The 2026 ACII Dyadic Conversations (DaiKon) Workshop & Challenge

The benchmark's actual novelty sits in formalizing interdependence as a measurable property. Most prior work treats conversation participants as isolated affect sources; ACII-DaiKon instead operationalizes coupled dynamics (influence flows, coordination timing, rapport trajectories) as first-class evaluation targets, not post-hoc analysis.

This connects directly to the Harvard diagnostic study from early May, which showed LLMs outperforming human clinicians on high-stakes judgment tasks. Both represent a shift from treating AI as a tool that augments human decision-making to treating it as a system that must model relational or contextual complexity at human-competitive granularity. The deepfake detection benchmark from the same period also reflects this pattern: as AI capabilities mature, evaluation frameworks must become more sophisticated to catch failure modes that simple metrics miss. ACII-DaiKon extends this logic to dialogue systems, where therapeutic AI and customer-facing agents now require benchmarks that capture whether the system actually understands the other person's state and adapts to it, not just whether it generates coherent text.

If any of the three ACII-DaiKon tasks show that current dialogue models (GPT-4, Claude, Llama) score below 0.65 F1 on directional influence detection, that signals a genuine capability gap requiring new architectures. If instead scores exceed 0.80, the benchmark may be measuring surface-level patterns rather than true relational reasoning, and the field should expect rapid saturation.