RESEARCH · RESEARCH · #522
Paper proposes a confidence-gated hybrid (ensemble + LLM) for cost-effective emotion recognition in conversation
arXiv:2609.17977v1 compares a low-cost stacked ensemble, off-the-shelf LLM prompting (including GPT-4o-mini), and a confidence-gated hybrid that escalates low-confidence ensemble predictions to an LLM for dialogue-context emotion recognition. Across IEMOCAP, MELD, and CMU-MOSI the hybrid Pareto-dominates both pure systems (improving weighted F1 and cutting LLM spend by routing most turns to the near-zero-cost ensemble), with reported costs of roughly $10–85 vs $99–170 per million utterances for hybrid vs LLM-only pipelines.
KEY POINTS
- arXiv:2609.17977v1 compares a low-cost stacked ensemble, off-the-shelf LLM prompting (including GPT-4o-mini), and a confidence-gated hybrid that escalates low-confidence ensemble predictions to an LLM for dialogue-context emotion recognition.
- Across IEMOCAP, MELD, and CMU-MOSI the hybrid Pareto-dominates both pure systems (improving weighted F1 and cutting LLM spend by routing most turns to the near-zero-cost ensemble), with reported costs of roughly $10–85 vs $99–170 per million utterances for hybrid vs LLM-only pipelines.
- This demonstrates a practical, auditable deployment recipe for CCaaS and conversational-AI platforms to reduce LLM spend while improving or matching ERC accuracy by routing only uncertain turns to costly LLMs.
WHY IT MATTERS
This demonstrates a practical, auditable deployment recipe for CCaaS and conversational-AI platforms to reduce LLM spend while improving or matching ERC accuracy by routing only uncertain turns to costly LLMs.