New opportunity

Generative AI Researcher (3D, Video, and Agentic Intelligence)

Tencent · Singapore

About this role

Lightspeed Tech Center, an R&D department within Tencent's LIGHTSPEED STUDIOS, seeks a Generative AI Researcher to develop AI technologies for next-generation game development. The full-time role is listed as onsite at Singapore-CapitaSky.

Responsibilities

Develop AI systems for 3D object and scene generation, video generation and world modeling, human motion capture and generation, multimodal vision-language models, and agentic reinforcement learning.
Conduct original research in computer vision, computer graphics, and embodied intelligence to advance generative AI pipelines for virtual world creation.
Work with game developers, artists, and designers to integrate AI-generated content into gameplay. Explore advances in 3D generation, video understanding, multimodal reasoning, and simulation-based reinforcement learning to improve creative and production workflows.

Required qualifications

PhD in Computer Vision, Computer Graphics, Machine Learning, Robotics, or a related field, plus a strong research background in at least one of 3D object or scene generation, video generation or world modeling, human motion capture or generation, vision-language models, or reinforcement learning and agentic AI.
Publications in top-tier conferences or journals, such as CVPR, ICCV, ECCV, SIGGRAPH, NeurIPS, ICML, or ICLR; proficiency with deep learning frameworks such as PyTorch or TensorFlow; and experience in large-scale model training or evaluation.
Strong teamwork and communication skills in a fast-paced research environment.

Preferred qualifications

Experience with 3D modeling or rendering tools such as Blender, Unreal Engine, or Unity; coding and prototyping in Python, C++, or similar languages; and familiarity with datasets and benchmarks for 3D, motion, or video generation.
Ability to translate research into interactive applications, with creativity and an interest in AI-driven content creation tools.

For full details about this role, click “Apply on company site” to view the employer’s original posting.

Skills for this role

Generative AI3D object generation3D scene generationVideo generationWorld modelingHuman motion captureHuman motion generationVision-language modelsReinforcement learningAgentic AIComputer visionComputer graphicsEmbodied intelligenceMultimodal reasoningDeep learningLarge-scale model trainingModel evaluationPyTorchTensorFlowBlenderUnreal EngineUnityPythonC++ResearchPrototypingTeamworkCommunication

Checking your profile…

Preparation suggestions, not the employer’s actual interview questions. Always check the original posting for current requirements.

Understand the role

What matters most

Original research for virtual worlds
The role calls for original research in computer vision, computer graphics, and embodied intelligence. The posting also seeks a strong research background in at least one listed area and top-tier publications.
Build game-focused generative AI systems
Develop systems spanning 3D objects and scenes, video and world modeling, human motion, vision-language models, and agentic reinforcement learning, tailored to next-generation game development.
Train and evaluate models at scale
The posting requires experience in large-scale model training or evaluation and proficiency with a deep learning framework such as PyTorch or TensorFlow.
Translate research into gameplay and workflows
Collaborate with developers, artists, and designers to integrate AI-generated content into gameplay. The role also explores research advances that could improve production efficiency and creativity; interactive applications are a preferred qualification.

Likely challenges to think through

Depth across a broad research remit
Preparation hypothesis: the listed domains are wider than any one research specialty, while the requirement asks for depth in at least one. Prepare to explain a focused contribution and how its methods might transfer to adjacent domains, without claiming expertise across all of them.
Research quality versus interactive usability
Preparation hypothesis: a promising generative result may be difficult to use in gameplay. Prepare an example of how you would test outputs with developers and artists, balancing visual or motion quality with responsiveness, consistency, and creative control.
Fast exploration versus credible evaluation
Preparation hypothesis: applying new research quickly may compete with the time and compute needed for large-scale evaluation. Prepare to describe how you would use relevant benchmarks and targeted prototypes to decide whether an advance merits further training or integration.

Practice your answers

Technical practice

How would you design a 3D object or scene generation pipeline for a game?

What to coverSuggested preparation: Explain your choice of representation and generation approach, then how you would assess geometry, appearance, controllability and suitability for an interactive scene. Discuss trade-offs between visual quality, editability and production cost.

How would you evaluate a video-generation model intended for world modeling?

What to coverSuggested preparation: Address temporal consistency, motion plausibility and whether generated sequences preserve the conditions needed for a coherent world. Explain where quantitative benchmarks help and where human review or downstream tests remain necessary.

What would you check when captured or generated human motion looks unnatural?

What to coverSuggested preparation: Walk through possible issues in input data, pose representation, temporal modeling and animation output. Explain how you would isolate the cause and evaluate improvements in both motion quality and usability.

How could a vision-language model support virtual-world creation?

What to coverSuggested preparation: Describe a plausible task connecting visual content with language-based reasoning or direction from creators, without presenting it as an existing Tencent system. Discuss grounding, failure cases and how you would verify that outputs serve the intended creative task.

How would you formulate an agentic reinforcement-learning problem in a simulated game environment?

What to coverSuggested preparation: Define observations, actions, objectives and an evaluation setup for a plausible agent. Discuss trade-offs involving reward design, exploration and transfer from simulation to interactive use.

How would you establish that a generative-AI research result is an original contribution?

What to coverSuggested preparation: Frame a clear research question, identify relevant prior work and explain the ablations or comparisons needed to support a claim. Distinguish a publishable advance from a change that mainly improves a particular production workflow.

How would you diagnose a large-scale model training run that is not improving?

What to coverSuggested preparation: Describe checks on data quality, loss behavior, implementation and evaluation before increasing compute. Explain how a small controlled experiment could distinguish an optimization problem from a modeling or measurement problem.

How would you select datasets and benchmarks for 3D, motion or video generation?

What to coverSuggested preparation: Choose the domain closest to your research and explain how coverage, data quality and benchmark metrics relate to the intended game-development use. Discuss potential gaps between benchmark performance and the quality creators observe.

How would you turn a research model into an interactive content-creation prototype?

What to coverSuggested preparation: Outline how you would expose useful controls, inspect outputs and test an iteration loop with artists or designers. If relevant to your approach, discuss where Python or C++ prototyping and tools such as Blender, Unreal Engine or Unity would help; these are listed as preferred experience, not a mandated stack.

How would you decide whether a new generative method is worth integrating into gameplay?

What to coverSuggested preparation: Compare research quality with creator control, interactive performance and effects on production efficiency or creativity. Propose a limited test with game developers, artists and designers, and explain what evidence would justify further integration.

Behavioural practice

Tell me about a time you collaborated with people outside your research specialty.

What to coverSuggested preparation: Use a real situation involving different goals or vocabulary, the actions you took to align on a technical direction, and the result. The role specifically calls for collaboration with game developers, artists and designers, so focus on how you incorporated others' feedback.

Describe a research effort where an initial approach failed or produced inconclusive results.

What to coverSuggested preparation: State the research situation, how you diagnosed the problem or revised the experiment, and what resulted. Explain what the evidence changed about your next decision rather than portraying the first approach as a success.

Give an example of communicating a complex research finding to a non-specialist.

What to coverSuggested preparation: Identify the audience and decision at stake, the explanation or demonstration you used, and the result. Show how you conveyed limitations as well as potential value.

Tell me about balancing a promising research direction against a practical delivery need.

What to coverSuggested preparation: Describe the competing goals, the criteria and conversations behind your choice, and the outcome. Connect the example to the posting's emphasis on original research alongside production efficiency and interactive applications.

Describe a time you evaluated a new research trend before applying it.

What to coverSuggested preparation: Use a real situation, the tests or comparisons you carried out, and the resulting decision. Explain how you separated a compelling demonstration from evidence that the approach would help the intended workflow.

Sharpen your knowledge

Topics to prepare

3D object and scene generation
The role develops generation systems for virtual worlds and names 3D generation as one possible research specialty. Suggested exercise: compare two approaches to representing and generating a scene, then explain how you would assess visual quality and usability in a game.
Video generation and world modeling
The posting names video generation, world modeling, and video understanding. Suggested exercise: outline a model evaluation for a generated sequence, including how you would check temporal consistency and whether the sequence represents a coherent world.
Human motion capture and generation
Motion capture or generation is another named research specialty. Suggested exercise: trace a motion pipeline from captured input to generated movement, identifying where you would check plausibility and continuity.
Vision-language models and multimodal reasoning
The role includes multimodal VLMs and applying advances in multimodal reasoning. Suggested exercise: design a vision-and-language task involving a game scene, then explain what evidence would show that the model interpreted both inputs.
Agentic reinforcement learning
The role includes agentic RL systems and exploration of simulation-based RL. Suggested exercise: define an agent's observations, actions, and objective in a simulated game environment, then explain how you would evaluate its behavior.
Original research for virtual-world pipelines
The posting calls for original research in computer vision, computer graphics, or embodied intelligence and lists top-tier publications. Suggested exercise: revisit one research contribution in your strongest listed specialty and articulate its question, method, evidence, limitations, and potential relevance to virtual-world creation. The posting requires depth in at least one listed specialty, not all of them.

Skills to revise

Large-scale training and evaluation
Proficiency with a deep learning framework such as PyTorch or TensorFlow and experience in large-scale training or evaluation are required. Suggested exercise: walk through a training or evaluation workflow you can substantiate, including its data checks, measures, and failure analysis.
Research prototyping
Coding and prototyping in Python, C++, or a similar language are preferred. Suggested exercise: sketch a small prototype for your strongest listed research area, specifying inputs, outputs, and a test that would reveal a failure.
3D tool workflow
Hands-on use of Blender, Unreal Engine, or Unity is preferred. Suggested exercise: use a tool you know to inspect how a generated 3D asset or scene could be reviewed and brought into an interactive experience; note any integration constraints you encounter.
Benchmark interpretation
Familiarity with datasets and benchmarks for 3D, motion, or video generation is preferred. Suggested exercise: review a benchmark relevant to your research area and explain what its measures capture, what they miss, and how you would inspect qualitative results.
Cross-disciplinary research communication
The role requires teamwork and communication with game developers, artists, and designers to integrate generated content into gameplay. Suggested exercise: prepare a brief explanation of a research result for each audience, covering the creative use, technical limitations, and evidence needed before integration.

Your preparation checklist

Select a research project to walk through
Prepare a concise walkthrough in one advertised specialty—such as 3D generation, video world modeling, motion generation, VLMs, or agentic RL—covering the research question, your contribution, results, limitations, and relevance to game development.
Organize evidence of research qualifications
Have your PhD field, relevant publications, and specific contributions ready to discuss. The posting asks for a PhD in a related field and publications in top-tier conferences or journals.
Explain training and evaluation decisions
Prepare one example of large-scale model training or evaluation using PyTorch, TensorFlow, or another framework you have used. Note the data, compute tradeoffs, metrics, baselines, and failure cases.
Practice a virtual-world design exercise
Sketch how you would prototype and evaluate one AI-generated asset, scene, motion, or world-modeling capability for a game. This is a preparation exercise, not a stated interview task.
Show how research could reach gameplay
If you have a relevant example, prepare a short demo or walkthrough of an interactive prototype, including any Python or C++ work and Blender, Unreal Engine, or Unity use. These tools and application experience are preferred, not required.
Rehearse a cross-disciplinary collaboration example
Prepare a specific account of working with developers, artists, or designers: the differing needs, decisions you made together, and how you communicated research limitations. The role explicitly involves this collaboration.
Prepare questions about research priorities
Ask which of the advertised areas are most pressing, how researchers work with game teams, and how the team evaluates generated content for gameplay and production use. These are questions to clarify, not assumed team practices.

Availability last checked: 4 Nov 2026. Status can change after a check; confirm on the employer’s website. Report an issue · Our checking policy

KEEP EXPLORING

Similar AI jobs

Related skills and specializations in Singapore. Matched to this role, not your profile.

Explore more jobs

Finding similar opportunities…