Qwen 3.8 27B running locally identifies games from raw neural network weights

The paper by Rikiya Takehi, Jayden Teoh, Alex Zhang and Omar Khattab reports that frontier models can interpret and modify the raw weights of a machine learning model, which are presented as text full of floating point numbers. One of its examples asks a model to say which of four simplified grid games a policy network was trained on, with no code execution or other tools allowed.

Willison ran the Pac-Man prompt against Qwen 3.8 27B in LM Studio, using a long context window and a raised timeout because the default setting failed. The run took 46 minutes and 43,506 reasoning tokens, and the model answered with a Pac-Man-style game at 72% confidence. It also identified Pong at 75% and Breakout at 88%, but it did not identify Freeway.

Kimi K3.1 Pro, accessed through OpenRouter, correctly guessed Breakout after 9 minutes and 41 seconds. Willison says he is not ready to accept the wider claims made about the paper, and he treats the no-tools question as separate from work where models write and run code.

Developers can run a 27B open-weight model locally for some reasoning over raw parameter data, but expect runs of tens of minutes and set long timeouts and large context windows in agent pipelines. The results come from a few prompts run by one author, so they should not be treated as dependable weight analysis yet.