LLMs are Frequency Pattern Learners in Natural Language Inference

2025-05-27

AI Summary

This work investigates the implicit mechanisms acquired by large language models (LLMs) during fine-tuning for natural language inference (NLI). We find that models do not develop robust deep semantic reasoning capabilities; instead, they heavily rely on predicate frequency disparities between premises and hypotheses—reducing NLI to superficial frequency-pattern recognition. Using predicate frequency statistics, bias consistency analysis, adversarial evaluation, and WordNet-based hyponymy modeling, we systematically demonstrate, for the first time, a strong statistical association between lexical frequency bias and textual entailment—and show this bias is markedly amplified post-fine-tuning, causing over 30% performance degradation on adversarial examples. Our contributions include: (1) exposing an implicit shortcut mechanism in NLI fine-tuning; (2) establishing a quantifiable link between lexical frequency and semantic hierarchy; and (3) proposing a novel diagnostic and mitigation paradigm for reasoning biases in LLMs.

Technology Category

• Natural Language Processing: Sentence-level Semantics, Textual Inference, etc.
• Machine Learning: Large Multimodal Models (LMMs)
• Cognitive Modeling & Cognitive Systems: Conceptual Inference and Reasoning

Application Category

• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs
• Graph Algorithms and Modeling for the Web: Foundation models and LLMs for Web-related graphs
• Search and Retrieval-Augmented AI: Web learning to rank, online learning, and counterfactual learning for ranking

Abstract

While fine-tuning LLMs on NLI corpora improves their inferential performance, the underlying mechanisms driving this improvement remain largely opaque. In this work, we conduct a series of experiments to investigate what LLMs actually learn during fine-tuning. We begin by analyzing predicate frequencies in premises and hypotheses across NLI datasets and identify a consistent frequency bias, where predicates in hypotheses occur more frequently than those in premises for positive instances. To assess the impact of this bias, we evaluate both standard and NLI fine-tuned LLMs on bias-consistent and bias-adversarial cases. We find that LLMs exploit frequency bias for inference and perform poorly on adversarial instances. Furthermore, fine-tuned LLMs exhibit significantly increased reliance on this bias, suggesting that they are learning these frequency patterns from datasets. Finally, we compute the frequencies of hyponyms and their corresponding hypernyms from WordNet, revealing a correlation between frequency bias and textual entailment. These findings help explain why learning frequency patterns can enhance model performance on inference tasks.

Problem

• LLMs exploit frequency bias in NLI datasets
• Fine-tuned LLMs rely heavily on frequency patterns
• Frequency bias correlates with textual entailment performance

Innovation

• Analyzing predicate frequencies in NLI datasets
• Evaluating LLMs on bias-consistent and adversarial cases
• Correlating frequency bias with textual entailment