single neuron safety https://arxiv.org/pdf/2605.08513
task vectors: https://arxiv.org/pdf/2212.04089
qwen evil https://x.com/_lyraaaa_/status/2113833330429083662
interp and aligment https://www.goodfire.com/blog/we-can-and-must-solve-alignment
1D features, not in LLMs https://arxiv.org/abs/2405.14860
manifold steering good fire https://arxiv.org/abs/2605.05115
jacobian lens - ila fete https://arxiv.org/pdf/2610.31955
impossibility results in ai safety https://arxiv.org/pdf/2610.27373