The work paired a human-written math reasoning dataset with a generator-verifier pipeline. A language model generated many candidate solutions, while a separately trained verifier scored those solutions and selected the most promising answer.
OpenAI introduced GSM8K, a dataset of roughly 8,500 grade-school mathematics word problems, and showed that a learned verifier selecting among multiple generated solutions substantially improved large-model accuracy over finetuning alone.
GSM8K became one of the most reused small-scale mathematical reasoning benchmarks, and its verifier-based inference pattern anticipated later outcome- and process-verification methods used throughout LLM reasoning research.
Significance rationale: GSM8K became a durable benchmark for multi-step language-model mathematical reasoning, while the accompanying learned-verifier method established an influential generate-and-rank pattern for improving solution accuracy. Both the dataset and verification idea remained central reference points as reasoning models evolved.
—
The arXiv research record introduced GSM8K and the verifier-based solution-selection method.
The canonical record links the archived official OpenAI dataset repository, making the benchmark data and example evaluation code directly inspectable. E3 does not imply independent verification of every reported model result.
No corrections recorded.
