🔬 NLA — read Qwen3.6-27B's mind, one token at a time
A natural-language autoencoder for Qwen3.6-27B layer-42 activations: click a token and the actor verbalizes its activation into a salience-ordered list of lines, while the critic reconstructs the vector from each line-prefix — the bars show how much of the vector (FVE, fraction of variance explained) the first k lines recover. Checkpoints — the matryoshka rl_av_lora_iter400 verbalizer (ordering-trained to front-load by salience), co-trained rl_critic_step400 reconstructor.
Matryoshka explanations are independent lines, so the critic can encode an edited subset without going out-of-distribution. Delete or rewrite lines below; the critic re-encodes your version, the resulting vector is patched into the model at the clicked token (norm-matched), and the text after it is regenerated — side-by-side with the unedited reconstruction (control) and the true activation.