Trick or Neat: Adversarial Ambiguity and Language Model Evaluation
Abstract
Detecting ambiguity is important for language understanding, including uncertainty estimation, humour detection, and processing garden path sentences. We assess language models’ sensitivity to ambiguity by introducing an adversarial ambiguity dataset that includes syntactic, lexical, and phonological ambiguities along with adversarial variations (e.g., word-order changes, synonym replacements, and random-based alterations). Our findings show that direct prompting fails to robustly identify ambiguity, while linear probes trained on model representations can decode ambiguity with high accuracy, sometimes exceeding 90%. Our results offer insights into the prompting paradigm and how language models encode ambiguity at different layers. We release both our code and data:
Study specs
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.
- Authors
- A Karamolegkou,O Eberle,P Rust,C Kauf,A Søgaard
- Institution
- Aleph Alpha,Massachusetts Institute of Technology
- Discipline
- Natural Language Processing
- Study Type
- Evaluation Study
- Year
- 2025
- Human Data Platform
- Prolific
- Source
- View Source Google Scholar
Measured Outcomes
Language models' ability to detect ambiguity, including syntactic, lexical, and phonological types, as well as performance under adversarial variations.
Peer Review & Critical Discussion
Potential Selection Bias in 2023 Cohort
The participant pool shows a concerning overrepresentation of users from high-income demographics. Looking at Table 3, we can see that 78% of respondents had annual incomes above $75k, which significantly limits the generalizability of these findings to broader populations.
Non-naive Participants Issue
I've noticed a methodological concern regarding participant naivety. Given that Prolific users often complete multiple studies, there's a real risk that participants had prior exposure to similar experimental paradigms, which could confound the results.
RLHF Applicability to This Study Design
The implications for RLHF training pipelines are understated. If we accept the authors' conclusions about preference stability, this has direct consequences for how we should structure reward model training. The temporal decay effect described in Section 4.2 is particularly relevant.
Verify your expertise to join discussion
Create an account and verify your credentials to participate in peer discussions.