Huanxing Chen

Synonymix: Unified Group Personas for Generative Simulations

Accepted to CHI Extended Abstract // 2026

With Aditesh Kumar

‘High-fidelity’ generative simulations must choose between seeding agents with (a) rich, individual personas that compromise participant privacy or (b) reductive population-level statistics that smooth over individual heterogeneity. We explore whether abstracting to the group might offer a middle ground - retaining individual nuance while preserving privacy, striking a balance between the rich-contextuality of the individual-micro and the abstracted-scalability of the societal-macro. Synonymix operationalizes this via the "unigraph" - a unified graph representation merged from individual life stories graphs - that is then traversed to generate a synthetic agent bank with agents that are composite of many but copy of none. Our evaluation shows meaningful behavioral signal preservation with preliminary privacy guarantee.

The ‘unigraph’ opens up 2 questions: Is the group just a collection of discrete individuals? And, what could simulations be used for, beyond prediction?

Manuscript

WorldAgents: Survey-Anchored Agents for Cross-Cultural LLM Simulation

Preparing for PNAS Journal Submission

With Jonne Kamphorst, Michael Bernstein, Rob Willer

Large language models (LLMs) are increasingly used to simulate human populations for social science and policy research. However, growing evidence indicates that LLMs encode a systematic Western bias rooted in the composition of their training data. Applied to population simulation, this bias would be expected to produce both lower accuracy and artificially reduced diversity in the responses of simulated individuals from non-Western societies. Here, we first show that the standard approach — prompting LLMs with basic demographic characteristics such as age, gender, and education — indeed produces both distortions, and that both worsen as societies become more culturally distant from the United States. We then introduce survey-anchored agents, a new approach that conditions LLMs on representative survey responses rather than demographics alone. Using the World Values Survey, which covers 51 societies and 275 questions across 13 thematic domains, we compare survey-anchored agents against country and demographic baselines. Survey-anchored agents systematically achieve higher predictive accuracy and produce opinion diversity closer to that observed in human populations, with the largest gains in the most culturally distant countries. This work provides a scalable framework for building culturally robust generative agents, and suggests that anchoring artificial populations in representative surveys can help ensure that LLM-based research and applications generalize across diverse social contexts.

Generativity as a Design Principle for Human-AI Ideation

Under Review at ACM Creativity & Cognition Poster // 2026

With Mickey Mcmanus

LLM-assisted ideation tools help users explore solution spaces to overcome design fixation, but exploring more solutions is not the same as understanding the problem differently. For wicked problems, where no definitive problem formulation exists, the latter is crucial to creative cognition. Current tools employ descriptive structure - structure that organizes outputs for evaluation within presupposed problem frames. We propose generativity as a design principle for productive structure - scaffolding that engages users at the level of the dimensions through which problem and solution are jointly structured, shifting both the mode of cognitive work from evaluation to interpretation and its site from solution to problem frame itself. Grounded in morphological analysis, design co-evolution theory, and improvisational practice, we extract three design principles and propose an improvisation partner interaction metaphor - offering a theoretical foundation for human-AI thought partnership adequate to wicked, societal challenges.

Human-Simulation Interaction: From Prediction to Exploration in LLM Agent Simulation for Policy

Accepted to CHI PoliSim Workshop // 2026

Agent-based models have historically served as tools for generative explanation, constructing testbeds in which candidate micro-level behavioral rules can be tested for their capacity to produce observed macro-level phenomena. The integration of Large Language Models into agent-based simulation has expanded what these models can represent, but it has also introduced an unexamined shift in how users engage them. We argue that current generative agent-based models (GABMs) inherit the dominant interaction metaphor of conversational LLM interfaces - a question-answer pattern that positions users as consumers of system output rather than explorers of a possibility space. In the context of policy, where problems are wicked and ground truth is unknowable in advance, this metaphor produces a trust deficit that cannot be resolved through improved model accuracy alone. We open a design space we call human-simulation interaction, and argue that warranted trust requires interaction metaphors that restore the exploratory capacity simulation has historically supported.

Designing 'Human' Digital Experiences

January 30, 2024

The most 'human' experiences are those that are the most non-scalable.

How do we design for not just digital 'rooms' or 'galleries', but digital 'pubs'?

What does it mean for an experience to feel 'human'?

Read on Substack

Lower the Resolution

MAR 11, 2026

Generativity: The ability of a system to create unbound novelty from some initial seed of simplicity

Can we leverage it to make visible the unknown unknowns?

Generative system vs Generative human system

Read on Substack

In Praise of the Meso

FEB 03, 2026

An exploration of the meso-level — the undertheorized middle between the individual and the population. Drawing on complexity science, this essay argues that the most interesting phenomena in social systems emerge at intermediate scales, where local interactions give rise to global patterns without being reducible to either.

Read on Substack

Can LLMs Innovate?

JUL 14, 2025

===
*Suppose we’re living in the world before the introduction of behavioural economics as a distinct discipline.

Could an LLM, trained on diverse text corpora that doesn’t contain any explicit links made between the two fields, ever independently ‘connect the dots’ to propose ‘behavioural economics’?*
===

Elon bashes reasoning by analogy in favor of first principle reasoning, but what he's bashing is actually reasoning by precedent

Aside: are there really 'first principles' of human behavior?

Read on Substack

Personal Website

March 2026

Operationalize bottom-up design (composability) - admin panel allows for easy addition of new nodes (influences or outputs)

Tree-seed-fruit metaphor; PKMS inspired networked thought UX

Influence seeds can recombine to create new 'fruit' nodes (academic, project, substack)

Inspired by "A City is not a Tree" by Christopher Alexander

Amara

March 2025

How do we balancing intentionality with feeling of 'genuineness' for those who wish to be more intentional towards their social relationship?
- Too intentional, and curation feels transactional (professional)

Design principles: Ambient Intimacy & Intentional Serendipity

Seasonal tree metaphor for visualizing relationship health
- Cyclical -> Even if the user's not in a good state of their social life (winter), metaphor signals that this is but temporary (spring will come)

Weekend prototype

Popup Cities (eg. Zuzalu)

Novel form factor of social organization
- Conference of 1000 people over 2 days -> we get breadth of connection but lacks depth
- Hacker house of 6 people over a year -> We get depth but lacks breadth
- What if we can get a few hundred people together, for eg. a month?
--> We bring our 'whole self'

Individuation - Jung

Carl Jung
Might parallel buddhism's no-self as end point?

Chat with Mateo 19/03/2026

Ontological Design

Our design shapes what we CAN think, and what we think shapes what we design

Must be careful about the implicit assumptions we bring into the design of systems
Eg. in simulation: what constitutes an 'agent'; is time 'real'; is a group 'just' an aggregate of individuals

Surfacing the 'waters'/'frames'

Gift-Giving (Anthro)

What makes a gift a gift, instead of just an object of transaction?

Generativity

Ability of a system to create unbound novelty from some initial seed of simplicity.
Examples:
- Evolution (Simple rules -> complexity)
- Zwicky's morphological grid in engineering (simple rule forces combinatorial output)
- Chomsky's Generative Grammar (Simple rules produce infinite variation)
- LootRPG (Simple text loot spawns whole co-imagined game world)

A City is Not a Tree - Christopher Alexander

Artificial cities fail because planners organize them as trees (strict hierarchies); Natural cities form semi-lattices where elements belong to multiple groups simultaneously.

Rejection of reductionism in a sense; Also parallels the shift from the hierarchical knowledge organization (Evernote/Notion) to 'Networked Thoughts' paradigm (Roam, Obsedian, Logseq)

'Wholeness'

Why it matters: This essay fundamentally shaped how I think about representation in simulation systems. The tree vs. semi-lattice distinction maps directly onto the problem of demographic personas: collapsing people into non-overlapping categories (trees) versus preserving the overlapping, cross-cutting nature of real identity (semi-lattices).

Unfolding - Christopher Alexander

Bottom-up. Feedback loop between form and context; let the fit emerge rather than pre-specifying what the output ought to be

LootRPG

One of the few cool NFT projects back in the days. Users co-imagine an entire fantasy world from simple, procedurally-generated text-based loot boxes.

Maybe it's precisely because it's simple that this generativity was possible

Source

Locket App

Ambient Intimacy. Basically Snapchat but 1-to-1 + homescreen widget on partner's phone. Great for LDR

Slowly Penpal

Really well-executed penpalling app. Design retained the 'essence' of penpalling in a digital medium (i.e. a rare case where digital medium actually augmented a pre-existing experience)

'A new letter is arriving' + time lock proportional to distance from penpal = preservation of the 'slowness' that invites depth

Also fun to collect stamps

This is Water - David Foster Wallace

Best Commencement Speech.

===
There are these two young fish swimming along and they happen to meet an older fish swimming the other way, who nods at them and says “Morning, boys. How’s the water?”
And the two young fish swim on for a bit, and then eventually one of them looks over at the other and goes
“What the hell is water?"
===

Structure & assumptions in life that are invisible yet shapes - IS - our reality.
Only way to see the water is to jump out of it.

Cubist Art

Overlaying multiple pespectives on a single frame to show the 'full structure' of things as they are.
Abstraction as a means of distilling to the 'essence'

Essence ~ Relations. Can have relations of different things (eg. Structural Mapping Theory -> cares about conceptual relations)

Cool concept to play with for representing different ontologies of personhood in simulation

Structural Mapping Theory

Relational essence between concepts
Structural Similarity > Surface Similarity

Elon musk bashes reasoning by analogy BUT what he's actually bashing is reasoning by PRECEDENTS. Analogical reasoning done right (with SMT) yields combinatorial novelty

2x2 Grid

Juxtapose 2 axis which is nothing new individually, and all of a sudden we get 4 conceptual buckets - some categorizations we have never thought of.

Good way to do design research. In-depth categorization of the painpoints of users

Also used in generativity to make visible outputs we wouldn't have arrived at ourselves -- eg. Zwicky's morphological analysis

PKMS - Roam, Obsidian, Logseq

Networked thoughts. Bottom-up knowledge organization vs top-down tradition (Notion/Evernotes)

Versatile but bottleneck is in setting up the system

Picbreeder

Reject objective-orientation. Process > Outcome.
Parallels Christopher Alexander's Unfolding concept in some sense

Consequentialist vs Dialectical

Consequentialist activities = what we do as a means to an end
Dialectical activities = activities as means in and of themselves; value can only be felt through the act of doing them (eg. being a good friend; parenting)

Zhang 2024 argues that computers is the ultimate consequentialist machine, and if we're not careful, we might inadvertently bring this orientation into the realm of the dialectical

Improvisation

Temporal generativity.
Minimal structure: 'yes, and' rule.
Interpretation & co-constructed meaning as time unfolds
Sawyer's 'Retroactive Interpretation' (linked)

Source

Unknown Unknowns

Classic 2x2 grid - still very useful in thinking

Design Metaphors

Metaphors shape our interaction with computing systems

Push towards problem-framing

Qns: What would you do if you only have *2 years* to live?

Best life advice: live like you only have 2 years left to live
1 day's too short -> hedonist
10 year's too long -> procrastination
2 years ~ just right (meso)