TOOL / 2026
rerpg — Abyssal Tactics
A graduation research platform that compares LLM, reinforcement-learning, and human decisions inside a JRPG battle with lethal attacks and strict action-point constraints.
- 01RL, LLM, and human comparison
- 02DP oracle and academic metrics
- 03Japanese, English, and Chinese docs
Problem & goal
The research asks how LLMs, reinforcement learning, and humans make different decisions under limited actions and lethal risk, requiring a shared environment and evaluation protocol.
Technical challenge
The battle had to work as a game and as a reproducible experiment, including detection of reward hacking, arithmetic errors from LLM agents, and mismatches between measured data and reports.
Solution
The project integrates a Gymnasium environment, PPO training, LLM autoplay, a dynamic-programming oracle, and a React observation UI. Multiple reward designs share tests, metrics, and TensorBoard-based reporting.
Result & learning
The result is a reproducible research pipeline linking battle logic, training, visualization, and academic documentation, with metrics that explain why a policy succeeds rather than reporting wins alone.