STEM and research roles ask graduate-level experts to write hard problems with verifiable answers, solve them step by step, and find reasoning errors in model output. This group covers chemistry, biology, mathematics, statistics and other sciences. Postings mention PhD-level problems and exact, checkable solutions, and some ask scientists to judge AI safety in areas such as chemical or radiological risk. These practice questions are about rigour and clear explanation.

Practice questions

These are practice questions written for this site to help you prepare. No platform has said it asks these exact questions.

  1. Write a problem in your field that a strong model would likely get wrong, with a full solution.

    A good answer covers: A problem with a single exact answer, several reasoning steps and no ambiguity, plus a clean step-by-step solution and why it is hard.

  2. Find the error in this model's step-by-step solution.

    A good answer covers: Point to the first wrong step, explain why it is wrong, and show the correct continuation.

  3. How do you make sure a problem's answer is verifiable?

    A good answer covers: Numeric or symbolic answer with units, one correct answer, no reliance on outside data that may change.

  4. Explain a core concept of your field to a first-year undergraduate.

    A good answer covers: Correct, simple, one worked example, and a common misconception.

  5. How would you grade a partially correct solution?

    A good answer covers: Use a rubric: credit correct setup and steps, deduct for the error and its effects, and explain the score.

  6. Which sub-fields can you judge at expert level?

    A good answer covers: An honest list that matches your degree and research.

  7. How do you avoid writing a problem whose answer is easy to find online?

    A good answer covers: Original setups, new numbers or combinations, and checking that it is not a known textbook problem.

  8. What makes a reasoning chain convincing but wrong?

    A good answer covers: Confident tone, skipped steps, unit errors, sign errors, misapplied theorems. Give one example.

What the platforms say about the assessment

In the 9 October 2026 data, roles in this group came from Meridial, SME Careers, Welo Data (Welocalize), micro1, Terac and Innodata. Each platform runs its own process, and steps can differ by role. Below is what each platform's public pages or postings say, followed by lines quoted from current postings in this group.

Meridial

  • Meridial postings describe the work and requirements but do not describe the screening steps.
  • Postings say contractors supply their own secure computer and internet connection, and recording roles add a high-quality microphone and recording setup. Company benefits such as health insurance do not apply.
  • We could not confirm Meridial's selection steps from a public page; treat the posting itself as the guide.

Source: boards-api.greenhouse.io

SME Careers

  • Many SME Careers postings state: 'Selection process involves an AI interview, a domain-specific task, and an interview with a recruiter.'
  • Many postings also say there is no immediate project for the role, and that qualified applicants will be contacted first when relevant opportunities arise.

Sources: sme.careers, api.sme.careers

Welo Data (Welocalize)

  • Welo Data's join page lists these steps: apply, create a Welo Works account, take a Language and Cultural Assessment, optional study materials, a qualification exam, then payment setup and agreement.
  • Many postings say there is no formal interview and the process is automated. Several mention learning modules and a quality test designed by the client; translation roles may include a translation test.
  • Welo Data says it only contacts applicants from @welocalize.com, @weloglobal.com, @hire.lever.co and @welodata.ai addresses and never asks for fees.

Source: welodata.ai

micro1

  • micro1's experts page describes four steps: apply for opportunities, take the AI interview, get onboarded, start earning.
  • Postings commonly list screening questions, then an AI interview, then in some roles a technical assessment, then a review by the hiring manager. Interview lengths stated in postings range from about 10 to 30 minutes.
  • Some postings ask you to complete the interview and assessment without outside assistance.
  • micro1's FAQ says certification does not guarantee immediate placement on a project.

Source: www.micro1.ai

Terac

  • Terac's expert docs describe signing in with LinkedIn, verifying your phone, connecting LinkedIn or adding a resume, and a short voice conversation of about 10 minutes that the docs say you can skip and take later.
  • Each opportunity then has its own steps, usually screening questions followed by the tasks themselves. Research postings describe remote video or voice interviews of 30 to 90 minutes.

Source: terac.com

Innodata

  • Innodata postings describe the work and the hourly range but do not describe the screening steps.
  • Postings state that Innodata will never ask for payment, banking details or sensitive personal information during the application process.

Source: boards-api.greenhouse.io

Quoted from current postings

"Apply to the role and complete the screening questions."

Materials Science Expert, micro1

"Selection process involves an AI interview, a domain-specific task, and an interview with a recruiter."

Mathematics Quality Assurance Lead (QAL), sme

"Complete an AI interview of approximately 30 minutes."

Materials Science Expert, micro1

Steps were read from public pages and postings, not tested first-hand. Passing a screening does not guarantee a project, hours or pay.

Terms to know

Stump the model

A task where you write a hard question the model gets wrong and record the correct answer, often used in expert STEM and coding projects.

Chain of thought

The step-by-step reasoning a model writes before its final answer. Some tasks ask you to check each step, not only the result.

Process supervision

Judging each step of a model's reasoning rather than only the final answer, common in maths, science and coding work.

Ground truth

The correct answer or label that model output is compared against.

Sources