Reproduce “Reducing LLM Hallucinations using Epistemic Neural Networks” on OpenResearch

About this paper

Shreyas Verma, Kien Tran, Yusuf Ali, Guangyu Min

Integrating Epistemic Neural Networks (ENNs) with a Llama-2 7B model enhanced by DoLa's contrastive decoding was explored to reduce LLM hallucinations through uncertainty estimation for next token prediction. Although ENNs were successfully trained for this task, their direct application resulted in a decrease in factual accuracy on the TruthfulQA benchmark compared to the DoLa baseline, attributed to overfitting on the constrained training dataset.

Key insights

• Epistemic Neural Networks can be successfully trained for the next token prediction task, demonstrating their foundational applicability in a generative context previously unexplored.
• Architectural choices and hyperparameter tuning for ENNs, such as using a 2-layered MLP with 1024 neurons and 50 epistemic samples, are critical for achieving training efficiency and convergence.
• The generalization performance of the ENN-integrated system is highly sensitive to the scale and diversity of its training data, with observed overfitting on limited datasets leading to reduced factual accuracy on evaluation benchmarks.

Summary by alphaXiv.