Fine-tuning pre-trained Large Language Models (LLMs) for specialized tasks incurs substantial computational and data costs. While model merging offers a training-free solution to integrate multiple task-specific models, existing methods suffer from safety-utility conflicts where enhanced general capabilities degrade safety safeguards. We identify two root causes: \textbf{neuron misidentification} due to simplistic parameter magnitude-based selection, and \textbf{cross-task neuron interference} during merging. To address these challenges, we propose \textbf{LED-Merging}, a three-stage framework that \textbf{L}ocates task-specific neurons via gradient-based attribution, dynamically \textbf{E}lects critical neurons through multi-model importance fusion, and \textbf{D}isjoints conflicting updates through parameter isolation. Extensive experiments on Llama-3-8B, Mistral-7B, and Llama2-13B demonstrate that LED-Merging reduces harmful response rates(\emph{e.g.}, a 31.4\% decrease on Llama-3-8B-Instruct on HarmBench) while preserving 95\% of utility performance(\emph{e.g.}, 52.39\% accuracy on GSM8K). LED-Merging resolves safety-utility conflicts and provides a lightweight, training-free paradigm for constructing reliable multi-task LLMs.

Regulatory texts are inherently long and complex, presenting significant challenges for information retrieval systems in supporting regulatory officers with compliance tasks. This paper introduces a hybrid information retrieval system that combines lexical and semantic search techniques to extract relevant information from large regulatory corpora. The system integrates a fine-tuned sentence transformer model with the traditional BM25 algorithm to achieve both semantic precision and lexical coverage. To generate accurate and comprehensive responses, retrieved passages are synthesized using Large Language Models (LLMs) within a Retrieval Augmented Generation (RAG) framework. Experimental results demonstrate that the hybrid system significantly outperforms standalone lexical and semantic approaches, with notable improvements in Recall@10 and MAP@10. By openly sharing our fine-tuned model and methodology, we aim to advance the development of robust natural language processing tools for compliance-driven applications in regulatory domains.

Large language models (LLMs) present novel opportunities in public opinion research by predicting survey responses in advance during the early stages of survey design. Prior methods steer LLMs via descriptions of subpopulations as LLMs' input prompt, yet such prompt engineering approaches have struggled to faithfully predict the distribution of survey responses from human subjects. In this work, we propose directly fine-tuning LLMs to predict response distributions by leveraging unique structural characteristics of survey data. To enable fine-tuning, we curate SubPOP, a significantly scaled dataset of 3,362 questions and 70K subpopulation-response pairs from well-established public opinion surveys. We show that fine-tuning on SubPOP greatly improves the match between LLM predictions and human responses across various subpopulations, reducing the LLM-human gap by up to 46% compared to baselines, and achieves strong generalization to unseen surveys and subpopulations. Our findings highlight the potential of survey-based fine-tuning to improve opinion prediction for diverse, real-world subpopulations and therefore enable more efficient survey designs. Our code is available at https://github.com/JosephJeesungSuh/subpop.

Coreset Selection (CS) identifies a subset of training data that achieves model performance comparable to using the entire dataset. Many state-of-the-art CS methods, select coresets using scores whose computation requires training the downstream model on the entire dataset and recording changes in its behavior on samples as it trains (training dynamics). These scores are inefficient to compute and hard to interpret as they do not indicate whether a sample is difficult to learn in general or only for a specific model. Our work addresses these challenges by proposing an interpretable score that gauges a sample's difficulty using human-understandable textual attributes (concepts) independent of any downstream model. Specifically, we measure the alignment between a sample's visual features and concept bottlenecks, derived via large language models, by training a linear concept bottleneck layer and compute the sample's difficulty score using it. We then use this score and a stratified sampling strategy to identify the coreset. Crucially, our score is efficiently computable without training the downstream model on the full dataset even once, leads to high-performing coresets for various downstream models, and is computable even for an unlabeled dataset. Through experiments on CIFAR-10, CIFAR-100, and ImageNet-1K, we show our coresets outperform random subsets, even at high pruning rates, and achieve model performance comparable to or better than coresets found by training dynamics-based methods.

We present RapidPen, a fully automated penetration testing (pentesting) framework that addresses the challenge of achieving an initial foothold (IP-to-Shell) without human intervention. Unlike prior approaches that focus primarily on post-exploitation or require a human-in-the-loop, RapidPen leverages large language models (LLMs) to autonomously discover and exploit vulnerabilities, starting from a single IP address. By integrating advanced ReAct-style task planning (Re) with retrieval-augmented knowledge bases of successful exploits, along with a command-generation and direct execution feedback loop (Act), RapidPen systematically scans services, identifies viable attack vectors, and executes targeted exploits in a fully automated manner. In our evaluation against a vulnerable target from the Hack The Box platform, RapidPen achieved shell access within 200-400 seconds at a per-run cost of approximately \$0.3-\$0.6, demonstrating a 60\% success rate when reusing prior "success-case" data. These results underscore the potential of truly autonomous pentesting for both security novices and seasoned professionals. Organizations without dedicated security teams can leverage RapidPen to quickly identify critical vulnerabilities, while expert pentesters can offload repetitive tasks and focus on complex challenges. Ultimately, our work aims to make penetration testing more accessible and cost-efficient, thereby enhancing the overall security posture of modern software ecosystems.

Fine-tuning pre-trained transformers is a powerful technique for enhancing the performance of base models on specific tasks. From early applications in models like BERT to fine-tuning Large Language Models (LLMs), this approach has been instrumental in adapting general-purpose architectures for specialized downstream tasks. Understanding the fine-tuning process is crucial for uncovering how transformers adapt to specific objectives, retain general representations, and acquire task-specific features. This paper explores the underlying mechanisms of fine-tuning, specifically in the BERT transformer, by analyzing activation similarity, training Sparse AutoEncoders (SAEs), and visualizing token-level activations across different layers. Based on experiments conducted across multiple datasets and BERT layers, we observe a steady progression in how features adapt to the task at hand: early layers primarily retain general representations, middle layers act as a transition between general and task-specific features, and later layers fully specialize in task adaptation. These findings provide key insights into the inner workings of fine-tuning and its impact on representation learning within transformer architectures.

Unfortunately, these models are closed and can only be accessed through the user interfaces, tools or application programming interfaces provided by the companies that developed the models. Their parameters and implementation details are not publicly available and even if they were, their huge size would make their execution on commodity computing devices unfeasible. A different approach has been taken by some large companies such as Meta, i.e. the code as well as the parameters or weights of LLMs such as LLaMa

Language Models (LMs) have demonstrated impressive capabilities in solving complex reasoning tasks, particularly when prompted to generate intermediate explanations. However, it remains an open question whether these intermediate reasoning traces represent a dynamic, evolving thought process or merely reflect sophisticated pattern recognition acquired during large scale pre training. Drawing inspiration from human cognition, where reasoning unfolds incrementally as new information is assimilated and internal models are continuously updated, we propose to delve deeper into the mental model of various LMs. We propose a new way to assess the mental modeling of LMs, where they are provided with problem details gradually, allowing each new piece of data to build upon and refine the model's internal representation of the task. We systematically compare this step by step mental modeling strategy with traditional full prompt methods across both text only and vision and text modalities. Experiments on the MathWorld dataset across different model sizes and problem complexities confirm that both text-based LLMs and multimodal LMs struggle to create mental representations, questioning how their internal cognitive processes work.

This enables the identification of the text for which AI assistance has been used. How AI was used AI tools can be used for many different tasks: summarizing, translation, paraphrasing, finding related work and citations, etc. So, it is important to have information on how AI tools were used in the paper. For example, we can encode in the metadata that GPT -4 (so the "which") was used to summarize (the "how") and write the abstract (the "where").UNDERSTANDING THE IMP ACT OF AI IN ACADEMIC WRITING Let us consider now that we have a large corpus of papers and we want to know how many of them have used AI to summarize the abstract. Without metadata, all papers look the same (Figure 2, left), so we have to extract the text and either try to detect the use of AI in the abstract or find a disclosure of the authors that states the use of AI in the abstract. Instead if the proposed metadata has been added to the paper, we can just look at the how (summarizing) and where (abstract) to find the papers. The papers are now marked and can be easily identified (Figure 2, right). The metadata can be used to analyze many aspects of the use of AI in academic writing, for example, we can analyze: 1) The adoption of the different AI tools and their variations over time.

Solving complex long-horizon robotic manipulation problems requires sophisticated high-level planning capabilities, the ability to reason about the physical world, and reactively choose appropriate motor skills. Vision-language models (VLMs) pretrained on Internet data could in principle offer a framework for tackling such problems. However, in their current form, VLMs lack both the nuanced understanding of intricate physics required for robotic manipulation and the ability to reason over long horizons to address error compounding issues. In this paper, we introduce a novel test-time computation framework that enhances VLMs' physical reasoning capabilities for multi-stage manipulation tasks. At its core, our approach iteratively improves a pretrained VLM with a "reflection" mechanism - it uses a generative model to imagine future world states, leverages these predictions to guide action selection, and critically reflects on potential suboptimalities to refine its reasoning. Experimental results demonstrate that our method significantly outperforms several state-of-the-art commercial VLMs as well as other post-training approaches such as Monte Carlo Tree Search (MCTS). Videos are available at https://reflect-vlm.github.io.