Research Engineer / Scientist, Post-training & Reinforcement Learning - London

H exists to push the boundaries of superintelligence with agentic AI. By automating complex, multi-step tasks typically performed by humans, AI agents will help unlock full human potential. H is hiring the world’s best AI talent, seeking those who are dedicated as much to building safely and responsibly as to advancing disruptive agentic capabilities. We promote a mindset of openness, learning, and collaboration, where everyone has something to contribute. The Models team builds the foundational models that power our cutting-edge agentic technology. We focus on training techniques to optimize model capabilities specifically for agent applications. This allows us to achieve the best performance at a given inference cost.

Responsibilities

Develop and train advanced LLMs and VLMs, including multimodal architectures; Research and implement training methods for enhanced capabilities like instruction following and tool use; Design and optimize data pipelines and training systems for large-scale distributed training; Collaborate with cross-functional teams to integrate models into agentic AI systems; Evaluate model performance and communicate findings to stakeholders; Stay current with advancements in LLMs, VLMs, and related fields

Requirements

Strong research engineer / scientist mindset with experience training and improving large language models at different scales in distributed computing settings whether that’s modelling, data collection, experimenting and ablating, implementing SOTA; Strong programming skills in Python, Rust, or similar; Strong software engineering fundamentals building performant and reliable systems; Proficiency in deep learning frameworks (PyTorch, JAX, TensorFlow); Ability to work on different layers of the stack from low-level training backends, data ingestion to ML/RL algorithmic design and implementation; Knowledge of when and where to be rigorous and slower versus when to break and iterate quickly; Experience training LLMs/VLMs with techniques such as SFT, DPO, RLHF/RLVR, reward modelling, offline RL, distillation, etc.; Experience with offline and online reinforcement learning in or outside of the context of language models; Good communicator, collaborative and low-ego; Ability to handle a controlled-chaotic environment with a high-degree of between-teams dependencies and collaboration; Go-do attitude and ability to balance personal conviction/interests with wider team needs; Not shying away from hard research or engineering problems

Nice to have

Publications in top-tier AI conferences (e.g., NeurIPS, ICML, CVPR, ACL, ICCV, AAMAS); Advanced degree (PhD or MSc) in a relevant field (e.g., ML, DL, NLP, CV); Experience with large-scale distributed training and inference (multi-node, large models, MoE, parallelism strategies); Experience training models for computer use or other multi-turn and/or multimodal agentic settings; Extensive experience with reinforcement learning with sparse rewards; Experience with multi-domain training, data mixture design, curriculum learning, model merging, distillation

Benefits

Join the exciting journey of shaping the future of AI, and be part of the early days of one of the hottest AI startups; Collaborate with a fun, dynamic and multicultural team, working alongside world-class AI talent in a highly collaborative environment; Enjoy a highly competitive salary; Unlock opportunities for professional growth, continuous learning, and career development