Research

The questions we're living with

Four research areas, each anchored by durable questions rather than deadlines. Below: what we're asking, how we're approaching it, and what remains genuinely open.

AREA 01

Computer Vision

Modern vision systems recognize with remarkable fluency — yet recognition is not comprehension. Our vision research asks what it would mean for a machine to understand a scene: its structure, its affordances, its ambiguities, and the limits of what can be known from pixels alone.

What does a vision model actually "know" about a scene, versus what it can merely label?
How should perceptual systems represent uncertainty — and communicate it to whatever acts on their output?
What breaks when vision models encounter the long tail: rare objects, strange lighting, adversarial clutter?
Can we characterize what a model was never shown, so its blind spots become legible instead of silent?

Approaches

Probing vision models with carefully constructed edge cases rather than only aggregate benchmarks.
Studying uncertainty representation — calibrated confidence, abstention, and "I don't know" as a first-class output.
Error analysis as a craft: taxonomies of failure modes built from close reading of mistakes.

Open problems

Scene understanding that survives occlusion, viewpoint change, and clutter without brittle shortcuts.
Principled ways to audit a vision system's training provenance and its consequences.
Evaluation that rewards genuine comprehension over dataset-shaped pattern matching.

AREA 02

Agent Enhancement

Agents that plan, use tools, and work over long horizons are becoming practical — and their failure modes are becoming consequential. We study how to make agents more capable and more legible: able to do more, while remaining inspectable, correctable, and honest about their limits.

How do agents maintain coherent plans over long horizons without drifting, looping, or quietly abandoning the goal?
What should an agent remember — and what should it deliberately forget — across a long task?
How can tool-using agents verify their observations before acting on them, especially from visual interfaces?
What does trustworthy autonomy look like: not just success, but success for the right reasons?

Approaches

Building small, instrumented agents in constrained environments where every decision can be traced and studied.
Experimenting with memory architectures — episodic, semantic, and working memory — and their trade-offs.
Designing verification habits: grounding, cross-checking, and explicit uncertainty before action.

Open problems

Recovery: agents that notice their own mistakes and revise course, rather than compounding them.
Calibration of self-assessment — agents that know when to ask for help or stop.
Evaluation of long-horizon behavior that captures process quality, not just final outcomes.

AREA 03

Multimodal Perception

The world doesn't arrive in one modality. Agents that see, read, listen, and act must fuse these streams into a single coherent picture — and decide what to do when the streams disagree. We study the seams between modalities, where the most interesting failures live.

When vision and language conflict, which should an agent trust — and how should it decide?
How do agents build and maintain a unified world model from partial, asynchronous sensory streams?
What is lost in translation when continuous perception is compressed into tokens for reasoning?
Can cross-modal grounding make hallucinations rarer — or at least more detectable?

Approaches

Constructing scenarios where modalities genuinely disagree, then studying resolution strategies.
Probing how much spatial and temporal detail survives the journey from pixels to tokens to actions.
Cross-modal consistency checks as a built-in habit of perceptual systems.

Open problems

Principled fusion: knowing when to combine modalities and when to keep them separate.
Temporal coherence across long multimodal streams — video, dialogue, and action over time.
Grounding language in perception tightly enough that descriptions stay checkable.

AREA 04

Evaluation Methodology

A field that can't measure itself can't improve itself. Benchmarks shape research directions — sometimes productively, sometimes pathologically. We study evaluation itself: how to design measurements that reward genuine capability, resist gaming, and stay meaningful as systems change.

What do our benchmarks actually measure — and what important capabilities do they systematically miss?
How do we evaluate open-ended agent behavior where there is no single right answer?
How can evaluations stay valid when the systems being evaluated train on the internet that hosts the evaluations?
What would it mean to evaluate process — the quality of reasoning — rather than just outcomes?

Approaches

Dissecting existing benchmarks to understand what they reward and what they overlook.
Prototyping small, opinionated evaluations for specific capabilities we care about.
Studying human evaluation protocols — rubric design, rater calibration, and their limits.

Open problems

Contamination-resistant evaluation in an era of web-scale training data.
Measuring robustness and graceful degradation, not just peak performance.
Evaluations that remain interpretable as systems grow more capable and more opaque.

From questions to experiments

See how these areas become projects

Each research area spawns concrete lines of inquiry — documented with their status, questions, and methods.

Browse the projects