Do standard NLP benchmarks hide LLM ambiguity failures?

Standard NLP benchmark curation assumes single gold-standard interpretations. When annotators disagree, the practice is to filter out the ambiguous examples — treating disagreement as annotation noise rather than evidence of genuine interpretive multiplicity.

The consequence is systematic: benchmarks cannot evaluate what they have excluded. LLM ambiguity failure — the inability to recognize that sentences have multiple valid interpretations and to disentangle them — is invisible in standard evaluation because the test items that would reveal it are removed before evaluation begins.

This is not a minor calibration issue. Ambiguity management is central to human language understanding. The ability to anticipate misunderstanding, ask clarifying questions, revise interpretations, and use context to select among readings is what distinguishes robust language comprehension from pattern matching. A benchmark that excludes all ambiguous instances evaluates only the easy cases.

The methodological insight from AMBIENT (Blevins et al. 2023): by specifically targeting and including ambiguous examples (with diverse ambiguity types and multiple valid interpretations per example), the evaluation reveals a 32% vs. 90% accuracy gap between GPT-4 and humans that standard benchmarks are blind to.

This connects to Can models pass tests while missing the actual grammar? — both identify evaluation designs that allow LLMs to succeed without demonstrating the underlying competence being measured. The surface pattern passes; the structural capability is absent.

The NLI domain provides direct evidence. "Lost in Inference" (Bittermann et al.) analyzes annotation disagreement patterns across NLI benchmarks and finds that performance is not saturated: the best models still fail to match human performance on contested cases, and human annotators continue to disagree in structured ways. The disagreement isn't noise — it reflects genuine interpretive multiplicity. Since standard benchmarks adjudicate this disagreement away before evaluation, models never have to confront the hard cases. The practical implication: progress on standard NLP benchmarks may systematically overestimate language understanding for the specific capability that most distinguishes human communication from pattern completion.