A disconfounded probing benchmark, multi-model multi-dataset bootstrap, refusal-channel attribution, and sparse-autoencoder decomposition for small instruction-tuned LMs (Qwen3-1.7B, Gemma 2 2B, Llama ...
Hear a painting. See a song. SYN-1 reads artwork as a grid of light and color and plays it as music, then listens to music and paints it back. There are twenty translation methods, and all of them run ...
Some results have been hidden because they may be inaccessible to you
Show inaccessible results