Hugging Face Trending Papers

ALIGNBEAM : Inference-Time Alignment Transfer via Cross-Vocabulary Logit Mixing

Read the original on Hugging Face Trending Papers →

Domain fine-tuning degrades the safety of large language models: fine-tuned specialists readily comply with harmful prompts framed in domain language. Existing inference-time defenses that mix logits from a safe anchor model require both models to share a vocabulary, which rules them out for the cross-family specialists where safety is most degraded.

Summary generated by The Flow from the publisher's feed. The full article lives at Hugging Face Trending Papers.