According to a paper accepted to EMNLP 2026 Findings on arxiv.org, researchers have developed a knowledge-weighted fine-tuning approach to help large language models (LLMs) better express uncertainty. The method addresses hallucinations that “often arise from knowledge misalignment between pre-training and fine-tuning,” according to the abstract.
The approach estimates “a fine-grained, instance-level knowledge score via multi-sampled inference” and scales learning signals based on the model’s existing knowledge while encouraging explicit “I don’t know” responses for out-of-scope queries. According to the researchers, experimental results show this allows models to “explicitly express uncertainty when it lacks knowledge, while maintaining accuracy on questions it can answer.”
Separately, research accepted to CVPR 2026 Findings introduces HiViS (Hiding Visual Tokens from the Drafter for Speculative Decoding in Vision-Language Models), which addresses computational challenges in vision-language models. According to arxiv.org, the framework is “motivated by the observation” that “visual tokens in large VLMs are highly redundant, and most of them can be removed without compromising generation quality.”
HiViS uses the target VLM as a semantic fusion model, allowing the drafter to obtain visual information without explicitly processing visual tokens. According to the abstract, “extensive experiments across representative VLMs and benchmarks demonstrate that HiViS achieves significant improvements in average acceptance length and speedup ratio.”