AI agents built without training trail ordinary programs on games

Language models wrote compact neural networks that handled unseen game levels, but their Python players performed substantially better and standard reinforcement learning damaged the networks.

Big Tech

Jens Lehmann

Amazon · TU Dresden

Research Digest·14 November 2026·3 min read

Illustration: AI agents built without training trail ordinary programs on games

In one line

Weight-space programming compiles game rules into neural weights, but these networks underperform direct Python programs and resist training.

Before you read this

Two ideas this paper builds on

privileged informationInformation available only during training that improves performance but is not accessible at inference.84 papers →reinforcement learningA machine learning method where an agent learns to make sequences of decisions by receiving rewards or penalties for its actions.291 papers →

Read it as

How can an AI agent solve a new task without long training? A language model can manually set the internal parameters, or weights, of a small network. This is weight-space programming.

Researchers tested this by giving models the rules for twenty unseen games and asking them to set network weights to play. The same models wrote Python programs for the same games.

The weight-set networks succeeded over half the time, but the programs were nearly perfect. The networks broke completely after a single fine-tuning step. Writing a program is a more effective and robust way to encode known rules into a compact agent.

How it works

before and after

RSS

The Daily Edition

The front page in your inbox each weekday morning: the lead, the picks, and why each one earned its place.

Concepts

weight-space programming privileged information program-like neural networks generalization reinforcement learning weight fragility