To understand the risks posed by a new AI system, we must understand what it can and cannot do. Building on prior work, we introduce a programme of new "dangerous capability" evaluations and pilot them on Gemini 1.0 models. Our evaluations cover four areas: (1) persuasion and deception; (2) cyber-security; (3) self-proliferation; and (4) self-reasoning. We do not find evidence of strong dangerous capabilities in the models we evaluated, but we flag early warning signs. Our goal is to help advance a rigorous science of dangerous capability evaluation, in preparation for future models.

In robotics, the use of Large Language Models (LLMs) is becoming prevalent, especially for understanding human commands. In particular, LLMs are utilized as domain-agnostic task planners for high-level human commands. LLMs are capable of Chain-of-Thought (CoT) reasoning, and this allows LLMs to be task planners. However, we need to consider that modern robots still struggle to perform complex actions, and the domains where robots can be deployed are limited in practice. This leads us to pose a question: If small LMs can be trained to reason in chains within a single domain, would even small LMs be good task planners for the robots? To train smaller LMs to reason in chains, we build `COmmand-STeps datasets' (COST) consisting of high-level commands along with corresponding actionable low-level steps, via LLMs. We release not only our datasets but also the prompt templates used to generate them, to allow anyone to build datasets for their domain. We compare GPT3.5 and GPT4 with the finetuned GPT2 for task domains, in tabletop and kitchen environments, and the result shows that GPT2-medium is comparable to GPT3.5 for task planning in a specific domain. Our dataset, code, and more output samples can be found in https://github.com/Gawon-Choi/small-LMs-Task-Planning

We investigate whether modern AI can emulate expert creativity in complex scientific endeavors. We introduce novel methodology that utilizes original research articles published after the AI's training cutoff, ensuring no prior exposure, mitigating concerns of rote memorization and prior training. The AI are tasked with redacting findings, predicting outcomes from redacted research, and assessing prediction accuracy against reported results. Analysis on 589 published studies in four leading psychology journals over a 28-month period, showcase the AI's proficiency in understanding specialized research, deductive reasoning, and evaluating evidentiary alignment--cognitive hallmarks of human subject matter expertise and creativity. These findings suggest the potential of general-purpose AI to transform academia, with roles requiring knowledge-based creativity become increasingly susceptible to technological substitution.

Large Language Models (LLMs) have become very popular and have found use cases in many domains, such as chatbots, auto-task completion agents, and much more. However, LLMs are vulnerable to different types of attacks, such as jailbreaking, prompt injection attacks, and privacy leakage attacks. Foundational LLMs undergo adversarial and alignment training to learn not to generate malicious and toxic content. For specialized use cases, these foundational LLMs are subjected to fine-tuning or quantization for better performance and efficiency. We examine the impact of downstream tasks such as fine-tuning and quantization on LLM vulnerability. We test foundation models like Mistral, Llama, MosaicML, and their fine-tuned versions. Our research shows that fine-tuning and quantization reduces jailbreak resistance significantly, leading to increased LLM vulnerabilities. Finally, we demonstrate the utility of external guardrails in reducing LLM vulnerabilities.

Transformers have emerged as the underpinning architecture for Large Language Models (LLMs). In generative language models, the inference process involves two primary phases: prompt processing and token generation. Token generation, which constitutes the majority of the computational workload, primarily entails vector-matrix multiplications and interactions with the Key-Value (KV) Cache. This phase is constrained by memory bandwidth due to the overhead of transferring weights and KV cache values from the memory system to the computing units. This memory bottleneck becomes particularly pronounced in applications that require long-context and extensive text generation, both of which are increasingly crucial for LLMs. This paper introduces "Keyformer", an innovative inference-time approach, to mitigate the challenges associated with KV cache size and memory bandwidth utilization. Keyformer leverages the observation that approximately 90% of the attention weight in generative inference focuses on a specific subset of tokens, referred to as "key" tokens. Keyformer retains only the key tokens in the KV cache by identifying these crucial tokens using a novel score function. This approach effectively reduces both the KV cache size and memory bandwidth usage without compromising model accuracy. We evaluate Keyformer's performance across three foundational models: GPT-J, Cerebras-GPT, and MPT, which employ various positional embedding algorithms. Our assessment encompasses a variety of tasks, with a particular emphasis on summarization and conversation tasks involving extended contexts. Keyformer's reduction of KV cache reduces inference latency by 2.1x and improves token generation throughput by 2.4x, while preserving the model's accuracy.

The integration of ChatGPT as a supportive tool in education, notably in programming courses, addresses the unique challenges of programming education by providing assistance with debugging, code generation, and explanations. Despite existing research validating ChatGPT's effectiveness, its application in university-level programming education and a detailed understanding of student interactions and perspectives remain limited. This paper explores ChatGPT's impact on learning in a Python programming course tailored for first-year students over eight weeks. By analyzing responses from surveys, open-ended questions, and student-ChatGPT dialog data, we aim to provide a comprehensive view of ChatGPT's utility and identify both its advantages and limitations as perceived by students. Our study uncovers a generally positive reception toward ChatGPT and offers insights into its role in enhancing the programming education experience. These findings contribute to the broader discourse on AI's potential in education, suggesting paths for future research and application.

Several companies--such as Google, Microsoft, and OpenAI--have deployed techniques to watermark AI-generated content to enable proactive detection. However, existing literature mainly focuses on user-agnostic detection. Attribution aims to further trace back the user of a generative-AI service who generated a given content detected as AI-generated. Despite its growing importance, attribution is largely unexplored. In this work, we aim to bridge this gap by providing the first systematic study on watermark-based, user-aware detection and attribution of AI-generated content. Specifically, we theoretically study the detection and attribution performance via rigorous probabilistic analysis. Moreover, we develop an efficient algorithm to select watermarks for the users to enhance attribution performance. Both our theoretical and empirical results show that watermark-based detection and attribution inherit the accuracy and (non-)robustness properties of the watermarking method.

This paper presents the MasonTigers entry to the SemEval-2024 Task 8 - Multigenerator, Multidomain, and Multilingual Black-Box Machine-Generated Text Detection. The task encompasses Binary Human-Written vs. Machine-Generated Text Classification (Track A), Multi-Way Machine-Generated Text Classification (Track B), and Human-Machine Mixed Text Detection (Track C). Our best performing approaches utilize mainly the ensemble of discriminator transformer models along with sentence transformer and statistical machine learning approaches in specific cases. Moreover, zero-shot prompting and fine-tuning of FLAN-T5 are used for Track A and B.

With the rapid advancement of artificial intelligence (AI) in various domains, the education sector is set for transformation. The potential of AI-driven tools in enhancing the learning experience, especially in programming, is immense. However, the scientific evaluation of Large Language Models (LLMs) used in Automated Programming Assessment Systems (APASs) as an AI-Tutor remains largely unexplored. Therefore, there is a need to understand how students interact with such AI-Tutors and to analyze their experiences. In this paper, we conducted an exploratory case study by integrating the GPT-3.5-Turbo model as an AI-Tutor within the APAS Artemis. Through a combination of empirical data collection and an exploratory survey, we identified different user types based on their interaction patterns with the AI-Tutor. Additionally, the findings highlight advantages, such as timely feedback and scalability. However, challenges like generic responses and students' concerns about a learning progress inhibition when using the AI-Tutor were also evident. This research adds to the discourse on AI's role in education.

For a viewpoint-diverse news recommender, identifying whether two news articles express the same viewpoint is essential. One way to determine "same or different" viewpoint is stance detection. In this paper, we investigate the robustness of operationalization choices for few-shot stance detection, with special attention to modelling stance across different topics. Our experiments test pre-registered hypotheses on stance detection. Specifically, we compare two stance task definitions (Pro/Con versus Same Side Stance), two LLM architectures (bi-encoding versus cross-encoding), and adding Natural Language Inference knowledge, with pre-trained RoBERTa models trained with shots of 100 examples from 7 different stance detection datasets. Some of our hypotheses and claims from earlier work can be confirmed, while others give more inconsistent results. The effect of the Same Side Stance definition on performance differs per dataset and is influenced by other modelling choices. We found no relationship between the number of training topics in the training shots and performance. In general, cross-encoding out-performs bi-encoding, and adding NLI training to our models gives considerable improvement, but these results are not consistent across all datasets. Our results indicate that it is essential to include multiple datasets and systematic modelling experiments when aiming to find robust modelling choices for the concept `stance'.