Physics-Bench extends our bottom-up knowledge-graph reasoning evaluation to physics. The full suite contains 20,000 validated multiple-choice problems — a fully balanced grid of 5 hop depths × 20 topical clusters × 200 questions each. Every item is grounded in an explicit knowledge path (a chain of physics equations that composes the derivation from givens to the numerical answer) and ships with a step-by-step thinking trace that walks through the full derivation.
To keep each visit fresh, we rotate the pool weekly: 100 questions per hop per week for a 40-week cycle that fully exhausts the 4,000-item per-hop pool exactly once. On week 41 the pool reshuffles and the cycle starts over.
Single-equation derivation (e.g. v = v₀ + at)
Two-equation composition (e.g. compute velocity then momentum)
Three-step derivations combining kinematic, dynamical, or energetic quantities
Four-step chains spanning multiple sub-topics within a cluster
Deep five-step compositions that stress multi-concept integration
4,000 per hop · 200 per cluster · 20 clusters spanning classical → quantum
Physics-Bench spans 20 clusters across classical mechanics, thermodynamics, electromagnetism, waves & optics, and modern physics:
Choose any of the 1-Hop through 5-Hop bins. Each selection runs this week’s 100-question session for that hop, drawn across all 20 clusters.
Each item is a four-option physics question with numerical answer choices grounded in a real equation chain.
After each answer we visualise the k-step equation chain, show the exact numerical solution, and expand a full step-by-step thinking trace that walks through the derivation.
A fresh set of 100 questions per hop rotates in each week; a full 40-week cycle exhausts the 4,000-item per-hop pool, then it reshuffles.
Pick a k-hop reasoning level. Each selection runs a 100-question session for that hop, drawn from this week's rotation across 20 physics clusters (Kinematics, Electromagnetism, Thermodynamics, Quantum, and more).
Week 1 of 40 · cycle 0