Hallucination Detox: Sensitivity Dropout (SenD) for Large Language Model Training

TL;DR

This paper introduces Sensitivity Dropout (SenD), a new training method for large language models that reduces hallucination variance and improves factual accuracy without harming task performance.

Contribution

We propose SenD, a novel training protocol that deterministically drops high-variance embeddings, and EES, an efficient hallucination detection metric, enhancing LLM reliability.

Findings

SenD reduces hallucination variance during training.

EES provides fast, unsupervised hallucination detection.

SenD improves factual accuracy by up to 17% in tested models.

Abstract

As large language models (LLMs) become increasingly prevalent, concerns about their reliability, particularly due to hallucinations - factually inaccurate or irrelevant outputs - have grown. Our research investigates the relationship between the uncertainty in training dynamics and the emergence of hallucinations. Using models from the Pythia suite and several hallucination detection metrics, we analyze hallucination trends and identify significant variance during training. To address this, we propose Sensitivity Dropout (SenD), a novel training protocol designed to reduce hallucination variance during training by deterministically dropping embedding indices with significant variability. In addition, we develop an unsupervised hallucination detection metric, Efficient EigenScore (EES), which approximates the traditional EigenScore in 2x speed. This metric is integrated into our training protocol, allowing SenD to be both computationally scalable and effective at reducing hallucination variance. SenD improves test-time reliability of Pythia and Meta's Llama models by up to 17% and enhances factual accuracy in Wikipedia, Medical, Legal, and Coding domains without affecting downstream task performance.