My ongoing research spans the following three directions. For published work, see Publications.

🧬 Multimodal Reasoning & Domain Agents

How can multimodal agents reason reliably with domain knowledge as evidence and environments change?

Species Identification and Discovery

Studies whether an agent can jointly use image, DNA, and text evidence to identify known species and discover unknown ones.

Automated Species Knowledge Base

Asks whether an expert-grade species knowledge base can be built and kept accurate automatically, with little manual curation.

Polymer Mechanism Discovery

Studies whether an agent can predict organic-polymer properties and discover their underlying mechanisms, and keep improving as it reads new literature.

Navigation in Dynamic Environments

Investigates how an agent can follow language instructions to navigate environments that keep changing around it.

🔁 Recursive Self-Improvement

How can self-evolving agents improve both their capabilities and the processes that drive further improvement?

VisualEvolve: Self-Evolving Visual Reasoning

Studies whether a vision-language model can improve its own visual reasoning without any additional training.

Open-Ended Self-Evolution

Asks whether an agent can keep evolving open-endedly over long horizons, exploring many solution paths instead of settling early.

Experience Transfer for GUI Agents

Studies whether a GUI agent can carry experience from one task to the next, so skills learned in one workflow help in others.

🤝 Trustworthy Multi-Agent Systems

How can agents collaborate reliably over long horizons, including in adversarial settings?

Long-Horizon Multimodal Benchmark

Studies how to build a hard, automatically checkable benchmark for long-horizon, interactive multimodal reasoning.

Reliable Long-Session Reasoning

Studies whether an agent can stay reliable while reasoning over long, multi-session interactions without forgetting earlier context.

Decentralized Agents across Game-Theoretic Settings

Asks whether decentralized agents can use memory to evolve interaction strategies across cooperative, mixed-motive, and adversarial settings.

Privacy-Preserving Multi-Agent Safety

Studies whether malicious agents can be detected while keeping each agent’s own messages private.

💬 Get in Touch

I’m always glad to discuss any of these directions. Feel free to reach out at jiaweiwang@nus.edu.sg. Students interested in joining are welcome to see the internship opportunities.