Trick or Neat: Adversarial Ambiguity and Language Model Evaluation

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 randombased 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:coastalcph/lm_ambiguity.1 Except, of course, if the dress has a mirror.Such exceptions illustrate how pragmatics may override semantics (Morris, 1946), but LLMs must be sensitive to distinctions between syntactic ambiguities that support multiple conventional readings and those typically resolved by context. 2 spaces/lmsys/chatbot-arena-leaderboard 3 jndiogo/LLM-chat-templates 4 We define consistency here as the ability to make reliable decisions in semantically similar contexts demonstrating a systematic capacity to generalize across language variations" (Elazar et al., 2021).saahowwwnem ttwahimitonteehmaww w tntaihotnemh i wttttah h maehinntllehtew w s sttceecitihtolhte e heep stttelcheeh.oelseepcto steecp.loes.lepsec.ope.cope.sub sju_bsysonym | t suj_nbjyn_raonndyomThm | eJe mtneoatcheer ssa omhea ntee wle tsteelec ihcloess t tpheco.oppee