Machine reasoning has made great progress in recent years owing to large language models (LLMs). In the clinical domain, however, most NLP-driven projects mainly focus on clinical classification or reading comprehension, and under-explore clinical reasoning for disease diagnosis due to the expensive rationale annotation with clinicians. In this work, we present a ``reasoning-aware'' diagnosis framework that rationalizes the diagnostic process via prompt-based learning in a time- and labor-efficient manner, and learns to reason over the prompt-generated rationales. Specifically, we address the clinical reasoning for disease diagnosis, where the LLM generates diagnostic rationales providing its insight on presented patient data and the reasoning path towards the diagnosis, namely Clinical Chain-of-Thought (Clinical CoT). We empirically demonstrate LLMs/LMs' ability of clinical reasoning via extensive experiments and analyses on both rationale generation and disease diagnosis in various settings. We further propose a novel set of criteria for evaluating machine-generated rationales' potential for real-world clinical settings, facilitating and benefiting future research in this area.

Understanding long, real-world videos requires modeling of long-range visual dependencies. To this end, we explore video-first architectures, building on the common paradigm of transferring large-scale, image--text models to video via shallow temporal fusion. However, we expose two limitations to the approach: (1) decreased spatial capabilities, likely due to poor video--language alignment in standard video datasets, and (2) higher memory consumption, bottlenecking the number of frames that can be processed. To mitigate the memory bottleneck, we systematically analyze the memory/accuracy trade-off of various efficient methods: factorized attention, parameter-efficient image-to-video adaptation, input masking, and multi-resolution patchification. Surprisingly, simply masking large portions of the video (up to 75%) during contrastive pre-training proves to be one of the most robust ways to scale encoders to videos up to 4.3 minutes at 1 FPS. Our simple approach for training long video-to-text models, which scales to 1B parameters, does not add new architectural complexity and is able to outperform the popular paradigm of using much larger LLMs as an information aggregator over segment-based information on benchmarks with long-range temporal dependencies (YouCook2, EgoSchema).

Sequential planning in large state space and action space quickly becomes intractable due to combinatorial explosion of the search space. Heuristic methods, like monte-carlo tree search, though effective for large state space, but struggle if action space is large. Pure reinforcement learning methods, relying only on reward signals, needs prohibitively large interactions with the environment to device a viable plan. If the state space, observations and actions can be represented in natural language then Large Language models (LLM) can be used to generate action plans. Recently several such goal-directed agents like Reflexion, CLIN, SayCan were able to surpass the performance of other state-of-the-art methods with minimum or no task specific training. But they still struggle with exploration and get stuck in local optima. Their planning capabilities are limited by the limited reasoning capability of the foundational LLMs on text data. We propose a hybrid agent "neoplanner", that synergizes both state space search with queries to foundational LLM to get the best action plan. The reward signals are quantitatively used to drive the search. A balance of exploration and exploitation is maintained by maximizing upper confidence bounds of values of states. In places where random exploration is needed, the LLM is queried to generate an action plan. Learnings from each trial are stored as entity relationships in text format. Those are used in future queries to the LLM for continual improvement. Experiments in the Scienceworld environment reveals a 124% improvement from the current best method in terms of average reward gained across multiple tasks.

Sparse attention as a efficient method can significantly decrease the computation cost, but current sparse attention tend to rely on window self attention which block the global information flow. For this problem, we present Shifted Cross Chunk Attention (SCCA), using different KV shifting strategy to extend respective field in each attention layer. Except, we combine Dilated Attention(DA) and Dilated Neighborhood Attention(DNA) to present Shifted Dilated Attention(SDA). Both SCCA and SDA can accumulate attention results in multi head attention to obtain approximate respective field in full attention. In this paper, we conduct language modeling experiments using different pattern of SCCA and combination of SCCA and SDA. The proposed shifted cross chunk attention (SCCA) can effectively extend large language models (LLMs) to longer context combined with Positional interpolation(PI) and LoRA than current sparse attention. Notably, SCCA adopts LLaMA2 7B from 4k context to 8k in single V100. This attention pattern can provide a Plug-and-play fine-tuning method to extend model context while retaining their original architectures, and is compatible with most existing techniques.

Text-to-image generative models offer many innovative services but also raise ethical concerns due to their potential to generate unethical images. Most publicly available text-to-image models employ safety filters to prevent unintended generation intents. In this work, we introduce the Divide-and-Conquer Attack to circumvent the safety filters of state-of-the-art text-to-image models. Our attack leverages LLMs as agents for text transformation, creating adversarial prompts from sensitive ones. We have developed effective helper prompts that enable LLMs to break down sensitive drawing prompts into multiple harmless descriptions, allowing them to bypass safety filters while still generating sensitive images. This means that the latent harmful meaning only becomes apparent when all individual elements are drawn together. Our evaluation demonstrates that our attack successfully circumvents the closed-box safety filter of SOTA DALLE-3 integrated natively into ChatGPT to generate unethical images. This approach, which essentially uses LLM-generated adversarial prompts against GPT-4-assisted DALLE-3, is akin to using one's own spear to breach their shield. It could have more severe security implications than previous manual crafting or iterative model querying methods, and we hope it stimulates more attention towards similar efforts. Our code and data are available at: https://github.com/researchcode001/Divide-and-Conquer-Attack

Secure and reliable information systems have become a central requirement for the operational continuity of the vast majority of goods and services providers [42]. However, securing information systems in a fast-paced ecosystem of technological changes and innovations is hard [3]. New technologies in cybersecurity have short life cycles and constantly evolve [13]. This exposes information systems to attacks that exploit vulnerabilities and security gaps [3]. Hence, cybersecurity practitioners and researchers need to stay updated on the latest developments and trends to prevent incidents and increase resilience [14]. A common approach to gather cured and synthesized information about such developments is to apply bibliometrics-based knowledge entity extraction and comparison through embedding similarity [10, 50, 61] - recently boosted by the availability of entity extractors based on large language models (LLMs) [17, 46]. However, it is unclear how appropriate this approach is for the cybersecurity literature. We address this by emulating such an entity extraction and comparison pipeline, and by using a variety of common entity extractors - LLM-based and not -, and evaluating how relevant embeddings of extracted entities are to document understanding tasks - namely classification of arXiv documents as relevant to cybersecurity (https://arxiv.org). While LLMs burst into public attention in late 2022 - in large part thanks to public trials of conversationally fine-tuned LLMs [40, 4, 31]-, modern large language models pre-trained on large amounts of data trace their roots back to ELMo LLM, first released in 2018 [45].

Recent developments in Generative Artificial Intelligence (GenAI) have created a paradigm shift in multiple areas of society, and the use of these technologies is likely to become a defining feature of education in coming decades. GenAI offers transformative pedagogical opportunities, while simultaneously posing ethical and academic challenges. Against this backdrop, we outline a practical, simple, and sufficiently comprehensive tool to allow for the integration of GenAI tools into educational assessment: the AI Assessment Scale (AIAS). The AIAS empowers educators to select the appropriate level of GenAI usage in assessments based on the learning outcomes they seek to address. The AIAS offers greater clarity and transparency for students and educators, provides a fair and equitable policy tool for institutions to work with, and offers a nuanced approach which embraces the opportunities of GenAI while recognising that there are instances where such tools may not be pedagogically appropriate or necessary. By adopting a practical, flexible approach that can be implemented quickly, the AIAS can form a much-needed starting point to address the current uncertainty and anxiety regarding GenAI in education. As a secondary objective, we engage with the current literature and advocate for a refocused discourse on GenAI tools in education, one which foregrounds how technologies can help support and enhance teaching and learning, which contrasts with the current focus on GenAI as a facilitator of academic misconduct.

Factual error correction (FEC) aims to revise factual errors in false claims with minimal editing, making them faithful to the provided evidence. This task is crucial for alleviating the hallucination problem encountered by large language models. Given the lack of paired data (i.e., false claims and their corresponding correct claims), existing methods typically adopt the mask-then-correct paradigm. This paradigm relies solely on unpaired false claims and correct claims, thus being referred to as distantly supervised methods. These methods require a masker to explicitly identify factual errors within false claims before revising with a corrector. However, the absence of paired data to train the masker makes accurately pinpointing factual errors within claims challenging. To mitigate this, we propose to improve FEC by Learning to Inject Factual Errors (LIFE), a three-step distantly supervised method: mask-corrupt-correct. Specifically, we first train a corruptor using the mask-then-corrupt procedure, allowing it to deliberately introduce factual errors into correct text. The corruptor is then applied to correct claims, generating a substantial amount of paired data. After that, we filter out low-quality data, and use the remaining data to train a corrector. Notably, our corrector does not require a masker, thus circumventing the bottleneck associated with explicit factual error identification. Our experiments on a public dataset verify the effectiveness of LIFE in two key aspects: Firstly, it outperforms the previous best-performing distantly supervised method by a notable margin of 10.59 points in SARI Final (19.3% improvement). Secondly, even compared to ChatGPT prompted with in-context examples, LIFE achieves a superiority of 7.16 points in SARI Final.

Due to the substantial scale of Large Language Models (LLMs), the direct application of conventional compression methodologies proves impractical. The computational demands associated with even minimal gradient updates present challenges, particularly on consumer-grade hardware. This paper introduces an innovative approach for the parametric and practical compression of LLMs based on reduced order modelling, which entails low-rank decomposition within the feature space and re-parameterization in the weight space. Notably, this compression technique operates in a layer-wise manner, obviating the need for a GPU device and enabling the compression of billion-scale models within stringent constraints of both memory and time. Our method represents a significant advancement in model compression by leveraging matrix decomposition, demonstrating superior efficacy compared to the prevailing state-of-the-art structured pruning method.

Foundation models, such as Large Language Models (LLMs), can respond to a wide range of format-free queries without any task-specific data collection or model training, creating various research and application opportunities for the modeling and operation of large-scale power systems. In this paper, we outline how such large foundation model such as GPT-4 are developed, and discuss how they can be leveraged in challenging power and energy system tasks. We first investigate the potential of existing foundation models by validating their performance on four representative tasks across power system domains, including the optimal power flow (OPF), electric vehicle (EV) scheduling, knowledge retrieval for power engineering technical reports, and situation awareness. Our results indicate strong capabilities of such foundation models on boosting the efficiency and reliability of power system operational pipelines. We also provide suggestions and projections on future deployment of foundation models in power system applications.