OmniBehavior: A Large-Scale Multi-Scenario Behavioral Trajectory Benchmark Dataset for Real-World User Behavior Simulation

The Institute of Computing Technology, Chinese Academy of Sciences (ICT CAS) jointly released the OmniBehavior dataset with Kuaishou. It covers 90 days of cross-scenario user behavior trajectories on the Kuaishou platform, aiming to evaluate the ability of large language models (LLMs) in long-term, cross-domain, and heterogeneous behavior simulation tasks, and provide an important resource for the research community.

With the improvement of large language model capabilities, their potential in user behavior modeling has attracted attention, but existing benchmarks have limitations: limited time span (difficult to evaluate long-term interest evolution), single scenario (lack of cross-scenario correlation), and homogeneous behavior types (unable to reflect real complexity). OmniBehavior, based on real Kuaishou data, fills these gaps.

• Long-term observation window: Covers 90 days from September 1 to November 30, 2025, supporting tracking of interest evolution, identification of behavior pattern changes, etc.
• Multi-scenario coverage: Integrates six core scenarios: short video browsing, live stream interaction, e-commerce shopping, ad interaction, customer service dialogue, and search behavior.
• Data structure: User-centric, including user profiles (basic features) and behavior history (temporal action sequences with scenario identifiers, timestamps, context, etc.).

• Long-term interest modeling: Explore interest formation/transition, impact of external events, and construction of long-term dynamic interest models.
• Cross-domain behavior analysis: Study correlations between different scenarios (e.g., relationship between video viewing and shopping behavior, correlation between live stream interaction and e-commerce conversion rate).
• User behavior simulation: Provide real trajectories as ground truth, supporting training and evaluation of simulation agents, testing model prediction performance, etc.

OmniBehavior has been open-sourced on GitHub, providing a demo dataset demo.json (partial data of a single user). The full dataset is expected to be released in May 2026, and is currently undergoing privacy compliance audits, available only for academic research. The related paper was published in April 2026, elaborating on the construction method, evaluation metrics, and benchmark experimental results.

OmniBehavior defines the research paradigm of 'real-world behavior simulation', posing new challenges for LLM research (grasping long-term patterns, cross-scenario correlations, intention evolution), promoting their development in the field of user modeling; at the same time, it provides a basis for long-term effect evaluation of recommendation systems, solving the problem of inconsistency between offline and online effects.

With the release of the full dataset, we can expect to see: more powerful behavior prediction models, more realistic virtual user simulation systems, cross-scenario unified recommendation frameworks, and privacy-preserving behavior modeling technologies. OmniBehavior may become the standard benchmark in this field, promoting breakthroughs in academia and industry.