New Model Predicts Why Agents Abandon Tasks and How to Optimize Rewards

A team of researchers from Japan (Yasunori Akagi, Hideaki Kim, Daichi Fushihara, Ryosuke Nakahama, Hiroyasu Miyazaki, Takeshi Kurashima) has published a paper on arXiv proposing a tractable continuous-time model for designing interventions for time-inconsistent agents. The core problem is that agents (humans or AI) often plan to complete a long-term task but later abandon it because, under non-exponential discounting, the perceived trade-off between immediate effort and delayed reward changes over time. The model focuses on deadline-constrained progress-based tasks, where the agent repeatedly chooses a future trajectory that minimizes perceived cost and then follows its infinitesimal initial direction. This yields a continuous-time dynamic behavior defined through a variational problem. The authors show that under generalized hyperbolic discounting—a broad class that includes exponential and hyperbolic discounting—the resulting trajectory has a concise analytical representation. They use this to characterize when the agent completes the task, abandons it immediately, or shows time-inconsistent abandonment after partial progress.

The paper then tackles two intervention design problems: optimal goal setting and optimal reward scheduling. For goal setting, they derive optimal goals both when exploitative rewards (rewards that take advantage of time inconsistency) are allowed and when they are prohibited, and identify conditions under which exploitative rewards are ineffective. For reward scheduling, they prove that, for a fixed number of stages, equal-length periods and equal rewards are optimal. Furthermore, finer reward splitting monotonically improves final progress up to a discount-independent limit. These results provide a continuous-time alternative to existing discrete-time models and offer practical insights for designing incentives in learning, exercise, project management, and AI alignment. The framework is mathematically rigorous yet computationally tractable, making it suitable for applications in behavioral economics, reinforcement learning, and automated coaching systems.