How to apply for ML Research Resident

About Elicit

Elicit is a public benefit company building an AI research assistant to help researchers make better decisions. Its product is designed as a scalable ML system that prioritizes systematicity and transparency, with supervision of process rather than outcomes. This makes it a fit for people who care about verifiable reasoning and epistemically sound AI.

About the role

As an ML research resident, you will spend 3 months developing computational procedures (operators) that reliably improve a knowledge state over thousands of iterations. The work focuses on making LLM reasoning transparent and verifiable, using unlimited test-time compute. You will design and test operators that help a research agent make genuine progress in understanding, not just produce plausible outputs.

A typical day

A typical day might involve reading papers on iterative reasoning and process supervision, then designing and coding a new operator to test on a structured knowledge state. You would run experiments over many iterations, inspect whether the knowledge state improved in a human-readable way, and discuss findings with the Elicit team. Since the role is remote and contract-based, you would manage your own schedule and communicate progress asynchronously.

Who Elicit is looking for

• Has hands-on experience with transformer-based models and understands why iterative application degrades quality.
• Can design and test operators that improve structured knowledge states, such as claims, evidence, logical connections, scratchpads, and search trees.
• Cares about epistemic soundness: separating inference from raw evidence, finding implicit assumptions, connecting evidence across sources, and identifying gaps in reasoning.
• Is comfortable working remotely on a contract basis and can operate independently on open-ended research problems for 3 months.

Tips for this application

• In your application, describe a specific project where you improved an LLM's reasoning over multiple iterations, not just a single forward pass.
• Explain how you would represent a knowledge state. Give a concrete example, such as a scientific paper as claims supported by evidence and connected through logical reasoning.
• Show that you understand the difference between outcome supervision and process supervision. Elicit explicitly prioritizes supervision of process, not outcomes.
• Propose one or two operator ideas you would test in the residency. For example, an operator that reorganizes information to better answer a question or finds an implicit assumption in an argument.
• Address how you would keep each improvement human-readable and epistemically sound. Elicit wants improvements that make the knowledge state more useful while remaining transparent.

What to cover in your cover letter

['Your concrete experience with transformers and iterative LLM degradation.', 'A specific operator or computational procedure you have designed or would design to improve a knowledge state.', 'How you separate inferences from raw evidence and ensure each step is human-readable.', 'Why you want to work on transparent, verifiable reasoning for a research agent rather than a typical ML system trained to do whatever works.']

Research before applying

• Read the Elicit product pages and any public writing on how their AI research assistant works.
• Look for Elicit's public statements on being a public benefit company and what that means for their ML system design.
• Study the concepts of test-time compute, iterative transformer degradation, and process supervision versus outcome supervision.
• Find examples of structured knowledge representations, such as claims with evidence, logical reasoning graphs, scratchpads, and search trees.

Likely interview topics

Based on the job description, expect questions about:

• How would you define a knowledge state for a scientific paper, and what operators would you apply to it?
• Why do transformers degrade when applied iteratively, and what computational procedures could improve reliability over thousands of iterations?
• How would you test whether an operator makes a knowledge state more useful while remaining epistemically sound?
• What does supervision of process, not outcomes, mean in practice for an ML research agent?
• How would you use unlimited test-time compute without sacrificing transparency and verifiability?

Common mistakes to avoid

• Submitting a generic ML research cover letter that does not mention knowledge states, operators, or iterative improvement.
• Focusing only on model training or fine-tuning without addressing test-time compute and transparent reasoning.
• Claiming you care about transparency without explaining how you would keep each operator human-readable and epistemically sound.