Search papers, labs, and topics across Lattice.
The paper investigates why supervised fine-tuning (SFT) increases factual hallucinations in LLMs, even regarding knowledge learned during pre-training. They find that SFT-induced hallucinations are primarily driven by interference among overlapping semantic representations. To mitigate this, they propose a self-distillation-based SFT method that regularizes output-distribution drift, and also show that freezing parameter groups can reduce hallucinations when new knowledge acquisition is unnecessary.
Fine-tuning LLMs can make them *more* likely to hallucinate facts they already knew, but self-distillation offers a surprisingly simple fix.
Large language models are prone to hallucinating factually incorrect statements. A key source of these errors is exposure to new factual information through supervised fine-tuning (SFT), which can increase hallucinations w.r.t. knowledge acquired during pre-training. In this work, we explore whether SFT-induced hallucinations can be mitigated using established tools from the continual learning literature, since they arise as a by-product of knowledge degradation during training. We propose a self-distillation-based SFT method that facilitates effective factual learning while minimizing hallucinations w.r.t. pre-existing knowledge by regularizing output-distribution drift. We also show that, in settings where new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups, can preserve task performance while reducing hallucinations. Lastly, we investigate the mechanism behind SFT-induced hallucinations through three hypotheses: capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, and that self-distillation succeeds by mitigating this interference.
