Improve Large Language Model Systems with User Logs

Researchers have proposed a framework called UNO that aims to improve large language model systems by extracting training signals from user interaction logs, a resource the authors argue is underused despite containing authentic human feedback and procedural knowledge [1][2]. The approach, detailed in a paper posted to arXiv and last revised on 17 June 2026, addresses what its authors describe as a growing constraint on traditional LLM development: the scarcity of high-quality training data and diminishing returns from escalating computational costs [1][2]. Large language models are neural networks trained on vast text corpora for tasks such as generation, summarization, and translation, and they underpin modern chatbots [3]. Their performance is typically measured through benchmark evaluations that assess reasoning, factual accuracy, and alignment [3]. The UNO framework — short for User log-driveN Optimization — is designed to convert unstructured and noisy user logs into semi-structured rules and preference pairs [2]. It then applies query-and-feedback-driven clustering to handle data heterogeneity and quantifies what the researchers call a cognitive gap between the model’s prior knowledge and the incoming log data [2]. That assessment allows the system to filter out noisy feedback and build separate modules for primary and reflective experiences extracted from the logs [2]. User-generated content has long been recognized as a shift from passive consumption to active participation online, with platforms from the BBC to CNN building systems to capture eyewitness accounts and audience contributions [6]. The UNO paper extends that principle to model improvement, treating user interactions not as static content but as a continuous feedback loop for LLM systems [2]. In experiments, UNO outperformed Retrieval Augmented Generation — a technique that supplies external knowledge to a model at inference time — and memory-based baselines, achieving what the authors call state-of-the-art effectiveness and efficiency [2]. The code has been open-sourced on GitHub [2]. The first version of the paper was submitted on 6 February 2026 as a 2,663 KB file; the second and third revisions, the last dated 17 June 2026, weighed in at 5,482 KB [1]. The work was led by Changyue Wang and appears under the Computation and Language category on arXiv, where it is hosted alongside experimental community projects developed through arXivLabs [1].