<p>With the emergence of advanced large language models (LLMs) such as ChatGPT, state-of-the-art social bots can now generate coherent text, dynamically adapt behavioral patterns, and mimic human user profiles with high fidelity. However, existing detection methods predominantly rely on static representations and unimodal decision-making schemes, thereby exhibiting significant limitations when confronting LLM-driven social bots. To address these limitations, we propose BotEvo, a novel bot detection framework grounded in behavioral evolution modeling and cross-modal fusion. BotEvo integrates a dynamic weighting scheme to quantify the contribution of individual posts, an implicit gating mechanism for deep fusion of text, behavioral, and sentiment features, and a two-stage threshold optimization strategy for adaptive, risk-stratified classification. Experimental evaluations on the Botsim and Twibot-20/SDAR datasets demonstrate that BotEvo attains F1 scores of 96.46% and 89.30%, respectively, consistently surpassing 14 state-of-the-art baselines.</p>