VAST Aims to Advance AI Research with Young Talent through V·STAR Top Talent Program

·

A recent trend in the AI industry is placing increasing emphasis on actual contributions rather than just potential, presenting a dilemma for young researchers: companies want people who have already produced results, but opportunities to participate in core projects are becoming scarce. This has introduced another criterion into the competition for talent: not only do companies evaluate who has already produced results, but young researchers also need to assess who can provide the conditions necessary for their next achievement.

VAST, an artificial intelligence company that secured approximately 5 billion yuan in funding in less than half a year, recently launched a new edition of its V·STAR Top Talent Program. The program invites young researchers to focus on three specific research tasks: 3D Model Generation, 3D Perception and Agents, and World Models.

These directions all point toward the same goal: moving AI from generating content to building interactive worlds. Being able to write code and generate images and videos does not necessarily mean it can build an environment that runs continuously. The team chose to explore specialized 3D training, enabling the model to learn about shape, topology, and component relationships.

The V·STAR Top Talent Program seeks young researchers who can help define problems, make technical judgments, and turn their explorations into tangible results. Cao Yanpei, Chief Scientist at VAST, conducted an experiment using Tripo P2.0 to generate a robotic spider, then had an LLM name the individual components, write and debug the motion logic, and ultimately achieve component-level animation control.

VAST’s research focuses on addressing unresolved issues in 3D generation, such as editable and manipulable objects, agents that understand space and execute tasks, and world models that maintain their state and rules. The team is working to shift some of the trial-and-error typically encountered during production to the training phase, making generated objects easier to recognize, manipulate, and further modify.

The Nexus framework underlies the Tripo P-series models and attempts to resolve issues in mesh generation by decoupling vertex generation from topology generation. This paradigm has been recognized by leading international conferences, including SIGGRAPH, where a paper on Nexus was accepted this year.

Discussion

No messages yet. Be the first to write.