Mint-Agent: Introducing Finance-Native Agentic Foundation Models

Mint-Agent Team; Kun Wang; Gavin Zhang; Yaze Geng; Lei Tang; Yaoyang Yi; Zonghan Wu; Yifan Hu

cs.CLAug 17, 2026

Mint-Agent: Introducing Finance-Native Agentic Foundation Models

Authors: Mint-Agent Team, Kun Wang, Gavin Zhang, Yaze Geng, Lei Tang, Yaoyang Yi, Zonghan Wu, Yifan Hu,

Organizations: Mint Agent Team · Nanyang Technological University · Shanghai AI-Finance School, East China Normal University · Tsinghua University

Abstract

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.

Figures & tables

Figure 1: Overview of the Mint Agent stack. Mint Agent is jointly developed along three complementary dimensions: data, harness, and algorithm. The data engine builds verifiable atomic and long-horizon financial tasks; the harness supports persistent context and multi-source interaction; and the algorithmic pipeline combines supervised fine-tuning, reinforcement learning with verifiable rewards, and on-policy distillation. Together, these components produce the Mint-Cu and Mint-Ag models.

Figure 2: Capability-guided construction of atomic financial tasks. A target capability is selected from knowledge, extraction, calculation, analysis, and verification. Primary financial sources are sampled and segmented into semantic chunks, from which structured, provenance-linked facts are extracted. A capability-conditioned constructor generates a question, reasoning trace, and answer, which are retained only after passing checks for evidential support, reproducibility, and answer uniqueness.

Source layer	Collection	Role
Disclosures	EDGAR and issuer archives	Grounded financial text and tables
Structured facts	XBRL records	Normalized reported values
Financial series	Exchange records and FRED	Market and macro observations
Definitions	Accounting taxonomies	Metric semantics and valid formulas

Table 1: Sources used to construct atomic financial tasks.

Figure 3: Construction and verification of long-horizon agentic tasks. The agent observes only a research objective and must recover the answer through iterative search, inspection, calculation, and synthesis. The underlying source pack and evidence graph remain hidden, preserving the task’s exploratory nature while enabling deterministic verification.

Figure 4: MintHarness overview. Given a financial task and task-specific configuration, MintHarness orchestrates a stateful execution loop in which the agent proposes actions, the harness validates and executes tool calls, and resulting observations update the execution state. Evidence, working memory, and the complete action–observation trajectory are maintained outside the bounded model context. Upon submission, budget exhaustion, or error, an isolated evaluator compares the prediction against hidden reference data and produces a score and auditable execution record.

Figure 5: Twin-specialist training and integration pipeline. Starting from a shared base policy, we train a financial-reasoning specialist on atomic tasks using supervised fine-tuning and reinforcement learning with solution-level verifiers, and an agentic-execution specialist on long-horizon tasks using supervised fine-tuning, critical-step on-policy distillation, and trajectory-level reinforcement learning. The two specialists are first merged with TIES and then integrated through multi-teacher on-policy distillation with origin-matched routing, producing a unified policy instantiated as Mint-Cu and Mint-Ag.

Model	BizFin- Bench	Finance Bench	RFC Bench	FinanceAgentBench		FinSearchComp
Model	BizFin- Bench	Finance Bench	RFC Bench	v1.1	v2	T2	T3
Frontier models
Gemini-3.5-Flash	50.71	85.33	92.33	62.00	$25.93_{\scriptscriptstyle\pm 0.09}$	36.53	44.19
Claude Opus 4.8	49.71	90.00	95.33	66.00	$55.56_{\scriptscriptstyle\pm 0.00}$	45.66	51.16
GPT-5.6-Sol	51.57	86.67	94.67	66.00	$56.79_{\scriptscriptstyle\pm 0.12}$	41.10	50.58
GLM-5.2	47.57	88.00	94.33	58.00	$51.85_{\scriptscriptstyle\pm 0.09}$	68.49	34.88

Table 2: Performance comparison on financial benchmarks. All scores are percentages. FinanceAgentBench v2 reports the mean of three runs; the subscripted

\pm

value is the variance. Green cells, from darkest to lightest, denote the best, second-best, and third-best result in each column.

Panel of Figure 6 — Figure 6: Cost–performance trade-offs on FinanceAgentBench. Correct outcomes are plotted against average API cost per question. The dashed vertical line marks the mean cost of the compared models, and the highlighted curve traces the empirical Pareto frontier.

Figure 7: Failure taxonomy on FinanceAgentBench. Correct outcomes and primary failure modes are shown for representative frontier models and Mint-Cu on v1.1 and v2.

Checkpoint	Finance Bench	BizFin Bench	FAB v1.1	FinSearch T2
Reasoning expert ( $\pi_{\mathrm{R}}$ )	88.00	53.86	—	—
Execution expert ( $\pi_{\mathrm{A}}$ )	—	—	66.21	72.15
TIES merge ( $\pi_{\mathrm{M}}$ )	88.67	50.00	64.00	66.67
MOPD ( $\pi_{\mathrm{U}}$ )	90.00	53.86	68.00	69.86

Table 3: Effect of expert integration for the 9B model. All scores are percentages; — denotes a specialist not evaluated outside its target domain.

Figure 8: The mechanism analysis of Mint-Agent from the cognitive architecture view.

Figure 9: TSMC forecast. Persistent evidence and compact working state carry valid guidance and monthly revenue records through 37 steps of source failures, scope corrections, and recomputation.

Figure 10: BKR acquisition. Multiple evidence streams are reconciled only after the unaffected closing price is confirmed in the definitive proxy, enabling reproducible premium and synergy calculations.

Appendix figures & tables1 asset

Supplementary material from the paper’s appendix.

Appendix

Figure 11: AMD guidance. A failed document lookup is replaced by broader SEC discovery, exhibit-level extraction, and a replayable calculation over four quarterly guidance records.

LLM Agent Training

Large language model agent training teaches models to plan, use tools, and learn from multi-step interaction with environments.

See all 586 papers

Herculean: An Agentic Benchmark for Financial Intelligence
May 14, 2026Xueqing Peng, Zhuohan Xie, Yupeng Cao +60AI Agent Benchmarks LLM Agent Evaluation
FinToolBench: Evaluating LLM Agents for Real-World Financial Tool Use
Mar 9, 2026Jiaxuan Lu, Kong Wang, Yemin Wang +9LLM Agent Evaluation Tool-Using Language Agents
Shanghai AI Laboratory · Hunan University · Xiamen University +3
FinReasoning: A Hierarchical Benchmark for Reliable Financial Research Reporting
Feb 25, 2026Yiyun Zhu, Yidong Jiang, Ziwen Xu +4Financial Reasoning in Language Models Financial Reasoning Benchmarks
School of Computer Science and Technology, Tongji University, Shanghai, China · Shanghai Artificial Intelligence Laboratory, Shanghai, China

May 14, 2026J/K move · Enter open · S save

Herculean: An Agentic Benchmark for Financial Intelligence

Xueqing Peng, Zhuohan Xie, Yupeng Cao, Haohang Li+59

AI Agent Benchmarks LLM Agent Evaluation Long-Horizon Agent Evaluation Financial Services Financial Decision-Making Finance+1

Open paper AskPDF arXiv

As AI agents advance, the central question is no longer whether they can solve isolated financial tasks, but whether their capabilities generalize across complete professional workflows. Existing benchmarks cannot answer this question because they evaluate either static competencies or agents within a single interaction structure. We introduce Herculean, a benchmark of end-to-end agentic financial intelligence that crosses agent architecture, backbone model, and four structurally distinct workflows: Trading, Hedging, Market Insights, and Auditing. Each workflow is implemented as an MCP skill that preserves its native state, actions, constraints, and outcome criteria, allowing intact agent systems to be evaluated under matched external conditions. Across 30 architecture--backbone configurations, we find that agent capability is not globally ordered. Workflow rankings are weakly aligned, changing the architecture frequently reverses performance under the same backbone, and each architecture concentrates failure in different forms of workflow control. These results establish that agent performance is an interaction between architecture and workflow demand, rather than an intrinsic property of the underlying model. Consequently, success on a single financial workflow provides little evidence of dependable, general financial intelligence.

Abstract

LLM Agent Training

Explore similar work

Herculean: An Agentic Benchmark for Financial Intelligence

FinToolBench: Evaluating LLM Agents for Real-World Financial Tool Use

FinReasoning: A Hierarchical Benchmark for Reliable Financial Research Reporting