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

arXiv:2607.19794 · cs.CL, cs.CE, cs.DB, cs.LG · Submitted 2026-07-22 · Read on arXiv

Listen

Transcript

Introduction to the show: ident: AI Radio. Generated commentary on the latest Artificial Intelligence papers.

Tom: Hi, I'm Tom.

Jane: And I'm Jane, here with Lu, Meng and our in-house large language model, Lalam.

Tom: We're looking at the paper: "TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis".

Jane: Here's the gist of the paper. The gist The TriAgent system presents a divergence-aware multi-agent committee that uses contextual granularity to route queries, leading to superior cost-efficiency and performance in financial sentiment analysis.

Tom: How it works The TriAgent framework is a three-tier committee stratified by three contextual granularities: 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 cross-family checks) The three tiers correspond to three contextual granularities: a word level lexicon (VADER), a sentence level domain transformer (FinBERT), and a cross sentence reasoner VADER fails on phrase-bound polarity reversals, FinBERT fails on multi clause numeric comparisons, and the LLM overconfidently invents reasoning where lexicon neutrality is correct.

Jane: Semantic Divergence Index (SDI) and Routing A three way Semantic Divergence Index (SDI) measures pairwise disagreement across granularities and routes each query accordingly The SDI is defined by three pairwise Semantic Divergence Indices: SDILE = sV − sF, SDILR = sV − sL, and SDIER = sF − sL.

Tom: Thresholding the pair (SDILE, SDIER) at (τLE, θER) = (0.3, 0.7) partitions samples into four behaviour quadrants: consensus skips L3, domain shift trusts F and routes to L2.

Title and authors: Tom: So we're looking at this paper today, "TriAgent: Divergence-Aware Multi-Agent Committees for Cost-Efficient Financial Sentiment Analysis." It sounds like they've tackled that problem where you have a lot of queries that are easy to sort but you still end up using expensive cloud reasoners for everything.

Jane: That’s the core idea, right? They’re looking at how to make sure you only spend money on the heavy reasoning when it actually needs it, not just because it's there.

Lu: The authors introduce a system with three tiers of agents based on different levels of context. They start with a word-level lexicon, then move to a sentence-level domain transformer, and then they add a cross-sentence reasoner for the final check.

Meng: Three tiers sounds complicated if you’re thinking about deployment, but the goal is clearly cost efficiency because most queries are just trivially classifiable.

Lalam: It sounds like they're building a committee that uses these different levels of analysis to decide what to do next.

The paper's summary: Tom: Okay, so the paper explains how this system works by using a three-way Semantic Divergence Index, or SDI, to route each query through these tiers.

Jane: They define this index by measuring how much disagreement there is between the different levels of analysis—word level versus sentence level versus cross-sentence reasoning.

Lu: They measure this with three specific indices: one comparing word and sentence level, one comparing word and reasoner, and another comparing sentence and reasoner.

Meng: So if the disagreement is too high in a certain way, that’s when the system decides which agent to trust or route to next.

Lalam: That routing decision determines whether it just skips a layer or routes to a different kind of agent interaction protocol, like voting or debating.

The paper's improvements: Tom: One of the big things they found is something called the critic plateau, which suggests that when you re-task the LLM as a critic over those smaller agents' outputs, its performance stabilizes around an F1 score of about zero point eight seven across Qwen models between one point five billion and seven billion parameters.

Jane: That’s interesting because it shows that the improvement isn't just about having a bigger model; it’s about how you structure the task for that model to evaluate the smaller ones.

Lu: They found that this plateau is driven by granularity, not just by having three different agents voting together, which they noted when comparing it to a same-size three-persona vote that only got an F1 of zero point six six.

Meng: If we’re looking at deployment, the idea of having a three-granularity edge predictor using XGBoost with an AUC of zero point eight five allows for fully on-device routing, which is something practical for cost control at scale <ref:2607.19794#pg1>.

Lalam: They also have this shared consensus dictionary that acts like a cross-lingual canoniiser, meaning Chinese queries can match English labels at F1 equal to zero point nine nine without running the committee again <ref:2607.19794#pg1>.

Conclusion: Tom: So to wrap up, the main thing is that this SDI single-stage strategy achieved the best risk adjusted return of three point five zero on a twenty ticker backtest, outperforming always-FinBERT at one point three six and alwaysLLM at zero point one one <ref:2607.19794#pg2>.

Jane: That's a significant performance gain when you look at the financial outcomes, showing that this layered approach really pays off compared to just picking one model type permanently.

Lu: The paper shows that this architecture hits all three FinLLM two thousand twenty-six CFP topic buckets simultaneously, which is a pretty broad implication for how we structure these committees <ref:2607.19794#pg2,hits all three FinLLM 2026 CFP topic buckets simultaneously>.

Meng: From a deployment standpoint, they saved about nine point three million dollars per year at ten million users compared to the GPT-4o-mini baseline, which really puts the cost savings into perspective.

Lalam: And that SDI signal also doubles as a post-hoc LLM hallucination detector with an AUC of zero point nine zero on FPB data, so it has this security benefit too <ref:2607.19794#pg1>.

Isabel Xu, *, Cynthia Xu, *, Rachel Ren2, Cong Guo3, Jiacheng Ding3,†

The Overlake School · Edwards Vacuum Inc. · The University of Memphis

cs.CL, cs.CE, cs.DB, cs.LG

Submitted: 2026-07-22

Updated: 2026-07-22

Journal ref: Proceedings of the 35th ACM International Conference on Information and Knowledge Management (CIKM '26), 2026

Code: https://github.com/graphuofm/TRIAGENT

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Importance score: 89/100

The gist: The gist The TriAgent system presents a divergence-aware multi-agent committee that uses contextual granularity to route queries, leading to superior cost-efficiency and performance in financial

Key concepts

TriAgent Framework: A three-tier committee system using different contextual granularities: a word-level lexicon (VADER), a sentence-level domain transformer (FinBERT), and a cross-sentence reasoner (Qwen2.5, Mistral, Phi-3.5). It routes queries based on how these agents disagree.
Semantic Divergence Index (SDI): A metric that measures pairwise disagreement between the outputs of the three agents at different granularities. By calculating indices like SDILE and SDIER, it determines which agent's prediction is most reliable for a specific query, guiding routing decisions.
Critic Plateau Mechanism: A strategy where an LLM is re-tasked as a critic to evaluate the outputs of smaller agents. This mechanism shows that when the LLM critiques these smaller outputs, performance stabilizes at high accuracy (F1≈0.87), proving granularity stratification is key, not just the number of agents.

Terminology

AUC — 1× in this paper · explained in 2 episodes
Hallucination — 1× in this paper · explained in 4 episodes
Semantic Divergence Index — 1× in this paper
Transformer — 1× in this paper · explained in 2 episodes

Summary

The gist The TriAgent system presents a divergence-aware multi-agent committee that uses contextual granularity to route queries, leading to superior cost-efficiency and performance in financial sentiment analysis.

How it works

The TriAgent framework is a three-tier committee stratified by three contextual granularities: 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 cross-family checks) The three tiers correspond to three contextual granularities: a word level lexicon (VADER), a sentence level domain transformer (FinBERT), and a cross sentence reasoner VADER fails on phrase-bound polarity reversals, FinBERT fails on multi clause numeric comparisons, and the LLM overconfidently invents reasoning where lexicon neutrality is correct.

Semantic Divergence Index (SDI) and Routing

A three way Semantic Divergence Index (SDI) measures pairwise disagreement across granularities and routes each query accordingly The SDI is defined by three pairwise Semantic Divergence Indices: SDILE = sV − sF, SDILR = sV − sL, and SDIER = sF − sL. Thresholding the pair (SDILE, SDIER) at (τLE, θER) = (0.3, 0.7) partitions samples into four behaviour quadrants: consensus skips L3, domain shift trusts F and routes to L2. The routing decision is formally defined by yˆ(x) = I(x; V, F, L) SDIER(x) > θER F(x) SDILE(x) > θLE V (x) otherwise.

Interaction Protocols and the Critic Plateau

The interaction protocol I ∈ 3 is determined by the routing quadrant, where Vote returns the confidence weighted majority over V/F/L, critic feeds the LLM the original sentence plus V’s and F’s predictions and asks for a final label, and debate performs a round-2 LLM call that sees all three round-1 outputs including its own rationale. A central finding is the critic plateau: when the LLM is re tasked as a critic over the smaller agents’ outputs, F1 plateaus at ≈0.87 across 1.5 B–7 B Qwen (bootstrap 95% CIs overlap), while a same size three persona vote drops to F1=0.66 which is given by granularity stratified diversity. This plateau is granularity driven, not multi agent vote driven.

Shared Consensus Dictionary (SCD) and Cross-Lingual Canonicalization

The Shared Consensus Dictionary (SCD) is an optional fourth component that caches committee decisions in multilingual sentence BERT embedding space. A query x is embedded to ϕ(x), k-NN searched against cached entries by cosine similarity σi = ⟨ϕ(x), ϕ(˜xi)⟩/(∥ϕ(x)∥∥ϕ(˜xi)∥), and the cached label is returned when σ∗ (x) = max i σi ≥ τ. The SCD doubles as a cross lingual canoniiser, allowing 95% of Chinese queries to match an English cached label at F1=0.99 without re running the committee.

Performance and Economic Outcomes

The SDI single stage strategy attains the best risk adjusted return (Sharpe=3.50) on a 20 ticker backtest, dominating both always-FinBERT (1.36) and always LLM (0.11). At 10 M user scale, TriAgent saves 9.3 M/year vs GPT-4o-mini baseline. The SDI signal also doubles as a post hoc LLM hallucination detector at AUC=0.90. Furthermore, the three granularity edge predictor (XGBoost AUC=0.85) enables fully on device routing.

Security and Deployment Benefits

The architectural choice that makes our committee token efficient also exposes a free trust signal and a route to cross border fairness. SDI as a post hoc LLM hallucination detector yields AUC=0.90 on FPB. Cross border fairness via the SCD shows 95% of Chinese queries receive the same canonical label as the English equivalent, eliminating a common source of cross language drift in multinational deployments.

Conclusion

TriAgent is a three tier financial sentiment committee whose key insight is architectural: when each agent is small and stratified by contextual granularity, namely word (lexicon), sentence (specialist), cross sentence (reasoner), the same Semantic Divergence Index that gates routing for cost also gates interaction for capability and detects anomaly for security. The SDI single stage strategy attains the best risk adjusted return (Sharpe=3.50) of all evaluated strategies, including always-FinBERT (1.36) and alwaysLLM (0.11). At 10 M user scale, TriAgent saves 9.3 M/year vs a GPT-4o-mini baseline. All source code, mined trigger lexicons, the SCD, and all committee predictions are released to support both academic replication and direct industrial deployment.

Improvements for AI systems

Adaptive Routing based on Semantic Divergence Index (SDI) This system can dynamically route queries to a three-tier committee (word-level lexicon, sentence-level domain transformer, cross-sentence reasoner) by computing pairwise divergence scores and choosing an interaction protocol such as vote, critic, or debate.
Cost Reduction via Token-Economic Pareto Frontier This system can select the optimal model tier based on cost constraints by tracing the SDI routing strategy to a Pareto frontier that maps tokens per query against F1 score, achieving significant savings compared to always-using expensive models like GPT-4.
Cross-Border Canonicalization through Shared Consensus Dictionary (SCD) This system can act as a cross-lingual canonicalizer by using sentence embeddings to match Chinese queries with cached English labels at high accuracy (F1 = 0.99) without incurring any additional inference cost on a cache hit.
LLM Hallucination Detection via SDI Signal This system can serve as a post-hoc LLM hallucination detector by using the SDI signal; specifically, when SDIER exceeds a threshold, it flags potential issues with an AUC of 0.90 on FPB data.
Enhanced Reasoning through Critic Plateau Mechanism The system can improve reasoning accuracy by re-tasking the LLM as a critic over smaller agents' outputs, where F1 plateaus at approximately 0.87 across Qwen models from 1.5B to 7B parameters, even when a same-size three-persona vote drops to F1=0.66.
Cross-Lingual Generalization via SCD for New Agents The Shared Consensus Dictionary can mitigate family-specific plateau gaps by allowing a new agent joining the committee to inherit the cache, ensuring that cross-lingual deployment benefits from established canonical answers without retraining.