acceptodds
Under review as a conference paper at ICLR 2027

AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining

Abstract

Large language model (LLM) based multi-agent systems can automate alpha factor mining, but their reliance on external APIs limits control over cost, availability, and confidentiality. Long research loops also tend to revisit a few successful economic mechanisms that lead to research path collapse. To address these limitations, we propose **AlphaDiverse**, a framework that integrates multi-agent alpha research system, diverse research path collection, and post-training for local agents. We let the research system generates complementary plan portfolios and vary research environments across loops to collect diverse research paths. Using these diverse traces, we warm-start local Planner and Realizer agents with supervised fine-tuning. Then, we propose a joint GRPO method to optimize both of them using predictive quality and diversity of contributions. Research feedback is confined to inner period data, while a frozen final model is evaluated on a later outer period data, thereby avoid test-set tuning. Experiments across four Chinese stock universes show that **AlphaDiverse** can combine competitive prediction with broader exploration.

Primary area:
neurosymbolic & hybrid AI systems (physics-informed, logic & formal reasoning, etc.)
Keywords:
alpha factor mining, LLM agents, post-training, multi-agent reinforcement learning, diverse exploration, quantitative finance

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_32XSOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
1
Read review
Paper summary

The paper replaces the API-based Planner and Realizer of an LLM alpha-mining loop with locally post-trained Qwen3.8-27B policies. The Planner proposes plans from a fixed 38-mechanism catalogue, the Realizer writes two structurally different factor specifications per plan, and a LightGBM predictor decides which factors are retained. The agents are trained by SFT on selected traces from 72 API research loops, then by joint GRPO with a shared reward U = 0.75Q + 0.25D, where Q is the predictive score gain and D rewards new directions of prediction change relative to earlier research; Planner and Realizer receive separate advantages. Research uses May 2022–Dec 2025 and the frozen model is evaluated once on Jan–Aug 2026 across four Chinese stock universes, where the method reports the best predictive metrics on CSI300 and the highest number of useful mechanisms in all four markets.

Critical strengths
  1. A well-chosen diversity objective. Rewarding new directions of prediction change, net of what earlier research already covers and calibrated against noise factors, targets repeated exploration of the same mechanism more directly than formula-level deduplication. Removing it lowers useful coverage (Table 4: U 23 → 18 on CSI300).

  2. A sensible evaluation protocol. Research feedback is confined to the inner period, the outer period is evaluated once, and agentic baselines share data and a matched 160-slot budget. The authors also report results that do not favor them (e.g., GLM-5.3 in Table 19).

Critical weaknesses

W1. The performance gains are not statistically established.

Each configuration is run once on an outer window of about eight months (~160 trading days). From the paper's own CSI300 numbers, the standard error of the mean daily IC is about 0.014 (0.0378 / 0.2128 / √160), while the IC gap to QuantaAlpha is 0.0038; the standard error of an annualized IR over this window is roughly 1.3, while the IR gap is 0.23, and QuantaAlpha has the higher ARR. In addition, the three intermediate variants in Table 23 improve by nearly equal steps in every market and on many metrics at once (e.g., CSI300 ARR 32.57 → 35.17 → 37.77; CSI1000 IR 0.492 → 0.677 → 0.862; A-share MDD 7.38 → 6.67 → 5.96), which is unusual for separately trained policies evaluated by independent backtests. (Q1, Q2)

W2. The comparisons do not isolate the value of the mining process or its practical value.

The ML/DL baselines receive only the 21 base factors (C.3), while the agents can use V, which includes Alpha158 and minute-level features (C.1). The gap to these baselines therefore largely reflects access to more information; notably, on A-share, Ridge and Transformer on the 21 factors already exceed the proposed method in RIC (0.0409 and 0.0405 vs. 0.0392). Separately, all returns are gross of costs for a daily-rebalanced portfolio with no turnover reported (C.2), so tradability cannot be judged, and the confidentiality motivation is weakened by the fact that the SFT traces were generated by external APIs on the same environments. (Q3, Q4, Q5)

Questions for authors

Q1. [Clarification] Table 23.

Is each of the "Planner only", "Realizer only" and "w/o diversity reward" rows an independent GRPO run followed by its own 20-round loop and outer evaluation? Please explain the near-constant increments and provide the daily outer prediction and return series for these variants.

Q2. [Requested analysis] Run-to-run variance.

Table 6 shows a full run costs $2–7 in API fees or about 1.4 H200-hours. Could you report 3 seeds of the proposed method and QuantaAlpha on CSI300 and A-share (mean ± std), with a paired test on the daily IC difference (e.g., Newey–West t-statistic)?

Q3. [Requested analysis] Zero-agent baseline.

What does the same LightGBM pipeline achieve when given B ∪ V directly, without any mining? This needs one fit per market and shows how much the 20 research rounds add beyond the features already available to the agents.

Q4. [Requested analysis] Tradability.

Please report average daily turnover of the top-quintile portfolio and the break-even one-way cost at which ARR falls to zero, for the proposed method and the strongest baselines. Both can be computed from existing backtests.

Q5. [Clarification] Confidentiality claim.

Since the SFT traces come from GPT-5.5, Grok-4.6 and GLM-5.3 running on the same data, can the local agents be trained without external APIs? If not, should the claim be limited to inference-time confidentiality?

Reviewer_x3o8OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper proposes a method called "AlphaDiverse," in which large language models (LLMs) are tasked with generating research hypotheses and formulating investment factors. The process begins by leveraging an external LLM to produce a large volume of research records; these records are then used to train a local model, which subsequently takes over the research workflow. During training, the model is optimized not only for predictive performance but also for exploring diverse research directions.

Critical strengths

This approach reduces reliance on external APIs, enhances data security, and offers greater stability in terms of both cost and availability.

The framework captures not only successful outcomes but also research records detailing pivots following failures, as well as the exploration of entirely new avenues.

The reward mechanism balances performance with diversity, incentivizing both improvements in predictive accuracy and the generation of novel, valuable contributions that differ from previous efforts.

Critical weaknesses

Research directions are constrained by a predefined list, preventing the truly unconstrained generation of entirely new ideas.

The joint training process approximates a long-term research trajectory using fixed, discrete tasks, making it difficult to precisely disentangle the contributions of hypothesis generation from those of implementation.

The knowledge embedded in the local model is derived entirely from the outputs of the external LLM, the paper does not address how performance might shift if the external model is updated or replaced.

Questions for authors

Please refer to the weaknesses.

Reviewer_L5V7OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

This paper presents AlphaDiverse, a framework for post-training local LLM agents for automated alpha factor mining. The system decomposes research into a Planner, which proposes portfolios of economically motivated research plans, and a Realizer, which maps these plans and retrieved capabilities into executable factor specifications. To mitigate research-path collapse, the framework enforces complementary mechanisms within each round and varies markets, feature masks, and evaluation settings across source research loops to collect diverse trajectories.

The collected trajectories are used for two-stage post-training of Qwen3.8-27B Planner and Realizer agents. SFT first provides role-specific initialization. The authors then jointly optimize the two agents with GRPO using a shared reward combining predictive quality with a log-determinant diversity term based on complementary changes in downstream model predictions. Planner and Realizer receive different conditional advantages intended to separate credit for choosing research directions from credit for implementing them. Training uses 589 Planner and 1006 Realizer SFT examples collected from 72 API research loops, followed by GRPO over 96 task states.

Evaluation is performed on CSI300, CSI500, CSI1000, and the broader A-share universe. Research and model selection use May 2022–Dec. 2025, while Jan.–Aug. 2026 is reserved for outer evaluation. The paper compares against conventional ML or DL predictors and several agentic factor-mining systems, and provides extensive module, training, and data ablations. The reported local AlphaDiverse model generally achieves strong performance and substantially greater useful-mechanism coverage than the tested alternatives.

Critical strengths
  • The joint training scheme assigns credit according to the decisions made by each role. Planner advantages compare portfolios of research plans, while Realizer advantages compare implementations conditional on the same plans. Both roles share a utility based on predictive quality and complementary contributions. The training ablations report greater useful-mechanism coverage when both roles are optimized and diversity is rewarded.

  • The evaluation protocol separates adaptive factor search from final performance assessment. Research feedback and model selection use the inner period, and the selected factors and predictor are frozen before outer evaluation. Chronological splits and boundary purging specify which observations can inform each stage. This separation is important for factor-mining systems that repeatedly adapt their proposals to measured performance.

Critical weaknesses
  • The methodological contribution appears incremental, chiefly combining Planner–Realizer decomposition, role-specific GRPO advantages, and a quality-weighted diversity reward.

  • The training objective may under-reward exploratory research directions whose value emerges only after later refinements. Joint GRPO evaluates sampled plan–implementation portfolios at fixed research states. Its reward measures their predictive and complementary contributions at those states rather than the value of the full sequence of subsequent research decisions. A direction that enables useful later refinements can therefore receive little credit when first proposed, weakening incentives for exploratory detours.

Questions for authors
  • Would it be possible for the authors to clarify which part of the joint objective and role-specific credit assignment constitutes the main methodological advance?

  • How could credit assignment capture the delayed value of an exploratory research direction, and what information from later research states would need to be carried back to the Planner?

  • Would it be reasonable to accept mechanisms and operators that are proposed by the evaluated agent but are not covered by the library? Would incorporating an accepted extension require retraining the policies?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.