AI
Early experiments suggest frontier models can directly inspect and edit neural-network weights
A research team reports early experiments in which GPT-6 Astra inferred behaviors from raw neural-network weights and edited weights to change model behavior without gradient-based training. In one test, a targeted edit reduced looping in Qwen3.5-0.8B and raised reported MATH-500 performance from 7.9% to 48.1%; the authors stress that this removed an obstacle rather than adding new mathematical ability. The work is exploratory, includes failures on Gemma, and releases experiment logs for replication.
Editorial summary · Publisher article
What matters
- Astra identified tasks or structures in small game, diffusion and language networks from weights alone.
- It generated small game-playing networks that outperformed the cited PPO baselines in the reported evaluations.
- The Qwen edit changed 0.016% of weights and also improved reported instruction-following results.
- The model could not reliably identify sycophancy from weights alone, but succeeded when given attention patterns.
Context & caveats
- These are early, author-reported experiments; the packet does not establish independent replication or broad reliability beyond the tested models and tasks.
Why this made the edition
The research post reports extensive experiments, quantitative results, released logs and explicit limitations. Its extraordinary findings are clearly presented as early authors’ results rather than established capability, with failed and inconclusive cases included.
Source material