Algorithm Research Intern — Agentic Vision

Conduct research and development on Agentic Vision for Hunyuan multimodal foundation models, focusing on visual reasoning post-training. Integrate visual perception, planning, tool and code execution, environment feedback, and self-correction to enable dynamic, interactive visual reasoning (Thinking with Images / Video).

Advance visual reasoning and spatial intelligence, focusing on fine-grained perception, mathematical and scientific diagram reasoning, 3D spatial understanding, complex visual logic (puzzles and games), and embodied spatial planning. Improve reasoning and decision-making in complex scenarios.

Develop multimodal post-training methods, including cold-start SFT, online reinforcement learning (PPO / GRPO), on-policy distillation (OPD), and verifiers and reward models (Verifier / PRM / Rubric RM). Address credit assignment, long-horizon context management, and effective use of environment feedback in multi-turn interactions.

Build verifiable data synthesis pipelines using rendering engines, symbolic systems, and code sandboxes to generate challenging visual reasoning data and multi-turn agent trajectories. Establish a continuously improving data flywheel through real-world data mining, rejection sampling, difficulty stratification, deduplication, contamination prevention, and data mixture optimization.

Build interactive sandboxes supporting code execution, visual tools, and search. Develop evaluation and error attribution methods covering perception accuracy, reasoning correctness, tool efficiency, and task completion, creating a continuous cycle of evaluation, data refinement, and model improvement.

Master’s degree or higher in Computer Science, Artificial Intelligence, Computer Vision, Machine Learning, Mathematical Statistics, or a related field. Strong foundations in deep learning and algorithm implementation, proficiency in PyTorch, and experience with distributed training.

Research or industry experience in multimodal foundation models, visual reasoning, multimodal agents, or reinforcement learning for post-training. Familiarity with post-training methods and in-depth practical experience in one or more of SFT, PPO / GRPO, OPD, RLVR, or reward modeling. Experience with verl, OpenRLHF, Megatron, or related frameworks is a plus.

Strong system-level design and end-to-end execution skills, with the ability to jointly design data, tool environments, reward mechanisms, training strategies, and evaluation protocols around target tasks, and continuously improve model capabilities through experimental analysis.

A keen interest in multimodal reasoning, Agentic Vision, and spatial intelligence, with the ability to follow frontier research, exercise independent technical judgment, and translate ideas into validated experiments and practical improvements.