Tag
An experiment analyzes why Qwen3.5-0.8B gets stuck repeating 'Wait' by using an interpretability lens on the 'Wait' token, showing doubt-related features dominating near stopping points, and applies a tiny weight edit (0.016% of weights) that converts 'Wrong/Nope/Hmm' into 'Verification/Completion/Result' while sharply improving MATH-500 performance; code and data are released.
A tweet criticizes a viral claim about 'Weight-Space Programming' where GPT-6 Astra supposedly edits Qwen3.5-0.8B weights to boost math performance, noting the results lack competent baselines and the Qwen result is mere bug fixing, while acknowledging that models reading/writing weights in toy problems is a real trend worth watching.
Neel Nanda comments on a paper where GPT-6 Astra edits Qwen3.5-0.8B weights to improve math, arguing the result shows LLMs are surprisingly good at classic mechanistic interpretability techniques rather than anything fundamentally new.
A blog post introducing "Weight-Space Programming," where frontier LLM GPT-6 Astra natively reads and edits neural network weights—changing only 0.016% of Qwen3.5-0.8B's weights to boost MATH-500 from 7.9% to 48.1% and tripling IFEval, without gradients or training.
该论文提出一种训练无关的权重编辑方法，通过少量对比提示对直接修改模型权重，降低指令微调模型对有害指令（如说谎、性别歧视）的服从率，并在五个模型上验证了其优于激活引导等方法的鲁棒性。
Decoy Direction Optimization (DDO) is a fast, post-hoc defense method that protects open-weight LLMs from refusal feature ablation attacks by injecting decoy signals into the network, achieving high robustness at lower cost than trained defenses.
The paper introduces AMRA, a weight-editing method to mitigate abliteration in large language models by obscuring the refusal signal, improving post-abliteration refusal scores on Llama-3-8B and Gemma-2-9B with minimal utility degradation.