PaperlayerView exact revision on arXiv

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

arXiv:2608.16386v1cs.CLcs.LG

TL;DR

Financial research agents can produce plausible steps without a traceable path from source to calculation to conclusion, making results difficult to trust or repair. Mint-Agent combines provenance-linked task construction, an evidence-first harness, and integrated reasoning and execution experts; across seven benchmarks, Mint-Ag attains the highest reported score on every benchmark, including 98.33% on RFC-Bench.

  • Problem

    Existing agents can produce plausible individual steps without a traceable source-to-calculation-to-conclusion chain, limiting trust and repair in financial research.

    Source: §1 · p. 2

  • Method

    Mint-Agent combines provenance-linked task construction, an evidence-ledger harness, and training that integrates separate financial-reasoning and agentic-execution experts.

    Source: §7 · p. 20

  • Results

    Across seven benchmarks, Mint-Ag attains the highest reported score on every benchmark, including 98.33% on RFC-Bench.

    Source: §1 · p. 3

  • Takeaways & Limitations

    Mint-Agent positions recoverable evidence and testable conclusions as a shared substrate for trustworthy financial-agent competence.

    Source: §7 · p. 21

Abstract

from arXiv · show

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 treats financial answers as claims backed by recoverable evidence and reproducible calculations, and builds finance-native agents around data, an evidence-first harness, and learning algorithms.

1 Introduction

  • Financial research can fail through non-authoritative sources, incorrect reporting periods, or mixed units, making traceable source-to-calculation-to-conclusion evidence necessary for trust and repair.The paper argues that task provenance, preserved execution state, and supervision of decisions—not only final text—must reinforce the same standard of financial correctness.Source: §1 · p. 2
  • Separate reinforcement learning with verifiable rewards develops atomic financial capabilities and long-horizon execution, which multi-teacher on-policy distillation integrates into Mint-Cu (9B) and Mint-Ag (27B).Both models treat financial research as a recoverable sequential decision process rather than isolated answer generation.Source: §1 · p. 3
  • 98.33% on RFC-Bench, 76.00% on FinanceAgentBench v1.1, and 60.49% on v2 accompany Mint-Ag’s highest reported score across all seven benchmarks.Mint-Ag also scores 55.71% on BizFinBench, 91.33% on FinanceBench, and 89.04% and 54.07% on FinSearchComp T2 and T3; Mint-Cu reaches 69.86% on FinSearchComp T2.Source: §1 · p. 3
  • Mint-Agent unifies provenance-grounded task construction, persistent stateful execution, and verifiable policy learning under a recoverable evidence contract.Its data engine uses real-world financial sources and evidence crossover for grounded multi-hop queries; MintHarness supports heterogeneous sources and extended search while prioritizing evidence collection.Source: §1 · pp. 2–3

2 Related Work

The related work traces a shift toward learned agent policies, finance tasks grounded in recoverable evidence, and deep-research systems that combine autonomous investigation with auditable execution. Mint-Agent applies this joint requirement to finance through a shared evidence contract for task construction, execution, verification, and specialist training.

2.1 Agentic Models

  • Agentic policies increasingly learn multi-turn tool use, trajectory-level planning, context management, recovery, and coordination, while specialist integration combines weight-space composition with student-state on-policy distillation.The emerging agent-native model still depends on a harness to preserve state, constrain actions, and expose failures for learning (Yadav et al., 2023; Niu et al., 2026; Ma et al., 2026a).Source: §2.1 · pp. 3–4

2.2 Financial Intelligence

  • Financial intelligence has shifted from domain adaptation and grounded answers toward analyst workflows that discover evidence, reconcile sources, and make conclusions reconstructable.Correctness includes entity, reporting period, scale, unit, and derivation; auditability requires reconstructing source authority, time boundary, financial semantics, and calculation (Wu et al., 2023; Yang et al., 2023; Xie et al., 2023; Liu et al., 2025; Hu et al., 2025b; Bigeard et al., 2025; Zhu et al., 2025; Haque et al., 2026; Li et al., 2026a; Cheng et al., 2026; Wang et al., 2026a; Luan et al., 2026; Lu et al., 2026; Huang et al., 2026; Pauli et al., 2026; Srivastava, 2026; Xiao et al., 2026).Source: §2.2 · p. 4

2.3 Deep Research

3 Data Engine

Mint’s data engine trains financial executability at two scales: bounded-evidence atomic tasks and long-horizon tasks that require agents to find and compose evidence through interaction. It grounds both in auditable sources and retains the provenance and answer structures needed for verification.

3 Data Engine

  • The data engine separates atomic finance tasks, which use bounded evidence, from long-horizon tasks that require agents to discover and compose evidence through interaction.The two streams are synthesized separately as Datom and Dlong, then combined as DMINT.Source: §3 · p. 5

3.1 Atomic Financial Capabilities

  • Atomic capabilities cover knowledge, extraction, calculation, analysis, and verification, with each task assigned one primary capability separately from its financial topic.These groups cover concepts and metrics, fact retrieval, formula-based calculation, evidence-based interpretation, and checking whether claims are supported.Source: §3.1 · p. 5
  • Atomic tasks use primary financial materials and provenance-linked facts to generate capability-conditioned questions, traces, and answers retained only when supported, replayable, and unambiguous.Sources include EDGAR and issuer archives, exchange records, FRED, and accounting taxonomies; facts retain exact document-span or structured-field locators.Source: §3.1 · pp. 5–6, Figure 2 · p. 6

3.2 Long-Horizon Agentic Execution

  • Long-horizon tasks withhold context until interaction, using a fixed hidden answer graph; their sources include time-stamped snapshots and analyst-written seed requests filtered against evaluation sets.The snapshots preserve evidence reproducibility when live pages or feeds change.Source: §3.2 · p. 7
  • Long-horizon tasks retain grounded fact-and-operation graphs with localized provenance and sufficient reasoning depth and source breadth, then audit replay, answer uniqueness, leakage, and time boundaries.Only the objective is exposed during execution; the answer and audit metadata remain hidden.Source: §3.2 · pp. 7–8

4 MintHarness

MintHarness turns bounded financial questions and open-ended research objectives into a common stateful execution process, normalizing tasks, exposing task-specific actions, recording interactions, and maintaining persistent state outside the model context.

4.1 Execution Loop

  • The harness maps each task to a visible input and admissible actions, then validates proposed actions against tool schemas and remaining budgets before execution.Atomic tasks permit submission only; long-horizon tasks also expose configured tools, while the policy retains responsibility for research decisions.Source: §4.1 · p. 9
  • At termination, MintHarness returns the prediction, score, and auditable execution record, while the evaluator receives the extracted prediction and isolated reference answer.The record collects the evidence ledger, trajectory, usage, and status.Source: §4.1 · p. 9

4.2 Module Designs

  • Tool results are distilled into typed, provenance-linked financial records, while the ledger distinguishes failed retrieval paths from supporting evidence.Failed calls add no evidence, but their error observations remain available to the next turn; records retain acquisition time, tool status, and artifact references.Source: §4.1 · p. 9, §4.2 · p. 10
  • Persistent working memory carries the investigation across turns, retaining plans, confirmed conclusions, unresolved questions, and intermediate calculations linked to ledger records.Recent turns remain verbatim, older interactions are compressed into working memory, and long documents or tables remain accessible through artifact references under the token budget.Source: §4.2 · p. 10

5 Training Pipeline

Mint-Agent trains 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 atomic solutions and long-horizon research in auditable evidence.

5.1 Financial Reasoning

  • Financial reasoning uses SFT to establish valid derivations, then RLVR rewards answers and derivations that pass capability-aware, evidence-grounding, and financial-consistency checks.The verifier permits alternative derivations when they are grounded, internally consistent, and reach the verified answer; GSPO trains on complete solutions with informative rewards.Source: §5.1 · p. 11
  • For financial-reasoning SFT, teacher traces must be correct, replayable against evidence, free of specified patterns, and selected for clarity before fine-tuning.The selected traces and audited data-engine traces supervise one SFT round from the common base policy.Source: §5.1 · p. 11

5.2 Financial Agentic Execution

  • Agentic execution uses SFT for stable interactions, critical-step OPD to repair consequential decisions in failed trajectories, and RLVR to optimize complete auditable research trajectories.Critical turns are selected when changing an action can restore a valid route to the answer graph without changing prior interaction; the student resamples from that state and a frozen teacher scores its continuation.Source: §5.2 · pp. 12–13
  • Agentic SFT retains answer-correct trajectories whose evidence replays against the source pack and answer graph, while excluding unproductive or invalid actions from the loss.Environment observations remain conditioning context, not prediction targets.Source: §5.2 · p. 12

5.3 Expert Integration

  • TIES merging initializes a student, and task-origin routing directs atomic tasks to the reasoning expert and long-horizon tasks to the execution expert for on-policy distillation.The student generates the states used for reverse-KL distillation, which is intended to repair behavioral interference left by merging while retaining both specialists’ capabilities.Source: §5.3 · pp. 13–14

6 Experiments

Mint-Agent performs strongly on bounded financial reasoning and long-horizon agentic benchmarks, while its cost, case studies, and model analyses illustrate how auditable evidence and specialist integration support execution. The experiments also identify evidence extraction and answer completion as persistent failure points.

6.1 Experimental Setup

  • The experiments evaluate atomic reasoning and agentic execution on complementary financial benchmarks, using public FinanceAgentBench tasks and benchmark-specific evaluation protocols.FinanceAgentBench v1.1 and v2 use 50 and 27 public tasks, respectively, because no held-out validation split was publicly available; FinSearchComp uses its released test data. Models are evaluated in a single turn on atomic benchmarks, with temperature 0, repetition penalty 1.1, a 128K-token context, and official answer matching; FinanceAgentBench v2 reports the mean and variance across three runs.Source: §6.1 · pp. 14–15

6.2 Overall Performance

  • Mint-Ag leads BizFinBench at 55.71, FinanceBench at 91.33, and RFC-Bench at 98.33, beating the strongest external results by 4.14, 1.33, and 3.00 points.Its joint strength spans business analysis, disclosure question answering, and counterfactual-misinformation classification.Source: §6.2 · p. 15, Table 2 · p. 15
  • Mint-Ag reaches 76.00 on FinanceAgentBench v1.1, 60.49 on v2, 89.04 on FinSearchComp T2, and the top 54.07 on T3; Mint-Cu scores 69.86 on T2.Against the strongest external scores, Mint-Ag leads by 4.00, 3.70, and 7.30 points on v1.1, v2, and T2, respectively.Source: Table 2 · p. 15, §6.2 · p. 16

6.3 Cost Analysis

  • Both Mint checkpoints lie on the empirical Pareto frontier in both FinanceAgentBench versions, with Mint-Cu anchoring lower-cost operation and Mint-Ag higher accuracy.On v1.1, Mint-Cu reaches 68.0% at $0.016 per task and Mint-Ag reaches 76.0% at $0.090; on v2, Mint-Cu costs $0.069 and Mint-Ag scores 60.49% at $0.213. For locally hosted checkpoints, API-equivalent costs use the respective base-model prices applied to measured token usage.Source: §6.3 · p. 16, Figure 6 · p. 16

6.4 Case Study

  • A 37-step TSMC forecast retained valid guidance and revenue records through source failures and scope corrections, then estimated quarterly revenue at NT$825,110 million.After restoring the 2022–2024 window, the agent recomputed the March growth rate; the final NT$8,010 million miss is reproducible from the ledger’s locked guidance midpoint and exchange-rate conversion. In the BKR case, the ledger reconciled the $171.65 unaffected close, a 22.3% premium, and $325 million synergy ratios against revenue and SG&A, retaining primary sources for reported quantities.Source: §6.4 · pp. 17–18, Figure 7 · p. 17, Figure 8 · p. 18

6.5 Model Analysis

  • Mint-Cu’s largest v1.1 failure mode is answer omission (M10, 18%); on v2, accuracy falls to 41.98% and evidence extraction (M5, 18.5%) dominates.On v1.1, misunderstanding, retrieval, evidence extraction, and calculation each account for 2%; on v2, misunderstanding, financial interpretation, and answer omission each account for 7.9%. The analysis associates v1.1 errors with executive closure and v2 errors with selecting and encoding relevant evidence at the correct scope.Source: §6.5 · pp. 19–20, Figure 9 · p. 19
  • MOPD improves the merged checkpoint by 1.33, 3.86, 4.00, and 3.19 points across four benchmarks, while TIES transfers unevenly.The results align with Mint-Cu’s Table 2 performance across both capability groups. The proposed mechanism is that MOPD consolidates reasoning and execution specialists into one policy, retaining their division of labor while reducing skill interference and coordination costs; the authors identify selective attention over evidence as a remaining improvement target.Source: Table 3 · p. 19, §6.5 · p. 20

7 Conclusion

Mint-Agent connects financial data, agent execution, and policy learning around a shared notion of correctness and auditable evidence. Its conclusion emphasizes both benchmark performance and the ability to trace and test how conclusions were reached.

7 Conclusion

  • Mint-Agent combines provenance-linked task construction, MintHarness evidence-ledger and context management, and training with SFT, critical-step OPD, and RLVR before integrating experts with TIES and multi-teacher OPD.These components connect what the model learns, how it acts, and how conclusions are verified.Source: §7 · p. 20
  • 98.33 on RFC-Bench, 76.00 on FinanceAgentBench v1.1, 60.49 on FinanceAgentBench v2, and 89.04 on FinSearchComp T2 are reported for Mint-Ag across seven benchmarks.Mint-Cu is presented as a compact 9B model with strong financial research performance at favorable inference cost, while Mint-Ag targets more demanding reasoning and execution; trajectory studies show evidence, calculations, and intermediate decisions preserved across multi-step investigations.Source: §7 · p. 20
  • A trustworthy financial agent must recover evidence, respect the question’s temporal boundary, expose fact-to-claim transformations, and revise conclusions when a link fails.The conclusion treats auditability as a shared substrate for data construction, agent execution, and policy learning—not as a reporting layer added after inference.Source: §7 · p. 21
  • Financial intelligence is not just producing a plausible answer; it is useful at scale when conclusions can be traced and tested.The paper frames Mint-Agent as a step toward financial agents whose competence includes whether each conclusion can be traced and tested.Source: §7 · p. 21

A Benchmark Verification Protocols

The appendix describes a benchmark verification protocol that uses each dataset’s released scoring materials and evaluation prompts to convert model outputs into scores.

A Benchmark Verification Protocols

A.1 BizFinBench · A.2 FinanceBench · A.3 FinSearchComp T2 and T3

The three benchmarks use their released or official evaluation procedures on extracted final answers, with task-specific rules governing answer normalization, matching, and scoring. FinSearchComp uses a shared binary-judge template for both tracks, while BizFinBench and FinanceBench use their official evaluators.

A.1 BizFinBench

  • BizFinBench uses the official v2 scripts, selecting a released comparison function according to each sample’s task type.The task types include anomaly tracing, event-logic reasoning, financial data description, multi-turn perception, quantitative computation, report analysis, stock-price prediction, and sentiment analysis.Source: §A.1 · p. 27
  • BizFinBench extracts the model-visible final answer, applies only the formatting required by the official script, and counts returned scores of at least 1.0 as correct.The normalized prediction and original sample metadata go to a deterministic evaluator that uses no LLM judge prompt.Source: §A.1 · p. 27

A.2 FinanceBench

  • FinanceBench evaluates extracted final answers against released reference answers using its official answer-matching procedure.The evaluator receives no additional task-solving context or external evidence, and the reported measure is question-level accuracy.Source: §A.2 · p. 27

A.3 FinSearchComp T2 and T3

  • FinSearchComp T2 and T3 use the official judge system and prompt templates stored with each example, with the same template for both tracks.The official prompt’s four demonstrations are retained during evaluation.Source: §A.3 · p. 27
  • FinSearchComp’s judge compares the extracted student answer with the reference answer and scoring points, awarding credit only when every required element is correct.The judge returns a binary answer_score; empty or error-valued answers receive zero.Source: §A.3 · p. 27
  • FinSearchComp accepts numerically equivalent formats and omitted units when the question specifies the unit.These answer-matching rules apply to both T2 and T3 under the shared official judge template.Source: §A.3 · p. 27

A.4 FinanceAgentBench v1.1 and v2

FinanceAgentBench v1.1 and v2 use the same rubric-based evaluation protocol: a sample is correct only if every criterion is met and no confident factual contradiction is detected. The judge evaluates the supplied question, expert answer, prediction, and ordered criteria without outside knowledge.

A.4 FinanceAgentBench v1.1 and v2

  • A sample is correct only when every correctness criterion is satisfied and no contradiction with the expert reference is detected.When no correctness criterion is provided, evaluation falls back to exact, substring, and numeric-overlap answer matching; the protocol applies to both v1.1 and v2.Source: §A.4 · p. 28
  • The judge assesses each criterion separately and independently checks whether the answer contradicts a specific fact in the expert reference.It must not solve the question or use outside knowledge, and returns ordered criterion decisions, a contradiction decision, and a short rationale as valid JSON.Source: §A.4 · p. 28
  • The judge receives the question, expert reference answer, agent’s final answer, and rubric criteria as its evaluation inputs.Source: §A.4 · p. 28

A.5 RFC-Bench Task 2

RFC-Bench Task 2 evaluates whether a model can classify the principal manipulation in a source-grounded counterfactual financial paragraph. Its reported result is Task-2 classification accuracy, not an aggregate over the full suite.

A.5 RFC-Bench Task 2

  • The evaluator extracts one of four labels from the final answer, normalizes limited aliases, and uses exact matching; ambiguous or invalid outputs count as incorrect.The reported result is Task-2 classification accuracy rather than an aggregate over the full RFC-Bench suite, and the benchmark uses no LLM judge.Source: §A.5 · p. 29
  • Task 2 asks models to identify the main manipulation in a financial paragraph using only the provided source page and manipulated paragraph, without outside evidence.The four labels are numerical, flipping, sentiment, and causal.Source: §A.5 · p. 29

B Additional Case Study

The AMD guidance case shows retrieval recovery and reuse of collected evidence in a four-quarter financial calculation. After an initial document lookup failed, the agent broadened its SEC search and used a ledger to reopen a source without repeating the search.

AMD Guidance

  • 10.53%, 8.96%, 8.00%, and 8.45% of midpoint are the reported quarterly AMD guidance-range widths, calculated from four primary filings.The 13-step trajectory broadens the SEC search after an initial accession guess fails, extracts EX-99.1 exhibits, and reopens the Q3 source through the ledger without repeating the full search.Source: B Additional Case Study · p. 29, Figure 10 · p. 30

Related papers

Loading 2608.16386v1…