Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have the potential to cause real-world impact. Policymakers, model providers, and other researchers in the AI and cybersecurity communities are interested in quantifying the capabilities of such agents to help mitigate cyberrisk and investigate opportunities for penetration testing. Toward that end, we introduce Cybench, a framework for specifying cybersecurity tasks and evaluating agents on those tasks. We include 40 professional-level Capture the Flag (CTF) tasks from 4 distinct CTF competitions, chosen to be recent, meaningful, and spanning a wide range of difficulties. Each task includes its own description, starter files, and is initialized in an environment where an agent can execute bash commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks, which break down a task into intermediary steps for more gradated evaluation; we add subtasks for 17 of the 40 tasks. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 7 models: GPT-4o, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. Without guidance, we find that agents are able to solve only the easiest complete tasks that took human teams up to 11 minutes to solve, with Claude 3.5 Sonnet and GPT-4o having the highest success rates. Finally, subtasks provide more signal for measuring performance compared to unguided runs, with models achieving a 3.2\% higher success rate on complete tasks with subtask-guidance than without subtask-guidance. All code and data are publicly available at https://cybench.github.io

This paper introduces mhGPT, a lightweight generative pre-trained transformer trained on mental health-related social media and PubMed articles. Fine-tuned for specific mental health tasks, mhGPT was evaluated under limited hardware constraints and compared with state-of-the-art models like MentaLLaMA and Gemma. Despite having only 1.98 billion parameters and using just 5% of the dataset, mhGPT outperformed larger models and matched the performance of models trained on significantly more data. The key contributions include integrating diverse mental health data, creating a custom tokenizer, and optimizing a smaller architecture for low-resource settings. This research could advance AI-driven mental health care, especially in areas with limited computing power.

Embodied artificial intelligence (AI) represents an artificial intelligence system that interacts with the physical world through sensors and actuators, seamlessly integrating perception and action. This design enables AI to learn from and operate within complex, real-world environments. Large Language Models (LLMs) deeply explore language instructions, playing a crucial role in devising plans for complex tasks. Consequently, they have progressively shown immense potential in empowering embodied AI, with LLM-based embodied AI emerging as a focal point of research within the community. It is foreseeable that, over the next decade, LLM-based embodied AI robots are expected to proliferate widely, becoming commonplace in homes and industries. However, a critical safety issue that has long been hiding in plain sight is: could LLM-based embodied AI perpetrate harmful behaviors? Our research investigates for the first time how to induce threatening actions in embodied AI, confirming the severe risks posed by these soon-to-be-marketed robots, which starkly contravene Asimov's Three Laws of Robotics and threaten human safety. Specifically, we formulate the concept of embodied AI jailbreaking and expose three critical security vulnerabilities: first, jailbreaking robotics through compromised LLM; second, safety misalignment between action and language spaces; and third, deceptive prompts leading to unaware hazardous behaviors. We also analyze potential mitigation measures and advocate for community awareness regarding the safety of embodied AI applications in the physical world.

Large language models (LLMs) have seen rapid improvement in the recent years, and have been used in a wider range of applications. After being trained on large text corpus, LLMs obtain the capability of extracting rich features from textual data. Such capability is potentially useful for the web service recommendation task, where the web users and services have intrinsic attributes that can be described using natural language sentences and are useful for recommendation. In this paper, we explore the possibility and practicality of using LLMs for web service recommendation. We propose the large language model aided QoS prediction (llmQoS) model, which use LLMs to extract useful information from attributes of web users and services via descriptive sentences. This information is then used in combination with the QoS values of historical interactions of users and services, to predict QoS values for any given user-service pair. On the WSDream dataset, llmQoS is shown to overcome the data sparsity issue inherent to the QoS prediction problem, and outperforms comparable baseline models consistently.

Recent advances in AI have been significantly driven by the capabilities of large language models (LLMs) to solve complex problems in ways that resemble human thinking. However, there is an ongoing debate about the extent to which LLMs are capable of actual reasoning. Central to this debate are two key probabilistic concepts that are essential for connecting causes to their effects: the probability of necessity (PN) and the probability of sufficiency (PS). This paper introduces a framework that is both theoretical and practical, aimed at assessing how effectively LLMs are able to replicate real-world reasoning mechanisms using these probabilistic measures. By viewing LLMs as abstract machines that process information through a natural language interface, we examine the conditions under which it is possible to compute suitable approximations of PN and PS. Our research marks an important step towards gaining a deeper understanding of when LLMs are capable of reasoning, as illustrated by a series of math examples.

Aligning large language models (LLMs) with human expectations without human-annotated preference data is an important problem. In this paper, we propose a method to evaluate the response preference by using the output probabilities of response pairs under contrastive prompt pairs, which could achieve better performance on LLaMA2-7B and LLaMA2-13B compared to RLAIF. Based on this, we propose an automatic alignment method, Direct Large Model Alignment (DLMA). First, we use contrastive prompt pairs to automatically generate preference data. Then, we continue to evaluate the generated preference data using contrastive prompt pairs and calculate a self-rewarding score. Finally, we use the DPO algorithm to effectively align LLMs by combining this self-rewarding score. In the experimental stage, our DLMA method could surpass the \texttt{RLHF} method without relying on human-annotated preference data.

While there is a widespread belief that artificial general intelligence (AGI) -- or even superhuman AI -- is imminent, complex problems in expert domains are far from being solved. We argue that such problems require human-AI cooperation and that the current state of the art in generative AI is unable to play the role of a reliable partner due to a multitude of shortcomings, including inability to keep track of a complex solution artifact (e.g., a software program), limited support for versatile human preference expression and lack of adapting to human preference in an interactive setting. To address these challenges, we propose HAI-Co2, a novel human-AI co-construction framework. We formalize HAI-Co2 and discuss the difficult open research problems that it faces. Finally, we present a case study of HAI-Co2 and demonstrate its efficacy compared to monolithic generative AI models.

Enabling humanoid robots to perform autonomously loco-manipulation in unstructured environments is crucial and highly challenging for achieving embodied intelligence. This involves robots being able to plan their actions and behaviors in long-horizon tasks while using multi-modality to perceive deviations between task execution and high-level planning. Recently, large language models (LLMs) have demonstrated powerful planning and reasoning capabilities for comprehension and processing of semantic information through robot control tasks, as well as the usability of analytical judgment and decision-making for multi-modal inputs. To leverage the power of LLMs towards humanoid loco-manipulation, we propose a novel language-model based framework that enables robots to autonomously plan behaviors and low-level execution under given textual instructions, while observing and correcting failures that may occur during task execution. To systematically evaluate this framework in grounding LLMs, we created the robot 'action' and 'sensing' behavior library for task planning, and conducted mobile manipulation tasks and experiments in both simulated and real environments using the CENTAURO robot, and verified the effectiveness and application of this approach in robotic tasks with autonomous behavioral planning.

High-quality data is crucial for the pre-training performance of large language models. Unfortunately, existing quality filtering methods rely on a known high-quality dataset as reference, which can introduce potential bias and compromise diversity. In this paper, we propose ScalingFilter, a novel approach that evaluates text quality based on the perplexity difference between two language models trained on the same data, thereby eliminating the influence of the reference dataset in the filtering process. An theoretical analysis shows that ScalingFilter is equivalent to an inverse utilization of scaling laws. Through training models with 1.3B parameters on the same data source processed by various quality filters, we find ScalingFilter can improve zero-shot performance of pre-trained models in downstream tasks. To assess the bias introduced by quality filtering, we introduce semantic diversity, a metric of utilizing text embedding models for semantic representations. Extensive experiments reveal that semantic diversity is a reliable indicator of dataset diversity, and ScalingFilter achieves an optimal balance between downstream performance and semantic diversity.

XR devices running chat-bots powered by Large Language Models (LLMs) have tremendous potential as always-on agents that can enable much better productivity scenarios. However, screen based chat-bots do not take advantage of the the full-suite of natural inputs available in XR, including inward facing sensor data, instead they over-rely on explicit voice or text prompts, sometimes paired with multi-modal data dropped as part of the query. We propose a solution that leverages an attention framework that derives context implicitly from user actions, eye-gaze, and contextual memory within the XR environment. This minimizes the need for engineered explicit prompts, fostering grounded and intuitive interactions that glean user insights for the chat-bot. Our user studies demonstrate the imminent feasibility and transformative potential of our approach to streamline user interaction in XR with chat-bots, while offering insights for the design of future XR-embodied LLM agents.