Ant Group and ZJU Propose TRACE: Dynamic Confidence Assessment Boosts Streaming Emotion Accuracy to 69.47%

1 reports archived1 independent sourcesupdated 10/8/2026, 21:03:00

Synthesis & Latest Updates

1 Sources Cross-Validated

Ant Group Security Lab and Zhejiang University introduced TRACE, a streaming emotion understanding framework for real-time human-AI interaction, addressing latency and high multimodal inference costs in offline models. By continuously tracking judgment changes and dynamically assessing reliability, TRACE significantly reduces unnecessary multimodal calls. On the custom StreamMER dataset, it improved online audio-only accuracy from 59.87% to 69.47%, with the paper accepted at EMNLP 2026.

LATEST/Ant Group Security Lab and Zhejiang University introduced TRACE, a streaming emotion understanding framework for real-time human-AI interaction, addressing latency and high multimodal inference costs in offline models. By continuously tracking judgment changes and dynamically assessing reliability, TRACE significantly reduces unnecessary multimodal calls. On the custom StreamMER dataset, it improved online audio-only accuracy from 59.87% to 69.47%, with the paper accepted at EMNLP 2026.

HEAT TRENDHourly heat curve

2 fully observed hours

Fewer than 3 fully observed hours; trend not plotted yet.

TIMELINECoverage timeline

Total 1 reports · Latest first

量子位 (微信公众号)T2·Oct 8, 21:0368 pts
Ant Group and ZJU Propose TRACE: Dynamic Confidence Assessment Boosts Streaming Emotion Accuracy to 69.47%
- The TRACE framework addresses misjudgments and high latency in streaming scenarios caused by incomplete information through dynamic reliability assessment.
- Constructed the StreamMER dataset based on Friends seasons 1-2, containing 8,926 utterance samples to support emotion evaluation in continuous dialogue contexts.
- Experiments show TRACE improved online audio accuracy to 69.47% on StreamMER while effectively reducing computational overhead for multimodal reasoning.
View Evidence Original Source