Trick or Neat: Adversarial Ambiguity and Language Model Evaluation

Authors: A Karamolegkou, O Eberle, P Rust, C Kauf, A Søgaard

Published: 2025

Publication: ArXiv

The paper assesses language models' sensitivity to ambiguity using an adversarial dataset and finds that direct prompting poorly identifies ambiguity, while linear probes achieve high accuracy in decoding ambiguity from model representations.

Methods: An adversarial ambiguity dataset was introduced with various types of ambiguities and transformations; models were tested using direct prompts and linear probes trained on internal representations.

Key Findings: Language models' ability to detect ambiguity, including syntactic, lexical, and phonological types, as well as performance under adversarial variations.

Institution: Aleph Alpha, Massachusetts Institute of Technology

Research Area: Adversarial Ambiguity, Language Model Evaluation, Artificial intelligence, Computation and Language, LLM, AI Evaluation, Red Teaming

Discipline: Natural Language Processing

Citations: 2