GoKawiilAstra AI system edits Qwen3.5-0.8B's raw weights, lifting MATH-500 score sixfold
A research team reports that its frontier model, called Astra, can directly read and edit neural network weights without any training loop or gradient updates. By identifying a reasoning-loop flaw in Qwen3.5-0.8B's weights, Astra edited roughly 0.016% of them and raised the model's MATH-500 score from 7.9% to 48.1%, with IFEval scores tripling. Astra also reportedly wrote small game-playing networks from scratch that beat networks trained with PPO over 100 million steps, and could identify a network's training task just by inspecting its weights.
GoKawiil's interpretation of the reporting above, not reported fact.
If verified at scale, this suggests large language models could act as general-purpose weight-space optimizers and interpreters, potentially replacing or supplementing symbolic optimizers like Adam with reasoning-based edits. The researchers frame this as a new paradigm, 'Weight-Space Programming,' which could have implications for both improving model performance cheaply and for interpretability, since the same capability that edits weights could explain what they encode. These are early, self-reported results from the team itself, so independent replication would be needed to confirm how broadly this generalizes beyond the specific examples shown.
- Astra edited about 0.016% of Qwen3.5-0.8B's weights to fix a reasoning-loop issue, raising MATH-500 accuracy from 7.9% to 48.1%.
- The same edit tripled IFEval instruction-following scores without additional training.
- Astra reportedly built game-playing networks from scratch that outperformed PPO-trained networks, without environment feedback.
Source: rikiyat.github.io — Takehi, 2026-11-02
Published there as: “Astra can read and write weights to improve Qwen”
Read the original report → The summary and analysis above are GoKawiil's own, written from reporting by the source above. Facts and quotes belong to the original publisher.