A new arXiv paper investigates the linear directions in a large language model's activation space that prior work associates with statement truth. It questions how universal or generalizable these truth directions are, building on earlier claims about their consistency across contexts. The work falls within ongoing research on interpreting and steering model internals.
A new arXiv paper examines safety alignment in language models, focusing on models fine-tuned to emit distinct refusal tokens that signal different categories of refusal before they answer. The authors investigate refusal directions tied to specific categories and how those directions might be discovered and steered. The abstract provided is truncated, so the full method and results are not available here.
Researchers describe a phenomenon they call Harmfulness Propagation Dynamics, in which the last-token hidden state of a harmful prompt projects increasingly onto a learned harm direction as depth increases. Benign prompts did not show this pattern, instead staying flat or fluctuating across layers. The finding points to layer-wise differences that could inform how models are monitored or steered for safety.