Alignbeam
1 mentions across 0 people
All mentions
Unknown speaker
Recommendedpaper · 2026-06-11
“We present ALIGNBEAM, a training-free method that lifts this restriction by translating anchor logits into the target model's vocabulary token-by-token at each decoding step”
Inference-Time Safety Transfer Across LLM Families via Cross-Vocabulary Logit Tr ↗