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

arXiv: https://arxiv.org/abs/2603.29396 · HTML: https://mlsanity.com/paper/2603.29396/

## Key facts

- No citing papers in the map yet (data through 14 Oct 2026).
- Placed in Reasoning benchmarks and evaluation, one of 17 directions in LLM reasoning.
- Builds on 9 of its 33 references that are in the map.
- First posted on arXiv 31 Mar 2026.

## Placed in

- LLM reasoning / [Reasoning benchmarks and evaluation](https://mlsanity.com/lr/reasoning-benchmarks-and-evaluation/)

## Builds on (most cited first, 9 of 9)

- [Qwen2.5 Technical Report](https://mlsanity.com/paper/2412.15115/) (2024)
- [On the Opportunities and Risks of Foundation Models](https://mlsanity.com/paper/2108.07258/) (2021)
- [Mistral 7B](https://mlsanity.com/paper/2310.06825/) (2023)
- [Gemma 3 Technical Report](https://mlsanity.com/paper/2503.19786/) (2025)
- [Token-Level Adversarial Prompt Detection Based on Perplexity Measures and Contextual Information](https://mlsanity.com/paper/2311.11509/) (2023)
- [LLMs with Industrial Lens: Deciphering the Challenges and Prospects -- A Survey](https://mlsanity.com/paper/2402.14558/) (2024)
- [TokenButler: Token Importance is Predictable](https://mlsanity.com/paper/2503.07518/) (2025)
- [Perplexed: Understanding When Large Language Models are Confused](https://mlsanity.com/paper/2404.06634/) (2024)
- [Fixing confirmation bias in feature attribution methods via semantic match](https://mlsanity.com/paper/2307.00897/) (2023)

## Abstract (arXiv)

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.