Adventures in Weight Space Programming

Weight-space programming uses a language model to inspect and directly modify another model’s weights, without training or gradients. Beaker, an autonomous agent improver, tested whether it could discover useful edits to Qwen3.5-0.8B, whose accuracy on MATH-500 was about 7% and whose answers often looped until they hit the token limit. In a run where Beaker could not execute the model, it identified a single neuron in layer 22 and doubled its output. The change made the token “Wait” less likely when the model was stuck in repetitive reasoning, raising accuracy to 37.4%. A more complex edit found with GPT-6 Astra reached 38.6%; allowing Beaker to inspect activations and token probabilities produced results between 34.8% and 39.5%.

Combining edits from several Beaker runs raised accuracy to 43.0%, chiefly by reducing answers cut off at the token limit. Adding Beaker’s neuron edit to the broader change from Takehi and colleagues - 64 output columns modified across two layers - reached 58.8%, compared with 51.1% for that edit alone. The experiments suggest that a small, targeted change can complement larger edits by addressing a different failure mode. A cheaper model, GPT-6.1 Sol, found the same key neuron and performed similarly at a fraction of the cost. Beaker could also make answers shorter, but failed to teach the model to say “I don’t know”; the results demonstrate promising but limited behavior changes on one small model and math benchmark.