💬 NLP

HumanLLM: Towards Personalized Understanding and Simulation of Human Nature

Motivated by the limitations of standard LLMs in capturing individual context, this paper introduces HumanLLM, a foundation model trained on a large-scale "Cognitive Genome Dataset" of real-world user logs that significantly improves the personalized simulation and prediction of human behavior, thoughts, and writing styles.

Original authors: Yuxuan Lei, Tianfu Wang, Jianxun Lian, Zhengyu Hu, Defu Lian, Xing Xie

Published 2026-01-23

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Yuxuan Lei, Tianfu Wang, Jianxun Lian, Zhengyu Hu, Defu Lian, Xing Xie

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The Big Idea: From "General Knowledge" to "Deep Empathy"

Imagine you have a brilliant student who has read every book in a massive library. This student (a standard Large Language Model, or LLM) is amazing at writing code, solving math problems, and summarizing news. They know everything about the world in a general sense.

However, if you ask this student to pretend to be you—to predict what you would think, say, or do in a specific situation—they often fail. They might sound like a generic person, not you. Why? Because they learned from disconnected snippets of text (like a library of random facts) rather than the continuous, messy, emotional story of a single human life.

HumanLLM is a new type of AI designed to fix this. It's not just a "smart encyclopedia"; it's a "human simulator" trained to understand the specific quirks, history, and personality of individual people.

The Problem: The "Library" vs. The "Diary"

Standard AI models are trained on the entire internet. Think of this as studying a Library of Random Facts. You learn that "people like coffee" or "people get angry when traffic is bad." But you don't know why a specific person, let's say "Sarah," gets angry when traffic is bad. Is it because she's late for her daughter's recital? Is she stressed about a deadline?

The paper argues that to truly simulate a human, the AI needs to read Sarah's Diary, not just the Library. It needs to see the continuous story of her life, her decisions, and her inner thoughts over time.

The Solution: The "Cognitive Genome"

To teach the AI how to read these "diaries," the researchers built a massive new dataset called the Cognitive Genome Dataset.

The Analogy: Imagine taking millions of real people's public posts from places like Reddit, Twitter, blogs, and Amazon reviews. These are like millions of scattered diary entries.
The Process: The researchers didn't just dump this data into the computer. They used a rigorous "cleaning and organizing" pipeline:
1. Filtering: They threw out bots, spam, and short, meaningless posts.
2. Synthesizing: They used a powerful AI to read these scattered posts and rewrite them into coherent stories. They created:
  - User Profiles: A summary of who the person is (their personality, values, history).
  - Scenarios: The specific situation the person was in (the "environment").
  - Social Q&A: Questions about what the person thought or did next, and the actual answer based on their real behavior.
3. Quality Control: They acted like strict editors, checking that the AI didn't make things up (hallucinations) and that the stories made sense.

The result is a dataset of 5.5 million high-quality user stories, effectively teaching the AI the "DNA" of human behavior.

How They Trained the AI: The "Specialist" vs. The "Generalist"

Once they had the data, they needed to train the model without breaking its existing smarts.

The Challenge: If you teach a genius student to act like a specific person, they might forget how to do math or write code. This is called "catastrophic forgetting."
The Trick: Instead of just retraining the model from scratch, the researchers used a technique called Model Merging.
- Analogy: Imagine you have a Generalist Chef (the original AI) who can cook anything. You also have a Specialist Chef (the new AI) who is an expert at cooking your favorite family recipes.
- Instead of firing the Generalist and hiring only the Specialist, they mixed their recipes together. They combined the weights of both models. The result is a chef who can still cook a perfect steak (general knowledge) but also knows exactly how you like your coffee (personalized understanding).

What Can HumanLLM Do? (The Results)

The researchers tested this new model in three main ways:

Predicting You (In-Domain Tasks):
- They asked the AI to guess what a user would buy next, how they would write a blog post, or how they would answer a social question.
- Result: HumanLLM was significantly better at guessing these things than the original AI. It could mimic a user's writing style and predict their choices much more accurately.
Understanding People (Out-of-Domain Benchmarks):
- They tested the AI on standard tests for "Social Intelligence" (like understanding motives, emotions, and beliefs) that it had never seen before.
- Result: Even on these new tests, HumanLLM performed better. This suggests it didn't just memorize the data; it actually learned how to think like a human.
Real-World Applications:
- Profile Generation: Creating realistic, detailed descriptions of people for simulations.
- Human Explainer: Explaining why a person acted a certain way in a specific situation (e.g., "She was angry because she felt her child was being disrespected, based on her past trauma").
- Personalized Writing: Writing emails or reviews that sound exactly like a specific person.

The Bottom Line

The paper claims that by feeding AI a "Cognitive Genome"—a massive, organized collection of real human stories and behaviors—they created a model that is much better at understanding and simulating individual human nature. It moves AI from being a "know-it-all encyclopedia" to being a "deeply empathetic observer" who understands the unique story of every person.

Key Takeaway: To make AI truly human-like, you can't just give it more facts; you have to give it the context of a human life.

Technical Summary: HumanLLM: Towards Personalized Understanding and Simulation of Human Nature

Problem Statement

While Large Language Models (LLMs) have achieved remarkable proficiency in objective tasks such as coding, mathematics, and text summarization, they exhibit significant limitations in simulating human behavior and cognition. Standard LLMs are pre-trained on vast, uncontextualized web corpora, which fail to capture the continuous, situated context of an individual's decisions, thoughts, and behaviors over time. Consequently, these models struggle to accurately predict motivations, infer inner states, or forecast future actions in complex social scenarios. Existing attempts to humanize LLMs often rely on limited-scale data from controlled psychology experiments, economic games, or fictional narratives, which do not fully reflect the complexity, spontaneity, and long-term dynamics of authentic human experiences in the real world.

Methodology

1. The Cognitive Genome Dataset

To address the data misalignment, the authors constructed the Cognitive Genome Dataset, a large-scale corpus derived from real-world user data across four diverse platforms: Reddit, Twitter, Blogger, and Amazon. The dataset construction follows a rigorous three-stage pipeline:

Data Filtering: Raw data undergoes rule-based filtering (e.g., removing bots, short posts, and advertisements) and model-based rewriting using Llama-3.3-70B-Instruct to remove noise, HTML tags, and sensitive information. A model-based scoring system further filters for quality and safety, retaining only high-quality, non-harmful content.
Data Synthesis: Inspired by Lewin's Equation ( $B = f(P, E)$ $B = f (P, E)$ ), the pipeline synthesizes data into three components:
- User ( $P$ ): Generates four levels of user profiles: User Persona (concise identity), User Profile (detailed background), User Stories (narratives of life events), and Writing Style summaries.
- Scenario ( $E$ ): Extracts structured triplets of background, characters, and plots from individual records to capture situational context.
- Social QA ( $B$ ): Constructs question-answer triplets focusing on user actions, inner mental states, and social reasoning (attribution of behavior) based on the specific user and scenario.
Data Quality Control: An automated evaluation using Qwen-2.5-72B assesses synthesized data on metrics including Hallucination, Coverage, Conciseness, Relevance, Fidelity, Novelty, and Leakage. Only samples meeting strict thresholds (overall score > 8, others > 7) are retained, resulting in a final dataset of over 5.5 million user logs.

2. HumanLLM Model Training

Based on the Cognitive Genome Dataset, the authors define six complementary learning tasks to cover the spectrum of human activity:

Profile Generation: Inferring detailed profiles from minimal personas.
Scenario Generation: Reconstructing complete social stories from characters and backgrounds.
Social Question Answering: Predicting user actions, thoughts, or reasoning given a persona and scenario.
Writing Imitation: Generating new text in a specific user's style.
Personalized Commenting: Generating context-aware responses to new posts.
Item Selection: Predicting user preferences from a history of interactions.

Training Strategy:
The authors fine-tune several open-source LLMs (including Qwen2.5, Qwen3, Phi-3, and Llama-3.1) on this dataset. To mitigate catastrophic forgetting of general capabilities—a common issue when fine-tuning on specialized data—the authors employ a model merging strategy. Instead of joint fine-tuning with general instruction data (which yielded suboptimal generalization), they perform a 1:1 weight merge between the fine-tuned model and the original instruction-tuned base model. This approach combines enhanced social intelligence with robust general capabilities.

Key Contributions

Cognitive Genome Dataset: The first large-scale, multi-platform dataset curated specifically for personalized human behavior simulation, containing hundreds of thousands of users and millions of synthesized scenarios and social QA pairs.
HumanLLM Foundation Model: A novel foundation model designed for personalized understanding and simulation of human nature, trained via a multi-task learning framework on the Cognitive Genome Dataset.
Model Merging Approach: A demonstration that model merging is more effective than joint fine-tuning with general instructions for preserving general intelligence while enhancing social reasoning.
Comprehensive Evaluation: Extensive validation across in-domain tasks, out-of-domain social intelligence benchmarks, and real-world application scenarios.

Experimental Results

In-Domain Performance

On five core tasks (Item Selection, Profile Generation, Scenario Generation, Social QA, and Writing Imitation), HumanLLM variants consistently outperformed their base models and existing human-behavior models (Centaur, BE.FM).

HumanLLM $_{Qwen2.5-7B}$ achieved the highest average score (27.38%), showing significant improvements over the base model (308.05% gain in average performance).
The model demonstrated superior capabilities in predicting user preferences, inferring life trajectories, and imitating writing styles compared to base models and even GPT-4o in specific in-domain contexts.

Out-of-Domain (OOD) Generalization

Evaluated on MotiveBench (motivational reasoning) and TomBench (theory of mind), HumanLLM showed enhanced generalization:

HumanLLM $_{Qwen3}$ achieved the highest average scores on both benchmarks among the tested open models.
The model showed consistent gains across different LLM backbones, indicating that the training approach improves social reasoning beyond the original training distribution.
In contrast, models like BE.FM showed unstable performance across benchmarks, suggesting limited generalization when trained on narrower data sources like economic games.

Real-World Applications

Qualitative and quantitative evaluations in three application scenarios confirmed HumanLLM's utility:

Profile Generation: Produced more faithful, realistic, and comprehensive user profiles than base models, avoiding demographic hallucinations.
Human Explainer: Generated explanations of human behavior that were more contextually relevant, logically consistent, and aligned with reference answers.
Personalized Writing: Successfully imitated user styles and reproduced critical details and reasoning patterns in generated content.

Significance and Claims

The paper posits that HumanLLM represents a significant step toward human-centric AI systems. By grounding LLMs in authentic, large-scale, real-world user data rather than disconnected text snippets or fictional narratives, the model achieves a more nuanced understanding of the interplay between individual identity, environment, and behavior.

The authors claim that HumanLLM serves as a superior social data generator, human explainer, and user simulator. This capability has profound implications for:

Advancing social science research through large-scale sociological simulations.
Developing personalized AI companions and emotional support agents.
Enhancing recommender systems with realistic user simulations.
Generating marketing intelligence through simulated customer feedback.

The work suggests that leveraging diverse, real-world user data is essential for bridging the gap between general-purpose LLMs and the specific, empathetic, and context-aware intelligence required for authentic human interaction.