More output does not mean more taste.
When generation is cheap, the scarce work is structuring the design space: what dimensions matter, what to vary, what to preserve, and how to justify a direction.
Design judgment + human-centered AI
Making AI tools that sharpen design judgment.
I build and study interfaces for an age of generative abundance: tools that help people explore alternatives, notice consequential differences, explain tradeoffs, and build taste over time.
Cognitive Science Ph.D. student at UC San Diego Advised by Prof. Steven P. Dow in ProtoLab.
Thesis thread
Scaffolding taste in an age of generative abundance.
Scaffolding for Taste asks a simple question: when AI can make ten plausible options in a minute, how do people learn to notice the difference that matters? I study interfaces that support reflective practice: making a move, seeing what changed, comparing alternatives, and building better reasons for choosing one direction over another.
Design principle
Design help should not just make more artifacts. It should make the next move easier to see, and the next judgment easier to explain.
Don Norman's Design Lab talk helped sharpen this framing: the next generation of designers needs tools that cultivate taste, not just tools that generate options.I am testing this frame across creative tools, learning environments, mixed reality, everyday robotics, and community or civic design settings: places where tools can help people notice viable options, weigh tradeoffs, and choose the next move.
Why now
Why this work matters now
AI makes it easy to produce variants; my question is how people learn to choose well.
Everyone is closer to the prototype.
Designers, researchers, PMs, and engineers all shape early artifacts now. Interfaces need to support shared reasoning before decisions harden.
Speed still needs reflection.
Fast cycles are useful only if teams keep room for critique, incubation, and explaining why a direction is better.
Research focus
Design, Evaluate, Situate.
I keep testing the same loop: how to make alternatives comparable, let evidence sharpen the next question, and fit assistance to the situation where judgment happens.
Showing the Design motion sketch. Vary the alternatives, compare them, then refine.
Design lens
Human-AI for design thinking
I study interfaces that support design thinking beyond artifact generation: dimensional scaffolds that make the design space visible, variation mechanisms that expose alternatives, structured comparison that helps people reason across options, and staged fidelity that lets ideas mature from rough intent to defensible direction.
See DesignWeaver Moves
- 01
Externalize useful dimensions so people can see what matters.
- 02
Use variation and structured comparison to reason across alternatives.
- 03
Move from coarse ideas to committed artifact choices through staged fidelity.
Evaluate lens
Evaluation for design thinking
I evaluate AI tools with evidence between controlled psychology tasks and outcome-only benchmarks: process traces, artifact changes, rationales, preservation and explanation behavior, and learning probes that show what users carry beyond the interface.
Read CHI 2026 workshop paper Evidence loop
- 01
Treat final artifact quality as one layer of evidence.
- 02
Connect traces and rationales to concrete design questions.
- 03
Measure explanation, preservation, revision, and transfer.
Situate lens
Interfaces fitted to context
I use situated work for settings where judgment depends on practice and medium: studio critique, mixed reality, everyday robotics, and civic or urban projects each change what help should do, when it should appear, and when it should stay out of the way.
Situations
- 01
Let practice shape what assistance should do.
- 02
Match the interface to the medium and setting.
- 03
Make help fit the situation instead of interrupting it.
Exploring
Adjacent contexts for the same loop
These threads are places I test the same concern with judgment, agency, and evidence in context.
Situate Everyday robotics Integrating intelligence into robots so they can better assist people in everyday living and working contexts. Design Learner-centered AI Scalable support that keeps agency and the learning process visible. Evaluate Post-deployment iteration Using traces from real use to improve benefit and catch harm. Situate Community-centered tools Coordination and sensemaking for groups with shared ownership.
Motion note Inspired by motion craft from Stripe's design team and Katie Dill at Stripe Sessions. Interaction rules grounded by Jeffrey Heer and collaborators on animated transitions and CMU's Data Interaction Group .
Live field The visible Dot Orbit layer uses Paper Shaders by Paper; the restraint also follows the founder's design walkthrough.
Selected publications
Papers worth opening first.
- Short form Workshop First author 0 citesIn Herding CATs: Making Sense of Creative Activity Traces (CHI 2026 Workshop), Apr 2026Workshop Position Paper
When this paper helps.
This paper helps when you can see what someone did in an AI tool but still cannot tell what that action meant to them.
Look for- You are designing lightweight feedback at the moment a creative workflow becomes ambiguous.
- You need a unit of analysis that joins a short trace, interface state, and a user's optional explanation.
It addsA proposal for trace-guided micro-episodes and a utility-for-rationale pattern that turns useful recovery controls into chances to learn why something happened.
Its limitIt is a position paper and research agenda, not an empirical validation of the interventions or their downstream effects.
I checked this against the paper and its official record on 2026-07-13.
Read the full evidence and citation context - CVPR
Full paper Coauthor 18 citesIn Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition (CVPR), Jun 2025HighlightWhen this paper helps.
This paper helps when an implicit-surface argument needs more than the usual eikonal condition—especially around convergence, stability, topology, or surface area.
Look for- You are comparing objectives for reconstructing signed distance functions from unoriented points.
- You need theoretical and experimental evidence about a screened-Poisson alternative.
It addsA heat-loss objective with convergence and stability analysis, plus 2D and 3D reconstruction evidence and a natural surface-area penalty.
Its limitThe reported gains belong to the evaluated reconstruction settings; sparse boundaries and parameter choices can still produce failure modes.
I checked this against the paper and its official record on 2026-07-13.
Read the full evidence and citation context - CHI
Full paper First author 41 citesIn Proceedings of the 2025 CHI Conference on Human Factors in Computing Systems, Apr 2025When this paper helps.
This paper helps when a blank prompt box is asking novices to supply design vocabulary they have not had a chance to learn.
Look for- You are designing dimensional scaffolds for text-to-image exploration.
- You want evidence about prompt language, visual diversity, novelty, and rising expectations.
It addsA palette that externalizes product-design dimensions, grounded in expert practice and evaluated with 52 novice designers.
Its limitThe study used one chair-design task with novices, and richer prompts did not reliably improve satisfaction or expectation alignment.
I checked this against the paper and its official record on 2026-07-13.
Read the full evidence and citation context - NeurIPS
Full paper Coauthor 187 citesIn Proceedings of the Neural Information Processing Systems Track on Datasets and Benchmarks, Dec 2021When this paper helps.
This paper helps when you need a reproducible benchmark for asking whether a model predicts physical outcomes in human-like ways.
Look for- You are comparing object-centric and non-object-centric visual models with human judgments.
- You need a common contact-prediction task across several simulated physical scenarios.
It addsEight scenario families, matched human and model evaluation, and released data and code for systematic comparison.
Its limitSynthetic scenes and binary contact prediction do not establish broad real-world physical understanding.
I checked this against the paper and its official record on 2026-07-13.
Read the full evidence and citation context
Updates
Recent signals.
Pinned Met the great Don Norman and even got a selfie + a “To Sirui” signed “Yellow Book”! 
Our CHI 2026 workshop position paper, What Happened and Why? Trace-Guided Micro-Episodes with Elicited User Explanations for Product Iteration, was accepted to Herding CATs - easily one of the best workshop names ever! The whole ProtoLab is going to Spain! Might be a UCSD party! 
HotSpot got selected as a CVPR 25 Highlight!
Research opportunities
For students who like messy questions.
I like working with curious, motivated, and kind undergraduate and master's students, especially people who have a question they cannot stop poking at.
You do not need to show up as a polished researcher. It helps if you like reading carefully, making small prototypes, testing claims, looking honestly at evidence, and writing clearly about what changed.
The best fit is someone with real stake in a domain or problem, plus enough patience to turn that interest into a concrete study.
Reading closely Prototyping Evaluation Analysis Communication
Start here
- Read the research-start post.
- Email s1tao@ucsd.edu with subject "UCSD Research Interest".
- Include a 1-page CV or resume, an optional portfolio link, and a 3-5 sentence note on the questions or domains you care about.
Connect
Find me around the web.
Advisor Lineage & Past Collaborations
Ph.D. Steven P. Dow HCI
Master's Steven P. Dow HCI Tzu-Mao Li Graphics
Undergrad Judith E. Fan Cognition & Intuitive Physics
Build ledger · since May 22, 2026 Site-build tokens and agent-hours are estimated from retained Codex logs. The commit count is exact from this repository's Git history. Energy is an estimated equivalence using a midpoint of 0.0006 Wh per token, not measured electricity use.
- 12.9B
Burnt ~$9.8K of Sam's money* *Site-build tokens at public API rates—not a real bill. - site-build tokens
- 824
- agent-hours
- 500
- site commits
- 7758
- est. kWh