Genome foundation models
01DNABERT-2
Efficient genomic representations and a multi-species benchmark for understanding DNA sequences.
Publications · MAGICS Lab
RESEARCH RECORD / 2023—2026The science behindscientific intelligence.
From the mathematical foundations of learning to genome models, molecular design, and scientific agents. Research by Han Liu, students, and collaborators at Northwestern and beyond.
Selected contributions
Foundations, models, and systems.
Four entry points into the research.
Foundations of intelligence
02In-context algorithm emulation
Constructive guarantees for emulating attention-implementable algorithms through prompts in frozen Transformers.
Generative molecular design
03Energy-aligned antibody design
A diffusion-based framework for antibody sequence–structure co-design with Pareto-aligned energy objectives.
Scientific research agents
04SciSciGPT
A multi-agent collaborator connecting literature, scholarly data, executable analysis, and evaluation.
Recent publications
48 core ML papers · 8 additional publications
2023–2026 · Updated September 2026
The default view brings together every NeurIPS, ICML, and ICLR entry from the 2023–2026 CV record. Switch to all venues to explore the broader body of recent work.
48 papers · NeurIPS, ICML & ICLR · 2023–2026
2026
23 papersLikelihood-Free Generative Policy Optimization
Generative models · Agents & embodimentFind paper ↗
On Differential Private ℓ₁, ℓ₂ and ℓₚᵖ Distance Queries
Learning foundationsFind paper ↗
Flow Matching Reinforcement Learning via SDE Inference
Generative models · Agents & embodimentFind paper ↗
On JEPA Isotropy
Learning foundationsFind paper ↗
A Theory on Flow Matching with Neural Networks
Generative modelsPaper ↗
Chain-of-Thought Gradient Descent
Learning foundationsPaper ↗
On Structured State-Space Duality
Learning foundationsPaper ↗
Learning Manifold Data with Flow Matching
Generative modelsPaper ↗
Hyperbolic neural population geometry benefits computation
Learning foundationsFind paper ↗
Universal Approximation with Softmax Attention
Learning foundationsPaper ↗
Decoupled Alignment for Robust Plug-and-Play Adaptation
Agents & embodimentFind paper ↗
Towards Sparse Video Understanding and Reasoning
Agents & embodimentFind paper ↗
SciSciGPT: advancing human–AI collaboration in the science of science
AI for science · Agents & embodimentPaper ↗
In-Context Algorithm Emulation in Fixed-Weight Transformers
Learning foundationsPaper ↗
Sci2Pol: Evaluating and Fine-tuning LLMs on Scientific-to-Policy Brief Generation
AI for science · Agents & embodimentPaper ↗
2025
13 papersPareto-optimal energy alignment for designing nature-like antibodies
Generative models · AI for sciencePaper ↗
Attention Mechanism, Max-Affine Partition, and Universal Approximation
Learning foundationsPaper ↗
In-Context Deep Learning via Transformer Models
Learning foundationsPaper ↗
Nonparametric Modern Hopfield Models
Learning foundationsPaper ↗
In-Context Learning as Conditional Associative Memory Retrieval
Learning foundationsPaper ↗
Computational Limits of Low-Rank Adaptation for Transformer Models
Learning foundationsPaper ↗
2024
10 papersOne-Layer Transformer Provably Learns One-Nearest Neighbor In Context
Learning foundationsPaper ↗
Global Convergence in Training Large-Scale Transformers
Learning foundationsPaper ↗
Outlier-Efficient Hopfield Layers for Large Transformer-Based Models
Learning foundationsPaper ↗
Uniform Memory Retrieval with Larger Capacity for Modern Hopfield Models
Learning foundationsPaper ↗
DOS: A Deployment Operating System for RoboOps
Agents & embodimentFind paper ↗
2023
2 papersNo matching papers.
Try another title, author, or research area.
About this publication record
Scope: 2023–2026, inclusive. The core ML count includes 48 unique NeurIPS, ICML, and ICLR entries; the all-venues view adds eight other recent publications. Featured papers are not counted twice. These are coauthored publications, not claims of sole authorship.
The source is Han Liu’s supplied 2026 Curriculum Vitae, Recent Articles R1–R56. Conference years and acceptance/presentation labels follow that record. Selected author and title inconsistencies were reconciled against linked publisher or paper records; this is not an independent verification of every acceptance. Research-area labels are editorial navigation aids. An asterisk denotes equal contribution where specified in the CV.
“Paper” links lead to located paper or publisher records. “Find paper” links perform an exact-title Google Scholar search where a direct record was not located. SciSciGPT is listed under its 2026 journal volume; its first online publication was December 2025. For the broader career record, visit Google Scholar or the earlier publications archive.
From papers to practice
Explore the systems behind the research.
Genome foundation models, research collaborators, protein-evolution policies, and executable environments.