Applied Scientist: Agentic RL & Post-Training

Vecna AI; Chicago (IL); On-site; USD 140,000 - 210,000; Full time

Job summary

Vecna AI seeks a researcher to own the end-to-end post-training and alignment loop for long-horizon, tool-using agents. You will design and execute SFT/DPO/GRPO with trajectory rewards, training models that remain coherent across hundreds of tool calls in unseen environments.

You will publish and contribute to the community. You will work directly with founders, shaping how models are trained and evaluated, not a static benchmark.

Qualifications

• PhD or equivalent research depth in ML with published work in post-training, RL, alignment, evaluation, or agentic systems.
• Owned a post-training pipeline end to end: data curation through post training to evaluation and deployment.
• Experience building evaluation infrastructure for LLMs with automated benchmarks and distribution-shift analysis.

Responsibilities

• Post-training and alignment: SFT, DPO, GRPO, and trajectory-level rewards over real agent runs.
• Agentic RL: live tool interactions and trajectory outcomes with reward shaping for coherence.
• Procedural environment generation: domain randomization, non-stationary dynamics, adversarial perturbations.
• Agent memory as model substrate: read/write memory, retrieval at inference, memory-aware prompts.
• Graph-based reasoning: path traversal, link prediction, multi-step trajectory planning.
• Evaluation infrastructure: offline benchmarks, red-teaming, regression testing.
• Inference optimization and serving: quantization, KV-cache management, deployment for long-context workloads.

Education

PhD or equivalent ML research depth

Tools

PyTorch

Distributed training

Serving frameworks

Job description

Vecna AI seeks a researcher to own the end-to-end post-training and alignment loop for long-horizon, tool-using agents. You will design and execute SFT/DPO/GRPO with trajectory rewards, training models that remain coherent across hundreds of tool calls in unseen environments.

You will publish and contribute to the community. You will work directly with founders, shaping how models are trained and evaluated, not a static benchmark.