Is my model perplexed for the right reason? Contrasting LLMs' Benchmark Behavior with Token-Level Perplexity
This addresses the interpretability gap in LLM evaluations for researchers, though it is incremental as it builds on existing perplexity and benchmark methods.
The authors tackled the problem of understanding whether LLMs' correct benchmark performance stems from appropriate linguistic reasoning or superficial heuristics, by introducing a token-level perplexity framework to analyze minimal sentence pairs; they found that linguistically important tokens influence behavior but never fully explain perplexity shifts, indicating reliance on other heuristics.
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.