TriAgent: A Committee of Small Models Instead of an Expensive LLM Saves $9.3M a Year

In July 2026, a group of researchers published a paper on arXiv about TriAgent, a multi-agent system for financial sentiment analysis that saves $9.3M a year at 10M users compared with a baseline built on GPT-4o-mini.

How the TriAgent committee works

TriAgent distributes queries across three levels of "vision" on the text, each more expensive than the last. The cheap tiers handle most of the simple cases, and the expensive LLM is called in only for disputed ones. The idea targets a structural cost trap: most queries are trivial to classify, yet a cloud reasoner processes all of them, so the bill grows linearly with the number of users.

• The VADER lexicon: word-level analysis
• The FinBERT transformer: sentence-level analysis
• Qwen2.5 (from 0.5B to 14B, 4-bit) as a reasoner across multiple sentences, with cross-checks from Mistral-7B and Phi-3.5-mini
• The Semantic Divergence Index (SDI) measures how far the tiers disagree and decides who gets the query
• Savings: $9.3M a year at 10M users versus GPT-4o-mini

"Most queries are trivial to classify, yet expensive cloud reasoners

process all of them" — from the paper's abstract on arXiv.

What is the critic plateau?

The paper's main finding is the "critic plateau": when an LLM is reassigned the role of critic over the outputs of smaller agents, the F1 score levels off at about 0.87 for Qwen models from 1.5B to 7B (the confidence intervals overlap). For comparison, a vote among three "personas" of the same size yields only F1=0.66.

According to the authors, the gap comes from diversity created by different granularities of analysis: the three tiers "see" the text differently, and that is more useful than three identical models with different roles.

Three side results of SDI

From the same SDI signal, the authors derive three further outcomes. A shared consensus dictionary built on multilingual sentence-BERT answers 95% of Chinese queries from the English cache at F1=0.99, meaning it transfers labels across languages at almost no extra cost.

SDI also works as a post-hoc detector of LLM hallucinations with AUC=0.90. And in a backtest on 20 tickers, the SDI-based strategy delivered the best risk-adjusted return: a Sharpe ratio of 3.50, versus 1.36 for the "always FinBERT" strategy and 0.11 for "always LLM."

What this means

TriAgent offers a practical recipe: instead of sending every query through an expensive model, route by difficulty. For a production system with millions of users, that is the difference between a linearly growing bill and savings of millions of dollars, without a loss in quality. The authors have released the code, the dictionaries, and the consensus dictionary as open source.

Frequently asked questions

How much does TriAgent save?

By the authors' calculations, at 10M users the system saves $9.3M a year compared with a GPT-4o-mini baseline, because the expensive LLM is called in only for disputed queries.

What is the SDI?

The Semantic Divergence Index (SDI) is a metric of pairwise divergence between the three analysis tiers (word, sentence, multiple sentences). It routes queries and also serves as an LLM hallucination detector with an AUC of 0.90.

Which models does TriAgent use?

The VADER lexicon, the FinBERT transformer, and a Qwen2.5 reasoner (0.5B–14B, 4-bit) with cross-checks from Mistral-7B and Phi-3.5-mini.