PHYSICS (Benchmarking Foundation Models on University-Level Physics Problem Solving)

1,297 PhD-qualifying-exam physics problems across six subfields, needing multi-step derivation; the best model in the 2025 paper solved only 59.9% of the test set.

Also known as: PHYSICS Benchmark

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
SubcategoryPhD-qualifying-exam-level physics problem solving across six subfields
Page statusactive
Metricaccuracy (SymPy symbolic equivalence plus LLM-judged answer validation)
Directionhigher_is_better
Unit%
Dataset size1297
Dataset licenceMIT, stated explicitly in the repository's README and confirmed via the GitHub API
PublisherYale University and New York University

What it measures

This page documents the "physics" hint that resolves to OpenCompass's `PHYSICS` dataset: Feng et al.'s 2025 benchmark of graduate, PhD-qualifying-exam-level physics problems. The name collides with an unrelated BIG-bench task also called "physics" -- a much smaller, high-school-level, multiple-choice task asking a model only to identify the correct formula for a word problem, authored by two individual contributors as part of BIG-bench's 2021-2022 crowdsourced task collection. This page is about the former; see Lineage for the latter. Feng et al.'s PHYSICS requires solving expert-annotated problems across six subfields -- classical mechanics, quantum mechanics, thermodynamics and statistical mechanics, electromagnetism, atomic physics, and optics -- each demanding multi-step mathematical derivation and reasoning to a final answer, not formula selection or fact recall. About 23% of problems include an accompanying figure.

Task format

Open-ended, free-response problem solving: the model works step by step and gives its final answer in LaTeX boxed format. Some problems bundle several sub-questions that must each be answered in order. Scoring combines SymPy-based symbolic equivalence checking with an LLM-based (originally GPT-4o) natural-language answer validator, since physics answers can be correct while differing in algebraic form.

Models reporting this benchmark

No model card in ModelSpec reports this benchmark yet.

Data

This page as JSON · Edit on GitHub