AI Paper Atlas

Research map / Multi-agent systems and games / LLM agents

Towards Professional-Grade Financial Agents: Benchmarking, Tooling, and Structured Reasoning

Cheng Huang, Jinghua Piao, Ranran Wang, Yong Li

ICML 2026 · Multi-agent systems and games › LLM agents · approach: Agent data synthesis and behavior studies

Read the paperCheck your idea against 137,217 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Introduces ProFinR, a benchmark of 528 expert-designed professional finance tasks across four domains, along with tooling and structured reasoning for LLM financial agents.

Problem

LLM agents lack holistic benchmarks for realistic professional financial workflows, and unstructured reasoning in them is fragile.

Approach

A holistic professional-finance benchmark (ProFinR) with expert-designed tasks, paired with tooling and structured reasoning to make financial agents more robust than unstructured reasoning.

Measured effects reported in the paper

What was changedWhereMeasuredEffect
wrap LLM with ProFinAgent (tools, DAG planner, case memory)GPT-5.2 · ProFinRTotal score delta vs native: +49.81%Improved
wrap LLM with ProFinAgentQwen3-4B-Instruct-2507 · ProFinRTotal score delta vs native: +27.84%Improved
add DAG parallel execution (ablation)ProFinAgent on DeepSeek-R1 · ProFinRtotal execution time: 42,579.9 s -> 22,547.2 s (47.1% reduction)Improved
add 53-tool library to ReAct without structure (ablation)DeepSeek-R1 ReAct · ProFinR L3accuracy: 29.45% -> 21.70%Hurt
add tool use to ReAct (ablation)Qwen3-32B · ProFinR L1accuracy: 5.28% -> 40.38%Improved
add DAG planner (ablation)DeepSeek-R1 ReAct · ProFinR L3accuracy: 29.45% -> 33.33%Improved
add case-based memory (ablation)DeepSeek-R1 · ProFinR L1accuracy: 43.75% -> 53.36%Improved
combine DAG and CBM versus either alone (ablation)DeepSeek-R1 · ProFinR L3accuracy: 33.33% / 31.78% -> 54.26%Improved
domain-specific financial fine-tuning (Fin-R1)native inference · ProFinRtotal score: 1.70% (Fin-R1) vs 1.70% (Qwen3-4B)No clear effect
use ProFinAgent instead of tree-search tool plannersGPT-4o · RestBench-TMDBavg: 74.50% -> 81.25%Improved
ProFinAgent vs ToolTree win rateGPT-4o-mini · RestBench-Spotifywin rate: 72.18% -> 71.93%Mixed

Earlier methods this paper compares with

Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.

Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.

Experimental setup

Datasets: ProFinR

Benchmarks: ProFinR

Closest papers by meaning

Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy