The increasing prevalence of texts generated by large language models (LLMs) drives a critical demand for reliable LLM text detection, where text watermarking has emerged as a promising proactive solution. Among existing methods, sentence-level watermarking is more robust to semantic-preserving adversarial attacks, but it still faces two key limitations: generation often requires iterative rejection sampling, and detection treats all sentence-level signals as equally reliable. To address these limitations, we propose CAD-Mark, a context-aware distributional sentence-level watermarking framework. During generation, CAD-Mark partitions a batch of candidate sentences with a context-derived adaptive boundary and samples the output sentence from the subset corresponding to the secret bits, embedding multiple bits per sentence without iterative rejection sampling. During detection, it estimates sentence-level reliability from re-sampled candidates and aggregates bit-level signals with a diversity-weighted z-test. Experiments on two benchmarks demonstrate that CAD-Mark consistently outperforms strong baselines in TPR@1%FPR while preserving generation quality with comparable cost.

Large language models (LLMs) in retrieval-augmented generation systems can still produce hallucinations, generating content that is unsupported or contradicted by the source texts and undermines reliability. Recent work addressed this problem by training span-level hallucination detectors using reinforcement learning (RL) and chain-of-thought (CoT) reasoning. In this work, we show through error analysis that incorrect predictions by existing reasoning-based detectors are strongly associated with CoT processes that lack explicit grounding in source evidence, particularly when verification steps do not quote or verify claims against the retrieved documents. This behaviour contrasts with human verification practices in benchmarks such as RAGTruth, where evidence quotation is a prerequisite for determining hallucinated spans. Motivated by this observation, we propose an evidence-grounded RL framework, namely RLSeek, to explicitly enforce active evidence seeking during CoT reasoning by requiring quotation of relevant source segments at each verification step. Experiments on the RAGTruth and NewsSum dataset demonstrate consistent improvements in hallucination span detection performance, with limited additional reasoning overhead and improved robustness in out-of-domain settings.