Dynamic Counting

A 250-trial BIG-bench task that asks which of four closing brackets finishes a Shuffle-4 prefix, testing counter-style counting rather than full Dyck hierarchy.

Also known as: Shuffle-n last parenthesis, BIG-bench dynamic_counting

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 Shuffle-4 last-closer multiple choice (250 trials)
Page statusunknown
Metricmultiple_choice_grade
Directionhigher_is_better
Unit%
Dataset size250
Dataset licenceApache-2.0
PublisherGoogle (BIG-bench collaboration)

What it measures

dynamic_counting generates a Shuffle-n string (the shuffle of n Dyck-1 languages over distinct bracket pairs), drops the last symbol, and asks the model which closer belongs there. Default n is 4 over (), [], <>, {}. Authors Mirac Suzgun, Sebastian Gehrmann, Yonatan Belinkov and Stuart M. Shieber argue that the last closer is the output of a real-time k-counter, not a stack that stores nested structure. The skill is counting opens and closes, not recognising a full Dyck language. It is not [dyck_languages](dyck_languages.md) (JSON Dyck-4 suffix) and not HELM [dyck_language](dyck_language.md).

Task format

Four-option multiple choice via cond_log_prob over the four closers. Preferred metric multiple_choice_grade. Zero-shot. Sequences drawn at evaluation time (seed 42, p=0.4, q=0.4, lengths 4-100, 250 unique trials). Canary GUID embedded.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub