SEEK: Steering LLM Reasoning for RAG via Internal Reasoning Sketches

A new framework called SEEK aims to reduce redundant knowledge retrieval in Retrieval-Augmented Generation (RAG) systems by using internal reasoning sketches to guide Large Language Models (LLMs), according to a paper posted on arXiv [1]. The paper, submitted on 14 Jan 2026 and revised on 5 Jun 2026, proposes SEEK as a solution to a known weakness in iterative RAG pipelines [1]. As reasoning trajectories lengthen, previously accumulated knowledge and earlier queries can interfere with later retrieval decisions, producing sub-queries with repetitive intents and redundant knowledge acquisition [2]. SEEK addresses this by first prompting an LLM to build a structured steering sketch composed of multiple groups of steering gists, each followed by a slot for knowledge filling [2]. The framework then iteratively retrieves and refines knowledge to populate those slots, and the completed sketch is used as contextual input for final answer generation [2]. Experimental results reported in the paper indicate that SEEK outperforms baseline models across multiple tasks [2]. Further analyses show the framework generates more diverse sub-queries, reduces redundant retrieval, and achieves a better balance between external knowledge utilization and internal knowledge conflict mitigation [2]. The code is publicly available on GitHub [2]. The work appears on arXiv, an open-access repository of electronic preprints that, as of November 2024, receives about 24,000 submissions per month [6]. arXiv hosts papers across mathematics, physics, computer science, and related fields, and has surpassed two million articles as of late 2021 [6]. The paper’s abstract page also features arXivLabs, a framework launched in 2020 that allows community collaborators to develop experimental tools directly on the site [5]. arXivLabs integrations include bibliographic explorers, code finders, and recommender systems, all operating under guidelines that require partners to uphold openness, community, excellence, and user data privacy [5]. Large language models, the underlying technology that SEEK aims to improve, are neural networks trained on vast text corpora for tasks such as generation, summarization, and translation [8]. Their reliability can be undermined by biased or inaccurate training data, a limitation that retrieval-augmented approaches seek to mitigate by incorporating external knowledge [8].