HumanLLM: Towards Personalized Understanding and Simulation of Human Nature

AI summary

Overview

• Research area: Natural Language Processing / large language models for social simulation, personalized user modeling, and computational social science.
• Technical level: Intermediate. The paper combines a large-scale data curation pipeline with supervised fine-tuning, model merging, and multi-task benchmark evaluation; it is readable without deep ML theory but assumes familiarity with LLM fine-tuning concepts.
• One-sentence scope: The paper introduces HumanLLM, a foundation model fine-tuned on a 5.5-million-log Cognitive Genome Dataset built from real user data across four online platforms, to better predict and simulate individual human behaviors, thoughts, and writing styles.

What This Paper Is About

Standard LLM pretraining uses large amounts of disconnected web text, which does not capture the continuous, situated context behind any one person's decisions and thoughts. The authors argue this is why LLMs are weak at simulating specific individuals, and they set out to build a foundation model trained on real, longitudinal user behavior instead. The goal is a model that can predict what a given person would do, feel, or write in a given situation.

Key Contributions

1. The Cognitive Genome Dataset. The authors state they are the first to leverage large-scale, real-world user data from multiple online platforms so that LLMs can learn personalized human behaviors and thoughts. The dataset is built through a three-stage curation pipeline (data filtering, data synthesis, data quality control) and contains hundreds of thousands of users, millions of scenarios, and social QA pairs, drawn from over 5.5 million user logs.
2. A multi-task learning design grounded in Lewin's equation. Six complementary training tasks—profile generation, scenario generation, social question answering, writing imitation, personalized commenting, and item selection—are designed around the idea that behavior is a function of person and environment, B = f(P, E).
3. HumanLLM, trained across multiple base models. The model is trained on five backbones (Qwen2.5-3B-Instruct, Qwen2.5-7B-Instruct, Qwen3-8B, Phi-3-mini-128k-instruct, Llama-3.1-8B-Instruct), with a 1:1 weight merge between the fine-tuned model and its original base to mitigate catastrophic forgetting.
4. Comprehensive evaluation. In-domain benchmarks, two out-of-domain public social intelligence benchmarks (MotiveBench and TomBench), and three real-world application studies, with code released at https://aka.ms/humanllm.

Main Findings

• In-domain performance improves substantially across all backbones. The best average score is achieved by HumanLLM_Qwen2.5-7B at 27.38%, versus 6.71% for its base Qwen2.5-7B-Instruct (a reported improvement of 308.05%). The reported improvement for HumanLLM_Llama over Llama-3.1-8B-Instruct is 234.72% and for HumanLLM_Qwen2.5-3B over Qwen2.5-3B-Instruct is 235.32%.
• Largest gains are on item selection and profile generation. The authors attribute this to improved understanding of user preferences and life trajectories, with writing imitation and scenario generation showing better modeling of style and context.
• GPT-4o underperforms HumanLLM on in-domain tasks. GPT-4o averages 10.38%, below HumanLLM_Qwen2.5-7B at 27.38%, which the authors interpret as evidence that targeted training makes the model more sensitive to real user behavior.
• Out-of-domain generalization improves but GPT-4o leads. On MotiveBench, HumanLLM_Qwen3 achieves the highest average among trained models at 0.7427 (GPT-4o: 0.8166). On TomBench, HumanLLM_Qwen3 also leads at 0.6589 (GPT-4o: 0.7378). The authors state that GPT-4o outperforming on OOD benchmarks is expected given its substantially larger capacity.
• Raw-text training gives little benefit. Training Qwen2.5-7B-Instruct directly on raw posts, tweets, blogs, and reviews (HumanLLM^raw_Qwen2.5-7B) yields only marginal in-domain gains (9.63% average) and a notable drop on TomBench (reported as -5.00% versus base).
• Model merging beats mixing in general instruction data. HumanLLM^gen_Qwen2.5-7B, trained with SmolTalk and Orca mixed in a 0.25:0.25:0.5 ratio, reaches comparable in-domain performance (27.12%) but falls short on OOD tasks, lagging behind even the base model on TomBench by 0.0783 (reported as -12.64%).
• Both task groups contribute. Training Qwen3-8B on Task Group 1 (profile generation, scenario generation, social QA) or Task Group 2 (writing imitation, personalized commenting, item selection) alone gives large gains on the corresponding tasks while degrading on unseen tasks, and OOD scores fall between Qwen3-8B and full HumanLLM.
• Baseline human-behavior models underperform. Centaur averages 2.93% in-domain, 0.0373 on MotiveBench, and 0.3750 on TomBench; BE.FM averages 11.49% in-domain, 0.6868 on MotiveBench, and 0.5969 on TomBench. The authors note BE.FM is unstable, beating Llama-3.1-8B-Instruct on MotiveBench but performing worse on TomBench.
• Application studies favor HumanLLM. Using GPT-4o as evaluator over 1,600 samples per application, HumanLLM outperforms its base model on faithfulness, realism, and coverage for profile generation. The remaining application results (human explainer and personalized writing) are truncated in the provided content.

Methodology in Plain English

The authors assemble raw user content from four platforms selected for having users with multiple records, a diverse population, and authentic everyday behavior: Reddit (47.2M submissions from 50 subreddits), Twitter (88.1M original tweets from CommonCrawl), Blogger (681K blogs), and Amazon (571.54M item reviews).

Stage one, filtering. Rule-based thresholds remove low-quality entries: Reddit posts need a score of at least 2, at least 2 comments, and at least 300 characters; tweets need at least 300 characters; blogs shorter than 300 characters or lacking first-person pronouns are excluded; Amazon reviews shorter than 100 characters are dropped, with a k-core filter keeping items with at least three purchase records and users with at least eight purchase histories. Llama-3.3-70B-Instruct then rewrites the text to strip HTML, URLs, non-English content, and formatting noise, and to replace sensitive information such as home addresses or ID numbers with fictitious data. A second Llama-3.3-70B-Instruct pass assigns a three-level quality label and a binary harmfulness label, retaining only high-quality, non-harmful text. The result is 2.8 million Reddit posts, 673,000 tweets, 368,000 blogs, and 1.7 million reviews.

Stage two, synthesis. Motivated by Lewin's equation, the authors distill three content types from each record: user (P), scenario (E), and social QA (B). User content comes in four forms—User Persona (under 100 words), User Profile (100–400 words), User Stories, and Writing Style. Each record is also converted into a structured scenario of background, characters, and plots, plus multiple scenario-question-answer triplets covering next actions, inner mental states, and social reasoning, with scenarios adjusted so the answer is not directly revealed.

Stage three, quality control. Because synthesized content still showed problems such as answer leakage and hallucinated profile details, Qwen-2.5-72B scores each sample on metrics including Hallucination, Coverage, Conciseness, Relevance, Fidelity, Novelty, Leakage, and Overall, each on a 1–10 scale. Samples are kept only if the overall score exceeds 8 and all other scores exceed 7, filtering out approximately 30% of users, 60% of scenarios, and 40% of social QA samples per source.

Training. Six tasks are used, and data is sampled so tasks and sources are roughly balanced, yielding 1,195,717 training samples and 132,869 test samples in ShareGPT format with next-token prediction and non-response positions masked from the loss. Training uses Llama-Factory on 8×40GB A100 GPUs, learning rate 5e-6, total batch size 64, max sequence length 8192, cosine scheduler, warm-up ratio 0.5, DeepSpeed Zero configured by model size, and 3 epochs; the 8B models took roughly 120 hours. Inference uses vLLM at temperature 0.7, with cost identical to the base model. To avoid catastrophic forgetting, the fine-tuned and original models are merged with 1:1 weights.

Evaluation. Open-ended generation tasks are converted into multiple choice: item selection offers 20 candidates (the actual purchased item is correct, 19 negatives sampled from the item pool), and the other four tasks are eight-way multiple choice, with seven negatives generated by GPT-4o, Llama-3.1-8B-Instruct, Llama-3.1-70B-Instruct, Qwen2.5-7B-Instruct, Qwen2.5-72B-Instruct, Phi-4, and Qwen3-8B.

Why This Matters

Impact on research. The paper argues that learning individual human behavior requires situated, longitudinal real-world data rather than psychology experiments, economic games, survey responses, or fictional role-play, which the authors say are limited in scale or captured in artificial settings emphasizing group averages. It offers a large, structured resource and a task taxonomy for building human-centric LLMs, and reports that these gains transfer to out-of-domain social intelligence benchmarks.

Real-world applications (as stated by the authors):

• Social simulation and large-scale sociological research.
• Personalized companions and emotional companions.
• Marketing intelligence through simulated customer feedback.
• More realistic user simulation for recommender systems.
• Additional stated uses include intelligent NPCs, personalized assistants, social data generation, and human behavior explanation.

Industry relevance. The practical value proposition is a social data generator, human explainer, and user simulator that runs at the same inference cost as its base model, and that can be built on several widely used open backbones, making personalized user modeling accessible without proprietary model training. The released code at https://aka.ms/humanllm supports replication and downstream adaptation.

Future Directions

• Extending the model merge recipe. The ablation shows merging beats mixing in general instruction data, but the authors do not report the optimal merge ratio beyond the 1:1 setting used by default, leaving the trade-off between social intelligence and general capability open.
• Closing the out-of-domain gap with top proprietary models. HumanLLM improves over its base models on MotiveBench and TomBench, but GPT-4o still leads both (0.8166 and 0.7378 versus HumanLLM_Qwen3's 0.7427 and 0.6589), raising the question of how much of the gap is capacity versus data.
• Broadening data sources and populations. The corpus covers Reddit, Twitter, Blogger, and Amazon; whether the approach transfers to other platforms, languages, or user populations is not reported.
• Improving answer leakage and hallucination control. Quality control still discards roughly 30% of users, 60% of scenarios, and 40% of social QA samples, suggesting room to improve synthesis rather than filter it away.

Target Audience

Researchers and practitioners in LLM training, computational social science, and user simulation; recommender-system and marketing-analytics teams interested in simulated user feedback; and anyone building personalized companions, social agents, or synthetic user data pipelines. Readers who need a replicable template for curating real-world user logs into supervised fine-tuning tasks will find the pipeline and ablation details most useful.

Note: the supplied paper content is truncated at the "Human Explainer" application section, so the full results for the human explainer and personalized writing studies are not reported here.

Authors’ abstract

Motivated by the remarkable progress of large language models (LLMs) in objective tasks like mathematics and coding, there is growing interest in their potential to simulate human behavior--a capability with profound implications for transforming social science research and customer-centric business insights. However, LLMs often lack a nuanced understanding of human cognition and behavior, limiting their effectiveness in social simulation and personalized applications. We posit that this limitation stems from a fundamental misalignment: standard LLM pretraining on vast, uncontextualized web data does not capture the continuous, situated context of an individual's decisions, thoughts, and behaviors over time. To bridge this gap, we introduce HumanLLM, a foundation model designed for personalized understanding and simulation of individuals. We first construct the Cognitive Genome Dataset, a large-scale corpus curated from real-world user data on platforms like Reddit, Twitter, Blogger, and Amazon. Through a rigorous, multi-stage pipeline involving data filtering, synthesis, and quality control, we automatically extract over 5.5 million user logs to distill rich profiles, behaviors, and thinking patterns. We then formulate diverse learning tasks and perform supervised fine-tuning to empower the model to predict a wide range of individualized human behaviors, thoughts, and experiences. Comprehensive evaluations demonstrate that HumanLLM achieves superior performance in predicting user actions and inner thoughts, more accurately mimics user writing styles and preferences, and generates more authentic user profiles compared to base models. Furthermore, HumanLLM shows significant gains on out-of-domain social intelligence benchmarks, indicating enhanced generalization.