Seer: Online Context Learning for Fast Synchronous LLM Reinforcement Learning

The paper presents ReFactX, a scalable method designed to enhance the reliability of Large Language Models (LLMs) by enabling them to access external knowledge without relying on additional models or services. This approach utilizes constrained generation with a prefix-tree index, allowing for efficient retrieval of factual information from a Knowledge Graph. The method aims to address persistent issues of knowledge gaps and hallucinations in LLM outputs.

Large Language Models (LLMs) have shown exceptional ability in text generation but often produce factually incorrect statements, known as 'hallucinations'. This study investigates hallucinations in conversational data across three low-resource languages: Hindi, Farsi, and Mandarin. The analysis of various LLMs, including GPT-3.5 and GPT-4o, reveals that while Mandarin has few hallucinated responses, Hindi and Farsi exhibit significantly higher rates of inaccuracies.

Large Language Models (LLMs) are advanced linguistic tools that can produce outputs that may sound plausible but are often factually incorrect, a phenomenon known as hallucination. This study introduces a mathematical framework to analyze, quantify, and mitigate these hallucinations. It employs probabilistic modeling and Bayesian uncertainty estimation to develop refined metrics and strategies, including contrastive decoding and retrieval-augmented grounding, aimed at enhancing the reliability of LLMs.

The paper introduces TASA (Teaching According to Students' Aptitude), a personalized mathematics tutoring framework that utilizes Large Language Models (LLMs) to adapt instruction based on students' evolving knowledge and cognitive retention. TASA integrates a structured student persona and event memory to enhance learning by addressing individual proficiency levels and forgetting patterns, aiming to improve the effectiveness of mathematics education.

The paper introduces Group Turn Policy Optimization (GTPO), a novel reinforcement learning algorithm aimed at enhancing the training of Large Language Models (LLMs) for multi-turn Tool-Integrated Reasoning (TIR). GTPO addresses limitations of existing methods like Group Relative Policy Optimization (GRPO) by implementing turn-level reward assignments, return-based advantage estimation, and self-supervised reward shaping, which collectively improve learning signals for complex interactions.

GlobalRAG is a proposed reinforcement learning framework aimed at enhancing global reasoning in multi-hop question answering (QA). It addresses limitations in current methods by decomposing questions into subgoals, coordinating retrieval with reasoning, and refining evidence iteratively. The framework introduces new rewards to encourage coherent planning and reliable execution of subgoals, aiming to improve the effectiveness of multi-hop QA systems.

ConInstruct is a benchmark designed to evaluate Large Language Models (LLMs) on their ability to detect and resolve conflicts in user instructions. While many existing assessments focus on adherence to instructions, ConInstruct addresses the often-overlooked scenarios where conflicting constraints arise. Initial evaluations show that proprietary LLMs generally perform well in conflict detection, with DeepSeek-R1 and Claude-4.5-Sonnet achieving the highest F1-scores.

The study presents a data-driven machine learning approach aimed at optimizing the performance of Large Language Model (LLM) adapters in GPU serving environments. It addresses the challenge of maximizing throughput while preventing request starvation by determining the optimal configuration of concurrent and parallel adapters. The introduction of a Digital Twin for LLM-adapter systems facilitates efficient training data generation, with experiments showing a throughput accuracy within 5.1% of real results.