The article introduces a new framework called Citation-aware Rubric Rewards (CaRR) aimed at improving reinforcement learning (RL) for deep search agents. Traditional RL methods mainly use binary rewards based on whether the final answer is correct, which can lead to problems like shortcut exploitation and hallucinations. These issues arise when agents reach answers without a thorough understanding or by relying on incomplete information. CaRR addresses these shortcomings by implementing a fine-grained reward system that emphasizes reasoning quality, factual grounding, and evidence connectivity. CaRR decomposes complex questions into smaller, checkable rubrics that agents must satisfy to receive rewards. These rubrics focus on three key aspects: identifying all relevant entities, ensuring statements are backed by cited content, and forming a clear evidence chain that leads to the final answer. Alongside CaRR, the article also presents Citation-aware Group Relative Policy Optimization (C-GRPO), which augments the existing Group Relative Policy Optimization by incorporating weighted rubric rewards. This dual approach promotes more comprehensive reasoning and improved accuracy in model training. Experiments using the Qwen3-4B-Thinking-2507 and Qwen3-30B-A3B-Thinking-2507 models, trained on the DeepDive dataset, show that C-GRPO consistently outperforms standard outcome-based GRPO on various deep search benchmarks. C-GRPO agents can utilize longer context budgets effectively, enhancing their performance in open-ended research tasks. The integration of APIs like Serper and Jina for web access further supports the robustness of these agents in real-world applications.