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 ...

Faced with a question it cannot answer, a small language model either refuses or confabulates: it confidently produces a wrong answer. Can the hidden state before the first generated token tell which ...