Overview
This forum post shares the paper GLARE: Generative Learning via Adversarial Reward Estimation (arXiv: 2609.12165), covering both the Meeting Dynamic Forecasting Benchmark (MDFB) and the proposed GLARE training method.
Field: Machine Learning Authors: Tenghao Huang, Zhaoxuan Tan, Muhao Chen, Jonathan May, Mengting Wan, Longqi Yang, Pei Zhou, Sihao Chen Posted: 2026-09-15
Key Points
- Meeting Dynamic Forecasting Benchmark (MDFB): constructed from 2,207 real-world meetings and 24,794 future-facing queries. Meeting continuation requires tracking the agenda, speaker roles, participant intentions, and disagreement across long multi-party discussions.
- Task setup: given a transcript prefix and an active question, a model generates a plausible multi-turn continuation in a single call.
- Evaluation: two criteria — *utility* (progress toward the question) and *human-likeness* (plausible conversational flow and role consistency) — without requiring exact reproduction of the observed future.
- GLARE method: an adaptation of adversarial imitation learning to conditional language generation. A discriminator ranks the observed continuation above samples from the current actor; its scores form a KL-regularized policy reward. The discriminator is retrained on negative samples from the current policy, so the reward landscape co-evolves with the actor.
- Results: in human evaluations, GLARE achieves a 0.66 utility win rate and 0.70 human-likeness win rate, outperforming SFT and SPIN, though still below the observed human continuations.
- Broader use: MDFB also serves as a social-reasoning arena for comparing general-purpose models, including closed-source systems, via reference-assisted judging — supporting both task-oriented learning and output-based evaluation of meeting behavior.
- Paper: https://arxiv.org/abs/2609.12165