Dongxi NLP Questions GPT-6 Astra's Weight Editing Method: Like an LLM's He Jiankui
Dongxi NLP questions the Weight-Space Programming proposed by Rikiya Takehi, arguing that the method has traces of activation steering: first find the activation direction corresponding to a certain tendency of the model, such as the loop problem of being unable to stop when solving problems, and then project out the corresponding components in the 128 columns of weights that write the most into that direction. He points out that evaluations such as MMLU-Pro, GPQA, and IFEval did indeed all improve, but these benchmarks were originally all affected by loops; once the loop is fixed, scores naturally improve across the board, and the real question is the side effects. The original post claims that GPT-6 Astra only changes 0.016% of Qwen3.5-0.8B's weights, improving math by up to 6x and instruction following by 3x, with no gradients or training required.
Why read this
AI score 58
The author compares the weight editing method with activation steering, and asks whether the benchmark improvements are merely a side effect of fixing the loop.
Topics
View original postWe asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq
View original
We asked GPT-6 Astra to edit Qwen3.5-0.8B weights to improve math. It directly changed only 0.016% of weights, fixing 0.8B's reasoning loops & improving math up to 6x and even instruction following 3x, without gradients or training. We call this Weight-Space Programming: frontier LLMs can now natively read/write neural network weights! 🧵 Blog: https://t.co/bOWCcFhCVq
Body · AI translation
Very thought-provoking! But... isn't this just like an He Jiankui for LLMs?
In the method I see echoes of activation steering: first find the activation direction corresponding to a certain tendency of the model, such as the loop problem in the article where the model can't stop doing math problems.
Rather than adding a steering vector to the model, the authors projected out the corresponding component in the 128 columns of weights that write most strongly into this direction.
The evaluations using MMLU-Pro, GPQA, IFEval, etc. did indeed all improve, but these benchmarks were originally all affected by the loop; once the loop is fixed, of course the scores improve across the board.
The question is, what about the side effects?