RESEARCHER, POST-TRAINING

Job description

**ABOUT THE COMPANY** We're building autonomous research agents for recursive self-improvement (multi-agent systems that propose, run, and analyze machine learning experiments). We're a small team based in San Francisco, on-site **ABOUT THE ROLE** You'll lead our work on model post-training: supervised fine-tuning, preference data, reinforcement learning from human and AI feedback, reward modeling, and the evaluation suites that tell us what's actually working. You'll own a research area that meaningfully shapes our model behavior and capability. This is a hands-on senior research role. You'll set direction, run experiments, and ship into production. You'll partner with the data, infrastructure, and engineering teams to make the post-training pipeline reliable and fast: improvements there compound into every model we ship. **WHAT YOU'LL DO** - Lead post-training research: SFT, RLHF/RLAIF, RLVR, DPO and successor methods, reward modeling, preference data design - Design and curate the data that goes into post-training (from sourcing, to filtering, to quality assessment) - Build and maintain the evaluation suites that measure what matters; resist Goodharting your own benchmarks - Run rigorous experiments (controls, ablations, statistical significance) and write up internal findings clearly - Scale data pipelines and the infrastructure team to scale training - Identify and characterize failure modes (reward hacking, distribution drift, eval saturation) and design experiments to address them - Stay current on the post-training literature; bring useful methods in, ignore the noise **WHAT WE'RE LOOKING FOR** - Strong track record of post-training research (SFT, RL, reward modeling) at a frontier-model lab or equivalent - 5+ years of hands-on ML research experience - Comfort with large-scale data curation and preference-data pipelines - Experience designing evaluation suites for capabilities that aren't easily benchmarked - Fluent in PyTorch or equivalent; comfortable at the scale of distributed training - Strong statistical instincts: you'd notice a flawed comparison before someone else points it out - Strong written communication **NICE TO HAVE** - PhD in ML, statistics, CS, or adjacent - Published research at NeurIPS, ICML, ICLR, COLM, RLC, or comparable venues - Experience with reward hacking detection, scaling reward models, or RLHF infrastructure - Synthetic data generation experience - Background in RL math (policy gradients, importance sampling, off-policy methods) - Open-source contributions to post-training infrastructure **THIS ROLE IS PROBABLY NOT FOR YOU IF** - You're primarily interested in pretraining (that's a different role)- You'd rather invent novel methods in isolation than ship them into a model that real users run - You prefer benchmarks that are stable to evaluation work where the right answer isn't yet defined