What does this research mean for the field?

Perplexity measures model fluency rather than task-distribution membership, failing to detect out-of-distribution inputs effectively. Novelty: ClaimNovelty.CONTRADICTORY. Consensus alignment: ConsensusAlignment.CHALLENGES_CONSENSUS.

What question did this study set out to answer?

This research aims to investigate how perplexity fails to accurately measure out-of-distribution detection in language models.

March 17, 2026Open Access

View Full Paper

The Geometric Blind Spot of Perplexity: When Low Loss Hides Out-of-Distribution

JMJesus Tabares Montilla

Key Points

This research aims to investigate how perplexity fails to accurately measure out-of-distribution detection in language models.
Evaluated LLaMA-3-8B and Mistral-7B models on a math-reasoning task.
Tested the models against five out-of-distribution domains.
Compared perplexity to intrinsic dimensionality and Mahalanobis distance.
Perplexity reduced to lower values for code snippets compared to in-distribution tasks.
Hidden-state intrinsic dimensionality reduced from 34 to 4, showing an 8.5× decrease.
Perplexity achieved AUROC scores of 0.352 for LLaMA and 0.150 for Mistral, indicating worse performance than random.
Intrinsic dimensionality and Mahalanobis distance achieved AUROC scores of 1.000 across all out-of-distribution categories.

Abstract

Perplexity is widely used as a proxy for out-of-distribution (ood) detection in largelanguage models, under the assumption that unfamiliar inputs yield higher loss. We show this assumption has a structural blind spot: domains where the model is fluent but the input does not belong to the task distribution. Concretely, we evaluate LLaMA-3-8B and Mistral-7B on a math-reasoning task with five ood domains. Code snippets produce perplexitylower than in-distribution math (2.57 vs. 2.76 for LLaMA), yet their hidden-state intrinsic dimensionality collapses from 34 to 4—an 8.5× reduction. Perplexity achieves auroc 0.352 (LLaMA) and 0.150 (Mistral) on code, worse than random. In contrast, intrinsic dimensionality and Mahalanobis distance achieve auroc 1.000 across all ood categories in both models. The dissociation is consistent across five transformer layers and robust to sample-size equalization. Our results demonstrate that perplexity measures model fluency, not task-distribution membership, and that geometric

AI에게 질문

Bookmark

View Full Paper

AI에게 질문

Bookmark

View Full Paper

The Geometric Blind Spot of Perplexity: When Low Loss Hides Out-of-Distribution

Key Points

Abstract

Cite This Study