Blog · Research · Oct 21, 2026 · 5 min
An agent that remembers what a person said predicted their answers four years later better than a persona, in every panel and on every question; an agent given only a life record did no better than a persona. And Living Worlds' crowd predicted how close real people came to a real robot in 7 of 8 recordings nobody had opened.
Two questions sit at the edge where a simulated life meets the real world. The first: if an agent is meant to stand for a particular person, what does it need to know about them to predict what they will say years later? A persona, a short description of who they are, is the common starting point. A life record, everything that has happened to them, sounds richer. What they have said themselves may matter more.
The second: when a machine works among people, how close do they come to it? People carrying boxes and buckets through a room do not move as they would in an empty corridor, and they adjust around a robot in ways a crowd model has to capture before it can say anything about that room.
Living Worlds tested both, on data nobody had read.
The agent tests used General Social Survey panels, which interview the same people several times over the years. Each test was registered before any test answer was read, on people held out from everything used to build the agent, and scored once. The robot test used two recording days nobody here had opened, registered before the data was read and scored once.
The reason is the same as everywhere in the register. An agent that has seen a person's later answers can repeat them; one scored once on answers that were sealed when the rule was written has to predict them. A crowd tuned on the recordings it is scored on describes them; one scored on days nobody had opened forecasts them.
An agent with a life record, against a persona gave a model everything that had happened to a person since their first interview and asked it to predict their 2020 General Social Survey answers. Across 912 held-out people it did no better than a persona alone: +0.0027 in the registered score, with an interval from -0.0073 to +0.0126. Not met.
What happened to someone, on its own, did not tell the model how they would answer. The next test added what the person had said.
An agent that remembers what a person said gave the agent a memory holding what happened to the person and what they had said about related things, never their answers to the questions scored. It then predicted their answers four years later.
It beat a persona by 2.5% in the registered score: 0.5965 against 0.6115, where lower is better, with an interval from 0.011 to 0.019. It was better in every panel and on every question. The test used 1,911 people in the 2006, 2008 and 2010 panels, never used before. Met.
This sits alongside other recent work on agents built from what people say. A 2025 Stanford brief describes agents built from two-hour interviews that replicated participants' General Social Survey answers 85% as accurately as the participants replicated their own answers two weeks later. The Living Worlds test asks a different, longer question: four years on, with a memory built from related answers rather than an interview, scored once on panels nobody had used.
People around a working robot, on days nobody had opened put a robot into the Living Worlds crowd. The data comes from THÖR-MAGNI, a room where visitors carried boxes, buckets and large objects around a working robot, cited with thanks.
With the robot on its recorded path, Living Worlds predicted how close real people came to it. Its people's closest approach to the robot was nearer the real one than for a crowd blind to the robot and for people who ignore it, in 7 of 8 recordings. Met (7 of 8).
The measure is simple and physical: how near real people came to a real machine, compared with how near the simulated people came.
Together these results say something consistent about what makes a simulated person useful. A description is a weak guide. A list of events is no better. What a person has said about related things, held in memory, predicts what they will say years later, on every question and in every panel tested. Around a machine, a crowd that knows the machine is there predicted how close people came better than one that ignores it, in 7 of 8 recordings on days nobody had seen.
The not-met result is part of the finding. Without it, a life record would look like the natural way to build an agent; the register shows it added nothing measurable on its own, and that what people said is what moved the score.
The memory agent's gain is 2.5% in the registered score, real and consistent across panels and questions, and modest in size. It was tested on General Social Survey panels in the United States, four years ahead; it says nothing yet about other surveys, other countries or other kinds of prediction. The life-record agent's interval spans zero.
The robot test predicts how close people come to a robot moving on its recorded path, in one room, across eight recordings. It measures closest approach and nothing more: it makes no claim about how the robot itself should move, and it covers two recording days.
Living Worlds gives one agent a whole life in a world, and that life travels with it into any body it works through, a robot, a digital persona, a game character or an assistant. What it does is written back to the same record. Its site states what is shown so far: an agent that remembers what a person said predicts their answers four years later better than a persona, in every panel and on every question. The crowds it simulates are checked against real measurements before they are relied on, and the register records each result, met and not met.
Blog
Physical AI now runs on open physics engines, world foundation models and synthetic demonstrations measured in hundreds of thousands. The newest benchmarks ask whether simulated scores predict real ones, and the part of the scene least checked against reality is the people.
Living Worlds' crowd against real pedestrians in Jülich experiments, a shopping centre in Osaka and six recorded street scenes, each scored once on recordings nobody had opened: met in the shopping centre, in four fresh Jülich experiments and with a learned walking layer; not met in the first Jülich set and in three of six street scenes.
World models now generate explorable places that hold together for minutes. The next frontier is behaviour over years, many lives at once, and results checked against a reality the model has never seen.