TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis

2026-07-23 · Source: cs.CL updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, FinTech & Digital Financial Services · Depth: Expert, extended

Summary

TriAgent is a multi-agent committee designed for cost-efficient financial sentiment analysis, addressing the high inference costs of large language models. It employs a three-tier architecture: a word-level lexicon (VADER), a sentence-level domain transformer (FinBERT), and a cross-sentence reasoner (Qwen2.5, 0.5 B–14 B-4bit, with Mistral-7B and Phi-3.5-mini checks). A Semantic Divergence Index (SDI) measures disagreement between these agents, routing queries to the appropriate tier. Key findings include a critic plateau where F1 ≈0.87 is achieved across 1.5 B–7 B Qwen models, outperforming same-size persona voting. TriAgent also features a Shared Consensus Dictionary for cross-lingual canonicalization, achieving F1=0.99 for 95% of Chinese queries from an English cache. The SDI doubles as an LLM-hallucination detector (AUC=0.90) and yields a Sharpe ratio of 3.50 in back-tests. This system can save \$9.3 M/year at a 10 M-user scale compared to a GPT-4o-mini baseline, with code released for deployment.

Key takeaway

For MLOps Engineers or AI Scientists deploying financial sentiment models, you can significantly reduce inference costs and improve accuracy by adopting a multi-agent, granularity-stratified architecture like TriAgent. Prioritize implementing a Semantic Divergence Index for intelligent routing and consider a Shared Consensus Dictionary for cross-lingual consistency, potentially saving millions annually. Validate the critic plateau for your chosen LLM family to optimize L3 agent size.

Key insights

Granularity-stratified multi-agent committees with divergence-aware routing enable cost-efficient, accurate financial sentiment analysis and hallucination detection.

Principles

• Agent diversity based on contextual granularity is crucial for committee performance.
• LLMs can effectively act as critics to resolve disagreements from smaller, specialized agents.
• Inter-agent disagreement signals LLM hallucination and query difficulty.

Method

TriAgent routes queries through a Shared Consensus Dictionary cache, then a three-tier committee (VADER, FinBERT, Qwen2.5), using a Semantic Divergence Index to trigger critic or debate protocols based on disagreement.

In practice

• Deploy a multi-tier sentiment system with word, sentence, and cross-sentence agents.
• Utilize a Semantic Divergence Index for intelligent query routing and hallucination detection.
• Implement a multilingual sentence embedding cache for cross-lingual consistency.

Topics

• Financial Sentiment Analysis
• Multi-Agent Systems
• Large Language Models
• Cost Optimization
• Semantic Divergence Index
• Cross-Lingual NLP
• LLM Hallucination Detection

Code references

• graphuofm/TRIAGENT

Best for: CTO, AI Architect, AI Engineer, AI Scientist, Machine Learning Engineer, MLOps Engineer

Related on AIssential

• TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis — Divergence-aware multi-agent committees offer cost-efficient and robust financial sentiment analysis by stratifying contextual granularity.
• MASTE: A Multi-Agent Pipeline for Zero-Shot Aspect Sentiment Triplet Extraction — Decomposing complex NLP tasks into multi-agent pipelines significantly enhances zero-shot LLM performance.
• A Training-Free Mixture-of-Agents Framework for Multi-Document Summarization using LLMs and Knowledge Graphs — A training-free agent framework combines LLMs and knowledge graphs for effective multi-document summarization.
• Automated Detection and Classification of Delusion-related Content in Naturalistic Audio Diaries Using Multi-Agent Language Models — The multi-agent LLM pipeline effectively detects delusion-related content in naturalistic speech transcripts using majority voting.
• OpenFinGym: A Verifiable Multi-Task Gym Environment for Evaluating Quant Agents — Fragmented evaluation of quant agents overstates competence; unified, multi-task environments are crucial for realistic assessment.
• MASCA: LLM based-Multi Agents System for Credit Assessment — MASCA employs a hierarchical LLM-based multi-agent system with contrastive learning to enhance credit assessment accuracy and address biases.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CL updates on arXiv.org.