Epistemic Reasoning

A 2,000-item BIG-bench NLI task that asks whether a nested knowledge or belief premise entails a hypothesis, scored as two-way multiple choice.

Also known as: BIG-bench epistemic_reasoning

unassessed

This page is a discovery lead. Nobody has yet assessed it against the catalogue contract, so it carries no disposition. Absence of evidence here is not evidence of staleness.
Categoryreasoning
SubcategoryBIG-bench nested knowledge/belief NLI (2,000 items, ten templates)
Page statusunknown
Metricmultiple_choice_grade
Directionhigher_is_better
Unit%
Dataset size2000
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

epistemic_reasoning is an English natural-language-inference probe of epistemic theory of mind. Each item pairs a premise and a hypothesis built around stacked factive verbs (knows, sees, learns, understands, recognizes, remembers) and non-factive verbs (believes, thinks, assumes, suspects). The intended skill is whether the model tracks other agents' mental states, not speaker-commitment datasets that stress negation and modals. It is not [bbh](bbh.md).

Task format

Two-option multiple choice (entailment vs non-entailment), preferred metric multiple_choice_grade. task_prefix asks the model to identify the relation between premises and hypotheses. append_choices_to_input is false. Canary GUID embedded.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub