GLARE: Generative Learning via Adversarial Reward Estimation for Social Meeting Continuation

Forum topic · 小凯 · 2026-09-15

Summary

This post introduces GLARE, a method adapting adversarial imitation learning to conditional language generation for meeting continuation, alongside the Meeting Dynamic Forecasting Benchmark (MDFB). MDFB is built from 2,207 real-world meetings and 24,794 future-facing queries: given a transcript prefix and an active question, a model must generate a plausible multi-turn continuation in a single call. Evaluation focuses on utility (progress toward the question) and human-likeness (plausible conversational flow and role consistency), without requiring exact reproduction of the observed future. GLARE trains a discriminator to rank observed continuations above samples from the current actor; its scores form a KL-regularized policy reward, and retraining on the current policy's negative samples lets the reward landscape co-evolve with the actor. 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 observed human continuations. The authors also demonstrate MDFB as a social-reasoning arena for comparing general-purpose models, including closed-source systems, via reference-assisted judging. Paper: arXiv 2609.12165.

Paper Overview

Field: Machine Learning Authors: Tenghao Huang, Zhaoxuan Tan, Muhao Chen, Jonathan May, Mengting Wan, Longqi Yang, Pei Zhou, Sihao Chen Published: 2026-09-15 arXiv: 2609.12165

Summary

Meeting continuation requires tracking the agenda, speaker roles, participant intentions, and disagreement across long multi-party discussions. The authors introduce the Meeting Dynamic Forecasting Benchmark (MDFB), constructed 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 one call.

Evaluation focuses on two axes:

Utility — progress toward answering the active question
Human-likeness — plausible conversational flow and role consistency

Neither axis requires exact reproduction of the observed future.

GLARE Method

The paper further presents GLARE (Generative Learning via Adversarial Reward Estimation), an adaptation of adversarial imitation learning to conditional language generation:

A discriminator ranks the observed continuation above samples drawn from the current actor.
The discriminator's scores constitute a KL-regularized policy reward.
The model is retrained on the current policy's negative samples, so the reward landscape co-evolves with the actor.

Results

In human evaluations, GLARE achieves:

0.66 utility win rate (vs. SFT and SPIN baselines)
0.70 human-likeness win rate

GLARE outperforms supervised fine-tuning (SFT) and SPIN, but still falls short of observed human continuations.

The authors also show MDFB can serve as a social-reasoning arena for comparing general-purpose models — including closed-source systems — using reference-assisted judging. Together, these results demonstrate the benchmark's usefulness both for task-oriented learning and output-based evaluation of meeting behavior.

---

*Auto-collected on 2026-09-15.*

GLARE: Generative Learning via Adversarial Reward Estimation for Social Meeting Continuation

Summary

Paper Overview

Summary

GLARE Method

Results

Tags