Applied Scientist Intern

Vecna AI•Chicago, IL

51d

About The Position

This internship focuses on working with the models that are central to the platform. These models execute multi-step task chains in complex environments, route to the right tools and reasoning strategies, and improve from real-world trajectories. The intern will collaborate with the research team on a specific aspect of the post-training, evaluation, or serving stack. The challenges in training agents involve ensuring coherence across numerous tool calls in unseen environments, reasoning over structured context that exceeds prompt limits, and recovering from plan failures. The company generates its own training data through procedurally generated synthetic gyms and self-mutating environments. Training occurs on real trajectories with trajectory-level rewards, and evaluation is performed against challenging environments. This is a research-intensive internship directly involving the founding team, where experiments and infrastructure contribute directly to model training and evaluation. The company encourages publishing and community contributions to advance the field.

Requirements

Pursuing an MS or PhD in ML, CS, or a related field, or equivalent research depth.
Demonstrated research experience in post-training, reinforcement learning, alignment, evaluation, or agentic systems (coursework, publication, preprint, or substantial project).
Proficiency with PyTorch and experience training or fine-tuning models.
Ability to reason from first principles.
Intellectual honesty about knowledge gaps.
Treating being wrong as information rather than failure.

Nice To Haves

Exposure to RLHF/DPO/GRPO.
Exposure to LLM evaluation.
Exposure to procedural environment or task generation.
Exposure to graph-structured context (retrieval over knowledge/memory graphs, GNNs).
Exposure to inference optimization.

Responsibilities

Take ownership of a focused research project within areas such as post-training and alignment (SFT, DPO, GRPO, trajectory-level rewards), procedural environment generation (synthetic gyms, self-mutating environments, adversarial perturbations), agentic online RL (learning from live interactions, reward shaping), bitemporal context graphs (reading, writing, reasoning over persistent graphs), graph-based reasoning (path traversal, link prediction, planning through graph-structured state), evaluation infrastructure (benchmarks, LLM-as-judge, red-teaming), or inference optimization and serving (quantization, speculative decoding, KV-cache management).