AI Researcher.

Remote

Study how model changes affect behavior, accuracy, and efficiency.

You’ll develop and evaluate edits to open-weight language models, then publish the methods, results, and modified weights. This is a practical research role centered on model code, controlled experiments, and careful interpretation of results.

How to apply

The role

Weights & behavior

Design targeted weight edits and study their effects on model behavior. Examine changes in answers, accuracy, and failure patterns.

Architecture & efficiency

Experiment with model structure to improve inference speed and memory use. Identify which capabilities each change preserves or compromises.

Evaluation

Build reproducible benchmarks with appropriate baselines. Inspect individual outputs alongside aggregate scores, investigate regressions, and test whether findings hold across tasks.

Model releases

Publish modified models on Hugging Face with model cards, evaluation results, reproducible code, and documented limitations.

Your experience

Practical work with open-weight language models and an understanding of transformer architectures.
Python and PyTorch or comparable tools, including reading and modifying model code and weight tensors.
Experimental design, baseline selection, and evaluation of model quality and inference performance.
Technical writing that makes methods, results, and limitations clear enough for others to reproduce the work.

How to apply

Complete the research exercise

Make a focused weight edit to IBM Granite 3.1 1B-A400M Instruct and test whether it produces more concise answers while retaining useful information. Evaluate the original and edited models locally through Ollama, include a prompt-only control, and publish your edited model on Hugging Face.

Allow 2–3 hours of active work; record downloads, unattended computation, conversion, and uploads separately. Stop at the time limit and report anything unfinished. We assess your method and analysis, including negative results. An improvement is not required.

Assignment & starter code Setup instructions, papers and tools, required benchmarks, and the submission checklist.

Our published Granite experiment is a reporting reference with weak, inconsistent results. Develop and explain your own experiment.

Send your application

Email hr@ovrlab.io with the subject AI Researcher Application — Granite Experiment — Your Name, replacing “Your Name”. Include:

A short introduction and your CV or professional profile.
Your experiment’s code repository or a ZIP of the code.
Your Hugging Face model link, including reloadable weights, a model card, and the exact GGUF you evaluated.
Raw evaluation results and a one-page research note covering your hypothesis, method, findings, limitations, time spent, and any AI assistance.

Our privacy policy explains how we handle application correspondence.

All open positions