300 policies (empty pool & bank) | 0.0497 (CSI 300), 0.0343 (S&P 500); vs. Frozen | Target Frozen-EvoMind: 0.0461, 0.0309 (64% / 68% of in-market gain) |
(With three runs per reward setting, we treat row 1 as a sensitivity check over the tested range; in row 3, source-market weights achieve 64% / 68% of the gain from training on the target market.)
All defaults were fixed before the first test evaluation (Section 5.1) and are not retuned. We will note in Section 6 that RL has been tested only on two smaller open-weight models, and report full results in new Appendices F, G, H, and K.
Response to W2 (Part 3/3): What Is the Contribution?
You are right that our GRPO-style objective, return-to-go credit, and multi-turn loop build intentionally on prior work (Section 2). Our contribution addresses a distinct problem: what signal should train an agent when tool feedback is noisy and overfittable? (CSI 300 numbers below are from Tables 2–3 unless marked new.)
1. Same backbone, loop, and GRPO objective—only the training signal changes.
Unlike math or coding with reliable verifiers, backtest metrics are noisy empirical estimates that RL readily overfits, and a growing factor pool makes utility non-stationary. Holding the backbone, loop, and optimizer fixed shows this is a training-signal question rather than a mere engineering implementation:
| Training signal variant (same backbone, loop, & GRPO) | CSI 300 Test RankIC | Note |
|---|---|---|
| Frozen-EvoMind (no RL) | 0.0461 | Untrained reference |
| Standard Outcome-GRPO | 0.0454 | Trails untrained agent on all 3 markets (Table 2) |
| + Gated process rewards (trajectory-level credit) | 0.0490 | Adds verifiable intermediate rewards |
| EvoMind: + Turn-level credit & bank | 0.0528 (0.0517 policy-only) | Full framework (Part 1) |
2. One tool-verifiable hierarchy driving reward, credit, and memory.
EvoMind scores each candidate through four tool-computed gated levels (Table 1: executability, predictive power, pool novelty, and hidden- stability), using this single hierarchy to address four domain failure modes:
| Domain difficulty | Design role of the hierarchy | If removed or altered (full: 0.0528) |
|---|---|---|
| Visible metrics can be overfit | Reward: heaviest weight placed on a later window hidden from the policy | No cross-period signal in training: 0.0482 (new; Table 3's 0.0468 also omitted admission screening). Signal kept, shown to agent (new): 0.0504 ( retention 89.6%→94.4%, test retention 63.2%→58.6%) |
| Individual criteria have shortcuts | Reward: levels gated by logical dependency | Ungated sum: 0.0472 |
| Noisy outcomes; good/bad turns mix | Credit: per-turn verifiable scores let GRPO compute turn-level advantages without a critic or extra rollouts | Trajectory-level advantage: 0.0490 |
| Falsified directions recur as pool grows | Memory: first failed level gives tool-typed failure records retrieved during RL | No bank: 0.0476 (repeated failures rise 7.8%→19.6%) |
Removing any of the four components drops RankIC by 0.0038–0.0056. Unlike learned/human process supervision, every level here is deterministic tool output; unlike AlphaAgentEvo (which collapses rewards into one visible-window trajectory scalar and trails the strongest frozen agent on all markets, Table 2), EvoMind credits individual turns using hidden cross-period stability.
3. Summary of contributions.
(1) Formulating factor mining as a multi-turn MDP with an explicit information boundary (Section 3); (2) unifying gated reward, turn-level credit, and typed memory under one verifiable hierarchy (Section 4), applicable to agent tasks with overfittable empirical feedback; and (3) open-sourcing our training infrastructure, prompts, and configs. We will sharpen the Introduction, Section 2, and Section 4 (adding the table above) accordingly. Thank you again for your review.
Response to W1-1 (Part 1/3): Original Backbones and a Frontier API vs. the Trained 4B Agent
Thank you for your careful review and for recognizing both our training infrastructure and the direction of training the policy itself. We had fixed a 4B backbone to isolate algorithmic effects from model scale; to address your concern directly, we evaluated the baselines with their original backbones as well as a frontier API, GPT-6.1 Sol (released September 2026), on CSI 300 (five runs, mean test RankIC; API pipeline SDs: 0.0014–0.0016).
1. Original backbones & frontier API comparison (CSI 300).
Each pipeline keeps its native prompts and loop while sharing Section 5.1's data windows, 640-evaluation mining budget, pool-admission rule, and downstream evaluation, swapping only the LLM (no warm start or RL; because OpenAI retired GPT-3.5-turbo in October 2026, AlphaAgent uses stronger GPT-5.2). For EvoMind, Table 2 reports the full system (trained weights + RL-accumulated experience bank); we also report policy-only (empty initial bank, matching every other row):
(Δ in the last three rows is relative to Frozen-EvoMind, 0.0461.)
2. Key head-to-head comparisons & takeaways.
Larger backbones do raise the baselines (+0.0018 to +0.0055), so we will clarify that Table 2 is a controlled same-backbone comparison; note also that GPT-6.1 Sol's April 2026 cutoff postdates our test window, which can only favor it. We will add this comparison and the policy-only clarification to Sections 5.1–5.2 and new Appendix L.