[PhD Proposal] Automated Calibration of LLM-Agent Based Educational Simulations скачать в хорошем качестве

[PhD Proposal] Automated Calibration of LLM-Agent Based Educational Simulations

We are moving from the "Classical Era" of simulations to the "Generative Era," but our methods for controlling AI agents are stuck in the past. "Off-the-shelf" LLM capabilities aren't enough—simulations need rigorous authorability to be scientifically valid. This proposal defense outlines a complete framework for moving from brittle manual prompting to robust, automated calibration. The Research Journey: 1. The Failure of Manual Authoring: Empirical evidence showing why hand-written prompts and schemas fail to scale or provide necessary cognitive scaffolding. 2. HypMix: A framework for modular, hypothesis-based calibration that allows for reusable behavioral priors. 3. The Consistency Gap: Results from the "Trust Game" revealing that LLM agents often fail to act consistently with their stated beliefs, proving that "roleplay" is not enough. 4. Proposed Work: Reframing calibration as an automated planning problem. By treating behavioral goals (like "balanced participation") as trajectory constraints, we can force agents to adhere to complex educational scenarios like pair programming.