We’re excited to announce the 2026 ACII Dyadic Contest (DaiKon) Workshop & Challenge, sponsored by Hume AI, a new machine learning benchmark focused on modeling interpersonal affect and social dynamics in two-person conversations.

### **Why DaiKon?**

Recent advances in affective computing have largely focused on **individual-level predictions**—what a single speaker is feeling or expressing. But real-world conversations are fundamentally **relational**. Emotions evolve between people, shaped by interaction, timing, and mutual influence.

DaiKon shifts the focus from isolated speakers to **dyadic dynamics**, capturing phenomena such as:

- Emotional contagion
- Directional influence between speakers
- Time-evolving affect within conversations

To our knowledge, this is one of the first large-scale efforts in the ML community, and the first at ACII, dedicated specifically to **interpersonal affect in dyadic settings**.

### **The Challenge**

DaiKon is structured around three core tasks:

#### **Influence**

Predict a speaker’s emotional state given conversational context.

Models estimate continuous intensity across **10 emotion dimensions** (e.g., joy, anxiety, curiosity).

#### **Turn-Taking**

Predict **who speaks next and when**.

This task captures conversational timing and coordination.

#### **Rapport**

Model how **interaction quality evolves over time**, predicting rapport continuously throughout a conversation.

### **The Dataset**

Participants will work with a large-scale, naturalistic dataset of dyadic conversations:

- **945 sessions**
- **743 hours of conversation**
- **5 languages and regions**
- **10 emotion dimensions**

All data were collected via Hume AI’s dual-channel recording platform, enabling clean separation of speakers and detailed analysis of conversational dynamics. Importantly, all participants provided **informed consent**, and the dataset reflects **diverse cultural and linguistic contexts**.

### **Baselines & Research Opportunities**

We provide multimodal baselines (audio, video, and combined), highlighting both the promise and difficulty of these tasks. Performance varies across modalities, with **multimodal approaches generally offering the strongest results**, though significant headroom remains.

Beyond the competition, we welcome workshop submissions on topics such as:

- Dyadic and multi-party interaction modeling
- Temporal modeling of affect
- Multimodal fusion (audio, video, text, physiology)
- Representation learning for social and emotional behavior
- Cross-cultural interaction modeling

### **Key Dates**

- **Challenge opens:** April 6, 2026
- **Baseline paper:** April 16, 2026
- **Test set release:** May 20, 2026
- **Submission deadline:** May 25, 2026
- **Workshop papers due:** May 30, 2026
- **Conference:** September 7–10, 2026 (Pueblo, Mexico)

### **Get Involved**

To participate, register your team by emailing **competitions@hume.ai** with:

- Team name
- Researcher(s)
- Affiliation
- Research goals

Participants will receive access to the dataset under a research EULA.

### **Looking Ahead**

DaiKon represents a step toward more **human-centered AI systems**—ones that understand not just what individuals feel, but how emotions emerge and evolve **between people**.

We’re excited to see what the community builds. For more information on our training datasets, see [here](/content/training-data/index.html).