Member of Technical Staff - Research Engineering

NextToken AI Inc.; Mountain View (CA); Sur place; USD 180 000 - 240 000; Plein temps

Résumé du poste

NextToken AI Inc. is hiring researchers and engineers to advance post-training methods for specialized language models and agents. You will design training recipes combining supervised fine-tuning, reinforcement learning, and distillation, and explore synthetic data generation for better signals.

You’ll work with founders, systems engineers, and customer teams to shape experiments, implement scalable evaluation tools, and document results for publication and open-source projects.

Qualifications

• Understanding of modern post-training methods like RL with verifiable rewards (RLVR).
• Depth in policy optimization (GRPO) or on-policy distillation (OPSD).
• Ability to reason about learning signals, tradeoffs, and failure modes.

Responsabilités

• Develop post-training methods for language models and agents.
• Design and run training recipes combining supervised fine-tuning, RL, and distillation.
• Investigate synthetic data generation and curation for learning signals.
• Build agent environments, reward models, and verifiers for training and evaluation.
• Collaborate with systems engineers to improve experiment and inference efficiency.
• Publish findings through technical writing and open-source contributions.

Connaissances

Experimental design

Communication of findings

Problem solving

Outils

PyTorch

JAX

Description du poste

NextToken builds specialized models and the systems that help them improve. We work with AI product teams to turn production traces, expert feedback, and real tasks into evaluations, training data, and more capable models. Our goal is to make it practical for companies to develop intelligence of their own: models that understand their work and improve as that work evolves.

Our founding team has built search infrastructure and post-trained models at frontier AI labs and on large-scale product teams. You’ll work directly with the founders on research and engineering, help shape our technical direction, and see your work deployed in customer products.

The role

You’ll develop the data, evaluations, and post-training methods that improve models on real product tasks. That means understanding where an agent fails, deciding what evidence would establish an improvement, and carrying an experiment through to a model people can use.

The work combines research judgment with hands‑on engineering. You’ll work with the founders, systems engineers, and customer teams, shaping both individual experiments and the tools that make the next experiment easier.

What you’ll work on

Develop post‑training methods for specialized language models and agents. Design and run training recipes that combine supervised fine‑tuning, reinforcement learning, and distillation, and investigate how they affect capability, generalization, and efficiency.

Research synthetic data generation and curation. Explore teacher selection, trajectory sampling, and data mixtures that produce useful learning signals from production tasks and feedback.

Develop agent environments, reward models, and verifiers for training and evaluation. Establish benchmarks that capture meaningful product improvements and expose weaknesses in model behavior.

Work with systems engineers to improve the efficiency of experiments and model inference. Connect algorithmic choices to their implications for rollout cost, memory use, and training stability.

Carry promising research into customer applications and turn successful approaches into reusable methods and tools. Use production behavior to identify the next research questions.

Communicate findings through technical writing, research publications, and open‑source contributions where appropriate.

What we’re looking for

Understanding of modern post‑training methods, including reinforcement learning with verifiable rewards (RLVR), policy optimization approaches such as GRPO, and on-policy distillation and self‑distillation (OPSD). You can reason about their learning signals, tradeoffs, and failure modes, with practical depth in at least one relevant area.

Familiarity with how frameworks such as PyTorch or JAX execute model training, and the ability to inspect research code, diagnose training problems, and assess whether an implementation matches the intended method.

Experience investigating model training, evaluations, agent environments, or related ML research. You can explain your contribution and the evidence supporting the result.

Strong experimental judgment: well‑chosen baselines, controlled comparisons, and attention to generalization, data leakage, and misleading rewards or metrics.

Ability to turn an open‑ended problem into a research plan, work through unfamiliar systems, and explain technical decisions to teammates and customers.

We welcome experienced research engineers and applied researchers, as well as strong engineers earlier in their careers.

Experience with agentic RL, reward modeling, synthetic data, distillation, or efficient inference is especially relevant. Research papers, open‑source contributions, and independently built systems are all useful ways to show your work.