$60–80/hr · Mercor · Part time
A physicist with deep expertise in theory, computation, or experiment who evaluates and designs training tasks for AI systems learning to reason about physics problems.
What you would do
- Author realistic physics problems drawn from your own research, specifying initial conditions, constraints, and the correct solution that AI attempts must match
- Evaluate AI-generated solutions for physical accuracy: do they violate conservation laws, mishandle boundary conditions, or make invalid simplifying assumptions?
- Provide critical feedback distinguishing between AI errors that stem from computational mistakes versus fundamental gaps in physics reasoning
- Review scientific writing and notation in AI outputs, assessing whether explanations reflect proper methodology and conceptual understanding
- Benchmark your assessments against peer reviewers to ensure consistency and rigor in evaluating AI performance
Who they want
- Postdoctoral or advanced professional experience in physics-theoretical, computational, or experimental domains with demonstrated research contributions
- Strong foundation in scientific programming, numerical methods, or experimental data analysis depending on specialization
- Ability to communicate clearly in technical writing, explaining physics concepts and AI reasoning gaps for interdisciplinary research teams
- Comfort working independently on a flexible schedule, typically 15-30 hours per week, with project start timelines as short as 48 hours
- Reliable internet and a quiet workspace suitable for focused technical work over multi-week engagements
Main skills
What the interview asks about
1.Problem design rigor
Evaluating AI requires carefully specified problems with clear reference answers; ambiguous or under-constrained problems make AI feedback unreliable.
For example: “Design a mechanics problem around energy conservation. How do you specify initial conditions, friction assumptions, and coordinate systems precisely enough that you can definitively judge whether an AI solution is correct?”
2.Identifying subtle physics errors
AI may produce outputs that pass surface-level checks but violate less obvious constraints or fail under parameter variations.
For example: “An AI solves a quantum mechanics problem correctly under one set of assumptions but makes an invalid approximation when you change the energy scale by an order of magnitude. How do you catch and articulate this failure?”
3.Computational versus analytical approaches
AI must learn when numerical simulation is appropriate versus when analytical methods are stronger-mismatches waste resources or introduce unnecessary error.
For example: “A problem could be solved analytically in closed form or numerically via finite elements. What factors determine which approach you'd expect an AI researcher to choose?”
4.Methodology assessment
AI-generated solutions often skip validation steps, ignore literature context, or propose computationally expensive alternatives to standard methods.
For example: “An AI proposes a novel numerical integration scheme for an N-body simulation. How would you evaluate whether this is a real advance or merely a computationally slower version of established methods?”
5.Domain boundary knowledge
Understanding where physics frameworks apply-when relativistic, quantum, or continuum assumptions are valid-is essential to catching AI overreach.
For example: “An AI applies classical mechanics to a problem involving particle velocities near light speed. Describe what's missing in its reasoning and how you'd explain the physics violation.”
A task you may get
Design a computational physics problem with spring constant, damping coefficient, and driving frequency. Specify statement, boundary conditions, and reference solution. Evaluate an AI solution for physical correctness.
How to prepare
- Identify three research problems from your own work that would translate well into physics training tasks with clear constraints and definitive correct answers.
- Gather computational code or simulation output from a recent project; prepare to discuss the modeling assumptions and parameter dependencies.
- Review a physics paper outside your narrow specialization; consider what computational challenges and conceptual gaps someone learning that subdomain must overcome.
- Think through a problem where AI reasoning diverges from rigorous physics-when approximations become invalid or how dimensional analysis constrains solutions.
The facts
- Pay
- $60–80/hr
- Hours
- Part time
- Where
- Remote
- Field
- Life, Physical, and Social Science
- Role type
- Talent network
- Posted
- 2/27/2026
We wrote this page from the public Mercor listing. It may be out of date, so read the full posting before you apply.