MIT Team Reports GPT-6 Astra Weight Edit Lifts Qwen3.5-0.8B Math Accuracy To 48.1%
A research team led by Rikiya Takehi at MIT's Computer Science and Artificial Intelligence Laboratory published a report on November 2, 2026, describing Weight-Space Programming, in which a large language model reads and edits another network's weights directly.
1 source 3h ago
Reporting from 1 source: GIGAZINE.
Why it matters
If a large model can read intent out of raw weights and rewrite them without training, the cost of adapting a model shifts from data and compute to a single inference pass, which is why the framing of the 7.9% to 48.1% result is being argued over so closely.
Key Facts
- Rikiya Takehi of MIT's Computer Science and Artificial Intelligence Laboratory led the research, published November 2, 2026.
- GPT-6 Astra changed 0.016% of the weights in Qwen3.5-0.8B, raising math benchmark accuracy from 7.9% to 48.1%.
- Astra changed 131,072 weights in Qwen3.5-0.8B, a model of about 800 million parameters.
- In a driving task, Astra rewrote 3.5% of a network's weights and the fixed network completed all 1,024 trials.
- Astra identified which of four Atari games a network was trained on, all sharing 17,253 weights, without being told the game names.
Using GPT-6 Astra, the team changed 0.016% of the parameters in Alibaba's Qwen3.5-0.8B, lifting math benchmark accuracy from 7.9% to 48.1% without additional training. The report has drawn criticism that its framing is misleading.
The headline number comes from 131,072 changed weights inside a model of about 800 million parameters. Astra found the cause of Qwen3.5-0.8B's repeated output in a specific direction written by a specific layer, after using up to 50 evaluation runs on math practice problems to locate it.
Earlier experiments in the same report had Astra identify which of four Atari games a network was trained on (Breakout, Pong, Pac-Man, Freeway), all sharing the same 17,253 weights, without being told the names. In a driving task, it rewrote 3.5% of weights so the car reversed before climbing, and the fixed network finished all 1,024 trials. Skeptics argue the report's presentation overstates what was shown.
Written with AI by Yomimono from the 1 cited source below, including Japanese-language reporting where cited. See our editorial policy.
Add Yomimono as a preferred source on Google