Dev turns Pokelike.xyz into an open-source LLM and RL benchmark harness

Logical_Delivery8331 · reddit · 2026-08-19

A data scientist built a full harness around the web game Pokelike.xyz, turning it into a dual LLM and reinforcement learning benchmark, now open source.

LLM side: the model receives a system prompt, tools, previous turns, and the current game state per turn — all configurable (strategy prompt, state representation, toolset). Starter bots are included in the repo.

Early results: GLM 5.2 at best reaches the second map before dying; Opus performs about the same. The author wants to isolate whether the bottleneck is strategy, missing information, state representation, or tooling — and hopes someone gets better results from smaller models via prompt/tool improvements.

Traditional RL agents run alongside on a shared leaderboard. Everything runs fully offline (game and assets downloaded once, simulated locally). Built by the author, a friend, and Claude Code.

Repo: https://github.com/pierpierpy/pokelike.xyz.bot

Original post →

More from coding & agent

coding & agent channel →