What if we treated personality the way CLIP treats images, embedding everything, unsupervised, at the resolution reality demands, and asked not "do these two people match?" but "what kind of relationship will they build?"
Hosted by The Relationship Singularity. Born from the Love Symposium and covered in The New York Times.
The tools we use to form relationships — swiping on photos, answering personality quizzes — operate at a resolution so low they're like trying to recognize a face from a 4×4 pixel grid. We have no serious research infrastructure for the representations that would change this.
Personality can be embedded the way CLIP embeds images: high-dimensional, unsupervised, learned from massive multimodal data. Relationship outcomes can be predicted as structured manifolds, not scalar compatibility scores. Alan Cowen's emotion-embedding work at Hume AI is the proof of concept; we propose extending it to the full personality and to dyadic outcome prediction.
Six research frontiers we want this workshop to interrogate.
High-dimensional, multimodal representations of the person, learned the way CLIP learned vision: unsupervised, from massive heterogeneous data. The Big Five is a 5-dimensional projection of something far richer; we want methods that recover the full manifold from text, voice, video, behavior, and digital footprints.
A relationship is not a scalar compatibility score — it is a trajectory through a high-dimensional outcome space. We want models that predict not "will these two match?" but "what kind of relationship will they build, when will it bend, and where?" Structured prediction over manifolds, not regression to a number.
LLM agents as Monte Carlo proxies for human social behavior — following the lineage of Park et al.'s Generative Agents and Generative Agent Simulations of 1,000 People. Can we run a relationship in silico, validate it against real couple data, and use it to design interventions before they touch a human?
Personality science is in the middle of a replication reckoning, and relationship science has barely started its own. We want held-out benchmarks, next-token personality prediction as an embedding-quality metric, and dataset releases — including the workshop's featured Keeper Labs corpus — that let claims be tested rather than asserted.
A black-box compatibility score is not actionable — and not trustworthy. We want methods that surface the reason two people are predicted to fit (or not), translate high-dimensional embeddings into legible narratives, and let users meaningfully consent to AI-mediated connection.
Romance is the dramatic case, but the broader prize is human connection at scale: friendship recommendation, AI interventions for social isolation, and rigorous measurement of relational wellbeing. Married people report happiness equivalent to a $50K raise — what's the upper bound on what relational AI could deliver?
Minimize conversational KL divergence.
At NeurIPS 2025, attendees took a personality assessment, were embedded into a high-dimensional vector space, and matched with their highest-cosine-similarity twin. 118 people showed up in person.
We projected the personality manifold live and discovered surprising axes: AI timeline beliefs, training habits, and a clear Theorist-to-Engineer gradient.
Proof of concept for the workshop thesis: unsupervised personality embeddings recover structure invisible to traditional instruments, and cosine similarity maps onto genuine interpersonal resonance. Validated at CVPR 2025.
Relationship Outcome Prediction Tech by Matthew Fisher — a two-hour tour through the research program: personality embedding spaces, emotion manifolds, LLM relationship simulation, and the path to the Relationship Singularity.
The Love Symposium is a curated weekend for earnest founders, experts, and intellectuals obsessed with solving our hardest relationship problems and proliferating healthy connection at scale. Talks, workshops, hackathons, and late-night conversations — aimed at sparking collaborations that go far beyond the room.
Keeper Labs partners with top researchers to study attraction, compatibility, and what makes relationships last. Match data, outcome metrics, custom surveys, and longitudinal tracking — privacy-first and IRB-friendly.
Available to workshop participants. Submissions to the Benchmarks & Datasets track are encouraged to use it.
Extended abstracts up to 4 pages in NeurIPS 2026 LaTeX format. Double-blind via OpenReview. Non-archival; concurrent submissions permitted. Three tracks: Research, Benchmarks & Datasets, and Position Papers.
At least one author must attend in person. Selected papers get oral slots; all accepted work presents posters.
All deadlines 23:59 AoE. Timeline extrapolated from NeurIPS 2025.
The full day-of program is still being assembled. We plan a mix of invited talks from our speakers and contributed paper presentations selected from the open call, with breaks for posters and discussion. Final schedule will be published after author notifications in late September 2026.
Sponsorship supports student travel grants, poster logistics, and invited speaker travel to Sydney.
Tiers from $5,000-$20,000. Contact us for the prospectus.
