Mitigating Intrinsic Hallucinations Using Orthogonal Fine-Tuning and Dirichlet Prior Calibration

Le résumé fourni par la source

Many large language models (LLMs) exhibit both intrinsic hallucination through internal representation drift and an unjustified level of confidence in their token predictions, even when inputs are clean and unambiguous. This paper is concerned with how to mitigate the issues related to intrinsic hallucination that occur within the Llama-3.1-8B models. Additionally, there is a focus on creating a fine-tuning approach that will help mitigate such hallucinatory behaviour through parameters that do not require a great computational resource expense for fine-tuning the entire model. Therefore, the goal is to create a fine-tuning technique to help suppress hallucinations, while not degrading any of the factual information that has been learned during the pretraining of the model. This objective will be met at the same time, by developing a way to generate factual representations and develop an awareness of uncertainty in the generative process. A hybrid approach will be proposed that is derived from the combination of Orthogonal Fine-Tuning (OFT) for maintaining knowledge subspaces and Dirichlet Prior calibration for smoothing and calibrating the output token probabilities. The amount of parameters being trained will be limited to those via Parameterefficient Fine-Tuning (PEFT) adapters on a single Graphics Processing Unit (GPU), which is only 1 % to 2 % of the total parameters. Therefore, by developing this new approach, it is possible to address both the issue of deteriorating representation drift and overconfidence due to probabilistic uncertainty within a single framework, via methodologically sound and resourceefficient means. Experimental results from the HaluEval and TruthfulQA benchmarks have shown that this approach reduces the rate of hallucination, creates more accurate predictions, and improves the accuracy of truthful answers, relative to fine-tuning methods that were used in the baseline experiment.

Ce résumé expose les affirmations des auteurs. BNTIC ne l’interprète pas comme une validation indépendante des résultats.