LLM Agent Research Intern

We are hiring a research intern to train LLM agents that play in-game characters. The agent has to stay in persona across multi-turn dialogue, use memory and game context, and keep relationship, tone, and world rules intact. The work sits on a live role-play NPC model: supervised fine-tuning today, with preference optimization and verifiable-reward reinforcement learning as the next training stages.

1. 01

   Train role-play agents with supervised fine-tuning (SFT), then post-train them with direct preference optimization (DPO) and reinforcement learning with verifiable rewards (RLVR). Iterate on persona consistency, multi-turn coherence, and instruction following.

2. 02

   Build and curate SFT, preference, and verifiable-reward data for character dialogue, covering memory injection, relationship stage, language switching, emotion, and out-of-character resistance.

3. 03

   Design checks that can be verified automatically, such as format, language, persona boundaries, and safety refusals, and use them both as RLVR rewards and as offline regression tests.

4. 04

   Work with game designers and engineers to bring the agent into a live game, and keep latency, cost, and player experience within production constraints.

1. 01

   Currently enrolled in a Master's or PhD program in Computer Science, NLP, Machine Learning, or a related field, or recently graduated. Available at least three days per week, with supervisor and university approval where required.

2. 02

   Fluent English (written and spoken) and strong communication skills are required. You will work closely with game designers, engineers, and researchers across regions, and must be able to explain experiment results and trade-offs clearly.

3. 03

   Hands-on experience with supervised fine-tuning (SFT) of LLMs, and with post-training methods including DPO and RLVR (or a closely related verifiable-reward RL method). You should be able to run the loop from data and training config through evaluation and the next iteration.

4. 04

   Experience training or evaluating role-play or character-dialogue models, including persona consistency, multi-turn dialogue, and staying in character.

5. 05

   Proficient in Python and PyTorch, and comfortable with the Hugging Face / Transformers training stack.

6. 06

   Preferred: experience with LLM agents (memory, planning, or tool use), LLM applications in games, or a publication at a major NLP/ML venue (ACL, EMNLP, NAACL, NeurIPS, ICLR, ICML).