跳转至

CLAIR-Fin：用于跨模态金融问答的声明级验证与自适应辩论对抗性多智能体框架

文章背景与核心概要

现有的检索增强生成（RAG）和多智能体系统在处理金融幻觉漏洞时往往力不从心：它们对冲突的数据模态一视同仁，评估的是聚合报告而非原子级声明（atomic claims），并且将验证推迟到了最终输出阶段。为了克服这些局限性，研究人员引入了 CLAIR-Fin，这是一个专为跨模态金融问答（QA）设计的强大的 九智能体框架。

通过将问题拆解为原子级声明并通过专用账本（ledger）进行追踪，CLAIR-Fin 在流水线的每个阶段都能确保严格的真实依据验证。在 BB-FinQA-X 跨模态金融数据集上的评估表明，该框架在事实忠实度方面显著优于标准的 RAG 基线以及 HyDE 和 Graph-RAG 等高级检索技术。

Existing retrieval-augmented generation (RAG) and multi-agent systems often struggle with financial hallucination vulnerabilities: they treat conflicting data modalities equally, evaluate aggregate reports rather than atomic claims, and defer verification until the final output stage.

To overcome these limitations, the researchers introduce CLAIR-Fin, a robust nine-agent framework designed specifically for cross-modal financial Question Answering (QA). By breaking questions down into atomic claims and tracking them via a specialized ledger, CLAIR-Fin ensures rigorous ground-truth verification at every phase of the pipeline. Evaluated on the BB-FinQA-X cross-modal financial dataset, the framework significantly outperforms standard RAG baselines and advanced retrieval techniques like HyDE and Graph-RAG in factual faithfulness.

CLAIR-Fin: An Adversarial Multi-Agent Framework for Claim-Level Verification and Adaptive Debate in Cross-Modal Financial QA

CLAIR-Fin：用于跨模态金融问答的声明级验证与自适应辩论对抗性多智能体框架

arXiv ID: arXiv:2608.13706 [cs.CL]
DOI: 10.48550/arXiv.2608.13706
Authors: Fatema Tuj Johora Faria, Mukaffi Bin Moin, Jubayer Al Mahmud, M. F. Mridha, Md. Alam Hossain
Submitted: August 13, 2026 (Last revised August 18, 2026)

Executive Summary

执行摘要

Existing retrieval-augmented generation (RAG) and multi-agent systems often struggle with financial hallucination vulnerabilities: they treat conflicting data modalities equally, evaluate aggregate reports rather than atomic claims, and defer verification until the final output stage.

To overcome these limitations, the researchers introduce CLAIR-Fin, a robust nine-agent framework designed specifically for cross-modal financial Question Answering (QA). By breaking questions down into atomic claims and tracking them via a specialized ledger, CLAIR-Fin ensures rigorous ground-truth verification at every phase of the pipeline. Evaluated on the BB-FinQA-X cross-modal financial dataset, the framework significantly outperforms standard RAG baselines and advanced retrieval techniques like HyDE and Graph-RAG in factual faithfulness.

Existing retrieval-augmented generation (RAG) and multi-agent systems often struggle with financial hallucination vulnerabilities: they treat conflicting data modalities equally, evaluate aggregate reports rather than atomic claims, and defer verification until the final output stage.

To overcome these limitations, the researchers introduce CLAIR-Fin, a robust nine-agent framework designed specifically for cross-modal financial Question Answering (QA). By breaking questions down into atomic claims and tracking them via a specialized ledger, CLAIR-Fin ensures rigorous ground-truth verification at every phase of the pipeline. Evaluated on the BB-FinQA-X cross-modal financial dataset, the framework significantly outperforms standard RAG baselines and advanced retrieval techniques like HyDE and Graph-RAG in factual faithfulness.

Key Core Mechanisms

核心机制

CLAIR-Fin introduces several architectural innovations to eliminate inter-agent errors and hallucinations:

  • Typed Financial Claim Ledger: Decomposes complex financial queries into atomic, manageable claims.
  • Asymmetric Evidence Authority: Conditions evidence trust based on the specific claim type rather than treating all data modalities as equally reliable.
  • Chain-of-Custody Verification: Actively checks grounding during hand-offs between drafting and adversarial review, rather than only testing at the final output stage.
  • Adaptive Rebuttal Cycle: Dynamically routes contested claims through targeted adversarial debates, scaling debate depth based on findings.
  • Terminal Entailment Audit & Hallucination Risk Index: Combines a final verification audit with a continuous risk index to differentiate rigorously scrutinized claims from untested ones.

CLAIR-Fin introduces several architectural innovations to eliminate inter-agent errors and hallucinations:

  • Typed Financial Claim Ledger: Decomposes complex financial queries into atomic, manageable claims.
  • Asymmetric Evidence Authority: Conditions evidence trust based on the specific claim type rather than treating all data modalities as equally reliable.
  • Chain-of-Custody Verification: Actively checks grounding during hand-offs between drafting and adversarial review, rather than only testing at the final output stage.
  • Adaptive Rebuttal Cycle: Dynamically routes contested claims through targeted adversarial debates, scaling debate depth based on findings.
  • Terminal Entailment Audit & Hallucination Risk Index: Combines a final verification audit with a continuous risk index to differentiate rigorously scrutinized claims from untested ones.

Evaluation & Results

评估与结果

The framework was tested on BB-FinQA-X, a 500-question cross-modal financial benchmark curated from Bangladesh Bank Annual Report materials (stratified by query type, format, and difficulty).

  • Faithfulness Improvement: Raised faithfulness scores from 0.780 (single-pass RAG baseline) up to 0.889.
  • Outperforms Advanced Strategies: Surpassed advanced retrieval-strategy baselines like HyDE and Graph-RAG (which scored \(\leq 0.874\) in faithfulness).
  • Safe Abstention: Intelligently abstains on 5.4% of queries when underlying evidence is deemed insufficient, preventing forced unsupported responses.

The framework was tested on BB-FinQA-X, a 500-question cross-modal financial benchmark curated from Bangladesh Bank Annual Report materials (stratified by query type, format, and difficulty).

  • Faithfulness Improvement: Raised faithfulness scores from 0.780 (single-pass RAG baseline) up to 0.889.
  • Outperforms Advanced Strategies: Surpassed advanced retrieval-strategy baselines like HyDE and Graph-RAG (which scored \(\leq 0.874\) in faithfulness).
  • Safe Abstention: Intelligently abstains on 5.4% of queries when underlying evidence is deemed insufficient, preventing forced unsupported responses.

资源链接与访问

License info: Creative Commons Attribution 4.0 International license icon

License info: Creative Commons Attribution 4.0 International license icon