[Paper Review] Are Larger Pretrained Language Models Uniformly Better? Comparing Performance at the Instance Level

Ruiqi Zhong, Dhruba Ghosh+2 more|arXiv (Cornell University)|May 13, 2021

Topic Modeling27 references4 citations

TL;DR

This paper investigates whether larger pretrained language models are uniformly better than smaller ones at the instance level, despite overall accuracy gains. Using statistically rigorous methods to account for pretraining and finetuning noise, the authors find that BERT-large performs worse than BERT-mini on at least 1–4% of instances across MNLI, SST-2, and QQP, challenging the assumption of uniform improvement with scale.

ABSTRACT

Larger language models have higher accuracy on average, but are they better on every single instance (datapoint)? Some work suggests larger models have higher out-of-distribution robustness, while other work suggests they have lower accuracy on rare subgroups. To understand these differences, we investigate these models at the level of individual instances. However, one major challenge is that individual predictions are highly sensitive to noise in the randomness in training. We develop statistically rigorous methods to address this, and after accounting for pretraining and finetuning noise, we find that our BERT-Large is worse than BERT-Mini on at least 1-4% of instances across MNLI, SST-2, and QQP, compared to the overall accuracy improvement of 2-10%. We also find that finetuning noise increases with model size and that instance-level accuracy has momentum: improvement from BERT-Mini to BERT-Medium correlates with improvement from BERT-Medium to BERT-Large. Our findings suggest that instance-level predictions provide a rich source of information; we therefore, recommend that researchers supplement model weights with model predictions.

Motivation & Objective

• To determine whether larger pretrained language models uniformly outperform smaller ones at the instance level, beyond aggregate accuracy.
• To address the challenge of prediction noise from random seeds in pretraining and finetuning, which obscures true model performance differences.
• To develop a statistically robust method for estimating instance-level accuracy that accounts for stochasticity in training.
• To investigate whether performance trends at the instance level exhibit momentum or variance patterns across model sizes.
• To recommend that researchers share model predictions alongside weights to enable deeper analysis of model behavior.

Proposed method

• Pretrain 10 BERT models of each size (mini, base, large) with different random seeds to account for pretraining noise.
• Fine-tune each pretrained model 5 times with different random seeds to account for finetuning noise, generating a 4D tensor of predictions (model size, instance, pretraining seed, finetuning seed).
• Estimate instance-level accuracy as the expected proportion of correct predictions across all random seeds, using Monte Carlo averaging.
• Develop a random baseline to estimate the false discovery rate in identifying instances where smaller models outperform larger ones.
• Apply an improved statistical procedure—superior to Benjamini-Hochberg with Fisher’s exact test—to upper-bound false discoveries in instance-level comparisons.
• Analyze variance components to attribute prediction variability to pretraining vs. finetuning seeds, and examine momentum in performance gains across model sizes.

Experimental results

Research questions

• RQ1 Are larger pretrained language models uniformly better than smaller ones at the instance level, or do they underperform on specific inputs?
• RQ2 How much of the observed performance difference between model sizes is due to random noise in pretraining and finetuning, rather than true model capability?
• RQ3 Does performance improvement from smaller to larger models exhibit momentum at the instance level, such that models that improve from mini to medium also tend to improve from medium to large?
• RQ4 How does the variance in model predictions scale with model size, and which source—pretraining or finetuning—contributes more to this variance?
• RQ5 What types of instances (e.g., rare, controversial, or ambiguous) are more likely to be misclassified by larger models?

Key findings

• BERT-large performs worse than BERT-mini on at least 1–4% of instances across MNLI, SST-2, and QQP, despite a 2–10% overall accuracy improvement.
• The proportion of instances where larger models underperform is significantly higher than expected by chance, as confirmed by a rigorous random baseline and statistical correction.
• Instance-level accuracy gains show momentum: models that improve from BERT-mini to BERT-medium are more likely to also improve from BERT-medium to BERT-large.
• Finetuning noise increases with model size, contributing more to prediction variance than pretraining noise, especially in larger models.
• The set of instances where larger models underperform includes both controversial or ambiguous labels and correctly labeled examples, indicating no simple pattern in failure cases.
• The study recommends that researchers release model predictions alongside weights to enable fine-grained analysis of model behavior at the instance level.