Competitive debate is a multi-turn, zero-sum adversarial task with opposing stances that requires debaters to anticipate opponents’ counterarguments in real-time interactions and select strategies yielding long-term advantages. Existing large language model debate methods predominantly adopt a greedy generation paradigm that optimizes only the immediate effectiveness of the current rebuttal while lacking forward-looking planning for the overall debate trajectory. To address this limitation, we propose MCTS-Debater, a Monte Carlo Tree Search framework that models competitive debate as a Markov Decision Process. Rather than selecting a single locally optimal response, MCTS-Debater systematically simulates multi-step adversarial interactions under diverse rebuttal strategies, estimates their long-term expected win rates via tree search, and synthesizes the top-performing strategies into a coherent final rebuttal through multi-strategy fusion. Experimental results on the ORCHID and IQ2 datasets demonstrate that MCTS-Debater significantly outperforms existing baselines in logical validity, factual accuracy, and overall persuasiveness. Source code is provided in the supplementary material.

Online harm detection is crucial for safeguarding digital ecosystems. However, current research is largely confined to single-dimensional classification paradigms or generative safety alignment under simplified and explicit adversarial prompts. To capture complex and highly covert threats in the real world, we introduce the Multi-dimensional Online Harm Detection Benchmark (MDHB). MDHB focuses on four core dimensions critically demanded by social platforms: misinformation, toxicity, sarcasm, and logical fallacies. Notably, we incorporate non-offensive implicit sarcasm (e.g., self-deprecation and veiled complaints) to propel content governance toward proactive psychosocial risk forecasting. Evaluations reveal that mainstream Large Language Models (LLMs), including GPT-5, suffer from severe cross-dimensional interference under both direct inference and ReAct paradigms. To address this, we propose the Self-Refining Sentinel (SRS), an error-driven self-evolving agent architecture. Without requiring parameter fine-tuning, SRS dynamically constructs and updates an experience repository from failure cases through an iterative self-evolution mechanism. Experimental results demonstrate that SRS significantly outperforms state-of-the-art baselines in overall performance, effectively mitigating cross-dimensional interference in multi-dimensional online threat detection.