Research Overview - Zhen Wang

My research sits at the emerging intersection of advanced AI reasoning and accelerated discovery across science and society. This space extends traditional data-driven discovery, calling for AI systems that actively explore hypotheses, simulate possibilities, and generate new, verifiable knowledge across disciplines.

I approach this field by identifying new foundational problems with rigorous, reproducible benchmarks and evaluation protocols (scBench, FIRE-Bench, LLM Reasoners), and by building systems that reason, simulate, and interact directly with complex data, code, tools, literature, and human collaborators. I validate these systems end-to-end across biology, materials science, machine learning research, and human-centered domains, with novel discovery results including new cancer driver genes, new materials, and new scientific insights.

My research agenda comprises three tightly connected pillars:

Reasoning

Reliable structured reasoning is the prerequisite for enabling autonomous discovery at scale. My contributions in this area include structured reasoning frameworks that facilitate probabilistic reasoning (ThinkSum, ACL'23), planning with textual world models (RAP, EMNLP'23; LLM Reasoners, COLM'24), simulation-based reasoning and strategic planning with language models (PromptAgent, ICLR'24), and dynamic, test-time optimization techniques (Nabla Reasoner).

Science

Science drives my research because its open-ended complexity offers the most rigorous testbed for reasoning systems and the clearest path toward socially meaningful progress. I operationalize this vision by training scientific foundation models (MutationProjector), building autonomous scientific copilots (scPilot, NeurIPS'25), creating agents capable of end-to-end research automation and insight rediscovery (FIRE-Bench), and pioneering systematic AI augmentation of scientific inquiry (TaijiChat in Multi-omics atlas for T-cell programming, Nature).

Society

True discovery is inherently human-centered, demanding scalable governance, human-aligned AI behavior, and deep societal simulation to ensure that AI-generated insights authentically serve societal progress. My research addresses these goals by developing scalable self-alignment and safe behavior (Dynamic Rewarding, EMNLP'24), decentralized and democratic evaluation frameworks (Decentralized Arena), and sophisticated human-behavior simulations for personalized, socially informed discovery (DeepPersona).

Research Keywords & Technical Expertise

My technical expertise spans the following interconnected areas:

Reasoning & Planning Methods

Reasoning with language models, planning as inference, Monte Carlo Tree Search (MCTS), textual world models, probabilistic reasoning, test-time compute optimization, structured reasoning frameworks, neuro-symbolic methods

Scientific Discovery & Automation

Automated scientific discovery, AI agents for science, foundation models for biology, single-cell transcriptomics analysis, computational biology, automated hypothesis generation, omics-native reasoning, scientific workflow automation, materials discovery

Adaptation & Efficient Learning

Parameter-efficient fine-tuning (PEFT), prompt optimization, multi-task learning, transfer learning, few-shot adaptation, tool-augmented reasoning, mixture-of-experts (MoE), test-time adaptation

AI Governance & Alignment

Scalable AI alignment, scalable oversight, decentralized evaluation, self-alignment (without human labels), AI safety benchmarking

Human-Centered AI & Social Modeling

Human behavior simulation, synthetic personas, social simulation with LLMs, LLM personalization, human-AI interaction, computational social science, agent-based modeling

Knowledge-Centric NLP & Graph Learning

Graph neural networks (GNNs), knowledge graphs, retrieval-augmented generation (RAG), knowledge-centric NLP, structured representations, knowledge acquisition, domain adaptation, biomedical NLP, graph embedding methods