Recent large language models (LLMs) have enabled the development of advanced agentic systems that can integrate various tools and APIs to fulfill user queries through function calling. However, the deployment of these LLMs on the edge has not been explored since they typically require cloud-based infrastructure due to their substantial model size and computational demands. To this end, we present TinyAgent, an end-to-end framework for training and deploying task-specific small language model agents capable of function calling for driving agentic systems at the edge. We first show how to enable accurate function calling for open-source models via the LLMCompiler framework. We then systematically curate a high-quality dataset for function calling, which we use to fine-tune two small language models, TinyAgent-1.1B and 7B. For efficient inference, we introduce a novel tool retrieval method to reduce the input prompt length and utilize quantization to further accelerate the inference speed. As a driving application, we demonstrate a local Siri-like system for Apple's MacBook that can execute user commands through text or voice input. Our results show that our models can achieve, and even surpass, the function-calling capabilities of larger models like GPT-4-Turbo, while being fully deployed at the edge. We open-source our dataset, models, and installable package and provide a demo video for our MacBook assistant agent.

This paper addresses user-specific dialogs. In contrast to previous research on personalized dialogue focused on achieving virtual user dialogue as defined by persona descriptions, user-specific dialogue aims to reproduce real-user dialogue beyond persona-based dialogue. Fine-tuning using the target user's dialogue history is an efficient learning method for a user-specific model. However, it is prone to overfitting and model destruction due to the small amount of data. Therefore, we propose a learning method for user-specific models by combining parameter-efficient fine-tuning with a pre-trained dialogue model that includes user profiles. Parameter-efficient fine-tuning adds a small number of parameters to the entire model, so even small amounts of training data can be trained efficiently and are robust to model destruction. In addition, the pre-trained model, which is learned by adding simple prompts for automatically inferred user profiles, can generate speech with enhanced knowledge of the user's profile, even when there is little training data during fine-tuning. In experiments, we compared the proposed model with large-language-model utterance generation using prompts containing users' personal information. Experiments reproducing real users' utterances revealed that the proposed model can generate utterances with higher reproducibility than the compared methods, even with a small model.

Social norm is defined as a shared standard of acceptable behavior in a society. The emergence of social norms fosters coordination among agents without any hard-coded rules, which is crucial for the large-scale deployment of AVs in an intelligent transportation system. This paper explores the application of LLMs in understanding and modeling social norms in autonomous driving games. We introduce LLMs into autonomous driving games as intelligent agents who make decisions according to text prompts. These agents are referred to as LLM-based agents. Our framework involves LLM-based agents playing Markov games in a multi-agent system (MAS), allowing us to investigate the emergence of social norms among individual agents. We aim to identify social norms by designing prompts and utilizing LLMs on textual information related to the environment setup and the observations of LLM-based agents. Using the OpenAI Chat API powered by GPT-4.0, we conduct experiments to simulate interactions and evaluate the performance of LLM-based agents in two driving scenarios: unsignalized intersection and highway platoon. The results show that LLM-based agents can handle dynamically changing environments in Markov games, and social norms evolve among LLM-based agents in both scenarios. In the intersection game, LLM-based agents tend to adopt a conservative driving policy when facing a potential car crash. The advantage of LLM-based agents in games lies in their strong operability and analyzability, which facilitate experimental design.

This study examines the political bias of chatbots powered by large language models, namely ChatGPT and Gemini, in the context of the 2024 European Parliament elections. The research focused on the evaluation of political parties represented in the European Parliament across 27 EU Member States by these generative artificial intelligence (AI) systems. The methodology involved daily data collection through standardized prompts on both platforms. The results revealed a stark contrast: while Gemini mostly refused to answer political questions, ChatGPT provided consistent ratings. The analysis showed a significant bias in ChatGPT in favor of left-wing and centrist parties, with the highest ratings for the Greens/European Free Alliance. In contrast, right-wing parties, particularly the Identity and Democracy group, received the lowest ratings. The study identified key factors influencing the ratings, including attitudes toward European integration and perceptions of democratic values. The findings highlight the need for a critical approach to information provided by generative AI systems in a political context and call for more transparency and regulation in this area.

Large Language Models (LLMs) frequently lack domain-specific knowledge and even fine-tuned models tend to hallucinate. Hence, more reliable models that can include external knowledge are needed. We present a pipeline, 4StepFocus, and specifically a preprocessing step, that can substantially improve the answers of LLMs. This is achieved by providing guided access to external knowledge making use of the model's ability to capture relational context and conduct rudimentary reasoning by themselves. The method narrows down potentially correct answers by triplets-based searches in a semi-structured knowledge base in a direct, traceable fashion, before switching to latent representations for ranking those candidates based on unstructured data. This distinguishes it from related methods that are purely based on latent representations. 4StepFocus consists of the steps: 1) Triplet generation for extraction of relational data by an LLM, 2) substitution of variables in those triplets to narrow down answer candidates employing a knowledge graph, 3) sorting remaining candidates with a vector similarity search involving associated non-structured data, 4) reranking the best candidates by the LLM with background data provided. Experiments on a medical, a product recommendation, and an academic paper search test set demonstrate that this approach is indeed a powerful augmentation. It not only adds relevant traceable background information from information retrieval, but also improves performance considerably in comparison to state-of-the-art methods. This paper presents a novel, largely unexplored direction and therefore provides a wide range of future work opportunities. Used source code is available at https://github.com/kramerlab/4StepFocus.

Generating long-term texts such as novels using artificial intelligence has always been a challenge. A common approach is to use large language models (LLMs) to construct a hierarchical framework that first plans and then writes. Despite the fact that the generated novels reach a sufficient length, they exhibit poor logical coherence and appeal in their plots and deficiencies in character and event depiction, ultimately compromising the overall narrative quality. In this paper, we propose a method named Extracting Excelsior and Expanding. Ex3 initially extracts structure information from raw novel data. By combining this structure information with the novel data, an instruction-following dataset is meticulously crafted. This dataset is then utilized to fine-tune the LLM, aiming for excelsior generation performance. In the final stage, a tree-like expansion method is deployed to facilitate the generation of arbitrarily long novels. Evaluation against previous methods showcases Ex3's ability to produce higher-quality long-form novels.

Large Language Models (LLMs) have demonstrated great capabilities in natural language understanding and generation, largely attributed to the intricate alignment process using human feedback. While alignment has become an essential training component that leverages data collected from user queries, it inadvertently opens up an avenue for a new type of user-guided poisoning attacks. In this paper, we present a novel exploration into the latent vulnerabilities of the training pipeline in recent LLMs, revealing a subtle yet effective poisoning attack via user-supplied prompts to penetrate alignment training protections. Our attack, even without explicit knowledge about the target LLMs in the black-box setting, subtly alters the reward feedback mechanism to degrade model performance associated with a particular keyword, all while remaining inconspicuous. We propose two mechanisms for crafting malicious prompts: (1) the selection-based mechanism aims at eliciting toxic responses that paradoxically score high rewards, and (2) the generation-based mechanism utilizes optimizable prefixes to control the model output. By injecting 1\% of these specially crafted prompts into the data, through malicious users, we demonstrate a toxicity score up to two times higher when a specific trigger word is used. We uncover a critical vulnerability, emphasizing that irrespective of the reward model, rewards applied, or base language model employed, if training harnesses user-generated prompts, a covert compromise of the LLMs is not only feasible but potentially inevitable.

The rapid advancements in AI technologies have revolutionized the production of graphical content across various sectors, including entertainment, advertising, and e-commerce. These developments have spurred the need for robust evaluation methods to assess the quality and realism of AI-generated images. To address this, we conducted three studies. First, we introduced and validated a questionnaire called Visual Verity, which measures photorealism, image quality, and text-image alignment. Second, we applied this questionnaire to assess images from AI models (DALL-E2, DALL-E3, GLIDE, Stable Diffusion) and camera-generated images, revealing that camera-generated images excelled in photorealism and text-image alignment, while AI models led in image quality. We also analyzed statistical properties, finding that camera-generated images scored lower in hue, saturation, and brightness. Third, we evaluated computational metrics' alignment with human judgments, identifying MS-SSIM and CLIP as the most consistent with human assessments. Additionally, we proposed the Neural Feature Similarity Score (NFSS) for assessing image quality. Our findings highlight the need for refining computational metrics to better capture human visual perception, thereby enhancing AI-generated content evaluation.

Not everything on the internet is true. This unfortunate fact requires both humans and models to perform complex reasoning about credibility when working with retrieved information. In NLP, this problem has seen little attention. Indeed, retrieval-augmented models are not typically expected to distrust retrieved documents. Human experts overcome the challenge by gathering signals about the context, reliability, and tendency of source documents - that is, they perform source criticism. We propose a novel NLP task focused on finding and summarising such signals. We introduce a new dataset of 6,709 "media background checks" derived from Media Bias / Fact Check, a volunteer-run website documenting media bias. We test open-source and closed-source LLM baselines with and without retrieval on this dataset, finding that retrieval greatly improves performance. We furthermore carry out human evaluation, demonstrating that 1) media background checks are helpful for humans, and 2) media background checks are helpful for retrieval-augmented models.

Nowadays, large-scale text-to-speech (TTS) systems are primarily divided into two types: autoregressive and non-autoregressive. The autoregressive systems have certain deficiencies in robustness and cannot control speech duration. In contrast, non-autoregressive systems require explicit prediction of phone-level duration, which may compromise their naturalness. We introduce the Masked Generative Codec Transformer (MaskGCT), a fully non-autoregressive model for TTS that does not require precise alignment information between text and speech. MaskGCT is a two-stage model: in the first stage, the model uses text to predict semantic tokens extracted from a speech self-supervised learning (SSL) model, and in the second stage, the model predicts acoustic tokens conditioned on these semantic tokens. MaskGCT follows the \textit{mask-and-predict} learning paradigm. During training, MaskGCT learns to predict masked semantic or acoustic tokens based on given conditions and prompts. During inference, the model generates tokens of a specified length in a parallel manner. We scale MaskGCT to a large-scale multilingual dataset with 100K hours of in-the-wild speech. Our experiments demonstrate that MaskGCT achieves superior or competitive performance compared to state-of-the-art zero-shot TTS systems in terms of quality, similarity, and intelligibility while offering higher generation efficiency than diffusion-based or autoregressive TTS models. Audio samples are available at https://maskgct.github.io.