Research, Mid Training

Thinking Machines Lab is building AI systems that extend human will and judgment through frontier model training and human-AI interfaces. The Research, Mid Training role owns late-stage training between pre-training and post-training, including synthetic data strategies, data mixture design, quality improvement, context extension, evaluation, and capability development across coding, mathematics, and reasoning.

Responsibilities:

• Own the data. Decide what the model needs to see for each capability and each area of knowledge, then source, curate, and synthesize it. Build the pipelines that filter, deduplicate, verify, and rewrite raw material into training-grade datasets. You'll be responsible for producing data that actually moves the model, and for making sure it reflects how people really use these models, not just clean benchmark-style tasks
• Improve what the model knows. Design and measure interventions that increase knowledge: targeted corpora, synthetic rephrasings and QA over source documents, knowledge-dense mixes. Characterize how knowledge scales with data and when it is retained through post-training versus forgotten
• Instill behaviors and set the prior. Introduce new behaviors during mid-training and measure them the right way: not just whether an eval goes up, but whether post-training becomes easier, more sample-efficient, and more stable as a result. Work closely with the post-training team to decide which behaviors belong in mid-training and which belong in RL
• Build the quality pipeline. Own the automatic filtering and scoring stack: train and calibrate quality classifiers and LLM-based judges, build verifiers for synthetic data, and use them to raise the quality bar of the mix by a lot, not a little. Give the team fine-grained control over data attributes (difficulty, domain, format, style, correctness) and measure the effect of each
• Develop and tune the recipe. Iterate on mid-training recipes: the collection of datasets, training stages, annealing schedules, and hyperparameters. Measure how recipe choices affect metrics, including downstream of post-training
• Iterate on evals. Mid-training involves a never-ending loop of defining a set of evaluations, optimizing them, and then realizing your existing evals don't capture what matters. You'll be responsible for both making numbers go up and making sure the numbers are meaningful
• Debug and understand. While tuning the details of a training configuration, we often observe results that don't quite make sense. You'll be responsible for both getting things to work and developing a deeper understanding we can bring to the next problem
• Scale and explore. Mid-training will involve a combination of scaling existing methodologies and developing new ones. We'll want to both measure how performance scales with dataset size and explore using completely different kinds of training data