FrontierPhysics benchmark evaluates AI agents on physics research
FrontierPhysics tests agents on 56 end-to-end tasks drawn from 42 physicists own research across nine domains.
Stories per day · 30 days
3 RSS
Relatedbenchmark2mathematics2openai2ai-agents1cosmology1disclosure1evaluation1exponential-time-hypothesis1huggingface1kakeya1kakeya-conjecture1llm-agents1math1neutrinos1pass-at-k1reasoning-models1reinforcement-learning1scientific-reasoning1
12 stories, newest first · one card per story
Stories tagged #physics
Researchers fault OpenAI for limiting access to claimed breakthroughs
OpenAI released an AI math write-up on Kakeya, will explain a computer-science claim through outsiders, and has not released the physics result.
1
OpenAI posts claimed math, computing, and physics breakthroughs
Outside groups will examine an ETH refutation and a QCD proof, while the math write-up is released unverified.
1
Outcome RL gains are mostly finishing and reliability, not reach
On four physics and math sets, newly solved problems add at most 2.1 points, and only 3.5% of 144 unique physics solves held up.
PhysClosureBench finds frontier agents often miss full physics research closure
Claude-Opus-5 scores 6.29/10 on average but reaches full closure in only 32.7% of attempts.
World Models' Last Exam in Physics
A 40-task physics benchmark finds video world models remain physically inconsistent, with the best scoring 57.76.
NVIDIA Researchers Introduce Physis-Lang: Self-Evolving Physical Language That Lifts Cosmos 3 Past Veo 3.1 on Physics Benchmarks
NVIDIA's Physis-Lang physics captions lift Cosmos 3 past Veo 3.1 on several physics video benchmarks.
6
VideoPhysEdit: Physical Counterfactual Video Editing via Rigid-Body Physical Scene Reconstruction
VideoPhysEdit reconstructs rigid-body physics so video edits produce physically consistent later motion.
Yes, Claude can do Nine Loops
Anthropic's Claude tackled a nine-loop N=4 super Yang-Mills calculation posed as a hard physics test.
Why Do Video Diffusion Models Violate Physics? Unveiling the Flaws in Attention Mechanisms
Researchers trace unphysical video-diffusion motion to RoPE attention decay and propose a frequency-scaling fix.
StudyBench: Can Self-Evolution Squeeze Textbooks for Olympiad Capability?
Researchers introduce StudyBench, a physics benchmark showing self-evolution gains on textbook problems rarely transfer to olympiad-level questions.
🔬“We have foundation models for language, not for physics” — Anima Anandkumar, Bren Professor of Computing
Caltech professor Anima Anandkumar discusses Neural Operators and FourCastNet for physics modeling, arguing inductive biases beat pure token scaling.
2