Their vast number of trainable parameters necessitates a wealth of data to achieve accuracy and mitigate overfitting. However, experimental measurements are often limited and costly to obtain in sufficient quantities for finetuning. To this end, we present a physics-based training pipeline that tackles the pathology of data scarcity. The core enabler is a physics-based modeling framework that generates a multitude of synthetic data to align the LLM to a physically consistent initial state before finetuning. Our framework features a two-phase training strategy: (1) utilizing the large-in-amount while less accurate synthetic data for supervised pretraining, and (2) finetuning the phase-1 model with limited experimental data. We empirically demonstrate that supervised pretraining is vital to obtaining accurate finetuned LLMs, via the lens of learning polymer flammability metrics where cone calorimeter data is sparse.

Prompt-based interfaces for Large Language Models (LLMs) have made prototyping and building AI-powered applications easier than ever before. However, identifying potential harms that may arise from AI applications remains a challenge, particularly during prompt-based prototyping. To address this, we present Farsight, a novel in situ interactive tool that helps people identify potential harms from the AI applications they are prototyping. Based on a user's prompt, Farsight highlights news articles about relevant AI incidents and allows users to explore and edit LLM-generated use cases, stakeholders, and harms. We report design insights from a co-design study with 10 AI prototypers and findings from a user study with 42 AI prototypers. After using Farsight, AI prototypers in our user study are better able to independently identify potential harms associated with a prompt and find our tool more useful and usable than existing resources. Their qualitative feedback also highlights that Farsight encourages them to focus on end-users and think beyond immediate harms. We discuss these findings and reflect on their implications for designing AI prototyping experiences that meaningfully engage with AI harms. Farsight is publicly accessible at: https://PAIR-code.github.io/farsight.

In this research short, we examine the potential of using GPT-4o, a state-of-the-art large language model (LLM) to undertake evidence synthesis and systematic assessment tasks. Traditional workflows for such tasks involve large groups of domain experts who manually review and synthesize vast amounts of literature. The exponential growth of scientific literature and recent advances in LLMs provide an opportunity to complementing these traditional workflows with new age tools. We assess the efficacy of GPT-4o to do these tasks on a sample from the dataset created by the Global Adaptation Mapping Initiative (GAMI) where we check the accuracy of climate change adaptation related feature extraction from the scientific literature across three levels of expertise. Our results indicate that while GPT-4o can achieve high accuracy in low-expertise tasks like geographic location identification, their performance in intermediate and high-expertise tasks, such as stakeholder identification and assessment of depth of the adaptation response, is less reliable. The findings motivate the need for designing assessment workflows that utilize the strengths of models like GPT-4o while also providing refinements to improve their performance on these tasks.

Text-to-SQL enables users to interact with databases using natural language, simplifying the retrieval and synthesis of information. Despite the remarkable success of large language models (LLMs) in translating natural language questions into SQL queries, widespread deployment remains limited due to two primary challenges. First, the effective use of text-to-SQL models depends on users' understanding of the model's capabilities-the scope of questions the model can correctly answer. Second, the absence of abstention mechanisms can lead to incorrect SQL generation going unnoticed, thereby undermining trust in the model's output. To enable wider deployment, it is crucial to address these challenges in model design and enhance model evaluation to build trust in the model's output. To this end, we introduce TrustSQL, a novel comprehensive benchmark designed to evaluate text-to-SQL reliability-defined as a model's ability to correctly handle any type of input question by generating correct SQL queries for feasible questions and abstaining from generating infeasible ones (e.g., due to schema incompatibility or functionalities beyond SQL). We evaluate existing methods using a novel penalty-based scoring metric with two modeling approaches: (1) pipeline-based methods combining SQL generators with infeasible question detectors and SQL error detectors for abstention; and (2) unified methods using a single model for the entire task. Our experimental results reveal that achieving high scores under severe penalties requires significant effort and provide a new perspective on developing text-to-SQL models for safer deployment. TrustSQL is available at https://github.com/glee4810/TrustSQL.

Although Large Language Models (LLMs) excel in NLP tasks, they still need external tools to extend their ability. Current research on tool learning with LLMs often assumes mandatory tool use, which does not always align with real-world situations, where the necessity for tools is uncertain, and incorrect or unnecessary use of tools can damage the general abilities of LLMs. Therefore, we propose to explore whether LLMs can discern their ability boundaries and use tools flexibly. We then introduce the Whether-or-not tool usage Evaluation benchmark (WTU-Eval) to assess LLMs with eleven datasets, where six of them are tool-usage datasets, and five are general datasets. LLMs are prompted to use tools according to their needs. The results of eight LLMs on WTU-Eval reveal that LLMs frequently struggle to determine tool use in general datasets, and LLMs' performance in tool-usage datasets improves when their ability is similar to ChatGPT. In both datasets, incorrect tool usage significantly impairs LLMs' performance. To mitigate this, we also develop the finetuning dataset to enhance tool decision-making. Fine-tuning Llama2-7B results in a 14\% average performance improvement and a 16.8\% decrease in incorrect tool usage. We will release the WTU-Eval benchmark.

Chat-based language models are designed to be helpful, yet they should not comply with every user request. While most existing work primarily focuses on refusal of "unsafe" queries, we posit that the scope of noncompliance should be broadened. We introduce a comprehensive taxonomy of contextual noncompliance describing when and how models should not comply with user requests. Our taxonomy spans a wide range of categories including incomplete, unsupported, indeterminate, and humanizing requests (in addition to unsafe requests). To test noncompliance capabilities of language models, we use this taxonomy to develop a new evaluation suite of 1000 noncompliance prompts. We find that most existing models show significantly high compliance rates in certain previously understudied categories with models like GPT-4 incorrectly complying with as many as 30% of requests. To address these gaps, we explore different training strategies using a synthetically-generated training set of requests and expected noncompliant responses. Our experiments demonstrate that while direct finetuning of instruction-tuned models can lead to both over-refusal and a decline in general capabilities, using parameter efficient methods like low rank adapters helps to strike a good balance between appropriate noncompliance and other capabilities.

In this dissertation I would like to guide the reader to the research on dialogue but more precisely the research I have conducted during my career since my PhD thesis. Starting from modular architectures with machine learning/deep learning and reinforcement learning to end-to-end deep neural networks. Besides my work as research associate, I also present the work I have supervised in the last years. I review briefly the state of the art and highlight the open research problems on conversational agents. Afterwards, I present my contribution to Task-Oriented Dialogues (TOD), both as research associate and as the industrial supervisor of CIFRE theses. I discuss conversational QA. Particularly, I present the work of two PhD candidates Thibault Cordier and Sebastien Montella; as well as the work of the young researcher Quentin Brabant. Finally, I present the scientific project, where I discuss about Large Language Models (LLMs) for Task-Oriented Dialogue and Multimodal Task-Oriented Dialogue.

Warning: This paper contains text examples that are offensive or harmful in nature. Despite careful safety alignment, current large language models (LLMs) remain vulnerable to various attacks. To further unveil the safety risks of LLMs, we introduce a Safety Concept Activation Vector (SCAV) framework, which effectively guides the attacks by accurately interpreting LLMs' safety mechanisms. We then develop an SCAV-guided attack method that can generate both attack prompts and embedding-level attacks with automatically selected perturbation hyperparameters. Both automatic and human evaluations demonstrate that our attack method significantly improves the attack success rate and response quality while requiring less training data. Additionally, we find that our generated attack prompts may be transferable to GPT-4, and the embedding-level attacks may also be transferred to other white-box LLMs whose parameters are known. Our experiments further uncover the safety risks present in current LLMs. For example, we find that six out of seven open-source LLMs that we attack consistently provide relevant answers to more than 85% malicious instructions. Finally, we provide insights into the safety mechanism of LLMs.

This study introduces ValueScope, a framework leveraging language models to quantify social norms and values within online communities, grounded in social science perspectives on normative structures. We employ ValueScope to dissect and analyze linguistic and stylistic expressions across 13 Reddit communities categorized under gender, politics, science, and finance. Our analysis provides a quantitative foundation showing that even closely related communities exhibit remarkably diverse norms. This diversity supports existing theories and adds a new dimension--community preference--to understanding community interactions. ValueScope not only delineates differing social norms among communities but also effectively traces their evolution and the influence of significant external events like the U.S. presidential elections and the emergence of new sub-communities. The framework thus highlights the pivotal role of social norms in shaping online interactions, presenting a substantial advance in both the theory and application of social norm studies in digital spaces.

Building carbon-free datacenters depends on effective load scheduling, such as suspend-and-resume [1, 12, 18] and wait-and-scale [5, This paper represents the first effort to quantify uncertainty in 16]. The core idea of these scheduling strategies is to adapt to renewable carbon intensity forecasting for datacenter decarbonization. We energy supplies based on carbon intensity forecasts. Inaccurate identify and analyze two types of uncertainty--temporal and spatial--and carbon intensity forecasts can not only fail to reduce carbon discuss their system implications. To address the temporal emissions but may even increase them [4]. While prior work has dynamics in quantifying uncertainty for carbon intensity forecasting, introduced various methods for carbon intensity forecasting such we introduce a conformal prediction-based framework. Evaluation as ARIMA models [3] and neural networks [9, 10], they focus on results show that our technique robustly achieves target point-based estimation, neglecting to account for their uncertainty coverages in uncertainty quantification across various significance levels. As prior studies point out, considering uncertainty is crucial levels. We conduct two case studies using production power traces, for effective scheduling [17].