Multi-agent llm diversity impacts emotion detection more than topology
Disentangling Topology and Diversity in Multi-Agent LLMs for Multilingual Low-Resource Emotion Detection
Computation and LanguageArtificial Intelligence
Summary
Detecting emotions in multiple languages with limited data is hard. This paper studies how using several large language model agents together can help, focusing on how these agents differ and how they connect. The authors find that making agents diverse through specialized training matters more than how they are linked. They test this on two popular models and nine languages and show their approach performs better than several existing methods.
What this means in practice
- •For multilingual chatbot developers: Improve emotion recognition in low-resource languages by using specialized multi-agent model calls to boost understanding of user feelings.
- •For social media monitoring teams: Enhance detection of emotional content across many languages by applying diverse multi-agent large language model setups tailored to scarce data.
Authors
Ulugbek Shernazarov, Charitha Ruwansiri Weerakon Basnayake, Abdelkhaleq El Jarjini, Noel Crespi, Praboda Rajapaksha
Abstract
Multi-agent LLM systems combine multiple inference calls, but prior work often confounds how calls are connected with how they are diversified. We study these factors independently: inference topology and source of inter-agent diversity. In a controlled $2 \times 3$ matrix, we cross parallel aggregation and sequential refinement with stochastic sampling, role prompting, and learned QLoRA specialization, under a fixed three-call budget and output protocol within each backbone. Using Qwen2.5-14B-Instruct and Llama-3.1-8B-Instruct, we evaluate all six configurations on multilingual low-resource emotion detection across nine languages. Parallel learned specialization is strongest on Qwen at 52.83 Macro-F1 and reaches 52.94 on Llama. On Qwen it also exceeds same-backbone zero-shot, few-shot, CoT, and seven-call self-consistency baselines. The preferred topology depends on diversity source: sequential refinement helps stochastic and prompted settings, while the learned Width advantage shrinks from 2.83 points on Qwen to 0.17 on Llama. Depth-wise analysis suggests that later learned specialists can overwrite correct early predictions, although the aggregate effect is backbone-dependent. Overall, how agents are differentiated produces larger performance shifts than topology, which should be evaluated jointly with specialization.