Large language models (LLMs) are commonly used to reduce documentation burden and cognitive load, yet many past studies report heterogeneous productivity gains. We study how LLM drafting assistance reshapes documentation work in two complementary experiments. Study 1 is an online experiment in which clinicians complete clinical documentation either by editing an LLM-generated draft or by writing it from scratch. Editing does not result in time savings or quality improvements on average: clinicians vary widely in how they interact with the provided draft, and accountability converts writing time into editing time. Study 2 is a randomized lab experiment in which 100 university students summarize recorded interviews with or without LLM drafts while we record keystrokes, self-reported workload, and physiological measures of fatigue (heart rate, skin conductance, pupil diameter, and eye movements). In this controlled setting the predicted gains materialize — editing makes students roughly 60% more productive, raises quality, and lowers self-reported workload — but physiological fatigue accumulates as quickly when editing as when writing, and in some cases faster. Working with LLM drafts transforms the author from writer to editor: time savings are setting-dependent, quality gains are expertise-dependent, and the apparent relief of editing can mask fatigue that registers in the body rather than in awareness.

Many decision-making settings hinge on recognizing when to act: correctly identifying critical moments where taking action has a large impact. This problem is especially hard in supervision-poor environments, where operations are richly observed but the decision-relevant state is not, with little ground truth about consequential moments. We develop a decision-aware learning framework for such settings by connecting the representation learner for the state space to the downstream decision-making problem, training the learned representation to be robust through adversarial reweighting of the most decision-relevant windows. We apply the framework to military aviation training, where wearable sensors track physiological and environmental conditions during training flights and managers seek to intervene before fatigue- or hypoxia-related risk becomes severe. Using data from more than 400 U.S. Air Force training flights, we find that the learned state space identifies coherent high-risk regions and that experienced aviators transition differently around those regions.

Digital learning platforms face an important trade-off: raising student workload can improve learning but might reduce engagement. We study this trade-off on a large Chinese language-learning platform with thousands of active students and roughly 6 million reviews per month, where each learned item generates four related skills (reading, writing, definition, and tone), so the platform must jointly manage forgetting and cross-task spillovers. We first estimate student-specific memory and transfer parameters in an offline memory model, then design a two-stage adaptive review policy that allocates daily review capacity via fast meta-adaptation and selects items using a spillover-aware bandit algorithm. Incorporating contextual transfer improves predictive accuracy for 85% of students; personalization lengthens review intervals by 15.8% without lowering predicted recall; and policy simulations reduce reviews to 48% of baseline while reducing projected attrition by 11%.