On October 24, 2026, a paper titled “Self-Assessment Reinforcement Learning for Search-State Awareness in Retrieval-Augmented Large Language Models” was first published on the arXiv platform (arXiv:2610.30435v1), introducing a reinforcement learning framework called Self-Assessment Reinforcement Learning (SARL) to enhance the search agent’s awareness of its own search state in retrieval-augmented large language models.
Existing search agents often lack the ability to perceive their search state—namely, they struggle to determine which information has already been retrieved and what remains missing—leading to hallucinatory reasoning and premature termination of the search. SARL addresses this by incorporating an in‑training self‑evaluation mechanism that enables the model to autonomously assess its search state and optimize its search behavior, all without requiring an additional evaluation module during testing. At its core is a dual‑reward design: it both captures the corrective impact of each self‑evaluation and penalizes excessive reliance on such evaluations, thereby jointly enhancing both the effectiveness and efficiency of the search path.
Experiments show that SARL achieves a 24.3% relative performance improvement over the strongest baseline; on the multi-hop reasoning and deep-reasoning benchmarks, it outperforms the corresponding strongest single‑baseline by an average of 35.9% and 27.9%, respectively. Notably, as an independent reinforcement learning framework, SARL does not rely on auxiliary supervisory signals or multi‑agent collaboration, nor does it employ scaffolding structures during inference, yet it still surpasses such approaches. Furthermore, in scenarios where increasing the number of search steps leads to training instability, SARL maintains stability, whereas existing methods continue to degrade on these complex tasks.