Is my model perplexed for the right reason? Contrasting LLMs' Benchmark Behavior with Token-Level Perplexity

TL;DR

This paper presents a new interpretability framework using token-level perplexity to analyze whether LLMs rely on linguistically relevant cues, revealing they often depend on heuristics beyond expected linguistic features.

Contribution

The authors introduce a simple, hypothesis-driven method based on token-level perplexity to interpret LLM behavior without unstable feature attribution techniques.

Findings

Linguistically important tokens influence model perplexity but do not fully explain shifts.

Models rely on heuristics other than the expected linguistic cues.

The method enables precise analysis of model reliance on specific tokens.

Abstract

Standard evaluations of Large language models (LLMs) focus on task performance, offering limited insight into whether correct behavior reflects appropriate underlying mechanisms and risking confirmation bias. We introduce a simple, principled interpretability framework based on token-level perplexity to test whether models rely on linguistically relevant cues. By comparing perplexity distributions over minimal sentence pairs differing in one or a few `pivotal' tokens, our method enables precise, hypothesis-driven analysis without relying on unstable feature-attribution techniques. Experiments on controlled linguistic benchmarks with several open-weight LLMs show that, while linguistically important tokens influence model behavior, they never fully explain perplexity shifts, revealing that models rely on heuristics other than the expected linguistic ones.