Active
Perceptual Grounding for Tool-Using Agents
Agents that operate visual interfaces often act on what they think they saw. This project studies grounding as a discipline: before an agent clicks, types, or submits, it must produce a checkable reference to the element it intends to act on — coordinates, labels, and a fresh verification pass. We're interested in where grounding still fails, and what those failures reveal about visual attention.
Research questions: Which verification steps actually reduce action errors, and which just add latency? How should an agent behave when the interface changes mid-task? What does "certainty about a UI element" even mean?
Active
Calibrated Visual Uncertainty
Most vision systems answer with the confidence of a witness who never blinks. We study uncertainty as a communicable quantity: how a perceptual system can express "I'm not sure what I'm looking at," and how downstream agents should modulate their behavior in response — slowing down, asking, or abstaining rather than guessing.
Research questions: Can uncertainty be calibrated without ground truth at runtime? How should agents trade off caution against task completion? What interfaces let humans read machine uncertainty at a glance?
Active
Memory Architectures for Long-Horizon Agents
Long tasks break agents at the seams of memory: they forget the goal, repeat themselves, or drown in their own context. This project compares memory designs — episodic traces, semantic summaries, working-memory discipline — and asks the harder question underneath: what should an agent remember, and what is it better off forgetting?
Research questions: How does memory design affect error recovery? When does remembering more make an agent worse? Can we characterize the "forgetting" that skilled human operators do naturally?
Active
A Failure-Mode Taxonomy for Visual Agents
Slow, close reading of agent mistakes — hundreds of them — to build a taxonomy of how visual agents fail. Misread states, phantom affordances, attention collapse, confident nonsense: naming failure modes precisely is the first step toward measuring them, and the second step toward preventing them.
Research questions: Which failure modes are fundamental versus incidental? Do different architectures fail in systematically different ways? Can a taxonomy guide evaluation design?
Exploratory
Embodied Observation
Early-stage exploration of perception coupled with movement: what changes about "seeing" when the observer can move to see better? We're probing active perception — choosing viewpoints, approaching ambiguous objects, and treating the camera as something the agent controls rather than something it passively receives.
Research questions: Does active viewpoint selection reduce ambiguity in practice, or just add complexity? What decision rules govern "look closer" versus "move on"?
Exploratory
Cross-Modal Alignment Audits
A young line of work asking an uncomfortable question: when a multimodal system describes what it sees, how often is the description actually grounded in the image? We're developing audit procedures that stress-test vision-language consistency — not to produce a leaderboard, but to understand the gap between seeing and saying.
Research questions: How do we detect ungrounded description without human review of every output? What systematic biases appear in cross-modal translation?
Archived
Sparse Vision Benchmarks, Revisited
An early project that re-examined how vision benchmarks sample their test distributions. The work concluded that the questions it raised belonged to evaluation methodology rather than benchmark construction — so the project was archived and its insights folded into our evaluation research. The notes remain worth reading.
Outcome: Reframed into the evaluation methodology research area; core observations published as lab notes.
Archived
Scene-Graph Sketches for Agent Planning
An investigation into whether explicit scene-graph representations help agents plan in visual environments. The finding: for the tasks we studied, the representation helped less than expected, and the maintenance cost outweighed the benefit. Archived as a negative result — which is to say, archived as a result.
Outcome: Negative result documented; lessons carried into the memory architectures project.