MIT CSAIL just showed me a weirder way to fix a model than retraining it. GPT-… · @harden_369741 · TwKit

@harden_369741

MIT CSAIL just showed me a weirder way to fix a model than retraining it. GPT-6 Astra read Qwen3.5-0.8B's raw weights, found the direction behind its looping, and edited 0.016% of the weights. MATH-500 went from 7.9% to 48.1%, GSM8K from 13.8% to 67.1%. My take: I read this as repair work, plain and simple. The same trick failed on Gemma 3 270M, which barely looped. If your small model is stuck in loops, I would hunt for the blocking direction before paying for more training. Edits unblock ability, they do not invent it. Source: https://venturebeat.com/technology/like-gene-editing-for-neural-networks-mit-researchers-used-gpt-6-astra-to-edit-qwens-open-weights-directly-boosting-math-performance #AI #MachineLearning

曝光 10 · 评论 0 · 点赞 0 · 书签 0