Zachary Bamberger

Problem. Inference-time methods like Tree-of-Thoughts diversify reasoning trajectories through temperature sampling, but temperature cannot control how branches differ, nor produce traces that explain which reasoning choices drove output quality.

Method. STATe replaces sampling with discrete, interpretable textual interventions. A controller selects actions from a structured action space (e.g., which discourse relation or subtopic to develop next), converts each into a text prefix that steers the generator's continuation, and an evaluator scores candidates to guide beam search. Each surviving branch carries a logged sequence of actions.

Result. STATe substantially improves both the diversity and quality of generated outputs over temperature-based baselines. In argument generation, sequential models over logged action sequences prove highly predictive of output quality, showing that the ordering and composition of rhetorical choices carries meaningful signal.

Problem. T5 and its variants are pre-trained with span masking, which teaches token-level and semantic recovery but does not encourage the model to learn how sentences relate to one another structurally. Discourse understanding therefore depends entirely on task-specific supervision.

Method. DEPTH augments T5's span-masking objective with a discourse-oriented pre-training signal. The model is additionally trained to reconstruct the correct ordering of shuffled sentences—learning to reason about how ideas connect across a document, directly from unlabeled text and without changing the architecture or fine-tuning procedure.

Result. Pre-training with DEPTH improves data efficiency and downstream performance on tasks requiring discourse understanding. These gains also transfer favorably to other linguistic capabilities, suggesting that discourse-aware pre-training instills broadly useful representations rather than narrow task-specific ones.