The research introduces GLARE, a model designed for forecasting meeting continuations. The system utilizes an adversarial imitation learning framework to improve its ability to generate plausible multi-turn conversations. A discriminator ranks generated continuations against observed continuations, providing a reward signal. This reward is then used to retrain the model, evolving the reward landscape through negative sampling.
Evaluation was conducted on the Meeting Dynamic Forecasting Benchmark (MDFB), comprising 2,207 real-world meetings and 24,794 future-facing queries. GLARE attained average human-evaluated win rates of 0.66 on utility and 0.70 on human-likeness. These results demonstrate the model’s ability to generate continuations that align with the question’s intent and maintain a consistent conversational flow.
Compared to standard supervised fine-tuning (SFT) and SPIN, GLARE’s performance remained below human continuation rates. This suggests a potential avenue for further improvement by incorporating human feedback more directly. The benchmark itself provides a valuable tool for comparing general-purpose models, including closed-source systems, through reference-assisted judgments.
The research highlights the use of the MDFB for both task-specific learning and output-based evaluation of meeting behavior. The system’s architecture and training methodology offer a novel approach to generative modeling within the context of social dynamics forecasting.
Source: https://arxiv.org/abs/2609.12165
