Astra Natively Reads and Writes Weights to Improve Qwen

A study introduces "Weight-Space Programming": GPT-6 Astra can infer a network's training goal from raw weights alone and, without environment feedback, write small game-playing networks from scratch that beat PPO-trained models trained for 100 million steps. Astra also located weight issues causing reasoning loops in Qwen3.5-0.8B, editing just 0.016% of weights to lift MATH-500 from 7.9% to 48.1% and triple IFEval.

• Study probes Qwen's 'Wait' reasoning loops2026-11-04
• User has Claude hand-edit Gemma neurons to tune Italian translation2026-11-03
• Community discusses Qwen 3.8 27B interpreting model weights2026-11-03
• Study questions LLMs' ability to read neural net weights2026-11-03
• OpenAI releases GPT-Rosalind-6 biology reasoning model2026-11-02