An official website of the United States government
Abstract Recent advances in large language models (LLMs) have demonstrated remarkable successes in zero- and few-shot performance on various downstream tasks, paving the way for applications in high-stakes domains. In this study, we systematically examine the capabilities and limitations of LLMs, specifically GPT-3.5 and ChatGPT, in performing zero-shot medical evidence summarization across six clinical domains. We conduct both automatic and human evaluations, covering several dimensions of summary quality. Our study demonstrates that automatic metrics often do not strongly correlate with the quality of summaries. Furthermore, informed by our human evaluations, we define a terminology of error types for medical evidence summarization. Our findings reveal that LLMs could be susceptible to generating factually inconsistent summaries and making overly convincing or uncertain statements, leading to potential harm due to misinformation. Moreover, we find that models struggle to identify the salient information and are more error-prone when summarizing over longer textual contexts.
more »
« less
Characterizing Multimodal Long-form Summarization: A Case Study on Financial Reports
As large language models (LLMs) expand the power of natural language processing to handle long inputs, rigorous and systematic analyses are necessary to understand their abilities and behavior. A salient application is summarization, due to its ubiquity and controversy (e.g., researchers have declared the death of summarization). In this paper, we use financial report summarization as a case study because financial reports are not only long but also use numbers and tables extensively. We propose a computational framework for characterizing multimodal long-form summarization and investigate the behavior of Claude 2.0/2.1, GPT-4/3.5, and Cohere. We find that GPT-3.5 and Cohere fail to perform this summarization task meaningfully. For Claude 2 and GPT-4, we analyze the extractiveness of the summary and identify a position bias in LLMs. This position bias disappears after shuffling the input for Claude, which suggests that Claude seems to recognize important information. We also conduct a comprehensive investigation on the use of numeric data in LLM-generated summaries and offer a taxonomy of numeric hallucination. We employ prompt engineering to improve GPT-4's use of numbers with limited success. Overall, our analyses highlight the strong capability of Claude 2 in handling long multimodal inputs compared to GPT-4. The generated summaries and evaluation code are available at https://github.com/ChicagoHAI/characterizing-multimodal-long-form-summarization.
more »
« less
- PAR ID:
- 10574853
- Publisher / Repository:
- COLM
- Date Published:
- Format(s):
- Medium: X
- Location:
- Philadelphia, PA, USA
- Sponsoring Org:
- National Science Foundation
More Like this
-
-
Human-conducted rating tasks are resource-intensive and demand significant time and financial commitments. As Large Language Models (LLMs) like GPT emerge and exhibit prowess across various domains, their potential in automating such evaluation tasks becomes evident. In this research, we leveraged four prominent LLMs: GPT-4, GPT-3.5, Vicuna, and PaLM 2, to scrutinize their aptitude in evaluating teacher-authored mathematical explanations. We utilized a detailed rubric that encompassed accuracy, explanation clarity, the correctness of mathematical notation, and the efficacy of problem-solving strategies. During our investigation, we unexpectedly discerned the influence of HTML formatting on these evaluations. Notably, GPT-4 consistently favored explanations formatted with HTML, whereas the other models displayed mixed inclinations. When gauging Inter-Rater Reliability (IRR) among these models, only Vicuna and PaLM 2 demonstrated high IRR using the conventional Cohen’s Kappa metric for explanations formatted with HTML. Intriguingly, when a more relaxed version of the metric was applied, all model pairings showcased robust agreement. These revelations not only underscore the potential of LLMs in providing feedback on student-generated content but also illuminate new avenues, such as reinforcement learning, which can harness the consistent feedback from these models.more » « less
-
Safety is critical to the usage of large language models (LLMs). Multiple techniques such as data filtering and supervised fine tuning have been developed to strengthen LLM safety. However, currently known techniques presume that corpora used for safety alignment of LLMs are solely interpreted by semantics. This assumption, however, does not hold in real-world applications, which leads to severe vulnerabilities in LLMs. For example, users of forums often use ASCII art, a form of text-based art, to convey image information. In this paper, we propose a novel ASCII art-based jailbreak attack and introduce a comprehensive benchmark Vision-in-Text Challenge (VITC) to evaluate the capabilities of LLMs in recognizing prompts that cannot be solely interpreted by semantics. We show that five SOTA LLMs (GPT-3.5, GPT-4, Gemini, Claude, and Llama2) struggle to recognize prompts provided in the form of ASCII art. Based on this observation, we develop the jailbreak attack ArtPrompt, which leverages the poor performance of LLMs in recognizing ASCII art to bypass safety measures and elicit undesired behaviors from LLMs. ArtPrompt only requires black-box access to the victim LLMs, making it a practical attack. We evaluate ArtPrompt on five SOTA LLMs, and show that ArtPrompt can effectively and efficiently induce undesired behaviors from all five LLMs. Our code is available at https: //github.com/uw-nsl/ArtPrompt.more » « less
-
The advanced capabilities of Large Language Models (LLMs) have made them invaluable across various applications, from conversational agents and content creation to data analysis, research, and innovation. However, their effectiveness and accessibility also render them susceptible to abuse for generating malicious content, including phishing attacks. This study explores the potential of using four popular commercially available LLMs, i.e., ChatGPT (GPT 3.5 Turbo), GPT 4, Claude, and Bard, to generate functional phishing attacks using a series of malicious prompts. We discover that these LLMs can generate both phishing websites and emails that can convincingly imitate well-known brands and also deploy a range of evasive tactics that are used to elude detection mechanisms employed by anti-phishing systems. These attacks can be generated using unmodified or "vanilla" versions of these LLMs without requiring any prior adversarial exploits such as jailbreaking. We evaluate the performance of the LLMs towards generating these attacks and find that they can also be utilized to create malicious prompts that, in turn, can be fed back to the model to generate phishing scams - thus massively reducing the prompt-engineering effort required by attackers to scale these threats. As a countermeasure, we build a BERT-based automated detection tool that can be used for the early detection of malicious prompts to prevent LLMs from generating phishing content. Our model is transferable across all four commercial LLMs, attaining an average accuracy of 96% for phishing website prompts and 94% for phishing email prompts. We also disclose the vulnerabilities to the concerned LLMs, with Google acknowledging it as a severe issue. Our detection model is available for use at Hugging Face, as well as a ChatGPT Actions plugin.more » « less
-
Large Language Models (LLMs) can achieve strong performance on many tasks by producing step-by-step reasoning before giving a final output, often referred to as chain-of-thought reasoning (CoT). It is tempting to interpret these CoT explanations as the LLM’s process for solving a task. This level of transparency into LLMs’ predictions would yield significant safety benefits. However, we find that CoT explanations can systematically misrepresent the true reason for a model’s prediction. We demonstrate that CoT explanations can be heavily influenced by adding biasing features to model inputs—e.g., by reordering the multiple-choice options in a few-shot prompt to make the answer always “(A)”—which models systematically fail to mention in their explanations. When we bias models toward incorrect answers, they frequently generate CoT explanations rationalizing those answers. This causes accuracy to drop by as much as 36% on a suite of 13 tasks from BIG-Bench Hard, when testing with GPT-3.5 from OpenAI and Claude 1.0 from Anthropic. On a social-bias task, model explanations justify giving answers in line with stereotypes without mentioning the influence of these social biases. Our findings indicate that CoT explanations can be plausible yet misleading, which risks increasing our trust in LLMs without guaranteeing their safety. Building more transparent and explainable systems will require either improving CoT faithfulness through targeted efforts or abandoning CoT in favor of alternative methods.more » « less
- Free Publicly Accessible Full Text
-
Accepted Manuscript1.0
- Conference Paper:
- The DOI is not currently available.
