GLARE Forecasts Social Dynamics in Multi-Party Meetings

Summary

The paper addresses meeting continuation, where a model must track the agenda, speaker roles, participant intentions, and disagreement across a long multi-party discussion. It introduces the Meeting Dynamic Forecasting Benchmark (MDFB), built from 2,207 real-world meetings and 24,794 future-facing queries. Given a transcript prefix and an active question, a model generates a plausible multi-turn continuation in a single call. MDFB evaluates progress toward the question, or utility, and plausible conversational flow and role consistency, or human-likeness, without requiring an exact match to the observed future. The authors also present GLARE, which adapts adversarial imitation learning to conditional language generation. A discriminator scores the observed continuation above samples from the current actor, and that score becomes a KL-regularized policy reward; current-policy negatives are used to retrain the discriminator as the actor changes. In human evaluations, GLARE reaches average win rates of 0.66 for utility and 0.70 for human-likeness, outperforming supervised fine-tuning (SFT) and SPIN but remaining below the observed human continuation. The benchmark also supports output-based comparison of general-purpose models, including closed-source systems, through reference-assisted judgments. The authors position MDFB as both a task-specific training resource and an arena for evaluating meeting behavior.