acceptodds
Under review as a conference paper at ICLR 2027

FinGym: Scaling Financial Agents via Grounded Environment Synthesis

Abstract

Financial agents have the potential to automate complex workflows, such as market analysis, trading, and payment processing. Training such agents requires environments that are diverse, scalable, and financially grounded. However, real financial environments are costly and risky to access, while general environment synthesis may overlook financial states, rules, and operational constraints, producing tasks that require infeasible actions. We introduce FinGym, a framework that translates natural-language requirements into customized financial environments grounded in domain knowledge and financial constraints. FinGym adopts reusable financial entities, market data, and tools into environments with explicit roles, constraints, and workflows. The environment specifications guide both task generation and rubric synthesis, providing executable training tasks for agentic reinforcement learning. Using FinGym, we construct 73 financial environments spanning ten scenarios and synthesize 5,040 verified agentic tasks to train FinGym-4B and FinGym-9B. Across five financial-agent benchmarks, these agents achieve relative improvements of 5.8%–441.9% over their corresponding base models and competitive performance against strong proprietary models. These results highlight FinGym as a scalable and reliable generator of environments and tasks for training financial agents to meet diverse financial requirements. The code and data are available at https://anonymous.4open.science/r/anony-0D833.

Primary area:
reinforcement learning
Keywords:
Financial agents, Agentic RL, Mock Environment

Reviews

Among:
Shape:
1234this paper, Top 49%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.33 over 3 reviews, above 51% of the others.
Reviewer_GxaCOpenReview ↗
Rating
3
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FinGym presents a framework for generating financially grounded environments and tasks for training financial agents. It combines financial knowledge, real-world data, and tools to create executable RL environments. Using 73 environments and 5,040 verified tasks, FinGym-trained agents achieve significant improvements over base models across five financial benchmarks.

Critical strengths
  1. The paper proposes a domain-specific solution targeting real problems in financial agent post-training.
  2. The trained models achieve relatively promising performance.
  3. The analysis experiments are relatively comprehensive, with interesting discussions on cross-scenario contributions and scaling.
Critical weaknesses
  1. Compared with general agent environment scaling methods, this paper seems to mainly add financial-domain grounding.
  2. The experimental comparison is not entirely fair: compared with EnvScaler-8B, AWM-8B, and TOUCAN-7B, this paper uses a more advanced base model and stronger models, and it does not compare the amount of training data used. Therefore, it is unclear whether the improvements over previous works come from the FinGym framework itself, or from the stronger base model, stronger models, or larger training data scale.
  3. Whether human involvement reduces environment scalability should be discussed more thoroughly.
Questions for authors
  1. What are the differences between the environments constructed by FinGym and real-world financial environments?
  2. How much does FinGym’s performance improvement come from the proposed environment synthesis framework itself, rather than the stronger base models or larger training data scale compared with previous environment scaling methods?
  3. How does the involvement of human experts affect the scalability and cost of constructing financial environments in FinGym?
4 comments
AuthorsOpenReview ↗

Response to Reviewer GxaC: first part

We thank the reviewer for recognizing the practical relevance of our domain-specific framework, the promising performance improvements, and the comprehensive experimental analysis, particularly regarding cross-scenario contributions and scaling behavior. We address your concerns and questions below:

Weakness 1 & Question 1: FinGym’s Contributions and Relation to Real-World Environments

Thank you for these questions. We would like to clarify the contributions of our work from three perspectives.

1. Methodological contribution: grounding task generation and reward construction through explicit financial environment specifications. Given a user requirement, FinGym composes reusable financial components into an environment specification defining participants, tools, objectives, financial constraints, and workflow dependencies. This specification guides task generation and, together with teacher execution evidence, informs task-specific rubrics. The financial knowledge thus shapes both what agents are asked to do and how their behavior is rewarded. Components and specifications can be reused to generate additional tasks as training requirements evolve.

2. Empirical contribution: demonstrating the value of financial knowledge integration and studying environment scaling. FinGym improves agent performance across five financial-agent benchmarks, supporting both the importance of financial knowledge and the effectiveness of our approach to incorporating it into tasks and rewards. We further investigate how performance changes as environment coverage expands, revealing overall gains from broader coverage and complementary training benefits across financial scenarios.

3. Resource contribution: enabling reuse, reproduction, and continued extension. We released our code, 73 financial environments, 5,040 verified tasks, and trained models. Users can directly use the released environments, tasks, and models, or apply our synthesis pipeline to construct additional environments and tasks tailored to their own financial requirements.

Regarding Question 1, FinGym preserves key requirements of real financial workflows. Agents gather evidence through tools, complete interdependent operations, and make decisions under financial constraints. These requirements apply to workflows such as fund information retrieval, product comparison, and authorization checks. In investment and trading scenarios, historical market data provide factual context where applicable, while the modeled constraints include available cash, holdings, transaction costs, and trading quantity limits.

The main difference from real-world financial environments is that these workflows run in a simulated, repeatable setting. Agents receive simulated tool responses without moving real funds or modifying real accounts, and tasks can be rerun from the same starting conditions. This supports repeated training and controlled comparisons while avoiding the operational risks of experimenting with real accounts. However, this controlled setting simplifies real-world complexity. The current task set may not fully capture rare market events, extreme trading conditions, or unexpected operational disruptions. We will clarify these contributions and the scope of the synthesized environments in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer GxaC: second part

Weakness 2 & Question 2: Performance Gains Attributable to FinGym’s Environment Synthesis

Thank you for raising this important question. We ran an additional experiment to test whether FinGym’s gains persist when all methods use the same base model and number of training tasks. We used AWM, TOUCAN, and EnvScaler to generate 5,040 tasks each, then trained Qwen3.5-4B model on each dataset and compared the resulting models with FinGym-4B.

Table A: Effect of environment synthesis methods on Qwen3.5-4B.

Synthesis methodFinMCP-Bench ↑Finova ↑FinVault ↓StockBench ↑InvestorBench ↑
AWM33.9459.3231.750.720.93
TOUCAN33.8358.9928.680.891.21
EnvScaler33.1459.7931.901.081.09
FinGym35.0761.4026.704.011.55

As shown in Table A, FinGym-4B outperforms all three baselines on all five benchmarks when the base model and number of training tasks are held fixed. Compared with the strongest baseline on each benchmark, FinGym-4B gains 1.13 points on FinMCP-Bench and 1.61 points on Finova, and lowers the FinVault aggregate score by 1.98 points. It also improves the StockBench Sortino ratio by 2.93 and the InvestorBench Sharpe ratio by 0.34.

We also observe that most retrained baselines outperform their corresponding released checkpoints in Table 4. For example, the retrained AWM model scores 33.94 on FinMCP-Bench and 59.32 on Finova, compared with 30.76 and 55.28 for the released AWM-8B checkpoint. This supports the reviewer’s point that differences in the base model and training setup can affect the comparison. We will include the experimental setup and results in the revised manuscript.

Weakness 3 & Question 3: Impact of Human Experts Involvement on Scalability and Cost

Human experts were primarily involved in two aspects of the framework. For rubric synthesis, they review financial rules informed by domain knowledge and failed teacher rollouts. These rules guide the construction of task-specific rubrics (Section 3.3). For environment synthesis, they review the specifications generated by the LLM and refine the selected components where needed to better reflect the user’s requirements (Section 3.2).

Importantly, the results of human review can be reused across tasks. Once reviewed, a financial rule can guide rubric synthesis for any task to which it applies. Similarly, a reviewed environment specification can be used to generate new tasks with similar requirements, supporting scalability while amortizing the initial human effort.

Moreover, the ablation results in Table 7 show that removing human environment calibration leads to only modest performance changes across all five benchmarks (only 0.24–0.97 points), indicating that the automated synthesis pipeline remains effective without human calibration, while expert involvement provides consistent additional gains.

AuthorsOpenReview ↗

Gentle Follow-up to Reviewer GxaC

Thank you again for your valuable comments. In our previous responses, we clarified FinGym's methodological and practical contributions, provided additional controlled comparisons against three environment synthesis baselines, and further analyzed the impact of human expert involvement on scalability.

We truly appreciate the time and effort you dedicated to reviewing our work, and your comments have been very helpful in improving the paper. We hope these clarifications and additional evidence have adequately addressed your concerns. If any concerns remain, we would be more than happy to discuss them further and provide any additional information needed to address them.

AuthorsOpenReview ↗

Gentle Follow-up to Reviewer GxaC

Thank you for your thoughtful review and comments

We have now incorporated the relevant revisions into the updated manuscript PDF, with all changes highlighted in blue\color{blue}{\text{blue}} and their locations summarized below:

  • W1, Q1 (FinGym’s methodological, empirical, and resource contributions, and its relation to real-world financial environments): Lines 210-212, 204-206, and Appendices C.1.3 and C.2.
  • W2, Q2 (Controlled comparisons and performance gains attributable to environment synthesis): Lines 387-389,and Appendix C.6.1.
  • W3, Q3 (Expert involvement, reuse, scalability, and amortization of human effort): Lines 204-206, and Appendix C.2.

We hope these revisions help address your concerns. We would be grateful if you could consider updating your evaluation score based on the additional experiments and clarifications presented in the revised manuscript.

Thank you again for your time and consideration.

Reviewer_NMh4OpenReview ↗
Rating
2
Confidence
3
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

FinGym turns a natural-language financial requirement into an executable training environment for agentic reinforcement learning. Its key idea is to ground task synthesis in explicit financial components, rules, and tools, so generated tasks are feasible and verifiable rather than merely plausible. The paper releases 73 environments and 5,040 tasks, trains FinGym-4B and FinGym-9B, and reports gains over base models and environment-scaling baselines on five financial-agent benchmarks, with competitive results against proprietary models on selected tasks.

Critical strengths

This work shows FinGym beating both LLM-only and AWM synthesis on all five benchmarks, and the gain is attributable to task synthesis rather than model scale. Rubric synthesis is a contribution, deriving rubrics from successful and failed trajectories plus environment specifications, and Figure 5 shows higher tool-call accuracy at every checkpoint with a widening gap. Financial grounding is measured rather than claimed: removing financial knowledge degrades all five benchmarks, and real-data constraints raise InvestorBench Sharpe. These results make FinGym a credible data-and-training resource for financial agents.

Critical weaknesses

The rubric judge that supplies the reward is never identified or checked against human experts, so it may be inconsistent, biased, or self-preferential if it is the same model as the teacher. Task validation also keeps only tasks with a successful teacher trajectory, and the teacher is much larger than the trained models. The rejection rate is not reported, so it is unclear whether the 5,040 tasks are representative or a teacher-solvable subset.

StockBench is an unexplained outlier and may reflect data overlap. FinGym-9B moves from 1.19 to 5.43, a +356% gain, far above the next largest relative gain. Training uses historical stock data via Qlib, and StockBench and InvestorBench also evaluate on historical market data, but the paper does not state training or evaluation windows and entities. Until overlap is ruled out, the gain is hard to attribute to capability.

Questions for authors

Which model serves as the rubric judge, and has its scoring been checked against human expert judgments?

What fraction of candidate tasks is rejected during teacher validation, and what are typical reasons for rejection?

Could you specify the time periods and entities used in the Qlib training data and in the StockBench and InvestorBench evaluations, and confirm whether they overlap?

5 comments
AuthorsOpenReview ↗

Response to Reviewer NMh4: first part

We thank the reviewer for recognizing the effectiveness of our financial environment synthesis framework, the contribution of rubric synthesis, and the comprehensive experimental evidence supporting the benefits of financial grounding and task synthesis. We address your concerns and questions below:

Weakness 1 & Question 1: Rubric Judge Reliability and Human Agreement

Thank you for raising this concern. We use GPT-5.5 as the rubric judge. To assess the reliability of its scoring, we conducted additional comparisons with human experts and two alternative LLM judges.

1. Agreement with human experts. We randomly sampled 100 evaluation trajectories, covering 500 rubric-level judgments. Three human experts independently assessed the trajectories against the corresponding rubrics. Their majority-vote decisions served as the reference for comparison with GPT-5.5.

Table a. Agreement between GPT-5.5 and human majority-vote judgments on rubric items.

ComparisonAgreement ↑Cohen’s κ ↑Macro-F1 ↑
GPT-5.5 vs. human majority vote93.4%0.8560.928

GPT-5.5 achieves 93.4% agreement with the human reference, with Cohen’s κ of 0.856 and Macro-F1 of 0.928. These results provide evidence that its rubric-level judgments closely align with human assessments on the sampled trajectories.

2. Consistency across LLM judges. We also evaluated Claude Opus 5 and DeepSeek V4 Pro on the same 500 rubric items, using the same trajectories, rubrics, and scoring instructions. Table b reports pairwise agreement among the three judges.

Table b. Pairwise agreement among LLM judges on the same rubric items.

ComparisonAgreement ↑Cohen’s κ ↑Macro-F1 ↑
GPT-5.5 vs. Claude Opus 593.2%0.8530.927
GPT-5.5 vs. DeepSeek V4 Pro93.8%0.8660.933
Claude Opus 5 vs. DeepSeek V4 Pro94.2%0.8750.937

Pairwise agreement ranges from 93.2% to 94.2%, with Cohen’s κ ranging from 0.853 to 0.875. This indicates substantial consistency in rubric-level scoring across the three models. Together with the human comparison, these results support the reliability of the reward judge on the evaluated samples. We will include the above results in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer NMh4: second part

Weakness 1 & Question 2: Task Rejection and Dataset Coverage

During teacher validation, approximately 35.6% of the 7826 generated candidate tasks were rejected. As described in Lines 219–221, validation considers three criteria: task executability, goal completion, and scenario constraint compliance. The corresponding rejection statistics and representative failure cases are detailed below.

1. Task Executability (1179 cases, 15.1%). This criterion assesses whether a candidate task can be completed given its initial state, available information, and accessible tools. For example, a task may require verifying a loan application, while all available tools require an application ID that is neither provided nor retrievable. Although the task description appears complete, no valid tool execution can be initiated. Such failures were identified in 1179 candidate tasks.

2. Goal Completion (459 cases, 5.9%). This criterion verifies whether the execution trajectory actually fulfills the user's objective, rather than merely completing tool calls successfully. For example, a task may require retrieving transaction records and completing settlement, but the generated trajectory only retrieves the records without executing settlement. Despite successful tool calls, the intended objective remains unfulfilled. This criterion identified 459 failures.

3. Scenario Constraint Compliance (1148 cases, 14.7%). This criterion examines whether task execution respects authorization requirements, trading restrictions, and domain-specific business rules. For example, a trajectory may attempt to sell shares purchased on the same day in a market subject to T+1 settlement restrictions. Even if the simulated tool reports a successful transaction, the trajectory is rejected for violating the applicable trading rules. This criterion identified 1148 failures.

These statistics illustrate the primary failure modes detected during teacher validation. By jointly examining executability, goal completion, and financial rule compliance, our validation procedure filters out trajectories that are infeasible, incomplete, or inconsistent with domain-specific constraints, rather than relying solely on successful tool execution.

AuthorsOpenReview ↗

Response to Reviewer NMh4: third part

Weakness 2 & Question 3: Training–Evaluation Data Separation and StockBench Performance

1. Training–evaluation data separation. Our Qlib-based training tasks use Chinese A-share data from 2019, whereas StockBench and InvestorBench evaluate U.S. stocks in later periods. The data coverage is summarized below.

Table c. Markets, assets, and data periods used in training and evaluation.

Data usageMarket and assetsPeriod
Qlib-based training tasksChinese A-sharesJanuary 2–December 31, 2019
StockBench evaluationU.S. stocks — 20 stocksMarch 3–June 30, 2025
InvestorBench (warm-up and evaluation)U.S. stocks — UVV, HON, JNJ, and MSFTJuly 1, 2020–May 5, 2021

InvestorBench uses July–September 2020 for warm-up, with evaluation beginning on October 1, 2020. The Qlib training-data period therefore strictly precedes both benchmark evaluation periods, including the InvestorBench warm-up. We additionally checked the training and evaluation stock pools and confirmed no overlapping securities. Thus, the Qlib-based training tasks do not directly reuse the evaluated stocks’ historical price series.

Within each training task, the agent receives an initial portfolio state and market information available at the specified decision point. Future market outcomes and subsequent trading trajectories are excluded from its decision context. To support independent verification, we have released the original training market data in our anonymous GitHub repository.

2. Interpreting the StockBench improvement. The reported 356.3% improvement is compared to the Qwen3.5-9B base model, whose Sortino ratio increases from 1.19 to 5.43 after FinGym training. This large relative gain partly reflects the low baseline score. One possible explanation for the baseline’s limited performance is the difficulty of long-horizon trading: StockBench and InvestorBench require sequential decisions with interdependent portfolio states, combining financial reasoning, portfolio tracking, and compliance with trading constraints. Early mistakes can affect subsequent decisions and accumulate over time, making these tasks particularly challenging for general-purpose base models.

Reviewer_v4Z5OpenReview ↗
Rating
2
Confidence
4
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

Authors proposed a pipeline for generating training environments for LLM agents in finance. An LLM takes a natural language requirement and selects components from a library: synthetic individuals, companies, regulators and institutions; about 1,500 tools; and market records from SEC EDGAR, Tiantian Fund and Qlib. From these it writes an environment specification of roles, tools, objectives, constraints and workflow.

Within each environment, the same LLM generates tasks, tool outputs and reference answers. A task is retained if a teacher model solves it and an LLM judges the solution. The judge then writes five rubrics per task. Qwen3.5-4B and Qwen3.5-9B are trained with GRPO, with rewards assigned by an LLM judge against these rubrics.

The authors report 73 environments and 5,040 tasks.

Critical strengths
  • The authors correctly identify that financial environments need careful, domain-specific design to be realistsi e.g .generic synthesis can produce tasks that violate market rules.
Critical weaknesses
  • The environments are generated, not data-backed. Tool outputs and reference answers are written by GPT-5.5 as “synthetic results” e.g Fig. 10, rules 9 and 12. The non-market components, including regulatory rules, are synthetic. Neither is checked against the cited data sources or for financial correctness.

  • Task verification establishes internal consistency, not correctness. A task is retained if GPT-5.5 can solve it against tool outputs that GPT-5.5 itself generated e.g Figs. 14 and 16. This confirms executability within a synthetic world, not that the task or its answer is financially valid.

  • Look-ahead bias is not controlled. Temporal integrity relies only on prompt instructions. The training data’s dates and assets are not reported, so overlap with the StockBench and InvestorBench test periods cannot be excluded.

  • The evaluation is circular. Task quality is scored by unspecified LLM judges on an author-defined rubric (Table 5), and the in-house test set is produced by the same pipeline as the training data (Fig. 3, Table 14).

  • Contribution of the teacher model not isolated. A task is retained only if GPT-5.5 solves it. The rubrics used as RL rewards are then written by an LLM from GPT-5.5’s own solution to that task (Fig. 16), so they reward the student for following GPT-5.5’s steps. Without a baseline that fine-tunes the student directly on these GPT-5.5 solutions, the contribution of the environments cannot be separated from imitation of GPT-5.5.

Questions for authors
  1. What proportion of tool outputs is taken directly from EDGAR, Tiantian Fund or Qlib rather than generated by GPT-5.5? How were the generated outputs and the regulatory components checked for accuracy?
  2. Was any sample of retained tasks and reference answers reviewed by human experts for financial validity? If so, what fraction was judged correct?
  3. What are the dates and assets of the market data used in training? Do the trading results on StockBench and InvestorBench hold when all training data strictly precede their evaluation periods?
  4. Which models scored task quality in Table 5, and how well do they agree with human experts? How do the trained models perform on a test set constructed independently of the FinGym pipeline?
  5. How does GRPO on FinGym compare with supervised fine-tuning on the same GPT-5.5 solutions at matched compute?
7 comments
AuthorsOpenReview ↗

Response to Reviewer v4Z5: first part

We thank the reviewer for recognizing the importance of domain-specific environment design in financial agent training, particularly the need to incorporate financial rules and constraints beyond generic environment synthesis. We address your concerns and questions below:

Weakness 1 & Question 1: Real-World Grounding and Financial Validity Verification

We would like to clarify that FinGym is not constructed solely from unconstrained LLM-generated outputs. Instead, its construction incorporates real-world financial data, human expert-curated financial rules, and multi-stage task validation to improve the financial validity of the synthesized environments and tasks. We provide further details below.

1. Real-world Data Grounding. FinGym incorporates financial records from three sources to provide factual context for task synthesis.

Table a. Real-world financial data sources and their coverage of generated tasks.

Data sourceInformation usedShare of generated tasks
SEC EDGARCorporate disclosures and financial statements11.2%
Tiantian FundHistorical fund information13.5%
QlibHistorical stock prices and market indicators39.7%

These sources provide factual grounding for financial information retrieval, corporate analysis, fund evaluation, and market trading tasks, rather than relying exclusively on GPT-5.5-generated financial information.

2. Human Expert-Curated Financial Rules. As described in Lines 226–228, the financial rule sets are curated by human experts based on failed teacher trajectories and domain-specific financial knowledge, rather than being generated solely by GPT-5.5. Specifically, the rule sets comprise 126 rules for financial information retrieval and interactive question answering, 244 rules for financial security tasks, and 111 rules for market trading tasks, totaling 481 expert-curated rules.

These rules are incorporated into rubric construction to guide task generation and evaluation. Overall, 96.5% of the generated tasks are influenced by these financial rules, with an average of 2.7 rubric criteria per task supplemented or corrected using expert knowledge. This provides additional domain-specific constraints beyond unconstrained LLM generation.

3. Multi-Stage Task Validation. As mentioned in Lines 219–221, our validation procedure considers three complementary criteria beyond successful tool execution. Approximately 35.6% of the 7826 generated candidate tasks were rejected.

  • Task Executability (1179 identified failures): This criterion examines whether the task can be completed given its initial state, available information, and accessible tools. For example, a loan verification task may require an application ID that is neither provided nor retrievable through the available tools. Although the task description appears complete, no valid execution can be initiated, and the task is therefore rejected.
  • Goal Completion (459 identified failures): This criterion verifies whether the execution actually fulfills the user's objective rather than merely completing tool calls. For example, a task may require retrieving transaction records and completing settlement, but the generated trajectory only retrieves the records without executing settlement. Despite successful tool calls, the intended objective remains unfulfilled, and the task is rejected.
  • Scenario Constraint Compliance (1148 identified failures): This criterion examines whether task execution complies with applicable financial and business constraints. For example, a trajectory may attempt to sell shares purchased on the same day in a market subject to T+1 trading restrictions. Even if the simulated tool reports a successful transaction, the trajectory is rejected for violating the corresponding trading rules. These checks help identify infeasible tasks, incomplete execution trajectories, and violations of financial constraints before the tasks are retained for training.

Overall, real-data grounding, expert-curated rules, and task validation help improve the financial validity of FinGym’s synthetic environments and tasks. We will include the above details in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer v4Z5: second part

Weakness 2 and Question 2: Task Verification and Financial Correctness

Thank you for your insightful comment. We agree that successful execution within a synthetic environment does not necessarily guarantee financial correctness. As detailed in our response to W1 and Q1, FinGym applies multi-stage task verification to assess executability, goal completion, and compliance with scenario-specific financial and business constraints.

To assess financial correctness beyond the internal consistency of synthetic environments, we conducted an additional evaluation involving 200 task trajectories, with 50 samples randomly selected from each of four financial task categories.

Three PhD students with backgrounds in both finance and AI independently reviewed the sampled tasks. They mainly checked whether (i) financial information, concepts, and indicator definitions in task descriptions and tool outputs were correct; (ii) calculations of monetary amounts, fees, returns, holdings, and account values were correct, with consistent units and calculation bases; and (iii) financial analyses, risk assessments, and related conclusions followed reasonable steps and were logically supported by the preceding analysis. Each expert assigned an overall valid/invalid label, with the final judgment determined by majority vote. The results are summarized below.

Table b. Human expert evaluation of financial validity across task categories.

Task CategoryEvaluated SamplesIdentified ErrorsError Rate ↓
Market Analysis5012%
User Asset Management5012%
Financial Security5036%
Market Trading5000%
Overall20052.5%

The results show that 97.5% of the sampled trajectories passed human expert verification, with an overall identified error rate of 2.5%. Notably, no errors were identified in the sampled market trading tasks, where verification outcomes were grounded in actual historical market data corresponding to the trading dates, rather than solely relying on LLM-generated outputs. The remaining issues were primarily associated with tool-selection ambiguities and domain-specific constraint handling. For example, an agent may select a functionally equivalent tool with a different name from the reference tool, resulting in a verification mismatch despite performing the intended operation.

Overall, LLM-based verification alone cannot guarantee financial correctness. Our multi-stage validation ensures task executability, goal completion, and constraint compliance, while independent human evaluation further assesses financial validity, mitigating reliance on GPT-5.5-generated outputs and reference answers.

Weakness 3 & Question 3: Temporal Integrity and Training–Evaluation Data Separation

Thank you for raising this concern. Our Qlib-based trading tasks use 2019 Chinese A-share data, whereas StockBench and InvestorBench evaluate U.S. stocks in later periods, with no temporal or asset overlap between the Qlib training data and either benchmark. We clarify the date ranges, asset coverage, and decision-time information below.

1. Temporal and asset separation. Table c summarizes the markets, assets, and data periods used for Qlib-based training and benchmark evaluation. Our trading tasks use Chinese A-share data from 2019, whereas both benchmarks use U.S. stock data from later periods.

Table c. Markets, assets, and periods of Qlib training data and our benchmark evaluations.

Data usageMarket and assetsPeriod
Qlib-based training tasksChinese A-sharesJanuary 2–December 31, 2019
StockBench evaluationU.S. stocks — 20 stocksMarch 3–June 30, 2025
InvestorBench (warm-up and evaluation)U.S. stocks — UVV, HON, JNJ, and MSFTJuly 1, 2020–May 5, 2021

The Qlib training-data period strictly precedes both benchmark evaluation periods. We also checked the training and evaluation stock pools and confirmed that they contain no overlapping securities. Thus, the Qlib-based training tasks do not reuse the evaluated stocks’ historical price series.

2. Point-in-Time Integrity. Our trading-related tasks are constructed as independent daily tasks rather than continuous trading trajectories. Each task specifies an initial portfolio state, including the user's available cash and existing holdings before the trading decision. The agent is then required to make trading decisions based on the market information available at that decision point, without access to future market outcomes or subsequent trading trajectories. This design limits the use of future information and prevents the training tasks from relying on long-horizon evaluation trajectories.

To further support reproducibility and independent verification, we have also released the original stock market data used for training in our anonymous GitHub repository. We will include these details in the revised manuscript.

AuthorsOpenReview ↗

Response to Reviewer v4Z5: third part

Weakness 4 & Question 4: Judge Reliability and Independent Evaluation

Thank you for this detailed comment. We provide further details on the task-quality assessment, agreement between LLM and human judgments, and performance on independent benchmarks below.

1. Evaluation Protocol for Synthesized Financial Tasks. Table 5 compares the quality of tasks generated by LLM Only, AWM, and FinGym. For each method, we randomly sampled 10 tasks from each of the 73 financial scenarios, yielding 730 tasks per method. All tasks were evaluated by Claude Opus 5 and DeepSeek-V4-Pro using the same prompt and scoring criteria. The criteria covers six dimensions: goal clarity, state–goal consistency, constraint consistency, information boundaries, tool sufficiency, and task feasibility, each scored on a 1–5 scale. For each task and dimension, we averaged the two judges’ scores, then averaged these scores across the 730 tasks to obtain the results reported in Table 5. We acknowledge that the original manuscript did not sufficiently describe this evaluation protocol. We will include the judge model names and versions, full evaluation prompt, and scoring procedure in the revised manuscript.

2. Human assessment and agreement with LLM judges. To assess the reliability of the LLM-based scores, we sampled a subset of 219 tasks, with three tasks from each of the 73 financial scenarios, and asked three human evaluators to score them using the same six-dimensional rubric. For each task and dimension, we compared the mean rating from the three human evaluators with the score from the LLM judge. The results are summarized below.

Table d. Agreement between LLM and human scores on FinGym-generated tasks.

Evaluation DimensionLinear Weighted κ ↑Spearman’s ρ ↑MAE ↓
Goal Clarity0.8420.8720.132
State–Goal Consistency0.8230.8640.152
Constraint Consistency0.8660.8800.107
Information Boundaries0.8290.8410.132
Tool Sufficiency0.8260.8400.151
Task Feasibility0.8390.8630.135

The results demonstrate consistently strong agreement between LLM-based scoring and human evaluation across all six dimensions, with linear weighted κ ranging from 0.823 to 0.866 and Spearman's ρ ranging from 0.840 to 0.880. These findings provide additional evidence supporting the reliability of our data quality assessment.

3. Performance on independently constructed benchmarks. Our main results in Table 4 are evaluated on five public benchmarks constructed independently of the FinGym pipeline: FinMCP-Bench, Finova, FinVault, StockBench, and InvestorBench. None of their evaluation instances were included in the FinGym training data. FinGym-trained models improve over their corresponding base models across all five benchmarks. Detailed results are provided in Tables 9–13.

We additionally conducted an out-of-distribution (OOD) evaluation in the supply-chain domain, which was absent from both the environment construction process and the real-world data sources used for training. As reported in Table 8, the FinGym-trained model achieves a final worth of 907.68, compared with 806.82 for Qwen3.5-4B, representing a 12.5% improvement. This result provides further evidence of the generalization capability and effectiveness of our proposed approach in an unseen domain.

Comments on the paper

1 comment
AuthorsOpenReview ↗

General Response to the Area Chair and Reviewers

General Response to the Area Chair and Reviewers

Dear Area Chair and Reviewers,

We sincerely thank all reviewers for their thoughtful feedback and constructive suggestions. During the discussion, we provided additional experiments and clarifications addressing the main concerns about FinGym’s contributions, comparison fairness, reward reliability, and evaluation validity. We summarize these responses below.

  • Positioning and Contributions. We clarified how FinGym incorporates financial knowledge into both task generation and reward construction. FinGym composes reusable financial components into explicit environment specifications covering participants, tools, objectives, financial constraints, and workflow dependencies. These specifications guide task synthesis and, together with execution evidence, inform task-specific rubrics. We also clarified the relationship between our simulated environments and real financial workflows.

  • Controlled Comparisons with Environment Synthesis Baselines. To separate the contribution of our synthesis framework from differences in model capacity and training data scale, we conducted additional comparisons with AWM, TOUCAN, and EnvScaler using the same Qwen3.5-4B base model and 5,040 training tasks per method. FinGym achieves the best results across all five benchmarks under this controlled setting. Compared with the strongest baseline on each benchmark, it improves FinMCP-Bench and Finova by 1.13 and 1.61 points, reduces the FinVault score by 1.98 points, and increases the StockBench Sortino ratio and InvestorBench Sharpe ratio by 2.93 and 0.34, respectively.

  • SFT Comparison. On Qwen3.5-9B, we compared rubric-guided GRPO with SFT using GPT-5.5 reference trajectories from the same FinGym training tasks. SFT improves over the base model, while GRPO achieves further gains across all five benchmarks.

  • Reward Judge Reliability. We identified the rubric judge and evaluated its judgments against human experts and two alternative LLM judges. Three human experts independently assessed 500 rubric items from 100 sampled trajectories, with majority-vote judgments serving as the reference. GPT-5.5 achieves 93.4% agreement, Cohen’s κ of 0.856, and Macro-F1 of 0.928 against this reference. Pairwise agreement among the three LLM judges ranges from 93.2% to 94.2%, providing additional evidence of scoring consistency on the evaluated examples.

  • Training–Evaluation Separation and Post-Release Evaluation. We documented the markets, assets, and periods used for training and evaluation: the Qlib-based training tasks use Chinese A-share data from 2019, whereas the trading benchmarks use later U.S. stock data, with no overlapping securities. To further investigate potential historical-data memorization, we trained Qwen2.5-3B and Llama3.2-3B and evaluated them on market data from periods after both base models were released. Training improves performance on both StockBench and InvestorBench for both models, providing further evidence that the gains persist on post-release market data.

  • Task Validation and Data Quality. We reported that approximately 35.6% of 7,826 candidate tasks were rejected during teacher validation and provided a breakdown of the rejection reasons. Validation checks task executability, goal completion, and compliance with scenario constraints, filtering out tasks with missing prerequisites, incomplete outcomes, or violations of financial rules. We also provided representative failure cases to clarify how these checks assess task validity beyond successful tool execution. An independent human review of 200 sampled task trajectories found that 97.5% were financially valid. Reviewer v4Z5 acknowledged that this human review addressed their concern about financial validity.

  • Human Involvement, Scalability, and Reproducibility. We clarified that expert involvement focuses on reviewing financial rules and environment specifications, which can be reused across tasks to reduce repeated review effort. The existing ablation shows that removing human calibration causes modest performance declines across all five benchmarks, while the automated pipeline remains effective. We also supplied additional evaluation data and execution logs through our anonymous repository to support independent verification.

We have incorporated these additional results and clarifications into the revised manuscript, with the new material highlighted in blue\color{blue}{\text{blue}}. We thank the reviewers again for helping us strengthen the empirical evidence and presentation of our work, and we hope this summary assists the Area Chair’s assessment.

Sincerely,
The Authors

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.