ACII-DaiKon 2026: A New Benchmark for Interpersonal Emotion Modeling in Bidirectional Dialogues

ACII-DaiKon 2026 introduces the first benchmark focused on modeling interpersonal emotions and social dynamics in bidirectional dialogues. It is built on the Hume-DaiKon dataset with 945 natural dialogues, covering three core sub-tasks: directional interpersonal influence prediction, turn transition prediction, and rapport trajectory prediction. This benchmark aims to address limitations of existing speaker-centric approaches and advance interaction-centric dialogue AI.

Despite rapid progress in dialogue emotion modeling, existing benchmarks are mostly speaker-centric, failing to capture coupled, time-evolving processes between dialogue partners, including:

• Directional interpersonal influence
• Dialogue time coordination
• Rapport development This perspective limits models' understanding of real-world interpersonal interaction dynamics.

The ACII-DaiKon benchmark is based on the Hume-DaiKon dataset, which includes:

• 945 bidirectional dialogues
• 743.4 hours of audio-video data
• 5 languages from natural collection scenarios

Its three core sub-tasks are:

1. Directional interpersonal influence prediction: Predict one party's emotional impact on the other
2. Turn transition prediction: Includes next speaker prediction and next utterance time interval prediction
3. Rapport trajectory prediction: Predict long-term changes in mutual rapport between dialogue partners.

The benchmark uses task-specific evaluation metrics:

• CCC (Consistency Correlation Coefficient) and Pearson correlation for continuous variable prediction
• Macro-F1 for classification tasks
• MAE (Mean Absolute Error) for time prediction

Baseline results show:

Task	Metric	Value
Influence Prediction	CCC	0.40
Influence Prediction	Pearson	0.50
Turn Transition	Macro-F1	0.66
Turn Time	MAE	1.50s
Rapport Trajectory	CCC	0.68
Rapport Trajectory	Pearson	0.70

Current methods capture coarse-grained bidirectional patterns but struggle with directional dependencies and long-term dynamics.

The benchmark's innovations include:

1. Multi-modal support: Provides audio, video, and text modalities to explore fusion strategies
2. Temporal reasoning focus: Task design emphasizes modeling time-evolving processes, requiring models to have temporal reasoning capabilities
3. Cross-context generalization: Fixed train/validation/test splits enable evaluating models' ability to generalize across contexts.

ACII-DaiKon advances research in:

• Interpersonal emotion computing
• Dialogue system modeling
• Multi-modal temporal learning
• Social signal processing

Practical applications include:

• Optimizing emotional interactions in smart customer service systems
• Evaluating dialogue quality in mental health counseling
• Enhancing naturalness of human-machine dialogue systems
• Supporting cross-cultural communication studies.

Beyond the technical benchmark, the ACII-DaiKon workshop fosters cross-disciplinary discussions on:

• Data validity verification methods
• Standardization of evaluation protocols
• Cultural-aware modeling strategies
• Theoretical frameworks for bidirectional interactions.

ACII-DaiKon 2026 establishes a new evaluation standard for interpersonal emotion and social dynamics modeling by introducing a large-scale multi-modal bidirectional dialogue dataset and three challenging sub-tasks. It is expected to drive the paradigm shift of dialogue AI from speaker-centric to interaction-centric.