A Relational-Latent World Model for Long-Horizon Dyadic Interaction: Results on the ACII-DaiKon 2026 Challenge

Abstract

We present a dyadic relational-latent world model and evaluate it on the ACII Dyadic Conversations (DaiKon) 2026 challenge. Rather than treating the three sub-challenges (Influence, Turn-Taking, and Rapport) as independent prediction targets, we frame them as probes of a single synthesized relational latent learned through dyadic world-model pretraining. A single backbone, fine-tuned end to end, exceeds the official Hume audio baseline on the challenge validation split by +0.105 CCC on Influence (0.495) and +0.047 CCC on Rapport (0.697), averaged over three seeds, and trails on Turn-Taking by 0.063 Macro-F1 (0.547), a gap consistent with the model’s relational rather than sub-relational timescale. A smooth-blend multi-turn Rapport head raises CCC from 0.668 to 0.697 while cutting seed-to-seed standard deviation roughly tenfold. On English (US) dyads, the largest validation slice, the model improves rapport CCC by +0.070 over the released challenge baseline in a like-for-like comparison (0.782 vs 0.712). Controlled ablations locate the backbone’s advantage in the session-level Rapport track and in the state tracker’s learned forward dynamics, which transfer across cross-domain and in-domain pretraining; the gains require end-to-end fine-tuning, and a matched-recipe frozen-backbone control does not reach the Hume baseline on any track, so the strong form of the world-model transfer claim does not hold for this benchmark. We report multi-seed results throughout, finding them robust at full supervision. The model is compact and deployable: at 7.49 M parameters it runs in real time at 21 ms per turn on a single GPU and 137 ms on one CPU core, within the ∼250 ms turn-latency budget of spoken dialogue and suited to on-device, edge deployment. We position the model in its intended downstream role as a long-horizon planner and critic that a conversational LLM calls as an external tool, predicting interaction trajectories from outside its context window; actively shaping those trajectories is the motivating goal and is left to future closed-loop work.