Topic
NF4
Latest news
- 16 OctPRISM framework studies transfer of activation interventions in tool-calling LMs
Researchers introduce PRISM, a framework to evaluate how activation interventions transfer across precision, languages, and model scales. They show unsafe tool calls can accompany textual refusals and find that lower-resource settings (quantized NF4 labels and cross-lingual transfer) can expose and improve alignment, with NF4 behavioural labels enabling stable intervention.
Research · 1 source
Articles
No articles about NF4 yet.