Safety alignment in large language models (LLMs) is achieved through fine-tuning mechanisms that regulate neuron activations to suppress harmful content. In this work, we propose a novel approach to induce disalignment by identifying and modifying the neurons responsible for safety constraints. Our method consists of three key steps: Neuron Activation Analysis, where we examine activation patterns in response to harmful and harmless prompts to detect neurons that are critical for distinguishing between harmful and harmless inputs; Similarity-Based Neuron Identification, which systematically locates the neurons responsible for safe alignment; and Neuron Relearning for Safety Removal, where we fine-tune these selected neurons to restore the model's ability to generate previously restricted responses. Experimental results demonstrate that our method effectively removes safety constraints with minimal fine-tuning, highlighting a critical vulnerability in current alignment techniques. Our findings underscore the need for robust defenses against adversarial fine-tuning attacks on LLMs.

MLLMs have recently become a focal point in the field of artificial intelligence research. Building on the strong capabilities of LLMs, MLLMs are adept at addressing complex multi-modal tasks. With the release of GPT-4, MLLMs have gained substantial attention from different domains. Researchers have begun to explore the potential of MLLMs in the medical and healthcare domain. In this paper, we first introduce the background and fundamental concepts related to LLMs and MLLMs, while emphasizing the working principles of MLLMs. Subsequently, we summarize three main directions of application within healthcare: medical reporting, medical diagnosis, and medical treatment. Our findings are based on a comprehensive review of 330 recent papers in this area. We illustrate the remarkable capabilities of MLLMs in these domains by providing specific examples. For data, we present six mainstream modes of data along with their corresponding evaluation benchmarks. At the end of the survey, we discuss the challenges faced by MLLMs in the medical and healthcare domain and propose feasible methods to mitigate or overcome these issues.

According to the Open Web Application Security Project (OWASP), Cross-Site Scripting (XSS) is a critical security vulnerability. Despite decades of research, XSS remains among the top 10 security vulnerabilities. Researchers have proposed various techniques to protect systems from XSS attacks, with machine learning (ML) being one of the most widely used methods. An ML model is trained on a dataset to identify potential XSS threats, making its effectiveness highly dependent on the size and diversity of the training data. A variation of XSS is obfuscated XSS, where attackers apply obfuscation techniques to alter the code's structure, making it challenging for security systems to detect its malicious intent. Our study's random forest model was trained on traditional (non-obfuscated) XSS data achieved 99.8% accuracy. However, when tested against obfuscated XSS samples, accuracy dropped to 81.9%, underscoring the importance of training ML models with obfuscated data to improve their effectiveness in detecting XSS attacks. A significant challenge is to generate highly complex obfuscated code despite the availability of several public tools. These tools can only produce obfuscation up to certain levels of complexity. In our proposed system, we fine-tune a Large Language Model (LLM) to generate complex obfuscated XSS payloads automatically. By transforming original XSS samples into diverse obfuscated variants, we create challenging training data for ML model evaluation. Our approach achieved a 99.5% accuracy rate with the obfuscated dataset. We also found that the obfuscated samples generated by the LLMs were 28.1% more complex than those created by other tools, significantly improving the model's ability to handle advanced XSS attacks and making it more effective for real-world application security.

As transformer-based large language models (LLMs) increasingly permeate society, they have revolutionized domains such as software engineering, creative writing, and digital arts. However, their adoption in cybersecurity remains limited due to challenges like scarcity of specialized training data and complexity of representing cybersecurity-specific knowledge. To address these gaps, we present Foundation-Sec-8B, a cybersecurity-focused LLM built on the Llama 3.1 architecture and enhanced through continued pretraining on a carefully curated cyber-security corpus. We evaluate Foundation-Sec-8B across both established and new cybersecurity benchmarks, showing that it matches Llama 3.1-70B and GPT -4o-mini in certain cybersecurity-specific tasks. By releasing our model to the public, we aim to accelerate progress and adoption of AI-driven tools in both public and private cybersecurity contexts.

Early detection of security bug reports (SBRs) is crucial for preventing vulnerabilities and ensuring system reliability. While machine learning models have been developed for SBR prediction, their predictive performance still has room for improvement. In this study, we conduct a comprehensive comparison between BERT and Random Forest (RF), a competitive baseline for predicting SBRs. The results show that RF outperforms BERT with a 34% higher average G-measure for within-project predictions. Adding only SBRs from various projects improves both models' average performance. However, including both security and nonsecurity bug reports significantly reduces RF's average performance to 46%, while boosts BERT to its best average performance of 66%, surpassing RF. In cross-project SBR prediction, BERT achieves a remarkable 62% G-measure, which is substantially higher than RF.

The perceived quality of the explanations accompanying e-government services is key to gaining trust in these institutions, consequently amplifying further usage of these services. Recent advances in generative AI, and concretely in Large Language Models (LLMs) allow the automation of such content articulations, eliciting explanations' interpretability and fidelity, and more generally, adapting content to various audiences. However, selecting the right LLM type for this has become a non-trivial task for e-government service providers. In this work, we adapted a previously developed scale to assist with this selection, providing a systematic approach for the comparative analysis of the perceived quality of explanations generated by various LLMs. We further demonstrated its applicability through the tax-return process, using it as an exemplar use case that could benefit from employing an LLM to generate explanations about tax refund decisions. This was attained through a user study with 128 survey respondents who were asked to rate different versions of LLM-generated explanations about tax refund decisions, providing a methodological basis for selecting the most appropriate LLM. Recognizing the practical challenges of conducting such a survey, we also began exploring the automation of this process by attempting to replicate human feedback using a selection of cutting-edge predictive techniques.

Prompt injection attacks have emerged as a serious threat in curr ent large language models (LLMs), where adversaries may alter model behav ior by injecting malicious instructions into the prompt [2]. Existing approach es - such as input sanitization, fixed prompt templates, and heuristic-based filtering - often mix trusted system instructions with untrusted us er inputs, leading to brittle defenses that are easily circumvented. For examp le, an adversary could include a cleverly worded request that causes the model to "forget its internal guidelines," thereby triggering unintended beh avior. Our PICO (Prompt Isolation and Cybersecurity Oversight) propos al circumvents these limitations, first of all, by architecturally segregat ing the system prompt and user input into distinct channels. In doing so, we ensure that the trusted instructions remain intact while only the untruste d user input is subject to adaptation. Furthermore, we augment the mode l with a dedicated Security Expert Agent and a Cybersecurity Knowledge G raph [4] to provide supplemental, domain-specific signals that reinforce the invariant. In what follows, we first present a mathematical formalization of th e PICO security strategy, and then we describe its concrete realiza tion, both via PICO-based retraining of transformer models from the bottom up, and via a more efficient if less ideal fine-tuning strategy. We flesh out theapproach by considering how it would be expected to handle two specific example situations, including a basic prompt injection and then a subtler Policy Puppetry attack.

--The rapid evolution of malware variants requires robust classification methods to enhance cybersecurity. While Large Language Models (LLMs) offer potential for generating malware descriptions to aid family classification, their utility is limited by semantic embedding overlaps and misalignment with binary behavioral features. We propose a contrastive fine-tuning (CFT) method that refines LLM embeddings via targeted selection of hard negative samples based on cosine similarity, enabling LLMs to distinguish between closely related malware families. Our approach combines high-similarity negatives to enhance discriminative power and mid-tier negatives to increase embedding diversity, optimizing both precision and generalization. Evaluated on the CIC-AndMal-2020 and BODMAS datasets, our refined embeddings are integrated into a multimodal classifier within a Model-Agnostic Meta-Learning (MAML) framework on a few-shot setting. Experiments demonstrate significant improvements: our method achieves 63.15% classification accuracy with as few as 20 samples on CIC-AndMal-2020, outperforming baselines by 11-21 percentage points and surpassing prior negative sampling strategies. Ablation studies confirm the superiority of similarity-based selection over random sampling, with gains of 10-23%. Additionally, fine-tuned LLMs generate attribute-aware descriptions that generalize to unseen variants, bridging textual and binary feature gaps. This work advances malware classification by enabling nuanced semantic distinctions and provides a scalable framework for adapting LLMs to cybersecurity challenges.

The advent of Large Language Models (LLMs) holds promise for revolutionizing various fields traditionally dominated by human expertise. Urban planning, a professional discipline that fundamentally shapes our daily surroundings, is one such field heavily relying on multifaceted domain knowledge and experience of human experts. The extent to which LLMs can assist human practitioners in urban planning remains largely unexplored. In this paper, we introduce a comprehensive benchmark, UrbanPlanBench, tailored to evaluate the efficacy of LLMs in urban planning, which encompasses fundamental principles, professional knowledge, and management and regulations, aligning closely with the qualifications expected of human planners. Through extensive evaluation, we reveal a significant imbalance in the acquisition of planning knowledge among LLMs, with even the most proficient models falling short of meeting professional standards. For instance, we observe that 70% of LLMs achieve subpar performance in understanding planning regulations compared to other aspects. Besides the benchmark, we present the largest-ever supervised fine-tuning (SFT) dataset, UrbanPlanText, comprising over 30,000 instruction pairs sourced from urban planning exams and textbooks. Our findings demonstrate that fine-tuned models exhibit enhanced performance in memorization tests and comprehension of urban planning knowledge, while there exists significant room for improvement, particularly in tasks requiring domain-specific terminology and reasoning. By making our benchmark, dataset, and associated evaluation and fine-tuning toolsets publicly available at https://github.com/tsinghua-fib-lab/PlanBench, we aim to catalyze the integration of LLMs into practical urban planning, fostering a symbiotic collaboration between human expertise and machine intelligence.

The post-training phase of large language models is essential for enhancing capabilities such as instruction-following, reasoning, and alignment with human preferences. However, it demands extensive high-quality data and poses risks like overfitting, alongside significant computational costs due to repeated post-training and evaluation after each base model update. This paper introduces $ParamΔ$, a novel method that streamlines post-training by transferring knowledge from an existing post-trained model to a newly updated base model with ZERO additional training. By computing the difference between post-trained model weights ($Θ_\text{post}$) and base model weights ($Θ_\text{base}$), and adding this to the updated base model ($Θ'_\text{base}$), we define $ParamΔ$ Model as: $Θ_{\text{Param}Δ} = Θ_\text{post} - Θ_\text{base} + Θ'_\text{base}$. This approach surprisingly equips the new base model with post-trained capabilities, achieving performance comparable to direct post-training. We did analysis on LLama3, Llama3.1, Qwen, and DeepSeek-distilled models. Results indicate $ParamΔ$ Model effectively replicates traditional post-training. For example, the $ParamΔ$ Model obtained from 70B Llama3-inst, Llama3-base, Llama3.1-base models attains approximately 95\% of Llama3.1-inst model's performance on average. $ParamΔ$ brings a new perspective on how to fully leverage models in the open-weight community, where checkpoints for base and instruct models are readily available and frequently updated, by providing a cost-free framework to accelerate the iterative cycle of model development.