Logical Deduction

1,500 BIG-bench questions that ask which object sits at a given position after a minimal set of ordering clues; also three 250-item BBH tasks.

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 (and BBH) object-ordering deduction from a minimal clue set
Page statusactive
Metricmultiple_choice_grade
Directionhigher_is_better
Unit%
Dataset size1500
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

logical_deduction describes three, five or seven similar objects in a natural order (books on a shelf, golfers in a ranking, fruit by price) and gives a minimal set of clues such as "X is left of Y" or "X is third." The model must assign higher probability to the true statement about which object occupies a queried position than to the false statements for the other objects. Authors James Simon and Chandan Singh generated the puzzles so each has exactly one solution and no spare clue. The task is meant to need multi-step deduction rather than pattern match.

Task format

Multiple-choice over N statements for an N-object puzzle, scored with `multiple_choice_grade`. BIG-bench ships three subtasks (three_objects, five_objects, seven_objects). A canary GUID is embedded. BIG-bench Hard keeps three separate 250-example files under the same names.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub