Riki Takehi says GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights to raise math scores
Article timeline
The articles in this story. Times are ET.
Vaibhav (VB) SrivastavXGPT-6 Astra edits Qwen3.5-0.8B weights to improve math without trainingAIOpenAI's GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B's weights, without gradients or training, to fix reasoning loops. The edits improved math performance up to 6x and instruction following 3x, according to researcher Rikiya Takehi, who calls the approach Weight-Space Programming.
Yuchen JinXLarge LLM edits small model weights to boost math and instruction followingAIYuchen Jin says using a large LLM to directly edit a smaller model's weights could dramatically improve its capabilities, calling it gene editing for neural networks. The linked post reports that GPT-6 Astra changed only 0.016% of the weights in Qwen3.5-0.8B without gradients or training. It says the edits raised math performance up to 6x and instruction following 3x.
Rohan PaulXGPT-6 Astra edits 0.016% of Qwen3.5-0.8B's weights, lifting math accuracy from 7.9% to 48.1%AIRohan Paul relays research from MIT CSAIL showing that GPT-6 Astra improved Qwen3.5-0.8B's math accuracy on MATH-500 from 7.9% to 48.1% without training. Astra edited only 0.016% of the weights after tracing a repetition loop to a signal written by two layers, then subtracting that signal.
Thomas WolfXThomas Wolf posts "what?" with no supporting contentAIThe main post from Thomas Wolf, owner of the Hugging Face account, contains only the single word "what?" and no further details. It quotes a post from @rikiyatakehi claiming that GPT-6 Astra edited 0.016% of Qwen3.5-0.8B weights, improving math up to 6x and instruction following 3x without gradients or training.
Jerry LiuXJerry Liu reacts to GPT-6 Astra editing Qwen3.5-0.8B weightsAIJerry Liu posts only "wtf" in response to a quoted post from Riki Takehi. The quoted post says GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B's weights without gradients or training, raising math performance up to 6x and instruction following 3x.
whXPost jokes about asking GPT-6 Astra to fix a model's weightsAIAccount @nrehiew_ posts a short joke, "Astra, my model is shit. Please fix. Make no mistakes.", addressed to GPT-6 Astra. The post quotes @rikiyatakehi's claim that GPT-6 Astra directly edited 0.016% of Qwen3.5-0.8B's weights, improving math up to 6x without training.
Ethan MollickXWeight-space programming lets GPT-6 Astra edit Qwen3.5-0.8B weightsAIRiki Yatake's team asked GPT-6 Astra to edit the weights of Qwen3.5-0.8B to improve math, changing only 0.016% of them. The team says the edits fixed the model's reasoning loops and raised math performance up to 6x and instruction following 3x, without gradients or training. Ethan Mollick, who is waiting for replication, calls the result a pretty big deal.
Dongxi NLPXPost questions whether weight-editing fix for Qwen3.5-0.8B loops has side effectsAIDongxi NLP compares the method to activation steering and argues that the reported gains on MMLU-Pro, GPQA, and IFEval may come from fixing reasoning loops that affected those benchmarks. The author removed the projection of 128 weight columns that write the loop direction rather than adding a steering vector. The post asks what side effects the edit causes.
Andrew CurranXGPT-6 Astra edits Qwen3.5-0.8B weights to improve math without trainingAIRiki Yatakehi says GPT-6 Astra directly changed only 0.016% of the weights in Qwen3.5-0.8B, without gradients or training. The edits reportedly fixed the model's reasoning loops and raised math performance up to 6x and instruction following up to 3x.
