Generative AI

A model that can generate plausible text, images, or molecules has learned something about the structure of its domain. Our generative AI research asks what, exactly, it has learned — and how that knowledge can be made controllable, efficient, and reliable.

Research topics

  • Foundation models
  • Large language models
  • Diffusion models
  • Multimodal generation
  • Generative representations
  • Model efficiency
  • Generative system reliability
Fig. 03Schematic — not experimental data
t=1.00t=0.75t=0.50t=0.25t=0.00
Fig. 03Schematic. Unstructured samples (left) are progressively transformed into an ordered representation (right), as in iterative generative processes.

Why it matters

Why it matters

Generative models are increasingly used as general interfaces to computation: they write code, summarize evidence, design candidate structures, and act as components inside larger systems. Their usefulness depends on properties that are not captured by fluency alone — factual grounding, consistency, controllability, and an honest account of their own uncertainty.

They are also scientific objects in their own right. A model trained to generate data implicitly represents a probability distribution over that data. Understanding the shape of that distribution tells us both what these systems know and where their knowledge runs out.

Key research questions

Key research questions

  1. Q1

    How do generative models represent compositional structure, and why does composition fail when it does?

  2. Q2

    Can we detect, at generation time, when a model's output is unsupported by what it has learned?

  3. Q3

    What are the trade-offs between autoregressive, diffusion, and hybrid approaches across modalities?

  4. Q4

    How can generation be steered toward constraints — physical, logical, or stylistic — without retraining?

  5. Q5

    Which parts of a large model's capacity are necessary, and what can be removed without loss?

Primary subfields

Primary subfields

Foundation models
Large models pretrained on broad data and adapted to many downstream tasks.
Diffusion and flow models
Generating data by learning to reverse a gradual corruption process.
Multimodal generation
Joint models of language, images, audio, and structured data.
Controllable generation
Conditioning, guidance, and constraint satisfaction during sampling.
Efficiency
Distillation, quantization, sparsity, and faster sampling.

Representative methodologies

Representative methodologies

  • M1

    Likelihood and sample-based evaluation

    Combining distributional measures with targeted human and automated review.

  • M2

    Mechanistic analysis

    Tracing how specific outputs arise from internal computation.

  • M3

    Constrained decoding

    Imposing verifiable structure on generation at inference time.

  • M4

    Compression studies

    Measuring which capabilities survive aggressive reduction in model size.

Relationships to other domains

Relationships to other domains

Working on questions in Generative AI?

We welcome conversations with researchers and organizations working on related questions — whether to compare notes, discuss a collaboration, or share work in progress.