Beyond Accuracy Optimization: Computer Vision Losses for Large Language Model Fine-Tuning

Anonymous

Abstract: Large Language Models (LLMs) have achieved promising performance on Math Word Problem (MWP) and Question Answering (QA) tasks. LLM fine-tuning is commonly based on cross-entropy loss minimization to perform accurate predictions. However, the standard cross-entropy function neither considers the underlying token distribution over training data nor weighs differently correct and misclassified samples. To address tasks such as closed-ended QA and step-by-step MWP resolution LLMs require advanced language reasoning capabilities. This prompts the adoption of established computer vision loss functions that optimize LLMs' performance rather than simple accuracy. This paper shows the higher effectiveness of combining cross-entropy with computer vision loss functions across MWPs and closed-ended QA datasets. We show relevant LLMs' performance improvements with equal model complexity and the same number of training samples or even fewer. We also demonstrate the efficacy of reproducing step-by-step reasoning on the MWP task.

Paper Type: long

Research Area: Question Answering

Contribution Types: NLP engineering experiment

Languages Studied: English