很酷！但是 sparse rewiring 又跟所谓的 weight space progmming 有什么分别。
可以确定的，做可解释性 LLM 的研究员，春天来了。
依然推荐阅读！感兴趣的，把 post 直接扔给 agent 解读！
Dongxi NLP asks how sparse rewiring differs from weight space programming
AIDongxi NLP says Sparse Rewiring, which Max Sloef reports, trains 89 of Qwen3-8B's 442,368 neurons to make an owl persona that matches full SFT. Dongxi NLP asks how this differs from so-called weight space programming, and says interpretability researchers should pay attention.
Dongxi NLPVerified on X
@dongxi_nlp
Finetuning changes all model weights, even when learning a simple behavior. Can we make that change interpretable? Yes! Sparse Rewiring learns new triggers for neurons, not new outputs. Training just 89 of Qwen3-8B's 442,368 neurons makes an owl persona that matches full SFT!
Source details
Dongxi NLP · x.comPublished · added here
Original title: 很酷！但是 sparse rewiring 又跟所谓的 weight space progmming 有什么分别。