Masked language models (e.g., UL2, T5, PaLM2) potentially learn inconsistent distributions of tokens. Inference-time ensembling can improve their accuracies.
Figure 1: We prompt the model for different versions of the same distribution to measure their inconsistency and ensemble them to improve accuracy.
Our distribution of interest is P(?|Tom bought 4 apples and ate 2. He now has). ? can be 2.
Figure 2: Ensembling different versions of the same distribution improves accuracy. Aggregated results from 2 models on 3 datasets.
Learning to predict masked tokens in a sequence has been shown to be a helpful pretraining objective for powerful language models such as PaLM2. After training, such masked language models (MLMs) can provide distributions of tokens in the masked positions in a sequence. However, this paper shows that distributions corresponding to different masking patterns can demonstrate considerable inconsistencies, i.e., they cannot be derived from a coherent joint distribution when considered together.
This fundamental flaw in MLMs can lead to self-contradictory behaviors during inference. On various benchmark datasets including MMLU, MLMs can give different predictions to the same input question. From BERT-base to UL2-20B, we show that such inconsistencies exist ubiquitously in MLMs of diverse sizes and configurations. In light of our observations, we further propose an inference-time strategy for MLMs called Ensemble of Conditionals. It jointly considers a selected range of inconsistent conditionals directly produced by the MLM for the final prediction, which often leads to considerable accuracy improvement.
Minimum system requirement is to have at least 60GB aggregated memory across your GPUs.
To set up environment:
conda create --name inconsistencies python=3.11.5
conda activate inconsistencies
pip install -r requirements.txt
To run experiments on MMLU and BigBench with UL2-20B and T5-11B, use
main.ipynb
@article{young2022inconsistencies,
title={On the Inconsistencies of Conditionals Learned by Masked Language Models},
author={Young, Tom and Chen, Yunan and You, Yang},
journal={arXiv preprint arXiv:2301.00068},
year={2023}
}
We thank Fuzhao Xue and Shenggui Li for the valuable discussions.