Large language models (LLMs) enable social simulation, yet faithfully reproducing real users requires grounding both structured platform decisions and language in sparse, noisy behavioral evidence. We introduce PsyMemBehav, which maps psychology‑informed user profiles and retrieved memories to target‑first platform behavior through a two‑level identity‑to‑action reasoning process.
To support this approach, we build the PsyMemBehav‑bench dataset containing 5.57 M observed actions from 5,862 real users across Reddit, Bluesky, and Twitter—approximately 950 actions per user. Each recorded engagement is modeled as target selection followed by a Comment, Like, or Repost, with an optional response generation step.
PsyMemBehav integrates 41‑dimensional user descriptors, the top‑5 BM25‑retrieved memories, reverse‑CoT supervision, and rubric‑guided DPO over complete reasoning‑action trajectories. Experiments show that the 27B‑parameter PsyMemBehav ranks first in 11 of 12 reported individual‑level metrics, raising average response quality from GPT‑5.4’s 54.28 to 59.68 (a 9.95% relative gain), and achieves the lowest group‑level MAE (0.081) and MSE (0.009) among simulated systems.
The benchmark remains conditional on recorded engagements and does not model no‑action decisions.
Review: PsyMemBehav demonstrates a promising direction by coupling psychological profiling with memory grounding, offering a concrete example of evidence‑based user modeling for social media simulation.