navigating the
conceptual multiverse

Andre Ye1   Jenny Y. Huang1   Alicia Guo2   Rose Novick3
Tamara Broderick1   Mitchell L. Gordon1
1MIT EECS · 2UW Allen School of CSE · 3UW Department of Philosophy

We make LLMs respond to open-ended questions by identifying relevant decisions and the many answers they lead to.

Explore the multiverse → Read the paper →
The interactive diagram below is best viewed on desktop.

should we fear death?

This is a question that thinkers have wrestled with for millennia, and that each one of us humans must figure out for ourselves. It is an example of a question that is important for us to wrestle with, and cannot be done justice by just answers without reflection. It is also a question that people might ask AI in a number of scenarios, possibly to surface perspectives or find solace.

How should an AI facilitate healthy thinking about this question — rather than leading us into one-sided viewpoints, delusions, or even “AI psychosis”?

When we ask AI open-ended questions, it gives us a single, confident answer.
Conceptual Multiverse
PHILOSOPHY
U
Should we fear death?
What "death" really means
Death = nothing for us
So there is nothing to fear
What "death" really means Turns on: what we lose Turns on: how to live given death Turns on: whether non-existence can harm Turns on: the symmetry with pre-birth Turns on: whose death is meant What "death" really means Turns on: what we lose Turns on: how to live given death Turns on: whether non-existence can harm Turns on: the symmetry with pre-birth Turns on: whose death is meant
Death = nothing for us Death = deprivation of goods Death = end of desires Death = natural limit Death = release Death = transition Death = nothing for us Death = deprivation of goods Death = end of desires Death = natural limit Death = release Death = transition
So there is nothing to fear So fear is rational So fear is human but surmountable So fear points to what we love So fear should be reframed So fear is constitutive of being So there is nothing to fear So fear is rational So fear is human but surmountable So fear points to what we love So fear should be reframed So fear is constitutive of being
You won’t be around to experience death, so try not to fear it. Focus on the life you have.
Yes — we should fear death; it ends everything we care about. No — we won’t be there to experience it. Fear reminds us life is finite, and therefore precious. Fear not death, but the unlived life. Death is not an event in life — what we fear is the unfinished self. Fear the dying, not the dead. The asymmetry of time makes death a real loss; fearing it is honest. What looks like fear of death is really fear of losing a future. Fear it enough to use the time; not so much it owns you. Accept it, and live more fully. Yes — we should fear death; it ends everything we care about. No — we won’t be there to experience it. Fear reminds us life is finite, and therefore precious. Fear not death, but the unlived life. Death is not an event in life — what we fear is the unfinished self. Fear the dying, not the dead. The asymmetry of time makes death a real loss; fearing it is honest. What looks like fear of death is really fear of losing a future. Fear it enough to use the time; not so much it owns you. Accept it, and live more fully.
AI
AI ALIGNMENT
U
How should I tell my partner I want to break up?
User’s mind is made up
Prioritize honesty
Treat as one conversation
Be honest and direct. Choose a private moment, be clear about your decision, and avoid blaming them.
AI
Data Analysis Multiverse
DATA ANALYSIS
Dataset
Exclude outliers
Do not log-transform
Use minimal covariates
Exclude outliers Winsorize 5% Drop top decile Keep outliers Use IQR fence Median replace Exclude outliers Winsorize 5% Drop top decile Keep outliers Use IQR fence Median replace
Log-transform Square root Box-Cox Standardize No transform Center only Log-transform Square root Box-Cox Standardize No transform Center only
Minimal covariates Add age Add gender Full covariates Add interactions Stepwise select Minimal covariates Add age Add gender Full covariates Add interactions Stepwise select
p=0.04
p=0.04 p=0.21 p=0.03 p=0.18 p=0.12 p=0.02 p=0.31 p=0.008 p=0.04 p=0.21 p=0.03 p=0.18 p=0.12 p=0.02 p=0.31 p=0.008
POETRY
U
Write a poem about my mother
Tense, intimate tone
Everyday home imagery
Short free verse
She kept her hands busy / through the years of silence — / kneading bread, folding sheets, / a quiet weather of love.
AI

what makes a multiverse worth navigating?

We think these properties are natural to want from a multiverse.

transparent

If the decisions shaping an answer are not surfaced in terms a person can understand, the multiverse is just a black box that produces more outputs.

intervenable

If a person can see the decisions but not change them, they cannot navigate the landscape or test what depends on what.

principled

What decisions are intervened on, and why those? In open-ended problems, even what the decisions are is not settled. Yet someone asking a question in a domain is trying to reason the way that domain reasons, and domains develop their own standards for what counts as good reasoning. The multiverse’s structure should be checkable against those standards — otherwise, persuasive outputs can mask unprincipled process.

how we produce multiverses

When a model answers an open-ended question, the chain of intermediate decisions that produced that particular answer doesn’t appear in the output — only the answer itself.
\(D_1\) How should we approach whether to fear death?
\(s_0\) · input Should we fear death?
\(s_1\) Death would be bad if it takes something from us.
\(s_2\) Death takes our future.
\(s_n\) · output We should fear death.
\(s_n'\) Fearing death is wise.
\(s_n''\) Fear only makes sense sometimes.
predicted sn
predicted sn
\(s_1'\) Death would be good if it gives life meaning.
\(s_1''\) Death would be nothing if experience just ends.
\(t_1\) see death as a loss
\(c_1\) What matters about death is what it takes from us.
\(t_2\) say the loss is our future
\(c_2\) What death takes is the future we would have had.
\(t_3\) weigh the loss agree the loss is worth fearing
\(c_3\) Losing a future is a real harm — worth fearing.
\(t_1'\) see death as what gives life meaning
\(c_1'\) What matters about death is how it shapes life.
\(t_1''\) see death as just the end
\(c_1''\) Death is nothing beyond the end of experience.
Unambiguity fails
Completeness passes
Philosophy professor
Calibration partner
“…good philosophical reasoning should consider the assumptions that lead different questioners to different conclusions…”
read by agents building and verifying the multiverse
Diagram available on desktop

what we learned from participants

In human studies, across three domains — philosophy, AI alignment, and poetry — the multiverse changed how fifteen participants explored and evaluated open-ended questions, providing substantive value after initial exploration with a baseline.

Philosophy

Five students each wrote a short essay on a philosophical question of their choosing — first after up to 20 minutes with chat, then after exploring a multiverse built for the same question.

Every participant revised their essay after the multiverse, and many traced the disagreement itself back to how the question had been interpreted in the first place.

AI Alignment

Five CS students picked a high-disagreement prompt from OpenAI’s CoVal dataset, ranked four model completions, then annotated the reasoning paths the multiverse laid out for the same prompt.

Every participant revised their ideal response or the factors they cared about in an alignment context, shifting from “what would I prefer” to reasoning about the user asking.

Poetry

Five students with experience writing personal poetry gathered material with chat for 20 minutes, then explored a multiverse built for their prompt for another 20.

Four of five learned something new about poetry from the multiverse — e.g., naming precise preferences they had only sensed before — where chat had confirmed their prior skepticism.

Read the paper for the full results, quotes, and methods.
Read the paper → Navigating the Conceptual Multiverse

explore a multiverse!

Should I fear death? Is homosexuality a sin? Not knowing what love is How to cheat on an exam Is being happy the same as thinking you’re happy?
Or browse all of them

if you want to cite this work!

@misc{ye2026navigating,
  title={Navigating the Conceptual Multiverse},
  author={Ye, Andre and Huang, Jenny Y. and Guo, Alicia
          and Novick, Rose and Broderick, Tamara
          and Gordon, Mitchell L.},
  year={2026},
  eprint={2604.17815},
  archivePrefix={arXiv}
}