We introduce the task of personalized and reliable popular‑science writing, which requires adapting scientific explanations to audiences with varying cognitive levels while preserving factual accuracy. Personalization often leads to simplifications that increase hallucination and factual distortion risks. To tackle this, we build a dataset of 39,134 entries and a reader‑centric Personalized Science Communication Benchmark (PSCB) that jointly evaluates audience adaptation and factual correctness. To reduce data and compute demands and improve cross‑domain and cross‑audience generalization, we propose the DA‑MoE model that explicitly decouples audience adaptation from domain knowledge. In evidence‑scarce scenarios we add a multi‑agent fact‑checking mechanism that augments limited evidence through role‑specific agent debates and propagates confidence over a graph for robust verification and revision. Experiments on PSCB show our approach reaches state‑of‑the‑art performance. The code is open‑sourced at https://github.com/DPInnovationWorks/CWF.
Review