New SFT method reduces language model hallucinations

Researchers have developed a method called knowledge-aligned SFT to reduce factual hallucinations in language models. This approach constrains training targets to the model's existing parametric knowledge, contrasting with standard SFT which may use external knowledge. New variants, Evidence Rewrite and Recall Rewrite, were introduced and tested on models like Qwen 3 4B and OLMo 3 7B. Results indicate that knowledge-aligned SFT can improve factuality and preserve general capabilities, with Recall Rewrite showing the most significant gains in factuality and refusal behavior. AI

IMPACT This research could lead to more factually accurate language models by improving the fine-tuning process.

RANK_REASON The cluster contains a research paper detailing a new method for supervised fine-tuning of language models. [lever_c_demoted from research: ic=1 ai=1.0]

Read on arXiv cs.AI →

• Biography
• Evidence Rewrite
• knowledge-aligned SFT
• language model
• OLMo 3 7B
• Qwen 3 4B
• Recall Rewrite
• Stick to What You Know: A Study of Knowledge-Aligned Supervised Fine-Tuning
• supervised fine-tuning
• UnknownBench
• WildHalu

• paper
• model release

AI-generated summary · Google Gemini · from 1 sources. How we write summaries →