Does fine-tuning on NLI teach inference or amplify shortcuts?

"LLMs are Frequency Pattern Learners in NLI" identifies a consistent frequency bias in NLI datasets: predicates in hypotheses are more frequent in training data than predicates in premises, for positive (entailment) instances. LLMs exploit this pattern. The disturbing finding: fine-tuning on NLI corpora increases reliance on frequency bias rather than decreasing it.

The mechanism connects to a real property of language. Hypernyms (more general terms: "animal") are more frequent than hyponyms (more specific terms: "dog") in natural text. Since upward entailment works from specific to general (SPRINT entails RUN), frequency can be a useful proxy for entailment direction. Fine-tuning teaches models to exploit this proxy more aggressively.

The problem: frequency is a statistical artifact, not a semantic relationship. It works often enough to appear as learning on standard benchmarks but fails on adversarial cases where the frequency pattern disagrees with the actual entailment label. After fine-tuning, LLMs perform significantly worse on adversarial instances than base models — they have learned the shortcut more deeply.

This is a general pattern in the vault: Can models pass tests while missing the actual grammar? shows that surface heuristics enable correct behavior on easy cases while degrading robustness on unusual ones. Fine-tuning amplifies this problem by rewarding the heuristic through the training signal. The model that appears to "learn inference" has learned to use training data statistics more efficiently.

What distinguishes this from the attestation bias (memorization of specific sentences): frequency bias operates at the corpus level — it is a statistical regularity learned from the distribution of natural text, not from specific memorized statements. Both are shortcuts that substitute for inference, but they originate from different levels of the training data.

Source: Natural Language Inference