Researchers report model edit improved its maths score

The researchers describe the method as “weight-space programming.”[2] They define it as an LLM directly reading and writing a network’s weights, rather than relying on an optimizer or gradient descent.[2] In the reported maths experiment, Astra was asked to make Qwen3.5-0.8B better at maths without gradients.[2] The researchers say Astra did not train the model and was not allowed to write training loops.[2] That makes the reported intervention different from an approach that updates a model through a training process.

The researchers report that Qwen3.5-0.8B’s MATH-500 score rose from 7.9% to 48.1% after Astra’s edit.[2] They say the comparison used the same sampling settings before and after the change.[2] The edit altered 0.016% of the model’s weights, according to the researchers.[2] They also report that the model’s IFEval score tripled.[2] These are reported benchmark results; the evidence does not establish how the edit would perform across other models or real-world tasks.

The researchers say Astra identified reasoning loops in Qwen3.5-0.8B’s weights as a problem.[2] They also say Astra can infer what a network was trained to do by examining its raw weights.[2] Their post describes the broader approach as a way to find a failure in a network’s weights, fix it, or write a new network.[2] The evidence does not specify the exact weight changes Astra made to address the reasoning loops, so readers cannot assess the edit’s mechanics from these findings alone.

In a separate experiment, the researchers say Astra wrote small game-playing networks from scratch that outplayed networks trained with PPO for 100 million steps.[2] That result is distinct from the Qwen maths-score comparison.[2] The researchers describe their findings as very early results.[2] For professionals, that caveat matters: the results are a reason to examine the method, not evidence that direct AI-led edits are ready for routine use or will improve other models.