cs.CL

PRISM: Probability Reallocation with In-Span Masking for Knowledge-Sensitive Alignment

arXiv:2604.01682v1 Announce Type: new
Abstract: Supervised fine-tuning (SFT) with token-level hard labels can amplify overconfident imitation of factually unsupported targets, causing hallucinations that propagate in multi-sentence generation. We stud…