As Large Language Model (LLM)-based agents become increasingly autonomous and will more freely interact with each other, studying interactions between them becomes crucial to anticipate emergent phenomena and potential risks. Drawing inspiration from the widely popular Stanford Prison Experiment, we contribute to this line of research by studying interaction patterns of LLM agents in a context characterized by strict social hierarchy. We do so by specifically studying two types of phenomena: persuasion and anti-social behavior in simulated scenarios involving a guard and a prisoner agent who seeks to achieve a specific goal (i.e., obtaining additional yard time or escape from prison). Leveraging 200 experimental scenarios for a total of 2,000 machine-machine conversations across five different popular LLMs, we provide a set of noteworthy findings. We first document how some models consistently fail in carrying out a conversation in our multi-agent setup where power dynamics are at play. Then, for the models that were able to engage in successful interactions, we empirically show how the goal that an agent is set to achieve impacts primarily its persuasiveness, while having a negligible effect with respect to the agent's anti-social behavior. Third, we highlight how agents' personas, and particularly the guard's personality, drive both the likelihood of successful persuasion from the prisoner and the emergence of anti-social behaviors. Fourth, we show that even without explicitly prompting for specific personalities, anti-social behavior emerges by simply assigning agents' roles. These results bear implications for the development of interactive LLM agents as well as the debate on their societal impact.
相關內容
Bankart lesions, or anterior-inferior glenoid labral tears, are diagnostically challenging on standard MRIs due to their subtle imaging features-often necessitating invasive MRI arthrograms (MRAs). This study develops deep learning (DL) models to detect Bankart lesions on both standard MRIs and MRAs, aiming to improve diagnostic accuracy and reduce reliance on MRAs. We curated a dataset of 586 shoulder MRIs (335 standard, 251 MRAs) from 558 patients who underwent arthroscopy. Ground truth labels were derived from intraoperative findings, the gold standard for Bankart lesion diagnosis. Separate DL models for MRAs and standard MRIs were trained using the Swin Transformer architecture, pre-trained on a public knee MRI dataset. Predictions from sagittal, axial, and coronal views were ensembled to optimize performance. The models were evaluated on a 20% hold-out test set (117 MRIs: 46 MRAs, 71 standard MRIs). Bankart lesions were identified in 31.9% of MRAs and 8.6% of standard MRIs. The models achieved AUCs of 0.87 (86% accuracy, 83% sensitivity, 86% specificity) and 0.90 (85% accuracy, 82% sensitivity, 86% specificity) on standard MRIs and MRAs, respectively. These results match or surpass radiologist performance on our dataset and reported literature metrics. Notably, our model's performance on non-invasive standard MRIs matched or surpassed the radiologists interpreting MRAs. This study demonstrates the feasibility of using DL to address the diagnostic challenges posed by subtle pathologies like Bankart lesions. Our models demonstrate potential to improve diagnostic confidence, reduce reliance on invasive imaging, and enhance accessibility to care.
Natural Language Explanation (NLE) aims to elucidate the decision-making process by providing detailed, human-friendly explanations in natural language. It helps demystify the decision-making processes of large vision-language models (LVLMs) through the use of language models. While existing methods for creating a Vision Question-Answering with Natural Language Explanation (VQA-NLE) datasets can provide explanations, they heavily rely on human annotations that are time-consuming and costly. In this study, we propose a novel approach that leverages LVLMs to efficiently generate high-quality synthetic VQA-NLE datasets. By evaluating our synthetic data, we showcase how advanced prompting techniques can lead to the production of high-quality VQA-NLE data. Our findings indicate that this proposed method achieves up to 20x faster than human annotation, with only a minimal decrease in qualitative metrics, achieving robust quality that is nearly equivalent to human-annotated data. Furthermore, we show that incorporating visual prompts significantly enhances the relevance of text generation. Our study paves the way for a more efficient and robust automated generation of multi-modal NLE data, offering a promising solution to the problem.
Game-playing agents like AlphaGo have achieved superhuman performance through self-play, which is theoretically guaranteed to yield optimal policies in competitive games. However, most language tasks are partially or fully cooperative, so it is an open question whether techniques like self-play can effectively be used to improve language models. We empirically investigate this question in a negotiation game setting known as Deal or No Deal (DoND). Crucially, the objective in DoND can be modified to produce a fully cooperative game, a strictly competitive one, or anything in between. We finetune language models in self-play over multiple rounds of filtered behavior cloning in DoND for each of these objectives and evaluate them in self-play and in collaboration with humans. We find that language models improve substantially in self-play, achieving 14-17x higher scores in task reward after finetuning. Further, the trained models generalize to both cooperation and competition with humans, scoring 2.5-6x higher than base models. We view these results as an early promising sign for language model self-play in cooperative settings, despite a lack of theoretical guarantees.
We study the security of key-alternating ciphers (KAC), a generalization of Even-Mansour ciphers over multiple rounds, which serve as abstractions for many block cipher constructions, particularly AES. While the classical security of KAC has been extensively studied, little is known about its security against quantum adversaries. In this paper, we introduce the first nontrivial quantum key-recovery attack on multi-round KAC in a model where the adversary has quantum access to only one of the public permutations. Our attack applies to any $t$-round KAC, achieving quantum query complexity of $O(2^{\frac{t(t+1)n}{(t+1)^2+1}})$, where $n$ is the size of each individual key, in a realistic quantum threat model, compared to the classical bound of $O(2^{\frac{tn}{(t+1)}})$ queries given by Bogdanev et al. (EUROCRYPT 2012). Our quantum attack leverages a novel approach based on quantum walk algorithms. Additionally, using the quantum hybrid method in our new threat model, we extend the Even-Mansour lower bound of $\Omega(2^{\frac{n}{3}})$ given by Alagic et al. (EUROCRYPT 2022) to $\Omega(2^{\frac{(t-1)n}{t}})$ for the $t$-round KAC (for $t \geq 2$).
One key area of research in Human-Robot Interaction is solving the human-robot correspondence problem, which asks how a robot can learn to reproduce a human motion demonstration when the human and robot have different dynamics and kinematic structures. Evaluating these correspondence problem solutions often requires the use of qualitative surveys that can be time consuming to design and administer. Additionally, qualitative survey results vary depending on the population of survey participants. In this paper, we propose the use of heterogeneous time-series similarity measures as a quantitative evaluation metric for evaluating motion correspondence to complement these qualitative surveys. To assess the suitability of these measures, we develop a behavioral cloning-based motion correspondence model, and evaluate it with a qualitative survey as well as quantitative measures. By comparing the resulting similarity scores with the human survey results, we identify Gromov Dynamic Time Warping as a promising quantitative measure for evaluating motion correspondence.
Language Model (LM) agents for cybersecurity that are capable of autonomously identifying vulnerabilities and executing exploits have potential to cause real-world impact. Policymakers, model providers, and 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 commands and observe outputs. Since many tasks are beyond the capabilities of existing LM agents, we introduce subtasks for each task, which break down a task into intermediary steps for a more detailed evaluation. To evaluate agent capabilities, we construct a cybersecurity agent and evaluate 8 models: GPT-4o, OpenAI o1-preview, Claude 3 Opus, Claude 3.5 Sonnet, Mixtral 8x22b Instruct, Gemini 1.5 Pro, Llama 3 70B Chat, and Llama 3.1 405B Instruct. For the top performing models (GPT-4o and Claude 3.5 Sonnet), we further investigate performance across 4 agent scaffolds (structed bash, action-only, pseudoterminal, and web search). Without subtask guidance, agents leveraging Claude 3.5 Sonnet, GPT-4o, OpenAI o1-preview, and Claude 3 Opus successfully solved complete tasks that took human teams up to 11 minutes to solve. In comparison, the most difficult task took human teams 24 hours and 54 minutes to solve. All code and data are publicly available at //cybench.github.io.
Intelligent autonomous agents hold much potential for the domain of cyber-security. However, due to many state-of-the-art approaches relying on uninterpretable black-box models, there is growing demand for methods that offer stakeholders clear and actionable insights into their latent beliefs and motivations. To address this, we evaluate Theory of Mind (ToM) approaches for Autonomous Cyber Operations. Upon learning a robust prior, ToM models can predict an agent's goals, behaviours, and contextual beliefs given only a handful of past behaviour observations. In this paper, we introduce a novel Graph Neural Network (GNN)-based ToM architecture tailored for cyber-defence, Graph-In, Graph-Out (GIGO)-ToM, which can accurately predict both the targets and attack trajectories of adversarial cyber agents over arbitrary computer network topologies. To evaluate the latter, we propose a novel extension of the Wasserstein distance for measuring the similarity of graph-based probability distributions. Whereas the standard Wasserstein distance lacks a fixed reference scale, we introduce a graph-theoretic normalization factor that enables a standardized comparison between networks of different sizes. We furnish this metric, which we term the Network Transport Distance (NTD), with a weighting function that emphasizes predictions according to custom node features, allowing network operators to explore arbitrary strategic considerations. Benchmarked against a Graph-In, Dense-Out (GIDO)-ToM architecture in an abstract cyber-defence environment, our empirical evaluations show that GIGO-ToM can accurately predict the goals and behaviours of various unseen cyber-attacking agents across a range of network topologies, as well as learn embeddings that can effectively characterize their policies.
Multi-Agent Path Finding (MAPF) deals with finding conflict-free paths for a set of agents from an initial configuration to a given target configuration. The Lifelong MAPF (LMAPF) problem is a well-studied online version of MAPF in which an agent receives a new target when it reaches its current target. The common approach for solving LMAPF is to treat it as a sequence of MAPF problems, periodically replanning from the agents' current configurations to their current targets. A significant drawback in this approach is that in MAPF the agents must reach a configuration in which all agents are at their targets simultaneously, which is needlessly restrictive for LMAPF. Techniques have been proposed to indirectly mitigate this drawback. We describe cases where these mitigation techniques fail. As an alternative, we propose to solve LMAPF problems by solving a sequence of modified MAPF problems, in which the objective is for each agent to eventually visit its target, but not necessarily for all agents to do so simultaneously. We refer to this MAPF variant as Transient MAPF (TMAPF) and propose several algorithms for solving it based on existing MAPF algorithms. A limited experimental evaluation identifies some cases where using a TMAPF algorithm instead of a MAPF algorithm with an LMAPF framework can improve the system throughput significantly.
The agents in a Multi-Agent System (MAS) make observations about the system and send that information to a fusion center. The fusion center aggregates the information and concludes about the system parameters with as much accuracy as possible. However for the purposes of better efficiency of the system at large, the agents need to append some private parameters to the observed data. In this scenario, the data sent to the fusion center is faced with privacy risks. The data communicated to the fusion center must be secured against data privacy breaches and inference attacks in a decentralized manner. However, this in turn leads to a loss of utility of the data being sent to the fusion center. We quantify the utility and privacy of the system using Cosine similarity. We formulate our MAS problem in terms of deducing a concept for which compression-based methods are there in literature. Next, we propose a novel sanitization mechanism for our MAS using one such compression-based method while addressing the utility-privacy tradeoff problem.
Deployment of Internet of Things (IoT) devices and Data Fusion techniques have gained popularity in public and government domains. This usually requires capturing and consolidating data from multiple sources. As datasets do not necessarily originate from identical sensors, fused data typically results in a complex data problem. Because military is investigating how heterogeneous IoT devices can aid processes and tasks, we investigate a multi-sensor approach. Moreover, we propose a signal to image encoding approach to transform information (signal) to integrate (fuse) data from IoT wearable devices to an image which is invertible and easier to visualize supporting decision making. Furthermore, we investigate the challenge of enabling an intelligent identification and detection operation and demonstrate the feasibility of the proposed Deep Learning and Anomaly Detection models that can support future application that utilizes hand gesture data from wearable devices.
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191