Large language models (LLMs) have achieved remarkable progress in reasoning tasks, yet the optimal integration of Supervised Fine-Tuning (SFT) and Reinforcement Learning (RL) remains a fundamental challenge. Through comprehensive analysis of token distributions, learning dynamics, and integration mechanisms from entropy-based perspectives, we reveal key differences between these paradigms: SFT induces coarse-grained global changes to LLM policy distributions, while RL performs fine-grained selective optimizations, with entropy serving as a critical indicator of training effectiveness. Building on these observations, we propose Supervised Reinforcement Fine-Tuning (SRFT), a single-stage method that unifies both fine-tuning paradigms through entropy-aware weighting mechanisms. Our approach simultaneously applies SFT and RL to directly optimize the LLM using demonstrations and self-exploration rollouts rather than through two-stage sequential methods. Extensive experiments show that SRFT achieves 59.1% average accuracy, outperforming zero-RL methods by 9.0% on five mathematical reasoning benchmarks and 10.9% on three out-of-distribution benchmarks.

Medical document analysis plays a crucial role in extracting essential clinical insights from unstructured healthcare records, supporting critical tasks such as differential diagnosis. Determining the most probable condition among overlapping symptoms requires precise evaluation and deep medical expertise. While recent advancements in large language models (LLMs) have significantly enhanced performance in medical document analysis, privacy concerns related to sensitive patient data limit the use of online LLMs services in clinical settings. To address these challenges, we propose a trustworthy medical document analysis platform that fine-tunes a LLaMA-v3 using low-rank adaptation, specifically optimized for differential diagnosis tasks. Our approach utilizes DDXPlus, the largest benchmark dataset for differential diagnosis, and demonstrates superior performance in pathology prediction and variable-length differential diagnosis compared to existing methods. The developed web-based platform allows users to submit their own unstructured medical documents and receive accurate, explainable diagnostic results. By incorporating advanced explainability techniques, the system ensures transparent and reliable predictions, fostering user trust and confidence. Extensive evaluations confirm that the proposed method surpasses current state-of-the-art models in predictive accuracy while offering practical utility in clinical settings. This work addresses the urgent need for reliable, explainable, and privacy-preserving artificial intelligence solutions, representing a significant advancement in intelligent medical document analysis for real-world healthcare applications.

Extreme activation outliers in Large Language Models (LLMs) critically degrade quantization performance, hindering efficient on-device deployment. While channel-wise operations and adaptive gradient scaling are recognized causes, practical mitigation remains challenging. We introduce Outlier-Safe Pre-Training (OSP), a practical guideline that proactively prevents outlier formation rather than relying on post-hoc mitigation. OSP combines three key innovations: (1) the Muon optimizer, eliminating privileged bases while maintaining training efficiency; (2) Single-Scale RMSNorm, preventing channel-wise amplification; and (3) a learnable embedding projection, redistributing activation magnitudes originating from embedding matrices. We validate OSP by training a 1.4B-parameter model on 1 trillion tokens, which is the first production-scale LLM trained without such outliers. Under aggressive 4-bit quantization, our OSP model achieves a 35.7 average score across 10 benchmarks (compared to 26.5 for an Adam-trained model), with only a 2% training overhead. Remarkably, OSP models exhibit near-zero excess kurtosis (0.04) compared to extreme values (1818.56) in standard models, fundamentally altering LLM quantization behavior. Our work demonstrates that outliers are not inherent to LLMs but are consequences of training strategies, paving the way for more efficient LLM deployment. The source code and pretrained checkpoints are available at https://github.com/dmis-lab/Outlier-Safe-Pre-Training.

Abstract--Generating reports for computed tomography (CT) images is a challenging task, while similar to existing studies for medical image report generation, yet has its unique characteristics, such as spatial encoding of multiple images, alignment between image volume and texts, etc. Existing solutions typically use general 2D or 3D image processing techniques to extract features from a CT volume, where they firstly compress the volume and then divide the compressed CT slices into patches for visual encoding. These approaches do not explicitly account for the transformations among CT slices, nor do they effectively integrate multi-level image features, particularly those containing specific organ lesions, to instruct CT report generation (CTRG). In considering the strong correlation among consecutive slices in CT scans, in this paper, we propose a large language model (LLM) based CTRG method with recurrent visual feature extraction and stereo attentions for hierarchical feature modeling. Specifically, we use a vision T ransformer to recurrently process each slice in a CT volume, and employ a set of attentions over the encoded slices from different perspectives to selectively obtain important visual information and align them with textual features, so as to better instruct an LLM for CTRG. Experiment results and further analysis on the benchmark M3D-Cap dataset show that our method outperforms strong baseline models and achieves state-of-the-art results, demonstrating its validity and effectiveness.

While large language models (LLMs) have shown remarkable capabilities to generate coherent text, they suffer from the issue of hallucinations -- factually inaccurate statements. Among numerous approaches to tackle hallucinations, especially promising are the self-correcting methods. They leverage the multi-turn nature of LLMs to iteratively generate verification questions inquiring additional evidence, answer them with internal or external knowledge, and use that to refine the original response with the new corrections. These methods have been explored for encyclopedic generation, but less so for domains like news summarization. In this work, we investigate two state-of-the-art self-correcting systems by applying them to correct hallucinated summaries using evidence from three search engines. We analyze the results and provide insights into systems' performance, revealing interesting practical findings on the benefits of search engine snippets and few-shot prompts, as well as high alignment of G-Eval and human evaluation.

In the era of large-scale artificial intelligence, Large Language Models (LLMs) have made significant strides in natural language processing. However, they often lack transparency and generate unreliable outputs, raising concerns about their interpretability. To address this, the Chain of Thought (CoT) prompting method structures reasoning into step-by-step deductions. Yet, not all reasoning chains are valid, and errors can lead to unreliable conclusions. We propose ECCoT, an End-to-End Cognitive Chain of Thought Validation Framework, to evaluate and refine reasoning chains in LLMs. ECCoT integrates the Markov Random Field-Embedded Topic Model (MRF-ETM) for topic-aware CoT generation and Causal Sentence-BERT (CSBert) for causal reasoning alignment. By filtering ineffective chains using structured ordering statistics, ECCoT improves interpretability, reduces biases, and enhances the trustworthiness of LLM-based decision-making. Key contributions include the introduction of ECCoT, MRF-ETM for topic-driven CoT generation, and CSBert for causal reasoning enhancement. Code is released at: https://github.com/erwinmsmith/ECCoT.git.

Large Language Models (LLMs) are widely used in sensitive domains, including healthcare, finance, and legal services, raising concerns about potential private information leaks during inference. Privacy extraction attacks, such as jailbreaking, expose vulnerabilities in LLMs by crafting inputs that force the models to output sensitive information. However, these attacks cannot verify whether the extracted private information is accurate, as no public datasets exist for cross-validation, leaving a critical gap in private information detection during inference. To address this, we propose PrivacyXray, a novel framework detecting privacy breaches by analyzing LLM inner states. Our analysis reveals that LLMs exhibit higher semantic coherence and probabilistic certainty when generating correct private outputs. Based on this, PrivacyXray detects privacy breaches using four metrics: intra-layer and inter-layer semantic similarity, token-level and sentence-level probability distributions. PrivacyXray addresses critical challenges in private information detection by overcoming the lack of open-source private datasets and eliminating reliance on external data for validation. It achieves this through the synthesis of realistic private data and a detection mechanism based on the inner states of LLMs. Experiments show that PrivacyXray achieves consistent performance, with an average accuracy of 92.69% across five LLMs. Compared to state-of-the-art methods, PrivacyXray achieves significant improvements, with an average accuracy increase of 20.06%, highlighting its stability and practical utility in real-world applications.

Prior work on input-token importance in auto-regressive transformers has relied on Softmax-normalized attention weights, which obscure the richer structure of pre-Softmax query-key logits. We introduce RCStat, a statistical framework that harnesses raw attention logits via Relative Contextualization (RC), a random variable measuring contextual alignment between token segments, and derive an efficient upper bound for RC. We demonstrate two applications: (i) Key-Value compression, where RC-based thresholds drive adaptive key-value eviction for substantial cache reduction with minimal quality loss; and (ii) Attribution, where RC yields higher-fidelity token-, sentence-, and chunk-level explanations than post-Softmax methods. Across question answering, summarization, and attribution benchmarks, RCStat achieves significant empirical gains, delivering state-of-the-art compression and attribution performance without any model retraining.

Log files provide valuable information for detecting and diagnosing problems in enterprise software applications and data centers. Several log analytics tools and platforms were developed to help filter and extract information from logs, typically using regular expressions (RegExes). Recent commercial log analytics platforms provide domain-specific languages specifically designed for log parsing, such as Grok or the Dynatrace Pattern Language (DPL). However, users who want to migrate to these platforms must manually convert their RegExes into the new pattern language, which is costly and error-prone. In this work, we present Reptile, which combines a rule-based approach for converting RegExes into DPL patterns with a best-effort approach for cases where a full conversion is impossible. Furthermore, it integrates GPT-4 to optimize the obtained DPL patterns. The evaluation with 946 RegExes collected from a large company shows that Reptile safely converted 73.7% of them. The evaluation of Reptile's pattern optimization with 23 real-world RegExes showed an F1-score and MCC above 0.91. These results are promising and have ample practical implications for companies that migrate to a modern log analytics platform, such as Dynatrace.

While Large Language Models (LLMs) possess significant capabilities in open-world agent tasks, they also face challenges in rapidly adapting to new, specialized tasks due to their reliance on static pre-trained knowledge. Traditional methods such as fine-tuning are often costly, data-intensive, and may lead to "catastrophic forgetting." Therefore, we present KnowMap, a novel approach that dynamically constructs a knowledge base from environmental and experiential data. KnowMap fine-tunes a small knowledge-embedding model to equip a larger LLM with valuable task-specific knowledge. Our experiments on the ScienceWorld benchmark demonstrate 17.71% improvement for the performance of gpt-4-turbo model. KnowMap not only provides an efficient and effective means for LLM task-adapting, but also highlights how integrating environmental and experiential knowledge can enhance LLMs' reasoning capabilities.