Build toward a Social World Model for Social Intelligence.逐步构建社会世界模型，走向社会智能。
Renlab is an independent research lab, incubated at Renmin University of China. We are building models of how people think and act, and training AI against them so it can work closely with people.Renlab 是一家独立的研究实验室，孵化于中国人民大学。我们为人如何思考、如何行动建模，并用这些模型训练 AI，让它能与人紧密协作。
Hypothesis 002 · falsifiable假设 002 · 可证伪
Models trained inside a simulation of real people will work with real people better than models trained on human ratings.在真实人群的模拟中训练出来的模型，与真人协作时，会比用人工评分训练的模型做得更好。
Wrong if, with real people, they do no better than the same models trained on ratings.如果面对真人时，它们并不比用评分训练的同一个模型更好，这个假设就不成立。
→ Social Intelligence→ 社会智能animation illustrative动画为示意
§ 02 / 06
02MindCast
first product · closed beta第一个产品 · 内测中
Product · 01 · ad creatives产品 · 01 · 广告素材● closed beta · open to partners● 内测中 · 开放合作
A model for ad creatives广告素材模型
MindCast
Trained to make ad creatives people are more likely to like.训练目标：让广告素材更可能被用户喜欢。
MindCast is trained with RL against a User Behavior Model we trained ourselves. The behavior model plays the audience an ad is for, and MindCast learns to make the creatives that audience is more likely to like.MindCast 以我们自研的用户行为模型为环境，用 RL 训练。行为模型扮演广告要面对的受众，MindCast 学着做出这些受众更可能喜欢的素材。
You give it你提供
The text of the ad you want to run.想投放的广告文字内容。
→
You get back你得到
An optimized prompt for generating the creative优化后的素材生成提示词
A delivery strategy to match配套的投放策略
Trained againstour own User Behavior Model, trained in-house训练环境我们自主训练的用户行为模型
Learns fromhow the behavior model, playing the audience, responds to each creative学习信号行为模型扮演受众时，对每条素材的反应
Statusclosed beta, open to partners · no results published yet状态内测中，开放合作 · 暂无公开效果数据
Become a beta partner成为内测合作伙伴
We reply by email, and use your address for nothing else. Or write to [email protected].我们会通过邮件回复，你的邮箱不会用于其他用途。也可以直接写信到 [email protected]。
FIG. 02 · how MindCast learns图 02 · MindCast 怎么学● illustrative · not a result● 示意图 · 不是结果
creative model → creatives → our user behavior model, as the audience → liked or skipped → reward → RL update → creative model素材模型 → 素材 → 我们的用户行为模型（扮演受众） → 喜欢或划走 → 奖励 → RL 更新 → 素材模型
§ 03 / 06
03The Social World Model社会世界模型
one person → many people → models that work with them一个人 → 许多人 → 与人协作的模型
Why this stack为什么这样搭
AI that works closely with people has to understand them, one person at a time and one group at a time.要与人紧密协作，AI 得先理解人：一个人一个人地理解，一个群体一个群体地理解。
Real people can't be the practice ground for every attempt. So we model them first: a User Behavior Model for each person, and a Social Simulator where many of them interact. Inside that model, we train models that work closely with people, starting from layer 01; the simulator is in design. Those three layers are the Social World Model.真人不能给每一次尝试当练习对象。所以我们先把人建模出来：为每个人建一个用户行为模型，再搭一个让许多人互动的社会模拟器。然后在这个模型内部，训练与人紧密协作的模型，先从第 01 层开始，模拟器还在设计中。三层合起来，就是社会世界模型。
Most simulations of people start from a persona or a demographic profile. Our first one did too, and it did no better than a plain prompt (§05). In our one-person study, only 3 of the 846 judgments came with a stated reason. A prompt can't supply reasons nobody writes down; a model has to learn them from what people do. The User Behavior Model behind MindCast is trained, not prompted.大多数对人的模拟，都从人设或人口统计画像出发。我们的第一版也是这样，结果并不比一个普通提示词更好（§05）。在这项单人研究里，846 个判断中只有 3 个附带了理由。没人写下来的理由，提示词给不了，只能让模型从人的行为里学。MindCast 背后的用户行为模型是训练出来的，不是提示词搭的。
The Social World Model社会世界模型layers 01 + 02 + 03第 01 + 02 + 03 层
Layer 01 · bottom第 01 层 · 底层
User Behavior Model用户行为模型
A person · built on a model of the human mind一个人 · 基于人类心智模型
It starts from what is going on inside a person: what they know, what they want, and how they feel. From there it models what they actually do, including what they choose, put off, and refuse. We train it in-house, and MindCast is trained against it.它从一个人的内在出发：知道什么、想要什么、感受如何。在此基础上，刻画这个人实际会怎么做，包括选什么、拖着什么、拒绝什么。我们自己训练它，MindCast 就以它为环境训练。
knows · wants · feels知道 · 想要 · 感受↓does行动
Layer 02 · middle第 02 层 · 中间层
Social Simulator社会模拟器
Many people · built from user behavior models许多人 · 由用户行为模型组成
Many behavior models share one setting, such as a class, a family chat, or a small team, and interact. The simulator is meant to show what the group does, and to be where the models that work with people train, against simulated people who push back.许多行为模型处在同一个场景里，比如一个班级、一个家庭群、一个小团队，彼此互动。模拟器将用来呈现这群人会怎么样，也将是与人协作的模型的训练场，练习对象是会反驳的模拟人。
01 × N↓a group一个群体
↓Trained inside it, with RL在模型内部用 RL 训练
Layer 03 · top第 03 层 · 顶层
Models that work with people与人协作的模型
Trained with RL inside the model · against the layers below在模型内部训练 · 以下面的层为环境
The models that work closely with people. Inside the Social World Model they practice on simulated people first, and are rewarded by how those people respond. The first is MindCast, for ad creatives (§02), trained against layer 01 and now in closed beta. The Teacher model and Renling come next, below.与人紧密协作的模型。它们在社会世界模型内部，先在模拟的人身上练，奖励来自这些人的反应。第一个是做广告素材的 MindCast（§02），以第 01 层为环境训练，目前内测中。Teacher 模型和小人儿在后面，见下方。
RL against 01 (+02)在 01（+02）上做 RL↓responses模拟反应
Next in layer 03第 03 层 · 接下来in development · not trained yet开发中 · 尚未训练
Teaches knowledge教授知识
Teacher modelTeacher 模型
Measured, not rated.看实测，不看打分。
A tutor to be trained inside the Social Simulator, against simulated learners built to disagree, stall, and resist flattery. Its reward will be what the learner can do afterwards. Teaching comes next because learning can be measured.一个将在社会模拟器里训练的辅导模型，练习对象是会反驳、会卡壳、不吃奉承的模拟学生。它的奖励将是学生学完之后能做到什么。教学排在下一个，因为学习效果可以测量。
小人儿 · a high-EQ assistant for small groupsRenling · 面向小群体的高情商助手
Renling小人儿
One assistant, the whole group.一个助手，照顾整个群。
Most assistants serve one person. Renling is meant to sit in a small group, such as a few friends, a family, or a five-person team, and keep several people in mind at once: what each one wants, and how they get along. It will practice on simulated groups before it joins a real one.大多数助手只服务一个人。小人儿要待在一个小群体里，比如几个朋友、一家人、一个五人小团队，同时把几个人放在心上：每个人想要什么，彼此怎么相处。它会先在模拟的群体里练，再加入真实的群。
FIG. 03 · Renling in a group chat图 03 · 小人儿在群聊里 · illustrative · synthetic · 示意 · 合成It noticed Saturdays kept falling through for Lei, and suggested Sunday to the group.它注意到阿磊的周六总是泡汤，就向大家提议了周日。
§ 04 / 06
04Research agenda研究议程
open questions · prior research开放问题 · 前期研究
What has to be true要成立，需要什么
Hypothesis 002 holds only if each layer holds. These are the questions we are working on, and where each one stands.假设 002 要成立，每一层都得成立。下面是我们正在研究的问题，以及各自的进展。
Q1Layer 01 · one person第 01 层 · 一个人
Can a model of one person match that person's calls more closely than a frontier model that is simply told who they are?一个人的模型，能否比只被告知「这人是谁」的前沿模型，更贴近这个人的决定？
Everything above it rests on this, so it is tested first.上面的一切都建立在这一层上，所以最先测它。
Answers Jun 2026 · scored Sep 20262026 年 6 月作答 · 9 月打分Not in our first test: tied with the baseline.第一次测试没有做到：和基线打平。4.39 vs 4.41 out of 10, on 18 real situations, with a prompted twin. The behavior model behind MindCast is trained, not prompted; we have not published a comparison for it yet.满分 10 分，4.39 对 4.41，18 个真实情境，测的是用提示词搭的分身。MindCast 背后的行为模型是训练出来的，不是提示词搭的；我们还没有公布它和基线的对比。→ §05 the n=1 study→ §05 n=1 研究
Q2Layer 02 · many people第 02 层 · 许多人
Does a group composed from individual behavior models act like the real group?由一个个行为模型组成的群体，表现得像真实的群体吗？
The simulator is only useful as a training ground if what works on simulated people also works on real ones.只有在模拟的人身上有效的做法在真人身上也有效，模拟器才能当训练场。
Not tested yet尚未测试The simulator is in design.模拟器还在设计中。
Q3Layer 03 · working with people第 03 层 · 与人协作
Do models trained inside the Social World Model do better with real people than models trained on ratings?在社会世界模型内部训练的模型，与真人协作时，是否比用评分训练的模型更好？
This is the hypothesis itself. MindCast is a first, partial test: it is trained against layer 01 only. Teaching comes next, because what a learner learns can be measured.这就是假设本身。MindCast 是第一次、也是部分的检验：它只以第 01 层为环境训练。接下来是教学，因为学生学到了什么可以测量。
In closed beta · no results published yet内测中 · 暂无公开结果MindCast, for ad creatives.MindCast，做广告素材。Trained with RL against our own User Behavior Model, and open to partners. A comparison with a ratings-trained model, on real viewers, is what would answer Q3.以我们自研的用户行为模型为环境用 RL 训练，开放合作。要回答 Q3，需要在真实观众上，和用评分训练的模型做对比。→ §02 MindCast
Prior research · papers by members of our team前期研究 · 团队成员参与的论文some of it from before Renlab was founded部分工作早于 Renlab 成立
20262026 年We shipped decision-support tools (/apps). The recurring bottleneck: the model didn't understand the person using it.我们做了一批决策辅助工具（/apps）。反复遇到的瓶颈：模型不理解用它的人。
May 20262026 年 5 月The n=1 study begins, from one person's previous month of work with AI coding assistants.n=1 研究启动，素材是一位成员此前一个月与 AI 编程助手协作的记录。
Jun–Sep 20262026 年 6–9 月The first test: 18 situations answered in June, every setup scored on one fixed judge in September. Tied with the baseline.第一次测试：6 月作答 18 个情境，9 月用同一个固定裁判给所有方案打分。和基线打平。
Nov 20262026 年 11 月The Social World Model: behavior model, simulator, and the models that work with people.确定社会世界模型：行为模型、模拟器，以及与人协作的模型。
Now现在MindCast, trained with RL against our own User Behavior Model. Closed beta, open to partners.MindCast，以我们自研的用户行为模型为环境用 RL 训练。内测中，开放合作。
Next下一步The Social Simulator, then the Teacher model trained inside it.先做社会模拟器，再在里面训练 Teacher 模型。
§ 05 / 06
05Evidence · the n=1 study证据 · n=1 研究
2026-05 → 09 · one person2026 年 5–9 月 · 一个人
First test of Q1 · a prompted twinQ1 的第一次测试 · 提示词搭的分身
Before modeling many, we measured one.建模许多人之前，我们先测了一个人。
We took one month of a team member's work with AI coding assistants (24 April to 24 May 2026) and extracted 846 judgments from 4,178 of their turns, each a situation and the call they made. With those, their meeting transcripts and their own writing, we built a behavior twin: a frontier model given a profile of the person and a retrieval index of their own words.我们取了一位团队成员一个月（2026 年 4 月 24 日至 5 月 24 日）与 AI 编程助手协作的记录，从 4,178 轮对话里提取出 846 个判断，每个都是「情境 + 本人做出的决定」。用这些判断、本人的会议记录和自己写的东西，我们搭了一个行为分身：给前沿模型一份这个人的画像，再加上一个检索本人原话的索引。
Then the person answered 18 real work situations, posed as interview questions, in their own words. A fixed LLM judge scored how close each setup came to those answers, three times per situation.然后，这位成员以访谈的形式，用自己的话回答了 18 个真实工作情境。一个固定的 LLM 裁判给每种方案打分，看它离本人的回答有多近，每个情境打三次。
Twin 4.39. Baseline, the same model told only who they are: 4.41. Out of 10, a tie.分身 4.39。基线，也就是只被告知「这人是谁」的同一个模型：4.41。满分 10 分，打平。
For this person, on these 18 situations, a profile and retrieval did not bring the model closer to their answers. A fine-tuned version scored lower, 3.17. The judge passes its basic checks: it gives the person's own answers 10.0, and repeat readings of an answer vary by a standard deviation of at most 0.08. Next: natural conversation data instead of interviews, a second person, and a blinded judge.对这个人、这 18 个情境来说，画像加检索并没有让模型更接近他的回答。一个微调版本的得分更低，3.17。裁判通过了基本检查：给本人的回答打 10.0 分，同一个回答重复打分的标准差不超过 0.08。下一步：用自然对话数据代替访谈，加第二个人，换成盲评的裁判。
Scope: a prompted twin of one person, not the trained User Behavior Model behind MindCast. Limits: one person and 18 situations; the judge is from the same model family as the twin and was not blinded; the rubric was designed, with an AI assistant's help, with the goal of beating the baseline; retrieval sometimes returned nothing. No benchmark, customer, or product result is claimed here.范围：测的是一个人的提示词分身，不是 MindCast 背后训练过的用户行为模型。局限：只有一个人、18 个情境；裁判和分身来自同一个模型家族，也没有盲评；评分细则是在 AI 助手的帮助下、以打败基线为目标设计的；检索有时什么都没返回。这里不声称任何基准、客户或产品结果。
FIG. 04 · the n=1 study图 04 · n=1 研究real numbers · one person真实数据 · 一个人
recorded → twin → 18 situations → judgetwin 4.39 · baseline 4.41 · out of 10记录 → 分身 → 18 个情境 → 裁判分身 4.39 · 基线 4.41 · 满分 10
846judgments extracted提取的判断one month of work一个月的工作
18real situations真实情境answered by the person本人作答
54judge readings裁判打分次数18 situations × 318 个情境 × 3
4.39behavior twin行为分身fidelity, out of 10保真度，满分 10
4.41baseline基线same model, told only who they are同一模型，只被告知这人是谁
−0.02twin minus baseline分身减基线out of 10 · a tie满分 10 · 打平
2026-05 → 09 · judge scores2026 年 5–9 月 · 裁判打分
§ 06 / 06
06Contact联系
research lab · Beijing研究实验室 · 北京
Talk to the lab.和实验室聊聊。
Renlab is an independent research lab, incubated at Renmin University of China. If you work on modeling people, social simulation, or RL for models that work with people, we'd like to hear from you.Renlab 是一家独立的研究实验室，孵化于中国人民大学。如果你在研究怎么给人建模、做社会模拟，或者为与人协作的模型做 RL，欢迎来聊。