Inference from Absence: When LLM Annotators Mistake Silence for Evidence in Relational Judgments

Under review. A controlled study showing that LLM annotators systematically over-assert a comparative asymmetry between two parties when a text leaves the second party's status unstated — treating silence as evidence rather than abstaining.

Authors: Iman YeckehZaare · Venue/status: ACM Collective Intelligence

Using minimal-pair stimuli whose correct labels are fixed by construction, this submitted manuscript isolates a reproducible 'inference from absence' failure: across many model configurations, most annotators over-assert asymmetry on the large majority of status-unstated items where the conservative answer is 'unclear.' The paper characterizes when the bias appears and what it implies for hybrid human-LLM annotation pipelines that treat model labels as evidence.

Public artifact boundary: this route exposes status, authorship, visual summary, citation metadata, and the contribution boundary; manuscript files are posted only when review and prepublication rules allow it.

The rendered manuscript page adds status, visual summary, review boundary, citation metadata, and contribution notes. Key links: home, systems, papers, manuscripts, Google Scholar, GitHub, LinkedIn, ORCID, MIT profile, and CV PDF.