The Role
You own the loop that turns real work into a better model.
The hard part of training agents is not fitting a curve to a benchmark. It is training models that stay coherent across hundreds of tool calls in environments they have never seen, reason over structured context that does not fit in any prompt, and recover when the plan breaks at step 173. Static benchmarks do not measure this. Static datasets do not teach it. So we generate our own: environments built programmatically rather than hand-authored, mutated between episodes, mechanically checked for validity before an agent ever touches them, with rewards grounded in system state rather than a judge's opinion. That work is open source.
Models we train are deployed into environments we do not control and do not get to inspect afterward, which constrains how everything upstream of that has to be designed. Your work sets the capability ceiling of the platform.
You will work directly with the founders. Your experiments and infrastructure feed into how our models get trained and evaluated, not a sandbox that gets shelved at the end.
What You Will Own
- Post-training and alignment. SFT, DPO, GRPO, and trajectory-level rewards over real agent runs, on a distributed training stack designed for long-horizon, tool-using behavior.
- Agentic RL. Models that learn from live tool interactions and trajectory outcomes rather than static datasets, with reward shaping that captures partial progress, recovery quality, and strategic coherence, and that resists reward hacking.
- Procedural environment generation. Domain randomization, non-stationary dynamics, and adversarial perturbations spanning task distributions no fixed benchmark could cover, with an admission gate that proves an environment is valid before it is used.
- Agent memory as model substrate. How models read from, write to, and reason over persistent memory that tracks both when a fact was true in the world and when the system learned it, including retrieval at inference time, memory-aware prompt construction, and training signals that reward grounded reasoning.
- Graph-based reasoning. Path traversal, link prediction, and next-hop selection across complex topologies, and how models learn to plan multi-step trajectories through graph-structured state.
- Evaluation infrastructure. Offline benchmarks, LLM-as-judge frameworks, trajectory scoring, red teaming, and automated regression testing that catches capability drift before it ships.
- Inference optimization and serving. Quantization, speculative decoding, KV-cache management, and deployment tuned for bursty, long-context agentic workloads, including our open-weight models.
Publishing
We create pathways for our team to publish and contribute to the community, because credibility in this field is earned by advancing it. Our environment generation work is open source, and our people speak at the conferences that matter in this space. Research does not go to die in an internal doc here.
Who Thrives Here
We care less about pedigree than evidence. Three traits are non-negotiable:
- Hacker mindset. You see systems as something to take apart and understand, not consume as designed. You read the source, poke at edges, and assume there is a faster or more correct way. The instinct to question how things work is the floor.
- Break things, then stabilize them. Robustness in long-horizon agent systems is not a feature you add at the end, it is the product. You push systems until they fail, study the failure honestly, and rebuild stronger. We do not punish breakage that produces learning, we punish hiding it.
- Try harder. When the obvious approach fails, you find the second, third, and fifth. You re-read the paper, instrument the system, ask a smarter question. Most ceilings here are self-imposed.
You Are Probably a Fit If You
- Have a PhD or equivalent research depth in ML, with published work in post-training, RL, alignment, evaluation, or agentic systems.
- Have owned a post-training pipeline end to end, from data curation through post training to evaluation and deployment.
- Have built evaluation infrastructure for LLMs, including automated benchmarks and honest statistical analysis of behavior under distribution shift.
- Are proficient with PyTorch and modern distributed training and serving frameworks, and understand the systems tradeoffs that decide whether a research idea scales.
- Ship research into production, reason from first principles, and are intellectually honest about what you do not know.
Useful, not required: inference optimization at production scale, training against graph-structured context, GNNs or hybrid LLM-graph systems, environment or simulation design.