The static nature of knowledge within Large Language Models (LLMs) makes it difficult for them to adapt to evolving information, rendering knowledge editing a critical task. However, existing methods struggle with challenges of scalability and retrieval efficiency, particularly when handling complex, multi-hop questions that require multi-step reasoning. To address these challenges, this paper introduces ALEX (A Light Editing-knowledge Extractor), a lightweight knowledge editing framework. The core innovation of ALEX is its hierarchical memory architecture, which organizes knowledge updates (edits) into semantic clusters. This design fundamentally reduces retrieval complexity from a linear O(N) to a highly scalable O(K + N/C). Furthermore, the framework integrates an Inferential Query Synthesis (IQS) module to bridge the semantic gap between queries and facts, and a Dynamic Evidence Adjudication (DEA) engine that executes an efficient two-stage retrieval process. Experiments on the MQUAKE benchmark demonstrate that ALEX significantly improves both the accuracy of multi-hop answers (MultiHop-ACC) and the reliability of reasoning paths (HopWise-ACC). It also reduces the required search space by over 80%, presenting a promising path toward building scalable, efficient, and accurate knowledge editing systems.

Recent work has shown that fine-tuning on insecure code data can trigger an emergent misalignment (EMA) phenomenon, where models generate malicious responses even to prompts unrelated to the original insecure code-writing task. Such cross-domain generalization of harmful behavior underscores the need for a deeper understanding of the algorithms, tasks, and datasets that induce emergent misalignment. In this work, we extend this study by demonstrating that emergent misalignment can also arise from narrow refusal unlearning in specific domains. We perform refusal unlearning on Cybersecurity and Safety concept, and evaluate EMA by monitoring refusal scores across seven responsible AI (RAI) domains, Cybersecurity, Safety, Toxicity, Bias, Sensitive Content, Medical/Legal, and Privacy. Our work shows that narrow domain unlearning can yield compliance responses for the targeted concept, however, it may also propagate EMA to unrelated domains. Among the two intervened concepts, Cybersecurity and Safety, we find that the safety concept can have larger EMA impact, i.e, causing lower refusal scores, across other unrelated domains such as bias. We observe this effect consistently across two model families, Mistral-7b-0.3v, and Qwen-7b-2.5. Further, we show that refusal unlearning augmented with cross-entropy loss function on a small set of retain data from the affected domains can largely, if not fully, restore alignment across the impacted domains while having lower refusal rate on the concept we perform unlearning on. To investigate the underlying causes of EMA, we analyze concept entanglements at the representation level via concept vectors. Our analysis reveals that concepts with higher representation similarity in earlier layers are more susceptible to EMA after intervention when the refusal stream is altered through targeted refusal unlearning.

Flaky tests that non-deterministically pass or fail waste developer time and slow release cycles. While large language models (LLMs) show promise for automatically repairing flaky tests, existing approaches like FlakyDoctor fail in industrial settings due to the context problem: providing either too little context (missing critical production code) or too much context (overwhelming the LLM with irrelevant information). We present FlakyGuard, which addresses this problem by treating code as a graph structure and using selective graph exploration to find only the most relevant context. Evaluation on real-world flaky tests from industrial repositories shows that FlakyGuard repairs 47.6 % of reproducible flaky tests with 51.8 % of the fixes accepted by developers. Besides it outperforms state-of-the-art approaches by at least 22 % in repair success rate. Developer surveys confirm that 100 % find FlakyGuard's root cause explanations useful.

Search queries with superlatives (e.g., best, most popular) require comparing candidates across multiple dimensions, demanding linguistic understanding and domain knowledge. We show that LLMs can uncover latent intent behind these expressions in e-commerce queries through a framework that extracts structured interpretations or hints. Our approach decomposes queries into attribute-value hints generated concurrently with retrieval, enabling efficient integration into the ranking pipeline. Our method improves search performanc eby 10.9 points in MAP and ranking by 5.9 points in MRR over baselines. Since direct LLM-based reranking faces prohibitive latency, we develop an efficient approach transferring superlative interpretations to lightweight models. Our findings provide insights into how superlative semantics can be represented and transferred between models, advancing linguistic interpretation in retrieval systems while addressing practical deployment constraints.

Language models generate functionally correct code that tends toward excessive verbosity, with elaborate documentation and defensive patterns that diverge from human baselines. Two prompting mechanisms have emerged for stylistic control: instruction based prompts that articulate abstract directives, and example based prompts that provide concrete code demonstrations. The core problem is whether stylistic constraints persist when models enhance initial implementations with additional features while maintaining high functional accuracy. Here we show that instruction-based, example-based, and combined prompts produce distinct patterns of initial control and expansion discipline over one enhancement turn. We manipulated system prompts across four conditions in a paired two-turn protocol where models first generated solutions to an intermediate Python task, then revised their code under general improvement directives, holding the user task fixed (N = 160 paired programs). Combined prompts produced the strongest initial compression and greatest expansion discipline. Instructions showed large initial effects and moderate expansion discipline. Examples showed modest initial effects with no expansion discipline. These results show that initial prompt effectiveness and expansion discipline are separate aspects of prompt design, and that combined approaches provide the most stable stylistic control in this two-turn workflow.

Training vision models to detect workplace hazards accurately requires realistic images of unsafe conditions that could lead to accidents. However, acquiring such datasets is difficult because capturing accident-triggering scenarios as they occur is nearly impossible. To overcome this limitation, this study presents a novel scene graph-guided generative AI framework that synthesizes photorealistic images of hazardous scenarios grounded in historical Occupational Safety and Health Administration (OSHA) accident reports. OSHA narratives are analyzed using GPT-4o to extract structured hazard reasoning, which is converted into object-level scene graphs capturing spatial and contextual relationships essential for understanding risk. These graphs guide a text-to-image diffusion model to generate compositionally accurate hazard scenes. To evaluate the realism and semantic fidelity of the generated data, a visual question answering (VQA) framework is introduced. Across four state-of-the-art generative models, the proposed VQA Graph Score outperforms CLIP and BLIP metrics based on entropy-based validation, confirming its higher discriminative sensitivity.

Purpose: Echocardiographic interpretation requires video-level reasoning and guideline-based measurement analysis, which current deep learning models for cardiac ultrasound do not support. We present EchoAgent, a framework that enables structured, interpretable automation for this domain. Methods: EchoAgent orchestrates specialized vision tools under Large Language Model (LLM) control to perform temporal localization, spatial measurement, and clinical interpretation. A key contribution is a measurement-feasibility prediction model that determines whether anatomical structures are reliably measurable in each frame, enabling autonomous tool selection. We curated a benchmark of diverse, clinically validated video-query pairs for evaluation. Results: EchoAgent achieves accurate, interpretable results despite added complexity of spatiotemporal video analysis. Outputs are grounded in visual evidence and clinical guidelines, supporting transparency and traceability. Conclusion: This work demonstrates the feasibility of agentic, guideline-aligned reasoning for echocardiographic video analysis, enabled by task-specific tools and full video-level automation. EchoAgent sets a new direction for trustworthy AI in cardiac ultrasound.

Despite the parallel challenges that audio and text domains face in evaluating generative model outputs, preference learning remains remarkably underexplored in audio applications. Through a PRISMA-guided systematic review of approximately 500 papers, we find that only 30 (6%) apply preference learning to audio tasks. Our analysis reveals a field in transition: pre-2021 works focused on emotion recognition using traditional ranking methods (rankSVM), while post-2021 studies have pivoted toward generation tasks employing modern RLHF frameworks. We identify three critical patterns: (1) the emergence of multi-dimensional evaluation strategies combining synthetic, automated, and human preferences; (2) inconsistent alignment between traditional metrics (WER, PESQ) and human judgments across different contexts; and (3) convergence on multi-stage training pipelines that combine reward signals. Our findings suggest that while preference learning shows promise for audio, particularly in capturing subjective qualities like naturalness and musicality, the field requires standardized benchmarks, higher-quality datasets, and systematic investigation of how temporal factors unique to audio impact preference learning frameworks.

The diminishing ability of large language models (LLMs) to effectively utilize long-range context-the "lost-in-the-middle" phenomenon-poses a significant challenge in retrieval-based LLM applications. To study the impact of this phenomenon in a real-world application setting, we introduce GM-Extract, a novel benchmark dataset meticulously designed to evaluate LLM performance on retrieval of control variables. To accurately diagnose failure modes, we propose a simple yet elegant evaluation system using two distinct metrics: one for spatial retrieval capability (Document Metric) and the other for semantic retrieval capability (Variable Extraction Metric). We conduct a systematic evaluation of 7-8B parameter models on two multi-document tasks (key-value extraction and question-answering), demonstrating a significant change in retrieval performance simply by altering how the data is represented in the context window. While a distinct U-shaped curve was not consistently observed, our analysis reveals a clear pattern of performance across models, which we further correlate with perplexity scores. Furthermore, we perform a literature survey of mitigation methods, which we categorize into two distinct approaches: black-box and white-box methods. We then apply these techniques to our benchmark, finding that their efficacy is highly nuanced. Our evaluation highlights scenarios where these strategies successfully improve performance, as well as surprising cases where they lead to a negative impact, providing a comprehensive understanding of their utility in a practical context.

Large Vision Language Models (LVLMs) demonstrate strong capabilities in multimodal reasoning and many real-world applications, such as visual question answering. However, LVLMs are highly vulnerable to jailbreaking attacks. This paper systematically analyzes the vulnerabilities of LVLMs integrated in Intelligent Transportation Systems (ITS) under carefully crafted jailbreaking attacks. First, we carefully construct a dataset with harmful queries relevant to transportation, following OpenAI's prohibited categories to which the LVLMs should not respond. Second, we introduce a novel jailbreaking attack that exploits the vulnerabilities of LVLMs through image typography manipulation and multi-turn prompting. Third, we propose a multi-layered response filtering defense technique to prevent the model from generating inappropriate responses. We perform extensive experiments with the proposed attack and defense on the state-of-the-art LVLMs (both open-source and closed-source). To evaluate the attack method and defense technique, we use GPT-4's judgment to determine the toxicity score of the generated responses, as well as manual verification. Further, we compare our proposed jailbreaking method with existing jailbreaking techniques and highlight severe security risks involved with jailbreaking attacks with image typography manipulation and multi-turn prompting in the LVLMs integrated in ITS.