Direct optimization for Dynamic_max forecasting

Develop an optimization objective that directly trains conversational derailment forecasting models for Dynamic_max evaluation, in which a single threshold-crossing prediction determines the classification of the entire dialogue, to better support moderation workflows prioritizing one proactive warning over continuous state tracking.

Background

The paper evaluates dynamic derailment forecasting using both mean aggregation and max aggregation. Under Dynamic_max, a conversation is classified as derailed if any timestep prediction exceeds a threshold, making the metric sensitive to isolated false-positive spikes but potentially aligned with moderation systems that issue a single warning.

The models in the paper are optimized independently at the dialogue-history level rather than directly for the sequence-level Dynamic_max objective. The authors therefore identify direct optimization for this evaluation setting as an unresolved direction that could better match single-intervention moderation workflows.