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, Qingsong Wen, Yilei Shao
cs.CL, cs.LG
2026-08-17
Mint-Agent trains finance-native agents on provenance-linked data and an evidence ledger. 27B Mint-Ag: 98.33 RFC-Bench, 76% FAB v1.1, about 22% of GPT-5.6-Sol cost.
A financial agent has to do more than recall jargon. It must compute over grounded evidence and keep a long research trail whose every conclusion can be replayed. Generic agents snap that chain: a source is relevant but not authoritative, the fiscal period is wrong, units get mixed. Without a recoverable path from source to calculation to claim, the answer cannot be trusted or repaired. The Mint team and Shanghai AI-Finance School at ECNU turn that contract into data, harness, and training.
The data engine has two streams. Atomic tasks cover knowledge, extraction, calculation, analysis, and verification, drawn from EDGAR, XBRL, exchanges, and FRED. Every fact keeps a locator; a candidate is kept only if the derivation replays and the answer is unique. Long-horizon tasks expose only the research objective and hide the evidence graph, so the agent must search, inspect, calculate, and synthesize; the graph is fixed at generation time for later audit. MintHarness stores an evidence ledger and working memory outside the model context, distills tool results into provenance-linked records, and scores submissions with an isolated evaluator. Training splits: a reasoning specialist gets SFT then RLVR with GSPO on atomic tasks; an execution specialist gets SFT, critical-step on-policy distillation, then trajectory-level RLVR. TIES merges the two, then multi-teacher OPD routes each task origin to its teacher. Bases are Qwen3.5-9B (Mint-Cu) and Qwen3.6-27B (Mint-Ag).
Mint-Ag leads all seven professional benches. RFC-Bench is 98.33%, above Claude Opus 4.8 at 95.33 and GPT-5.6-Sol at 94.67. FinanceAgentBench v1.1 is 76.00% and v2 60.49%, the latter 3.70 points above GPT-5.6-Sol. FinSearchComp T2 is 89.04%, above Cursor with Grok 4.5 at 81.74. The 9B Mint-Cu scores 69.86% on T2, above Agents-A1-35B at 47.03. Cost: Mint-Ag averages $0.213 on v2 versus $0.959 for GPT-5.6-Sol, about 22% of that bill. The public FinanceAgentBench sets are small: 50 v1.1 tasks and 27 v2 tasks. Failure review finds Cu's main leftover error on v1.1 is answer omission (18%); on v2 it shifts to evidence extraction (18.5%). TIES alone is uneven across four scores; multi-teacher OPD restores the 9B checkpoint to the reported Cu numbers.
| Model | RFC-Bench | FAB v1.1 | FinSearchComp T2 |
| GPT-5.6-Sol | 94.67 | 66.00 | 41.10 |
| Claude Opus 4.8 | 95.33 | 66.00 | 45.66 |
| Mint-Cu 9B | 96.67 | 68.00 | 69.86 |
| Mint-Ag 27B | 98.33 | 76.00 | 89.04 |
Vertical finance agents are often domain-tuned weights plus a generic tool scaffold. This stack treats provenance, replayable calculation, and an auditable trail as one correctness contract across task construction, execution, and RL rewards. A 27B model beating larger general systems on reliability and long-horizon execution, with a 9B already usable on search-style tasks, is a concrete recipe for teams building internal research agents. The eval surface is still thin: the public agent sets are tens of items.
FinanceAgentBench uses the 50+27 public tasks after the authors could not obtain an extra split, so overfitting risk is on the page. Atomic reasoning is single-turn; only long-horizon work uses the harness. API-equivalent cost applies Qwen base prices to local tokens, not a Mint product tariff. RFC evaluates Task 2 only. No independent third-party rerun. The failure shift from omission to evidence extraction says harder research items still stall when retrieved text must become correctly scoped financial evidence.