Retrieval-Augmented Generation (RAG) mitigates hallucinations by grounding generation in retrieved evidence, yet failures persist when models underutilize the retrieved context. While prior work attributes this to conflicts between external context and internal parametric knowledge encoded in Feed-Forward Networks (FFNs), our empirical analyses reveal that FFN-derived signals are unstable and insufficient for reliable diagnosis across diverse contexts. We therefore propose the Attention Escape Score (AES), a training-free diagnostic framework that provides a mechanistic lens on RAG hallucinations via dynamic attention routing failures. We models the failures as weakened Semantic Anchoring, increased Contextual Distraction, and heightened Generation Inertia, moving beyond static component-level analyses. Extensive experiments show that AES significantly outperforms state-of-the-art baselines with robust cross-model generalization. Furthermore, we introduce Pruning Escape Routes (PER), a lightweight inference-time intervention that suppresses hallucinations by attenuating identified escape pathways, achieving improved faithfulness with negligible degradation to complex reasoning and general capabilities.

Effective in-context learning (ICL) for complex reasoning relies on selecting the right demonstrations. Traditional retrieval methods based on surface similarity fail to capture the underlying problem-solving logic. Recent logic-based methods address this by matching predefined reasoning steps, but the rigid rules and exact-match criteria is improper to handle flexible or diverse reasoning processes. To address the problem, we propose SALA, a Semantic-Aware Logical Alignment framework. Instead of relying on a fixed inventory, SALA automatically learns task-specific reasoning operations. It then embeds these operations into a continuous semantic space and uses dynamic time warping (DTW) to align the reasoning sequences. This approach allows for soft, flexible matching of reasoning logic while remaining highly interpretable. Experiments across four reasoning benchmarks and three LLMs demonstrate that SALA outperforms existing demonstration selection methods. Further analysis confirming the roles of the operation induction and the logical semantic alignment.

Sparse autoencoders (SAEs) enable interpretability research by decomposing entangled model activations into monosemantic features. However, under what circumstances SAEs derive most fine-grained latent features for safety—a low-frequency concept domain—remains unexplored. Two key challenges exist: identifying SAEs with the greatest potential for generating safety domain-specific features, and the prohibitively high cost of detailed feature explanation. In this paper, we propose Safe-SAIL, a unified framework for interpreting SAE features in safety-critical domains to advance mechanistic understanding of large language models. Safe-SAIL introduces a pre-explanation evaluation metric to efficiently identify SAEs with strong safety domain-specific interpretability, and reduces interpretation cost by 55% through a segment-level simulation strategy. Building on Safe-SAIL, we train a comprehensive suite of SAEs with human-readable explanations and systematic evaluations for 1,758 safety-related features spanning four domains: pornography, politics, violence, and terror. Using this resource, we conduct empirical analyses and provide insights on the effectiveness of Safe-SAIL for risk feature identification and how safety-critical entities and concepts are encoded across model layers. All models, explanations, and tools are publicly released in an open-source toolkit at https://anonymous.4open.science/r/Safe-SAIL/.

Clinical agents promise to democratize access to electronic health records (EHRs), yet existing benchmarks fail to reflect the complexity of practical EHR analysis, e.g., often operating on idealized, clean EHRs via static SQL generation rather than interactive execution. In this work, we introduce **EHR-Complex**, a large-scale benchmark designed for interactive clinical database reasoning. Built on the large MIMIC-IV substrate (365K patients, 31 tables, 500M+ records), EHR-Complex comprises about 52K tasks spanning six clinical intents, supporting both patient-level and population-level queries, where each task requires an agent to interact with a sandboxed environment by executing SQL queries or Python code. Notably, EHR-Complex considers the real-world SQL task complexity for longitudinal multi-table aggregation and compositional reasoning, resulting in 31.93 SQL structural components per query on average. Evaluation results on EHR-Complex reveal the clinical difficulty of these EHR reasoning scenarios, with the top-performing model achieving only 62.3% exact-match accuracy. Pass^k consistency drops below 50% for nearly all evaluated models at k=4, exposing broad stochastic fragility. A fine-grained analysis of more than 3,800 failed trajectories for representative LLMs reveals three dominant failure modes: SQL logic errors, medical-code lookup failures, and semantic misunderstandings. EHR-Complex provides a rigorous testbed for clinical agents and highlights remaining gaps in robust reasoning for large-scale EHR analysis.

Model merging provides an efficient paradigm for constructing multi-task large language models (LLMs) without full model retraining, yet it remains challenged by parameter interference. While existing methods aim to preserve the capabilities of individual expert models and mitigate interference, they generally do not directly learn from the potentially degraded behaviors exposed by naive merging. In this paper, we propose a conflict-driven preference optimization framework for model merging (CoMerge), which reformulates model merging as a preference optimization problem. The approach utilizes a self-supervised, conflict-driven strategy that leverages the defects of naive merging methods (e.g., task arithmetic) as hard negative samples to construct preference pairs without external annotations. By applying preference optimization to refine lightweight, tensor-wise merging coefficients, CoMerge enables the model to mitigate parameter-space conflicts while preserving task-specific capabilities. Extensive experiments show that CoMerge achieves an average normalized performance of 0.9968 on MergeBench, outperforming all evaluated data-free and data-driven model-merging baselines. Furthermore, on Llama-3.1-8B-Instruct, CoMerge yields marked improvements on conflict-sensitive tasks such as instruction following and safety, while remaining highly competitive with full-parameter fine-tuning despite optimizing only 1,445 scalar coefficients.

Continual Pre-training (CPT) serves as a fundamental approach for adapting foundation models to domain-specific applications. Scaling laws for pre-training define a power-law relationship between dataset size and the test loss of an LLM. However, the marginal gains from simply increasing data for CPT diminish rapidly, yielding suboptimal data utilization and inefficient training. To address this challenge, we propose a novel perplexity-aware data scaling law to establish a predictive relationship between the perplexity landscape of domain-specific data and the test loss. Our approach leverages the pre-trained model’s own perplexity on domain data as a proxy for estimating the knowledge gap, effectively quantifying the informational perplexity landscape of candidate training samples. By fitting this scaling law across diverse perplexity regimes, we enable adaptive selection of high-utility data subsets, prioritizing content that maximizes knowledge absorption while minimizing redundancy and noise. Extensive experiments on both medical and general-domain benchmarks demonstrate that our method consistently identifies near-optimal training subsets, achieving superior performance with significantly reduced data consumption.

Methods for controlling large language models (LLMs), including local weight fine-tuning, LoRA-based adaptation, and activation-based interventions, are often studied in isolation, obscuring their connections and making comparison difficult. In this work, we present a unified view that frames these interventions as dynamic weight updates induced by a control signal, placing them within a single conceptual framework. Building on this view, we propose a unified preference-utility analysis that separates control effects into preference, defined as the tendency toward a target concept, and utility, defined as coherent and task-valid generation, and measures both on a shared log-odds scale using polarity-paired contrastive examples. Across methods, we observe a consistent trade-off between preference and utility: stronger control increases preference while predictably reducing utility. We further explain this behavior through an activation manifold perspective, in which control shifts representations along target-concept directions to enhance preference, while utility declines primarily when interventions push representations off the model’s valid-generation manifold. Finally, we introduce a new steering approach guided by this analysis that improves preference while better preserving utility.

Large language models (LLMs) have demonstrated remarkable capabilities across a wide range of tasks, yet they often refuse to answer legitimate queries—a phenomenon known as overrefusal. Overrefusal typically stems from over-conservative safety alignment, causing models to treat many reasonable prompts as potentially risky. To systematically understand this issue, we probe and leverage the models’ safety decision boundaries to analyze and mitigate overrefusal. Our findings reveal that overrefusal is closely tied to misalignment at these boundary regions, where models struggle to distinguish subtle differences between benign and harmful content. Building on these insights, we present RASS, an automated framework for prompt generation and selection that strategically targets overrefusal prompts near the safety boundary. By harnessing steering vectors in the representation space, RASS efficiently identifies and curates boundary-aligned prompts, enabling more effective and targeted mitigation of overrefusal. This approach not only provides a more precise and interpretable view of model safety decisions but also seamlessly extends to multilingual scenarios. We have explored the safety decision boundaries of various LLMs and construct the MORBench evaluation set to facilitate robust assessment of model safety and helpfulness across multiple languages. Code and datasets are available at https://github.com/Master-PLC/RASS.

Recently, the self-consistency decoding strategy has shown the ability to improve performance for complex reasoning tasks with large language models (LLMs). However, the costs may be high because the sampling process of the strategy generates some low-probability text, resulting in low-quality reasoning paths. As a consequence, it requires a relatively large sampling number to obtain good aggregation performance. In this paper, we propose an alternative strategy, self-para-consistency. It first generates multiple paraphrases for each test question, then generates reasoning paths for the original and all the paraphrased questions based on greedy decoding, and finally selects the most consistent answer. Since all the candidate paths have relatively high probabilities, the sampling number could be much smaller than the self-consistency strategy. Extensive experiments on complex reasoning datasets demonstrate the effectiveness of our method in reducing the sampling number.

Although the Conditional Variational Auto-Encoder (CVAE) model can generate more diversified responses than the traditional Seq2Seq model, the responses often have low relevance with the input words or are illogical with the question. A causal analysis is carried out to study the reasons behind, and a methodology of searching for the mediators and mitigating the confounding bias in dialogues is provided. Specifically, we propose to predict the mediators to preserve relevant information and auto-regressively incorporate the mediators into generating process. Besides, a dynamic topic graph guided conditional variational auto-encoder (TGG-CVAE) model is utilized to complement the semantic space and reduce the confounding bias in responses. Extensive experiments demonstrate that the proposed model is able to generate both relevant and informative responses, and outperforms the state-of-the-art in terms of automatic metrics and human evaluations.