Diffusion, flow, and diffusion language models for design search on a static dataset, including calibrated estimation, support-proximity regularization, multi-objective guided flows, and settings where labels or experiments are expensive.

Theme 02

Agentic Reasoning, Search, and Policy Optimization

Lead · Bowei He

Benchmarks, training methods, and runtime frameworks for language agents that retrieve evidence, use tools, and learn from process-level rewards, including search-integrated reasoning, branching policy optimization, and scaling from one device to collective systems.

Theme 03

Trustworthy Alignment and Learning from Human Feedback

Lead · Haolun Wu

Aligning single models and compound AI systems with human preferences, including system-level direct preference optimization, logit-only adaptation of closed models, internal-representation safeguards, and human-centered control of model behavior.

Theme 04

Personalized Retrieval and Human-centered Information Access

Lead · Haolun Wu

Representing evolving user interests and keeping retrieved content scrutable, including density-based user modeling, retrieval-augmented personalization, interpretable preference heads, and diversification in search and recommendation.

Theme 05

LLM-enhanced Recommendation, RAG, and Structured Understanding

Lead · Bowei He

Language models as rankers and sequential recommenders, grounded in entities, tables, and retrieved documents. This includes mutual augmentation between recommenders and LLMs, context-aware contrastive learning, embedding-based reranking, agentic table summarization, and retrieval-augmented generation.

Theme 06

Multimodal Understanding and Generation

Lead · Ye Yuan

Vision-language models and generative systems that read and produce across text, images, documents, and other discrete modalities, including multimodal retrieval, grounded multimodal agents, and diffusion models that move from tokenization through generation.

Theme 07

Computer-Using Agents and Interactive Environment Grounding

Lead · Jikun Kang

Autonomous agents operating directly across digital and desktop environments through multimodal perception, action space synthesis, and trajectory-level planning. This includes computer-using agents (CUAs), UI navigation, tool use, and learning grounded behavioral policies for complex interactive tasks.

Theme 08

Agent Memory Architectures and Reinforcement Learning

Lead · Jikun Kang

Stateful long-horizon memory representations, policy optimization, and rigorous evaluation harnesses for compound agentic systems. This includes episodic and working-memory routing, sample-efficient reinforcement learning, trajectory-level auditing, multi-step credit assignment, and life-long adaptation in open-ended dynamic environments.

03 / The community

People

Meet the researchers Join us

New Ground AI is a research collective spanning universities, labs, and industry. Members work at or have trained at Stanford University, McGill University, Mila, MBZUAI, MIT, the University of Cambridge, Imperial College London, City University of Hong Kong, and more, as well as at places such as Salesforce, RBC Borealis, Wayfair, and Next Pathway.

The group includes undergraduates, graduate students, postdoctoral researchers, and engineers. Together they study generative models, agentic systems, retrieval, alignment, and multimodal intelligence, and carry that work into production systems.

Meet our researchers, and Welcome to join us, to build advanced, analytical and adaptive intelligence together.

04 / Support

Technology & Community Support

Join us

Lambda
Hyper AI
Thinking Machines Lab
GreatRouter
Cohere Labs

Welcome to join us

Our community has received credits, resources, or program support from the organizations shown above. Inclusion does not imply formal endorsement or affiliation unless otherwise stated.