Lead Physics Researcher for Frontier Benchmark Solutions

Obsidian; New York (NY); Remote; USD 110,000 - 186,000; Full time

Job summary

Obsidian seeks expert physics researchers to author and verify golden reference solutions for the CritPt benchmark (arXiv:2509.26574v3). You will produce 100%-human-verified reference data used to evaluate large language models on frontier physics reasoning.

You may solve problems end-to-end with verifiable derivations and code, audit others' solutions, or adjudicate between parallel attempts, while documenting reasoning, tolerances, and verification tests.

Qualifications

• PhD or postdoc in the relevant subfield (senior PhD student minimum)
• Postdoc or junior professor in the relevant subfield (PhD minimum)
• Full professor or industry research PI in the relevant subfield (senior postdoc or junior professor minimum)
• Hands-on familiarity with at least two canonical methods of the target subfield
• 3–5 representative publications (arXiv ID or DOI)
• Working proficiency with LaTeX, Python, Jupyter, and SymPy
• Strong written English (B2/C1/C2 minimum; native or near-native preferred)

Responsibilities

• Solve research-level physics challenges end-to-end with verifiable derivations, code, and peer-reviewed references
• Decompose challenges into standalone checkpoint sub-problems that require genuine physical reasoning
• Author Python answer templates with auto-grading functions for symbolic or numerical answers
• Audit submitted solutions for correctness, scope, and method soundness; deliver actionable feedback across iterations
• Adjudicate between parallel solver attempts and decide which solution becomes the golden reference
• Document chain-of-thought reasoning, error tolerances, equivalent symbolic forms, and verification test cases

Skills

LaTeX

Python

Jupyter

SymPy

Strong English

Education

PhD

Postdoc

Tools

LaTeX

Python

Jupyter

SymPy

Job description

Obsidian seeks expert physics researchers to author and verify golden reference solutions for the CritPt benchmark (arXiv:2509.26574v3). You will produce 100%-human-verified reference data used to evaluate large language models on frontier physics reasoning.

You may solve problems end-to-end with verifiable derivations and code, audit others' solutions, or adjudicate between parallel attempts, while documenting reasoning, tolerances, and verification tests.