PaperlayerView exact revision on arXiv

Source-linked AI summary

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

Qingzhuo Wang, Zikun Wei, Zhihua Wei, Wen Shen

arXiv:2609.29014v1cs.AIcs.CEcs.MA

TL;DR

LLM-based alpha mining can revisit a few successful economic mechanisms, leaving much of the search space unexplored. AlphaDiverse trains local Planner and Realizer agents from diverse research traces and jointly optimizes predictive quality and useful diversity; across four stock universes, it reports competitive performance and sustained exploration.

  • Problem

    Alpha factor research systems can repeatedly revisit a few successful economic mechanisms, leaving much of the search space unexplored.

    Source: §1 · p. 1

  • Method

    AlphaDiverse collects diverse traces through complementary planning across varied environments, then uses supervised fine-tuning and joint GRPO to train local Planner and Realizer agents for predictive quality and useful diversity.

    Source: §1 · p. 2, §2.3 · pp. 5–6

  • Results

    Across four stock universes, AlphaDiverse leads all four predictive metrics, including IC of 0.0378 and RIC of 0.0407, while showing sustained exploration.

    Source: §3.1 · p. 7, §5 · p. 9

  • Takeaways & Limitations

    The results support effective alpha research with locally deployed agents.

    Source: §5 · p. 9

    The fixed mechanism catalogue and operator library bound expressible ideas, while finite inner-state evaluation only approximates plans’ delayed contributions in long research loops.

    Source: §5 · p. 10

Abstract

from arXiv · show

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 a multi-agent alpha research system, diverse research path collection, and post-training for local agents. We let the research system generate 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 avoiding test-set tuning. Experiments across four Chinese stock universes show that AlphaDiverse can combine competitive prediction with broader exploration.

1 INTRODUCTION

AlphaDiverse addresses external-API constraints and repeated exploration of a few economic mechanisms by collecting diverse research traces and post-training local agents. Its two-level exploration strategy and inner/outer evaluation split support broader exploration while reserving outer-period data for final evaluation.

1 INTRODUCTION

  • External APIs constrain cost, availability, latency, model behavior, and confidentiality in multi-agent alpha research.Long research loops can also revisit a few successful economic mechanisms, leaving much of the search space unexplored.Source: §1 · p. 1
  • AlphaDiverse combines multi-agent alpha research, diverse research-path collection, and post-training to replace API-based Planner and Realizer nodes with local agents.The research system first uses API-based LLMs to verify the workflow, then trains local agents from collected traces.Source: §1 · p. 2, Figure 1 · p. 2
  • Within each loop, planners propose plans grounded in complementary mechanisms; across loops, varying markets and features yields diverse research paths.High-quality complementary traces are selected and balanced for training.Source: §1 · p. 2
  • Two-stage post-training uses SFT for instruction-following and research knowledge, then joint GRPO rewards predictive quality and complementary contributions.Credit is assigned separately to planning and realization; the design is intended to improve exploration diversity and generalization beyond SFT paths.Source: §1 · p. 2
  • Research feedback uses only inner-period data, while the selected factors and model are frozen before outer-period evaluation.Across four Chinese stock universes, local post-training yields consistent improvements in research performance and explored-path diversity.Source: §1 · p. 2

2 METHODOLOGY

AlphaDiverse organizes alpha mining into iterative plan-and-factor research, then trains local Planner and Realizer agents on diverse traces. Joint optimization rewards predictive quality and complementary contributions.

2 METHODOLOGY

  • Each research loop works within an environment of stock universe, features, starting factors, operators, mechanisms, evaluation settings, and return target, while varying environments broadens research paths.A research path is the sequence of mechanisms and implementations explored over rounds.Source: §2 · p. 3

2.1 RESEARCH AGENT SYSTEM

  • The six-module loop screens factors, retains additions only on improving inner-fold evidence, and freezes factors and predictor before one outer-period evaluation.Screening checks coverage, predictive quality, and redundancy; retention requires weighted score improvement and improvement on a majority of folds.Source: §2.1 · pp. 3–4
  • Only inner-period data provide research feedback; the selected factors and predictor settings are frozen before the final outer-period evaluation.The final predictor is refit within inner data and evaluated once on outer data.Source: §2.1 · p. 4

2.2 TRAINING DATA CONSTRUCTION

  • Planner training samples emphasize model improvement, productive switches, and new directions, while Realizer samples pair plans with distinct valid implementations and repair examples.The Planner sample groups target a 50/30/20 mixture; Realizer repair targets retain the original validated specification.Source: §2.2 · pp. 4–5

2.3 LOCAL POST-TRAINING

  • SFT warm-starts the local Planner to propose research plans and the Realizer to produce valid factor specifications, with repair examples supervising only the final response.The policies share a frozen pretrained backbone and use separate LoRA adapters.Source: §2.3 · p. 5
  • Joint GRPO samples multiple plan portfolios and Realizer implementations per state, then rewards prediction-score gains and quality-weighted coverage beyond the research archive.High-quality plans receive more weight, while similar prediction changes receive diminishing reward; Planner and Realizer receive role-specific comparisons for credit.Source: §2.3 · pp. 5–6

3 EXPERIMENTS

Across four stock universes, the experiments evaluate predictive and portfolio performance, research diversity, ablations, and deployment costs. Research and model selection use the inner period; final evaluation uses the later outer period once.

3.1 MAIN RESULTS

  • 0.0378 IC and 0.0407 RIC lead all predictive metrics; AlphaDiverse also achieves IR 3.210, CR 7.641, lowest MDD 5.51%, and second-highest ARR 42.10%.The results are for CSI300; the paper reports the same overall pattern across all four markets.Source: Table 1 · p. 7, §3.1 · p. 7

3.2 RESEARCH DIVERSITY AND GENERALIZATION

  • Useful mechanisms increase from 4 to 23 and from 6 to 25 relative to SFT, while complete AlphaDiverse leads baselines on three diversity metrics.On CSI300, Single Synthesis produces nearly as many signal clusters but far fewer retained mechanisms.Source: Table 2 · p. 7, §3.2 · pp. 7–8
  • As budget grows, SFT saturates early in both markets, whereas AlphaDiverse continues to add mechanisms within a loop, with later proposals still altering the retained model.Source: §3.2 · p. 8

3.3 ABLATION STUDY

  • Removing complementarity concentrates research on 10 mechanisms; removing retrieval lowers useful coverage and predictive quality, while removing memory weakens prediction.The ablations support distinct roles for planning breadth, relevant inputs for implementation, and feedback for choices.Source: Table 3 · p. 8, §3.3 · p. 8
  • Joint post-training aligns plans with implementations, and the diversity reward raises useful coverage from 18 to 23 while reducing pair correlation.SFT improves prediction but narrows useful coverage; single-role training recovers part of the gain.Source: Table 4 · p. 8, §3.3 · pp. 8–9

3.2 RESEARCH DIVERSITY AND GENERALIZATION

  • In fixed-state tests, AlphaDiverse retains more useful mechanisms than API-based and SFT-only variants in both Familiar and Held-out groups.Each market uses 10 states per group, with eight independent single-round experiments from each state.Source: Figure 4 · p. 8, §3.2 · p. 8

3.3 ABLATION STUDY

  • Improvement-only data selection loses useful mechanisms, while removing paired supervision raises implementation correlation from 0.511 to 0.651 and reduces retained coverage.The results distinguish varied Planner decisions from distinct Realizer implementations.Source: Table 5 · p. 9, §3.3 · p. 9
  • Local AlphaDiverse incurs no external API charges and uses 1.436 H200 GPU-hours for a 20-round CSI300 run; the API alternatives cost $7.28, $5.57, and $2.19.The paper also reports stronger overall performance than all three API variants across four markets.Source: Table 6 · p. 9, §3.3 · p. 9

4 RELATED WORK

Related work applies LLMs to financial prediction and trading, including multi-agent research roles, alpha-generation policies, and external-memory-guided search. Research-path diversity work includes regularization, evolutionary search, redundancy control, and semantic planning.

LLM-Driven Financial Agents

  • LLM financial agents support prediction and trading through task training, role-separated multi-agent systems, alpha-generation policies, and external memory.The cited examples include planning, implementation, and evaluation assigned to distinct agents.Source: §4 · p. 9

Research-Path Diversity

  • Prior diversity approaches use expression regularization, evolutionary search, redundancy control, and semantic planning, while distinct outputs may still share research directions.AlphaDiverse requires different economic mechanisms within each round and collects traces across environments.Source: §4 · p. 9

5 CONCLUSION AND LIMITATIONS

AlphaDiverse combines diverse research traces, local-agent post-training, and joint optimization to support alpha mining with competitive performance and broader exploration. Its scope remains bounded by the mechanism catalogue and finite inner-state evaluation.

5 CONCLUSION AND LIMITATIONS

  • Across four stock universes, AlphaDiverse combines competitive financial performance with sustained exploration using locally deployed research agents.The conclusion presents this as support for effective alpha research with local agents.Source: §5 · p. 9
  • The fixed mechanism catalogue and operator library bound the ideas agents can express, while joint training uses a finite bank of inner states that approximates delayed contributions.The authors identify expanding the search language and assigning credit across evolving research states as directions for further work.Source: §5 · p. 10

A ADDITIONAL RELATED WORK

  • Traditional alpha-mining methods build predictors or factor collections, whereas AlphaDiverse targets the policy that conducts factor research.The related-work discussion distinguishes prediction and evaluation tools from the research-policy learning target.Source: A ADDITIONAL RELATED WORK · p. 15
  • Prior LLM research workflows coordinate proposals and feedback, while AlphaDiverse combines separated search feedback and frozen evaluation with post-training local planning and realization policies.The related-work discussion also notes evaluation leakage as motivation for separating search feedback from final evaluation.Source: A ADDITIONAL RELATED WORK · p. 15
  • Because distinct factor outputs can still concentrate on similar ideas, AlphaDiverse evaluates economic mechanisms and numerical signals together and collects traces through complementary plans and varied environments.The discussion contrasts this approach with prior methods for reducing formula redundancy or broadening factor pools.Source: A ADDITIONAL RELATED WORK · p. 15
  • Joint training compares Planner portfolios and Realizer implementations of the same plans, combining predictive quality with complementary prediction changes in a shared reward.The related-work discussion places this design in the context of cooperative and multi-agent reinforcement learning.Source: A ADDITIONAL RELATED WORK · p. 15

B METHODOLOGY DETAILS

  • Table 7 illustrates one initial research round, with Figure 5 showing two factor computations from an executed API example for one plan’s specifications.The figure’s response contains a realizations list with a plan ID and two specifications; the system, not the Realizer, assigns names, formulas, and data layers.Source: B METHODOLOGY DETAILS · p. 15
  • The economic-mechanism catalogue contains 38 mechanisms in six themes, with each environment retaining supported entries and recording exploration and performance in research history.Each catalogue entry records an economic explanation, required features and construction capabilities, and typical horizons.Source: B METHODOLOGY DETAILS · p. 15

B.1 SCREENING AND DATA SELECTION DETAILS

The appendix details how factors and training examples are screened, and how joint optimization measures quality-weighted diversity across plans and research states.

B.1 SCREENING AND DATA SELECTION DETAILS

  • Factor quality combines predictive and portfolio metrics, while unavailable metrics are omitted and empty or constant factors are rejected.The specified metrics include IC, RIC, ICIR, RICIR, and annualized top-quintile-minus-bottom-quintile return; the signed predictor score is distinct from the factor score’s use of absolute values.Source: §B.1 · pp. 15–16
  • Redundancy is the mean absolute daily cross-sectional correlation, and candidates are processed by decreasing factor score so redundant candidates yield to stronger retained factors.The text distinguishes research-loop acceptance thresholds from the stricter gate for Realizer training pairs.Source: §B.1 · p. 16, Table 9 · p. 17
  • Planner groups are assigned in a fixed priority order to remain disjoint, while each mechanism–implementation signature is retained at most twice and the complete portfolio remains the supervision target.The order is productive switch, model improvement, then new direction.Source: §B.1 · p. 16
  • Realizer training pairs preserve their original plan and retrieved contents, require both implementations to pass research acceptance, and exclude examples without recorded low-correlation evidence.Repair examples pair a faulty response and validation feedback with the validated target; examples from the same source run or identical prompt stay in one split.Source: §B.1 · p. 17

B.2 JOINT REWARD AND OPTIMIZATION DETAILS

  • Joint training uses 96 fixed task states combining four markets, three feature masks, and eight archive configurations, with archived plans supplying diversity histories and redundancy references.Each state is visited twice; archived plans are re-evaluated using the same evaluator as new plans, while archive factors are not added to the baseline model.Source: §B.2 · p. 17
  • Planner portfolios receive four plans each, and two independent Realizer responses implement each portfolio using the same plans and retrieved resources under a constrained JSON grammar.Each Realizer response contains two factor specifications per plan, eight specifications total; Planner generation is not grammar-constrained.Source: §B.2 · p. 18
  • Noise controls use independently sampled, date-standardized Gaussian factors and pass through the candidate admission and model-fitting procedure; they estimate gains from independent Gaussian inputs, not observed-factor dependence.Reward calibration uses the larger of the 75th percentile of absolute historical candidate gains and the 95th percentile of absolute noise-portfolio gains for σs, and the nonnegative 95th percentile of noise-plan gains for κs.Source: §B.2 · p. 18
  • The quality-weighted log-determinant diversity reward favors distinct high-quality prediction-change directions, with diminishing reward for similar directions and reduced gain when the archive covers them.With an empty archive and four unit-quality plans, four orthogonal vectors give D = 1, while four identical vectors give D = log(5)/(4 log 2) ≈ 0.58; sign-reversed vectors provide the same coverage as their originals.Source: §B.2 · p. 19

C.1 DATA AND EVALUATION PROTOCOL

The evaluation uses four Chinese stock universes, daily and intraday market inputs, and chronological inner research followed by a single outer evaluation. It reports predictive and long-only portfolio metrics under consistent portfolio definitions.

C.1 DATA AND EVALUATION PROTOCOL

  • The study covers CSI300, CSI500, CSI1000, and the broader Shanghai/Shenzhen A-share universe, using daily and minute-level price and volume data.Feature masking removes a fixed fraction of available features while retaining basic daily OHLC, volume, and amount, and leaves the initial factor set unchanged.Source: §C.1 · p. 19
  • All features use information available after close t, and the protocol excludes specified stocks and observations, including Beijing-listed and ST stocks, recent listings, suspensions, delistings, and opening-price-limit rejections.The A-share benchmark is the equal-weight return of eligible A-shares; index universes use their corresponding index returns.Source: §C.1 · p. 19
  • The main comparisons use May 2022–December 2025 inner data with chronological training, validation, and search segments in a 7:1:2 ratio; research feedback comes from inner results.Boundary observations are purged for the two-opening label lookahead; the multi-fold variant uses non-overlapping search segments.Source: §C.1 · p. 20
  • After research, factors and the model family are fixed, final settings are selected on the last 15% of inner observations, and the resulting model is evaluated once on outer data.Standalone ML/DL predictors instead fit on the earlier 85% of inner data and use the last 15% for selection; agentic results use all features and one inner fold.Source: §C.1 · p. 20

C.2 PREDICTIVE AND PORTFOLIO METRICS

  • Predictive metrics include date-averaged IC and RIC and their information ratios, while financial tables use an equally weighted long-only top-20% portfolio with annualized excess return, IR, MDD, and Calmar ratio.MDD is a positive loss magnitude, so lower is better; reported returns are gross of trading costs.Source: §C.2 · pp. 20–21

C.3 BASELINES IMPLEMENTATION

The appendix specifies the prediction and agentic baselines, their comparison budget, and the local models’ post-training and inference setup. It also reports the construction and size of the training data and optimization rollouts.

C.3 BASELINES IMPLEMENTATION

  • The ML/DL baselines span linear and tree models, feed-forward and recurrent networks, temporal architectures, and models that incorporate market or stock interactions.The appendix describes Ridge, LightGBM, MLP, GRU, LSTM, ALSTM, TCN, Transformer, PatchTST, iTransformer, MASTER, and StockMixer.Source: §C.3 · p. 21
  • ML/DL baselines receive the same base factors, sequence models use 20 trading days, and neural results are metric-wise medians across three seeds.Neural models use a shared 64-dimensional hidden representation and training settings, with model-specific adaptations; Ridge and LightGBM use validation-based selection or early stopping.Source: §C.3 · p. 21
  • The agentic baselines cover iterative feedback, originality and complexity checks, alternative mutation trajectories, and semantic plan selection with two realizations.Methods named are RD-Agent(Q), AlphaAgent, QuantaAlpha, and AlphaSchema; each receives 160 candidate-factor slots, and the paper reports one complete run per method–market configuration.Source: §C.3 · p. 21

C.4 LOCAL POST-TRAINING AND INFERENCE

  • Planner and Realizer use separate rank-64 LoRA adapters, while joint GRPO starts from their SFT policies with separate role references and quality/diversity weights of 0.75/0.25.Training uses 8×H200 GPUs; inference enables reasoning for local Planner, Realizer, and Analysis nodes, but Analysis uses the original model rather than a post-trained policy.Source: §C.4 · p. 22
  • The SFT datasets come from 72 API research loops across three backends, three feature-mask rates, four markets, and two fold settings, yielding 589 Planner and 1,006 Realizer examples.Realizer targets contain eight factor specifications for four plans, and the example counts include repairs.Source: §C.4 · p. 22
  • Training rollouts contain 768 Planner responses and 1,220 Realizer responses, fewer than the 1,536 maximum because invalid Planner outputs do not invoke the Realizer.Tables 12–14 report dataset sizes, source configurations, and generated response/token counts, respectively.Source: Table 12 · p. 22, Table 13 · p. 22, §C.4 · p. 23

C.5 DIVERSITY METRICS

The section defines mechanism-coverage metrics for generated and retained factors, plus signal-cluster and paired-plan correlation measures. External methods are mapped to the same catalogue, with the authors noting limits to cross-method comparability.

C.5 DIVERSITY METRICS

  • M counts distinct catalogue mechanisms among new factors, while U counts mechanisms represented by new factors retained in the final inner-selected model.Base factors are excluded, and a mechanism contributes once to U if at least one of its factors survives selection.Source: §C.5 · p. 23
  • C counts complete-linkage groups whose factors have pairwise mean absolute correlation of at least 0.9; larger values indicate more numerically distinct signals.Correlations are averaged over inner dates; sign-reversed copies are grouped together, while constant signals are excluded.Source: §C.5 · p. 23
  • Pair corr. is the median mean absolute correlation for the two factors implementing each plan, and is undefined for methods without paired plans.The measure is used only for AlphaDiverse variants with paired implementations.Source: §C.5 · p. 23
  • External methods are labelled against the shared catalogue by blinded annotators, with disagreements adjudicated; “other” and “uninterpretable” labels do not add to M or U.Differences in search spaces and factor representations can affect annotation, so the principal diversity comparison concerns changes within AlphaDiverse.Source: §C.5 · p. 23

D.1 ADDITIONAL FINANCIAL RESULTS

Across CSI500, CSI1000, and A-share, AlphaDiverse leads several predictive and risk-adjusted metrics, although other methods lead selected return metrics. The results extend the reported pattern to these additional stock universes.

D.1 ADDITIONAL FINANCIAL RESULTS

  • The comparison uses the same data splits, input settings, evaluation protocol, and eight predictive and portfolio metrics across all three additional markets.The experiments compare AlphaDiverse with ML, DL, and agentic baselines.Source: §D.1 · p. 24
  • AlphaDiverse has the highest IC, ICIR, RICIR, and IR and the lowest MDD across CSI500, CSI1000, and A-share, while other methods lead selected ARR and RIC comparisons.Other methods achieve the strongest ARR on CSI500 and CSI1000 and strongest RIC on CSI500 and A-share.Source: §D.1 · p. 24, Table 17 · p. 25, Table 18 · p. 25
  • 0.0364 RIC and 2.063 CR are AlphaDiverse’s highest values on CSI1000, alongside 16.45% ARR.On A-share, ARR is 19.04%, IR is 2.029, and MDD is 4.77%.Source: §D.1 · p. 25, Table 17 · p. 25, Table 18 · p. 25

D.2 RESULTS OF API-BASED ALPHADIVERSE

Across four markets, API-backend results vary: different backends lead on selected predictive, portfolio, and diversity measures. All three support diverse exploration, while performance depends on backend and market.

D.2 RESULTS OF API-BASED ALPHADIVERSE

  • The backend comparison assigns the same 20-round budget and evaluation protocol to GPT-5.5, Grok-4.6, and GLM-5.3 across all four markets.Each backend supplies the Planner, Realizer, and Analysis nodes.Source: §D.2 · p. 26
  • GPT-5.5 leads ARR on CSI300 and CSI500 at 35.97% and 6.12%, respectively, and leads IC and RIC in both markets.Grok-4.6 leads ARR and IR on CSI1000 and A-share, with ARR of 4.53% and 15.14%.Source: §D.2 · p. 26
  • GLM-5.3 explores the most mechanisms on CSI300 and CSI1000 and produces the most signal clusters on three markets, but its 28 useful CSI1000 mechanisms accompany negative outer ARR.This illustrates that inner-retained coverage and outer returns capture different outcomes.Source: §D.2 · p. 26

D.3 CROSS-METHOD RESEARCH DIVERSITY

Across four stock universes, AlphaDiverse retains more useful mechanisms than the compared methods, and its coverage continues to grow with candidate budget. It also retains stronger useful coverage on shared research states from source runs excluded from training-data construction.

D.3 CROSS-METHOD RESEARCH DIVERSITY

  • AlphaDiverse has the highest useful-mechanism coverage in all four markets and highest or joint-highest mechanism coverage; versus SFT-only, U rises from 4 to 23, 6 to 25, 18 to 29, and 16 to 28.Single Synthesis approaches AlphaDiverse’s CSI300 signal-cluster count (122 versus 129) but retains 8 rather than 23 mechanisms; GLM-5.3 has more clusters than the complete model on three markets but fewer useful mechanisms throughout.Source: §D.3 · p. 27

D.4 FOUR-MARKET COVERAGE AND TRANSFER

  • At 160 candidates, complete AlphaDiverse reaches 23 and 25 useful mechanisms on CSI300 and CSI500, versus 4 and 6 for SFT-only; it also leads final coverage on CSI1000 and A-share.On CSI1000 and A-share, complete AlphaDiverse reaches 29 and 28, compared with 18 and 16 for SFT-only; coverage is tracked at five candidate budgets.Source: §D.4 · p. 28, Figure 6 · p. 28, Table 21 · p. 28
  • On held-out source-run states, AlphaDiverse retains more useful alternatives than API-based GPT-5.5 and SFT-only variants in both groups across all four markets.Held-out coverage is 12.6 on CSI300, 13.2 on CSI500, 15.7 on CSI1000, and 13.0 on A-share; every variant has lower coverage in the held-out group.Source: §D.4 · p. 29, Figure 7 · p. 29

D.5 MORE ABLATION RESULTS

The ablations test how research-system modules, post-training objectives, and SFT-data selection affect predictive performance and research diversity. Further analyses examine factor retention and report the accounting scope for local deployment costs.

D.5 MORE ABLATION RESULTS

  • Removing complementarity cuts mechanism coverage to 7–10 across four markets, versus 24–28 for the complete workflow, while retrieval, memory, and single-agent synthesis also weaken useful outcomes.Removing complementarity nevertheless achieves a higher ARR on CSI300; removing retrieval lowers U and RIC, removing memory lowers RIC, and Single Synthesis retains fewer useful mechanisms.Source: §D.5 · p. 30
  • Adding the diversity reward to joint Planner–Realizer training increases U, reduces pair correlation, and improves ARR, IR, MDD, and CR across all four markets.On CSI300, U rises from 18 to 23 while pair correlation falls from 0.365 to 0.316; the ablations compare SFT, single-policy updates, joint training, and joint training without diversity.Source: §D.5 · p. 31, Table 23 · p. 31
  • Full-data SFT achieves the highest U and lowest pair correlation across markets; improvement-only selection and removing balancing reduce useful coverage or U.Removing paired supervision raises CSI300 pair correlation from 0.511 to 0.651 and lowers U from 4 to 1; improvement-only selection excludes productive-switch and new-direction examples.Source: §D.5 · p. 32, Table 24 · p. 32

D.6 FACTOR QUALITY

  • AlphaDiverse retains 56 factors, compared with 35 for GPT-5.5 and 8 for SFT-only AlphaDiverse.The untrained model produces more accepted factors than GPT-5.5 but retains fewer in the predictive model, motivating attention to downstream contribution when screening factors.Source: §D.6 · p. 33, Table 25 · p. 33

D.7 DEPLOYMENT COST DETAILS

  • The deployment accounting covers Planner, Realizer, and Analysis for one complete CSI300 research run with all features, one inner fold, and 20 rounds.In the local configuration, all three nodes use Qwen3.8-27B with reasoning enabled; the Planner and Realizer use jointly GRPO-trained policies, while Analysis uses original weights.Source: §D.7 · p. 33, Table 26 · p. 33
Loading 2609.29014v1…