Source-linked AI summary
Mint-Agent: Introducing Finance-Native Agentic Foundation Models
Mint-Agent Team, B. Zhang, Yaze Geng, Lei Tang, Yaoyang Yi, Zonghan Wu, Yifan Hu, Kun Wang, Qingsong Wen, Yilei Shao
TL;DR
Financial agents need reliable, auditable execution across multi-step research rather than isolated factual recall. Mint-Agent unifies provenance-grounded data, evidence-preserving interaction, and specialized training, with Mint-Ag achieving the highest reported score across seven professional benchmarks, including 98.33% on RFC-Bench.
Problem
Financial agents struggle to connect authoritative evidence, reporting periods, calculations, and conclusions into reliable, traceable financial research.
Method
Mint-Agent unifies provenance-grounded task construction, evidence-preserving execution, and training that integrates financial reasoning and agentic execution experts.
Results
Across seven professional benchmarks, Mint-Ag attains the highest reported score on every benchmark, including 98.33% on RFC-Bench.
Takeaways & Limitations
Mint-Agent demonstrates that financial agents can combine domain reliability, extended execution, and auditable evidence within one finance-native foundation.
Abstract
from arXiv · showhide
Financial agents must do more than recall domain knowledge: they must be both reliable, executing precise operations over grounded evidence, and executive, sustaining long-horizon research whose conclusions remain auditable. We present Mint-Agent, a family of finance-native agentic models designed around these two scales of financial intelligence. Mint-Agent is built upon three pillars: data, harness, and algorithm. Our data engine constructs clean, specialized tasks for atomic financial capabilities and long-horizon agentic execution from real-world financial sources. MintHarness enables stable interaction with open-ended environments and maintains auditable evidence trails across extended research trajectories. Our training recipe combines SFT, critical-step OPD, and RLVR to develop separate financial reasoning and agentic execution experts, which are then unified through model merging and multi-teacher on-policy distillation into compact, general-purpose financial agents. This pipeline yields two flagship models, Mint-Cu (9B) and Mint-Ag (27B). Across professional financial benchmarks, our models demonstrate two defining strengths: (1) Reliability: Mint-Ag achieves 98.33% on RFC-Bench, surpassing GPT-5.6-Sol and Claude-Opus-4.8 by 3.66 and 3.00 points; and (2) Executability: Mint-Cu reaches 69.86% on FinSearchComp T2, outperforming Agents-A1-35B and Nex-N2-mini by 22.83 and 12.78 points, while Mint-Ag achieves 76.00% and 60.49% on FinanceAgentBench v1.1 and v2, respectively. These results establish a path toward trustworthy financial intelligence in which domain expertise, long-horizon execution, and auditable evidence are jointly engineered as a unified foundation for frontier agentic models.
1 Introduction
Mint-Agent frames finance-native intelligence as auditable, evidence-backed research requiring reliable reasoning and executable long-horizon workflows. Its data, harness, and algorithm pillars produce unified agents evaluated across professional financial benchmarks.
- Motivation: Finance research requires contextual disclosure reading, period identification, cross-source comparison, and arithmetic that supports decisions.Existing agents may fail through non-authoritative sources, incorrect fiscal periods, or silently mixed calculations.
- Problem: Mint-Agent treats every answer as a claim backed by recoverable evidence and reproducible calculation.The framework preserves provenance and execution state so failures remain available for supervision.
- Method: Its three pillars combine real-world task construction, persistent stateful execution, and verifiable policy learning for auditable financial research.Training develops atomic financial capabilities and long-horizon execution separately before integrating them through multi-teacher on-policy distillation.
- Evaluation: 98.33% on RFC-Bench, 91.33% on FinanceBench, and 55.71% on BizFinBench are among Mint-Ag’s reported benchmark results.Mint-Ag also achieves 76.00% and 60.49% on FinanceAgentBench v1.1 and v2, plus 89.04% and 54.07% on FinSearchComp T2 and T3.
2 Related Work
Related work traces a shift from fixed workflows and financial knowledge toward learned agentic policies that execute, verify, and sustain research over explicit evidence. Mint-Agent unifies harness-based control with training-based investigation through a recoverable evidence contract for finance.
- Agentic models: Agentic models increasingly learn multi-turn tool policies that select, sequence, and verify actions, while managing context through learned retention, compression, recovery, and coordination.The harness continues to define the environment and expose actions, but post-training increasingly determines interaction trajectories.
- Financial intelligence: Financial intelligence has progressed from domain vocabulary and broad task coverage toward quantitative reasoning and verifiable problem solving.Financial pre-training and instruction tuning establish foundational knowledge, while post-training targets more explicit numerical and evidential boundaries.
- Deep research: Deep research treats intelligence as an extended loop of identifying unknowns, acquiring evidence, and revising conclusions until objectives are resolved.This broader research process centers on evidence acquisition and iterative conclusion refinement.
- Deep research: Two complementary deep-research lines use explicit harnesses for plans, state, parallel exploration, and evidence assembly, or training to internalize search, planning, reflection, and context management.Harness-based systems make trajectories observable and controllable, whereas training-based systems learn sustained search from long interactions with denser supervision.
- Mint-Agent: Mint-Agent joins these lines by grounding finance task construction, execution, verification, and specialist training in one recoverable evidence contract.The policy learns how to investigate while the harness retains sufficient evidence and computation state to audit conclusions.
3 Data Engine
The Mint data engine develops financial executability at atomic and long-horizon scales, using grounded, provenance-linked sources and separately synthesized task streams. Atomic tasks test bounded financial operations, while long-horizon tasks require discovering and composing hidden evidence through open-ended interaction.
- Atomic financial capabilities: Atomic tasks target five capabilities: knowledge, extraction, calculation, analysis, and verification.Each task receives one primary capability label κ ∈ K, while its financial topic is recorded separately.
- Atomic task data: The engine builds atomic tasks from primary financial materials, including EDGAR, issuer archives, XBRL, exchange records, FRED, and accounting taxonomies.Extracted facts retain links to exact document spans or structured fields.
- Atomic task construction: Atomic task generation samples semantic chunks and provenance-linked facts, then retains capability-conditioned questions, traces, and answers only when evidence support, replayability, and uniqueness checks pass.The retained task preserves source, evidence, reasoning trace, and capability metadata.
- Long-horizon execution: Long-horizon tasks withhold supporting context, requiring agents to locate and assemble information through extended interaction while training and evaluation use complete hidden answer graphs.The long-horizon corpus combines the grounded atomic corpus with reproducible time-indexed snapshots and overlap-free analyst-written seed requests.
- Long-horizon task construction: Evidence graphs encode extracted or derived facts, valid financial operations, and provenance, fixing nontrivial answers and complete derivations before query realization.Queries pass audit only when replay returns a unique answer, every leaf has localized provenance, and hidden routes remain unexposed.
4 MintHarness
MintHarness turns bounded financial questions and open-ended research objectives into a common stateful execution process with task-specific actions, persistent state, and auditable evidence. It separates durable execution state from bounded model context while carrying plans and grounded conclusions across turns.
- Task Setup: MintHarness normalizes each task, exposes its admissible action space, records interactions, and maintains persistent state outside the model context.Atomic tasks expose only SUBMIT, while long-horizon tasks additionally expose configured tools.
- Action Cycle: At each turn, PackB constructs a budgeted model-visible context, while Gateξ validates action types, tool schemas, and remaining execution budget.Validation determines whether a proposed action executes or produces an explicit error; the research decision remains with the policy.
- Evidence Ledger: MintHarness distills tool results into typed, provenance-linked financial records and merges them into a persistent evidence ledger.Failed calls add no evidence but preserve error observations, while the ledger separately records failed retrieval paths and supporting evidence.
- Completion: Termination occurs after submission, step-budget exhaustion, or an unrecoverable environment error, producing a prediction, score, and auditable execution record.The record includes the final ledger, trajectory, usage, and status.
- Working Memory: Working memory persists a compact semantic state containing plans, confirmed conclusions, unresolved questions, and intermediate calculations linked to ledger records.It is refreshed after each transition from the trajectory and evidence ledger.
- Context Management: The harness separates durable execution state from bounded context by retaining recent turns, compressing older interactions, and referencing long documents or tables as artifacts.The packing function selects the objective, relevant state, and ledger entries under the token budget.
5 Training Pipeline
Mint-Agent trains separate financial-reasoning and agentic-execution specialists from a common base policy, then unifies them through model merging and multi-teacher on-policy distillation. The pipeline grounds learning in verified answers, replayable evidence, valid interactions, and targeted correction of consequential execution errors.
- Specialist training: The pipeline trains a financial-reasoning specialist on Datom and an agentic-execution specialist on Dlong before integrating them into one policy.Both specialists start from the common base policy πθ0.
- Financial reasoning: SFT anchors financial reasoning to valid derivations, while RLVR rewards capability-aware answers whose premises are grounded and derivations replayable under financial constraints.The verifier checks answer equivalence, evidence locators, source support, and period and unit constraints; alternative grounded derivations receive credit.
- Agentic execution: Agentic execution uses SFT for stable interaction behavior, critical-step OPD to repair consequential decisions, and RLVR over complete MintHarness trajectories.Trajectory rewards require a verified final answer, recoverable evidence provenance, and valid actions respecting the collection cutoff.
- Expert unification: TIES merging combines the reasoning and execution experts by aligning their task-vector updates relative to the common base policy before adding the merged updates back.Low-magnitude coordinates and conflicting expert updates are discarded before aligned updates are averaged and scaled.
- Expert unification: Multi-teacher on-policy distillation routes atomic tasks to the financial-reasoning teacher and long-horizon interactions to the agentic-execution teacher.The student generates on-policy tokens conditioned on either bounded financial inputs or accumulated interaction histories and observations.
6 Experiments
Mint-Agent performs strongly on both bounded financial reasoning and open-ended agentic execution, with Mint-Ag leading several benchmarks and both checkpoints occupying favorable accuracy–cost positions. Trajectory analyses and failure reviews further examine persistent evidence use, source reconciliation, and remaining execution errors.
- Atomic Reasoning: Mint-Ag leads atomic reasoning with 55.71 on BizFinBench, 91.33 on FinanceBench, and 98.33 on RFC-Bench.These results exceed the strongest non-Mint results by 4.14, 1.33, and 3.00 points, respectively.
- Agentic Execution: 76.00 on FinanceAgentBench v1.1, 60.49 on v2, and 89.04 on FinSearchComp T2 make Mint-Ag the strongest external-comparison performer across these tasks.Mint-Ag exceeds the best external scores by 4.00, 3.70, and 7.30 points, while Mint-Cu reaches 69.86 on FinSearchComp T2 and Mint-Ag reaches 54.07 on T3.
- Pareto Efficiency: Both Mint checkpoints lie on the empirical Pareto frontier for FinanceAgentBench v1.1 and v2, with no evaluated alternative simultaneously more accurate and less expensive.On v1.1, Mint-Cu records 68.0% at $0.016 per task, while Mint-Ag records 76.0% at $0.090, 72.5% below the $0.327 comparison mean.
- Trajectory Analysis: A 37-step TSMC forecast trajectory survives blocked sources, wrong-quarter pages, scope correction, and recomputation while retaining valid guidance and recovered monthly revenue records.After restoring the 2022–2024 averaging window, the agent recomputes March 2025 growth and projects revenue.
- Trajectory Analysis: The BKR acquisition trajectory resolves an exact $171.65 closing price from the definitive proxy and reproduces a 22.3% premium and $325 million synergy ratios.The ledger preserves the primary source behind every reported quantity and treats blocked or ambiguous pages as discovery leads rather than evidence.
- Failure Modes: On v1.1, Mint-Cu solves 68% of tasks with answer omission at 18% as its largest residual failure, while v2 accuracy falls to 41.98% and the dominant error shifts.PhD researchers manually reviewed every trajectory and assigned its earliest consequential error.
- Effect of MOPD: MOPD improves the merged 9B checkpoint by 1.33, 3.86, 4.00, and 3.19 points across four benchmarks, restoring performance across both capability groups.TIES produces uneven transfer, whereas the resulting MOPD scores align with Mint-Cu results in Table 2.
7 Conclusion
Mint-Agent presents a full-stack finance-native foundation built around financial correctness, combining provenance-linked data, auditable execution, and integrated reasoning and execution training. Its complementary 9B and 27B models achieve strong results across financial benchmarks while treating auditability as foundational to trustworthy financial intelligence.
- System design: Mint-Agent combines atomic capabilities and long-horizon trajectories, recoverable evidence-ledger execution, context management, SFT, critical-step OPD, RLVR, TIES, and multi-teacher OPD.Its data engine uses provenance-linked sources, while the training stack integrates reasoning and execution experts.
- Model family: Mint-Cu is a compact 9B model designed to sustain strong financial research performance at favorable inference cost, while Mint-Ag targets more demanding reasoning and execution.The two models realize the same foundation at complementary scales.
- Benchmark results: 98.33 on RFC-Bench, 76.00 and 60.49 on FinanceAgentBench v1.1 and v2, and 89.04 on FinSearchComp T2 are reported for Mint-Ag across seven benchmarks.These results are presented as evidence of strong performance across financial benchmarks.
- Trustworthiness: Trustworthy financial agents must recover relied-on evidence, respect temporal boundaries, expose fact-to-claim transformations, and revise conclusions when links fail.The conclusion frames auditability as a shared substrate for data construction, agent execution, and policy learning rather than a post-inference reporting layer.
8 Contributors
The paper identifies four core contributors and three corresponding authors.
- Core Contributors: Core contributors are Yaze Geng, Lei Tang, Yaoyang Yi, and Zonghan Wu.
- Corresponding Authors: Corresponding authors are Kun Wang, Qingsong Wen, and Yilei Shao.
A Benchmark Verification Protocols … A.3 FinSearchComp T2 and T3
The benchmark verification protocol preserves each dataset’s released evaluation materials and applies benchmark-specific scoring procedures. BizFinBench, FinanceBench, and FinSearchComp are evaluated through official scripts, answer matching, or judge templates without a shared cross-dataset judge.
- A Benchmark Verification Protocols: Each benchmark retains its released evaluator, reference answers, or per-example rubrics, while prompts use released versions except for noted omitted demonstrations.The protocol avoids introducing a shared judge across datasets and applies only benchmark-specified evaluation materials.
- A.1 BizFinBench: BizFinBench uses the official v2 evaluation scripts and selects each sample’s released comparison function according to task_type.Covered task types include anomaly tracing, event-logic reasoning, financial data description, multi-turn perception, quantitative computation, report analysis, stock-price prediction, and sentiment analysis.
- A.1 BizFinBench: BizFinBench extracts the model-visible final answer and performs only formatting required by the relevant official script.The evaluator is therefore aligned with each task-specific comparison function rather than a shared scoring mechanism.
- A.2 FinanceBench: FinanceBench uses released reference answers and its official answer-matching procedure to report question-level accuracy.The evaluator receives the extracted final answer without additional task-solving context or external evidence.
- A.3 FinSearchComp T2 and T3: FinSearchComp T2 and T3 use the same released judge_system_prompt and judge_prompt_- template for both tracks.The judge compares the extracted student answer with the reference answer and accompanying scoring points.
- A.3 FinSearchComp T2 and T3: FinSearchComp awards 1 only when every required element is correct; otherwise, including empty or error-valued answers, it awards 0.Multi-part answers receive credit only when all required elements are correct.
- A.3 FinSearchComp T2 and T3: Numerically equivalent answer formats are treated as equal, and a unit may be omitted when the question already specifies it.The evaluator returns the binary answer_score together with a brief rationale in JSON.
A.4 FinanceAgentBench v1.1 and v2
FinanceAgentBench evaluates financial-agent answers with rubric-based correctness judgments and separate contradiction checks against expert references. Its binary sample score requires every criterion to be satisfied without factual contradiction.
- Evaluation protocol: Formatting, ordering, and writing style do not affect matching, while reasonable rounding is accepted when units and orders of magnitude align.Scoring is based on correctness rubrics released with each example.
- Evaluation protocol: The benchmark judge receives the question, expert reference answer, model prediction, and ordered correctness criteria.The judge evaluates the submitted answer against the provided reference and rubric inputs.
- Evaluation protocol: The judge makes criterion-level match decisions and separately checks whether confident factual claims contradict the expert reference.The procedure instructs the judge not to solve the question or use outside knowledge.
- Illustrative rubric example: 2/3 criteria were satisfied in the illustrative example, but the binary sample score was 0 because the operating-margin criterion failed and the prediction contradicted the reference.The prediction reported 14.0% operating margin versus the reference value of 14.3%.
A.5 RFC-Bench Task 2
RFC-Bench Task 2 evaluates source-grounded classification of the principal manipulation type in counterfactual financial paragraphs. It uses exact label matching rather than an LLM judge, so ambiguous or invalid outputs are incorrect.
- Task definition: Task 2 asks models to identify the principal manipulation type in a source-grounded counterfactual financial paragraph.The task provides source article or page content alongside a manipulated financial news paragraph.
- Evaluation: The evaluator extracts one of four labels—numerical, flipping, sentiment, or causal—from the model-visible final answer.It normalizes a small set of label aliases before evaluation.
- Evaluation: Exact matching against the reference perturbation type determines accuracy, with ambiguous or invalid outputs counted as incorrect.The benchmark uses no LLM judge, and reported results refer specifically to Task-2 classification accuracy rather than the full RFC-Bench aggregate.
- Task constraints: Models must use only the provided source page and manipulated paragraph, without browsing, retrieval, or outside evidence.The classification labels correspond to mechanisms including altered quantitative information or reversed direction or polarity.
B Additional Case Study
The AMD guidance case study demonstrates retrieval recovery and evidence reuse during a multi-period financial calculation. After an initial lookup failure, the agent broadens SEC search, extracts quarterly guidance data, and reopens prior evidence through an auditable ledger.
- AMD Guidance: 13-step trajectory sizes AMD revenue-guidance ranges across four quarters after a failed accession guess triggers broader SEC search and exhibit-level extraction.The agent locates relevant EX-99.1 exhibits and records each quarter’s midpoint and range endpoints.
- AMD Guidance: 10.53%, 8.96%, 8.00%, and 8.45% are the reported range widths for the four quarterly guidance records.The final answer is supported by four primary filings.
- AMD Guidance: The ledger reopens the Q3 source without repeating the full search, enabling evidence reuse in the replayable calculation.This case focuses on retrieval recovery and reuse of previously collected evidence during a multi-period financial calculation.