Large language models (LLMs) provide a powerful foundation for social media simulation, but their utility is fundamentally limited by the accuracy of user behavior modeling. Existing LLM agents exhibit behavioral homogenization and fail to capture users’ inherent identities and evolving preferences. To address this, we propose CLASP, a closed-loop behavior alignment framework for social media agents. CLASP leverages explicit, dynamically refined personas to bridge users’ historical behaviors and subsequent agent actions, ensuring behavioral consistency and temporal adaptability. Our framework adopts a decoupled dual-model design to separate persona refinement from action generation. This design enables continual behavior alignment driven by action prediction discrepancies. To facilitate training and evaluation, we further construct BlueTrack, a large-scale Bluesky dataset that covers long-horizon and heterogeneous social behaviors with complete action elements. Extensive experiments demonstrate that CLASP outperforms baselines in behavioral alignment with real users. It improves the overall alignment quality by up to 31.8% and reduces its variance by 32.1% during long-sequence simulation. This work provides a practical reference for further exploration in realistic social media simulation.

Although large language models (LLMs) have shown exceptional capabilities across a wide range of tasks, reliable evaluation remains a critical challenge due to data contamination, opaque operation, and subjective preferences. To address these issues, we propose League of LLMs (LOL), a novel benchmark-free evaluation paradigm that organizes multiple LLMs into a self-governed league for multi-round mutual evaluation. LOL integrates four core criteria (dynamic, transparent, objective, and professional) to mitigate key limitations of existing paradigms. Experiments on eight mainstream LLMs in mathematics and programming demonstrate that LOL can effectively distinguish LLM capabilities while maintaining high internal ranking stability (Top-k consistency = 70.7%). Beyond ranking, LOL reveals empirical findings that are difficult for traditional paradigms to capture. For instance, “memorization-based answering” behaviors are observed in some models, and higher in-family scores are found in the OpenAI model family (𝛥 = 9, p < 0.05). Finally, we make our framework and code publicly available as a valuable complement to the current LLM evaluation ecosystem.