Better Predictions, Fewer Ideas: The Alpha Agent's Exploration Problem

TL;DR for operators

An automated quantitative researcher can improve its predictions while becoming less effective at discovering different sources of predictive information. Better factor scores, more formulas, and broader productive research are not interchangeable measures of success.

AlphaDiverse addresses this problem by training local agents to select complementary economic hypotheses and implement them as executable factors. Its reinforcement-learning objective rewards both predictive improvement and contributions that existing signals do not already provide.

In the CSI300 experiments, supervised fine-tuning raises rank-based predictive correlation from 0.0291 to 0.0389 but reduces the number of useful economic mechanisms from 11 to 4. Adding joint, diversity-aware reinforcement learning raises these measures to 0.0407 and 23.

For quantitative research teams, the operational opportunity is to measure research agents by the marginal usefulness of their discoveries, distinguish planning quality from implementation quality, and preserve independent computational validation. Local deployment also reduces dependence on external model APIs.

These are historical benchmark findings, not evidence of profitable live trading. The search is restricted to 38 predefined mechanisms, and reported portfolio returns exclude trading costs.

A better predictor can become a narrower researcher

Consider a quantitative research team running an automated factor-discovery system. Each round generates candidate formulas, tests their predictive performance, and uses the results to guide the next round.

Over time, the agent learns which approaches have worked. It becomes better at refining those approaches, but increasingly concentrates on familiar trading ideas. The system continues producing valid factors without necessarily expanding the range of economic mechanisms contributing to its predictions.

That is the research problem investigated in AlphaDiverse: Post-Training Local Quantitative Research Agents for Diverse Exploration in Alpha Factor Mining, by Qingzhuo Wang and colleagues.¹

The paper’s CSI300 training ablation provides an unusually clear example. Supervised fine-tuning increases the rank information coefficient (RIC), a measure of agreement between predicted and realized return rankings, from 0.0291 to 0.0389. Yet the number of useful economic mechanisms falls from 11 to 4.

A mechanism counts as useful here when at least one factor representing that mechanism survives the inner predictive-model selection process. Merely proposing an idea, generating a valid formula, or passing an individual factor-quality screen does not qualify.

The result exposes a failure mode in automated research evaluation: optimizing predictive performance can improve the model’s immediate output while concentrating its successful discoveries in fewer directions.

Different formulas are not necessarily different discoveries

A natural response to repetitive research is to encourage the agent to generate more varied formulas. AlphaDiverse tests why that is insufficient.

On CSI300, a Single Synthesis baseline produces 122 distinct numerical signal clusters but only 8 useful economic mechanisms. Complete AlphaDiverse produces 129 clusters and 23 useful mechanisms (Section 3.2, Table 2).

The difference is not primarily the number of signals generated. It concerns whether those signals represent additional productive research directions.

Two factors might use different mathematical expressions yet change a prediction model’s outputs in almost identical ways. From the perspective of an existing investment model, their incremental contributions could therefore be redundant.

AlphaDiverse operationalizes diversity through changes to the baseline LightGBM predictor’s outputs.

For each research plan, it measures how the plan’s accepted factors change model predictions. Those changes are normalized and weighted by the plan’s predictive contribution. A log-determinant calculation then measures how much the new prediction-change directions expand the coverage already represented in a historical archive.

The shared reinforcement-learning utility is:

Here, $Q$ measures normalized predictive improvement, while $D$ measures incremental, quality-weighted diversity (Section 2.3, Equation 8).

A plan that adds no sufficiently positive predictive contribution receives little or no diversity weight. A useful plan whose prediction changes closely resemble previous discoveries contributes less additional diversity than one providing a complementary direction.

The 75/25 weighting is a design choice in the reported experiments, not an established optimal allocation for other markets or research tasks.

Crucially, prediction-change diversity measures functional complementarity, not proof of a distinct causal economic mechanism. The paper separately measures coverage of its predefined economic-mechanism catalogue.

The learning system separates choosing ideas from implementing them

AlphaDiverse organizes its research workflow into six modules: Planning, Retrieval, Realization, Execution, Validation, and Analysis.

The Planner proposes four research plans in each round, spanning at least three economic themes. Retrieval supplies plan-specific information about permitted features and operators. The Realizer then produces two structurally different factor specifications per plan.

Deterministic checks verify operator validity, allowed inputs, parameter constraints, and whether the required information would have been available at prediction time. Executed factors are screened before LightGBM evaluates their incremental contribution.

This division creates two learning problems.

First, the Planner must choose research directions that collectively contribute valuable information. Second, the Realizer must turn a particular research direction into effective, executable factors.

The authors address both through a two-stage post-training procedure.

Stage one: supervised learning from selected research histories.

Seventy-two API-generated research loops cover four Chinese stock universes, three API backends, different feature availability configurations, and alternative inner-fold settings.

The resulting dataset contains 589 Planner examples and 1,006 Realizer examples.

Planner examples are selected from episodes involving model improvement, productive changes of direction, and exploration of new directions, targeting a 50/30/20 mixture. Realizer training includes validated pairs of distinct factor implementations and additional repair examples.

Separate LoRA adapters are trained for the two roles on a shared frozen Qwen3.8-27B backbone.

Stage two: joint reinforcement learning with differentiated credit.

The authors use Group Relative Policy Optimization (GRPO) to refine the two adapters. Both share the quality-and-diversity utility, but each receives a different measure of relative performance.

• The Planner is evaluated by comparing the average outcomes of alternative research-plan portfolios.
• The Realizer is evaluated by comparing different implementations of the same research portfolio.

This conditional comparison reduces the risk of rewarding the Planner simply because an implementation happened to work well, or penalizing the Realizer because the selected research direction was weak.

The joint training uses 96 fixed research states, sampled twice, producing 768 Planner and 1,220 Realizer responses. The Analysis model remains unmodified.

The generalizable architectural idea is to assign learning feedback according to which decision an agent actually controls. The paper establishes its usefulness within this quantitative research workflow, not across arbitrary multi-agent systems.

What the CSI300 ablations actually establish

The post-training experiments isolate the roles of supervised imitation, joint reinforcement learning, and the explicit diversity reward.

Source: Section 3.3, Table 4. ARR is annualized historical excess return before trading costs; lower paired-factor correlation indicates less similar implementations.

The supervised-only result is particularly revealing. Predictive accuracy improves substantially relative to the untrained model, but useful-mechanism coverage contracts.

Joint reinforcement learning without the diversity reward recovers much of that coverage. Adding the diversity reward increases useful mechanisms from 18 to 23, reduces paired-factor correlation from 0.365 to 0.316, and raises reported ARR from 37.77% to 42.10%.

These differences support the contribution of the diversity objective within the joint-training design. They do not establish that the diversity term alone accounts for every improvement over supervised learning.

Other ablations address separate questions. Training only the Planner or only the Realizer improves performance over supervised-only policies, while jointly training both achieves the strongest reported CSI300 results. Removing paired Realizer supervision from the supervised dataset increases paired-factor correlation from 0.511 to 0.651 and reduces useful mechanisms from 4 to 1 (Section 3.3, Table 5).

That latter experiment evaluates training-data construction before reinforcement learning. It supports the role of paired examples in teaching implementation variety, rather than independently demonstrating the full joint-RL effect.

Exploration gains extend across four related stock universes

The main evaluation covers CSI300, CSI500, CSI1000, and the broader Shanghai–Shenzhen A-share universe.

All agentic methods receive a matched budget of 160 candidate-factor slots. The researchers separate iterative research and model selection during May 2022–December 2025 from a frozen outer evaluation during January–August 2026.

The full model’s useful-mechanism coverage improves over supervised-only training in every universe:

Source: Section 3.2, Table 2; Appendix D diversity comparisons.

These are substantial increases under the same candidate-generation budgets.

The authors also report that supervised-only exploration saturates relatively early, while the complete method continues adding useful mechanisms as the research budget expands. A separate held-out-state experiment tests research histories excluded from training-data construction. On CSI300, average useful-mechanism coverage reaches 12.6 for AlphaDiverse, compared with 5.9 for a GPT-5.5 implementation and 2.3 for supervised-only AlphaDiverse.

That is evidence of transfer to unfamiliar research states within the studied markets, not generalization to new countries or financial structures.

The financial comparisons are encouraging but require equally precise interpretation. On CSI300, AlphaDiverse records an information ratio of 3.210, maximum drawdown of 5.51%, and annualized excess return of 42.10%. QuantaAlpha achieves a higher ARR of 43.19%, so AlphaDiverse does not lead every metric.

Across all four universes, AlphaDiverse reports the strongest IC, ICIR, RICIR, and portfolio IR and the lowest maximum drawdown among compared methods.

These are historical, benchmark-relative results from an equally weighted, long-only top-quintile portfolio protocol, without trading-cost deductions.

What quantitative research teams can change

The findings suggest three operational design choices.

Measure retained contributions rather than generation volume. A research platform should distinguish executable factors, individually accepted factors, and factors retained after evaluation against an existing predictor. On CSI300, the complete agent generates 160 executable specifications, 118 accepted factors, and 56 model-retained factors. The supervised-only version retains just 8 (Appendix D.6, Table 25).

Separate research-direction decisions from factor construction. Teams implementing specialized agents can evaluate whether a proposed hypothesis complements the current research portfolio, then separately evaluate the quality of its implementation. This makes training feedback more closely aligned with the responsibility of each component.

Keep scientific evaluation outside unconstrained language generation. AlphaDiverse uses deterministic specification checks, conventional predictive modeling, chronological validation, and an outer evaluation period that is not directly reused during iterative research. These are reusable workflow-design principles, even for organizations that do not adopt its particular reinforcement-learning algorithm.

Local deployment presents another operational choice. In one complete CSI300 research run, the three tested API backends incur estimated charges of $7.28, $5.57, and $2.19. AlphaDiverse has no external model API charge but consumes 1.436 allocated H200 GPU-hours (Section 3.3, Table 6).

This supports reduced API dependence and greater deployment control. It does not establish lower total operating cost, because local compute, post-training, infrastructure, and maintenance are not comprehensively priced.

The boundary is productive exploration within a defined search language

AlphaDiverse’s research environment contains 38 predefined economic mechanisms organized into six themes, together with an available factor-operator library. The agent can search broadly within those constraints, but the experiments do not demonstrate discovery of entirely new economic theories.

The reinforcement-learning procedure also uses fixed sampled research states. This approximates the value of research decisions without fully resolving delayed contributions across long, evolving research histories.

The empirical boundary is similarly significant. Four Chinese stock universes provide variation in investable securities but not independent global market regimes. The principal benchmark comparisons use one complete mining run per method–market combination and do not report confidence intervals for those rankings. Cross-method mechanism-coverage comparisons also partly depend on LLM annotation.

Most consequentially for deployment, the portfolio simulations exclude commissions, slippage, and market impact. An eight-month outer evaluation cannot establish persistence through multiple future regimes, and the paper provides no live-trading validation.

The evidence is therefore strongest for a narrower proposition: within a bounded factor-research system, carefully designed reinforcement-learning rewards can improve productive exploration and historical prediction simultaneously.

Research quality needs its own objective

AlphaDiverse illustrates a specific weakness in training autonomous researchers solely to reproduce successful outputs. An agent can learn to predict more accurately while losing breadth in the ideas that contribute to its success.

Its contribution is a coordinated solution: collect examples of complementary research behavior, reward predictive changes that add information beyond previous discoveries, and assign learning credit separately to agents selecting research directions and implementing them.

For quantitative research organizations, this points toward a more demanding definition of research productivity. The question is not how many factors an agent can generate, nor even how attractive its best historical backtest appears. It is how reliably each additional research decision contributes valuable information beyond what the organization already knows.

That is a measurable target for bounded autonomous research systems. Whether those gains survive trading frictions and future market changes remains a separate empirical question.

Cognaptus: Automate the Present, Incubate the Future.