Inducing Epistemological Humility in Large Language Models: A Targeted SFT Approach to Reducing Hallucination

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. Since these errors arise as a by-product of knowledge degradation, we explore whether established continual learning tools can mitigate them. 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 when new knowledge acquisition is unnecessary, suppressing factual plasticity by freezing parameter groups preserves task performance while reducing hallucinations. Lastly, we investigate the mechanism, contrasting capacity limitations, behavior cloning, and localized interference. Our experiments show that a main driver is interference among overlapping semantic representations, which self-distillation mitigates and an associative-memory model explains: forgetting grows with the overlap between new and stored facts.

Hallucinations in multimodal large language models (MLLMs), where a model generates content inconsistent with the input image, pose significant risks in real-world applications. Existing hallucination benchmarks primarily test recognition, i.e., whether a model can select the correct answer among distractors. They overlook another capability that is essential for trustworthy AI: recognizing when none of the provided options is supported by the image and declining to commit to a false choice. We present HumbleBench, a hallucination benchmark that evaluates this false-option rejection behavior, a directly observable facet of epistemic humility, through multiple-choice questions that always include a "None of the above" (NOTA) option. HumbleBench is built from the Panoptic Scene Graph dataset: its object and relation annotations, together with candidate attribute cues, are used to prompt GPT-4-Turbo to generate questions, which are then manually verified, yielding 22,831 questions on object, relation, and attribute hallucinations. We evaluate 27 MLLMs, including general-purpose, reasoning-oriented, and proprietary models, and complement standard accuracy with NOTA-aware metrics, diagnostic stress tests, and a human evaluation. Even the strongest models select NOTA for only about half of the questions in which it is correct, cautious prompting shifts models toward NOTA without improving their ability to detect unsupported options, and humans reject unsupported options considerably more reliably than all evaluated models. These results show that recognition accuracy alone is insufficient for assessing hallucination robustness. Our code and dataset are publicly available at https://github.com/maifoundations/HumbleBench.

Large language models (LLMs) have seen widespread adoption across various domains, yet their reliability is frequently undermined by hallucinations - responses that are plausible-sounding but factually incorrect. In high-stakes domains, these errors can reduce trust and introduce real-world risk. To address this challenge, we present a parameter-efficient approach that uses soft prompts to mitigate hallucinated content and promote responsible abstention in generative question-answering (QA) tasks. Our method, called Responsible Contrastive Soft Prompting (RCSP), uses a composite loss to train soft prompts that balance three goals: suppressing hallucinatory content, encouraging abstention under uncertainty, and preserving or improving factual recall. To achieve these goals, we incorporate contrastive loss, curriculum learning, and KL regularization into our training mechanism. We evaluate our approach on five diverse generative QA datasets using an LLM-as-a-Judge framework. Experimental results on the Gemma 3 (12B) and Llama 3.1 (8B) backbones demonstrate that RCSP effectively balances factual recall with hallucination suppression and abstention, yielding a generally superior F-score over standard reasoning and instruction-based prompting baselines. Notably, these improvements are achieved by training only a fraction of the parameters required by other tuning techniques. Our results demonstrate that soft prompts provide a modular and computationally efficient path toward improving LLM reliability.