Large language models (LLMs) have achieved unprecedented performance in various applications, yet their evaluation remains a critical issue. Existing hallucination benchmarks are either static or lack adjustable complexity for thorough analysis. We contend that utilizing existing relational databases is a promising approach for constructing benchmarks due to their accurate knowledge description via functional dependencies. We propose ERBench to automatically convert any relational database into a benchmark based on the entity-relationship (ER) model. Our key idea is to construct questions using the database schema, records, and functional dependencies such that they can be automatically verified. In addition, we use foreign key constraints to join relations and construct multihop questions, which can be arbitrarily complex and used to debug the intermediate answers of LLMs. Finally, ERBench supports continuous evaluation, multimodal questions, and various prompt engineering techniques. In our experiments, we construct an LLM benchmark using databases of multiple domains and make an extensive comparison of contemporary LLMs. We observe that better LLMs like GPT-4 can handle a larger variety of question types, but are by no means perfect. Also, correct answers do not necessarily imply correct rationales, which is an important evaluation that ERBench does better than other benchmarks for various question types. Code is available at https: //github.com/DILAB-KAIST/ERBench.

Open-domain question answering (ODQA) has emerged as a pivotal research spotlight in information systems. Existing methods follow two main paradigms to collect evidence: (1) The \textit{retrieve-then-read} paradigm retrieves pertinent documents from an external corpus; and (2) the \textit{generate-then-read} paradigm employs large language models (LLMs) to generate relevant documents. However, neither can fully address multifaceted requirements for evidence. To this end, we propose LLMQA, a generalized framework that formulates the ODQA process into three basic steps: query expansion, document selection, and answer generation, combining the superiority of both retrieval-based and generation-based evidence. Since LLMs exhibit their excellent capabilities to accomplish various tasks, we instruct LLMs to play multiple roles as generators, rerankers, and evaluators within our framework, integrating them to collaborate in the ODQA process. Furthermore, we introduce a novel prompt optimization algorithm to refine role-playing prompts and steer LLMs to produce higher-quality evidence and answers. Extensive experimental results on widely used benchmarks (NQ, WebQ, and TriviaQA) demonstrate that LLMQA achieves the best performance in terms of both answer accuracy and evidence quality, showcasing its potential for advancing ODQA research and applications.

We explore how iterative revising a chain of thoughts with the help of information retrieval significantly improves large language models' reasoning and generation ability in long-horizon generation tasks, while hugely mitigating hallucination. In particular, the proposed method -- *retrieval-augmented thoughts* (RAT) -- revises each thought step one by one with retrieved information relevant to the task query, the current and the past thought steps, after the initial zero-shot CoT is generated. Applying RAT to GPT-3.5, GPT-4, and CodeLLaMA-7b substantially improves their performances on various long-horizon generation tasks; on average of relatively increasing rating scores by 13.63% on code generation, 16.96% on mathematical reasoning, 19.2% on creative writing, and 42.78% on embodied task planning. The demo page can be found at https://craftjarvis.github.io/RAT

Intent classifiers must be able to distinguish when a user's utterance does not belong to any supported intent to avoid producing incorrect and unrelated system responses. Although out-of-scope (OOS) detection for intent classifiers has been studied, previous work has not yet studied changes in classifier performance against hard-negative out-of-scope utterances (i.e., inputs that share common features with in-scope data, but are actually out-of-scope). We present an automated technique to generate hard-negative OOS data using ChatGPT. We use our technique to build five new hard-negative OOS datasets, and evaluate each against three benchmark intent classifiers. We show that classifiers struggle to correctly identify hard-negative OOS utterances more than general OOS utterances. Finally, we show that incorporating hard-negative OOS data for training improves model robustness when detecting hard-negative OOS data and general OOS data. Our technique, datasets, and evaluation address an important void in the field, offering a straightforward and inexpensive way to collect hard-negative OOS data and improve intent classifiers' robustness.

The performance of machine learning models heavily depends on the quality of input data, yet real-world applications often encounter various data-related challenges. One such challenge could arise when curating training data or deploying the model in the real world - two comparable datasets in the same domain may have different distributions. While numerous techniques exist for detecting distribution shifts, the literature lacks comprehensive approaches for explaining dataset differences in a human-understandable manner. To address this gap, we propose a suite of interpretable methods (toolbox) for comparing two datasets. We demonstrate the versatility of our approach across diverse data modalities, including tabular data, language, images, and signals in both low and high-dimensional settings. Our methods not only outperform comparable and related approaches in terms of explanation quality and correctness, but also provide actionable, complementary insights to understand and mitigate dataset differences effectively.

In-context learning (ICL) enables large language models (LLMs) to adapt to new tasks by conditioning on demonstrations of question-answer pairs and it has been shown to have comparable performance to costly model retraining and fine-tuning. Recently, ICL has been extended to allow tabular data to be used as demonstration examples by serializing individual records into natural language formats. However, it has been shown that LLMs can leak information contained in prompts, and since tabular data often contain sensitive information, understanding how to protect the underlying tabular data used in ICL is a critical area of research. This work serves as an initial investigation into how to use differential privacy (DP) -- the long-established gold standard for data privacy and anonymization -- to protect tabular data used in ICL. Specifically, we investigate the application of DP mechanisms for private tabular ICL via data privatization prior to serialization and prompting. We formulate two private ICL frameworks with provable privacy guarantees in both the local (LDP-TabICL) and global (GDP-TabICL) DP scenarios via injecting noise into individual records or group statistics, respectively. We evaluate our DP-based frameworks on eight real-world tabular datasets and across multiple ICL and DP settings. Our evaluations show that DP-based ICL can protect the privacy of the underlying tabular data while achieving comparable performance to non-LLM baselines, especially under high privacy regimes.

Although Large Language Models (LLMs) are showing impressive performance on a wide range of Natural Language Processing tasks, researchers have found that they still have limited ability to conduct induction. Recent works mainly adopt ``post processes'' paradigms to improve the performance of LLMs on induction (e.g., the hypothesis search & refinement methods), but their performance is still constrained by the inherent inductive capability of the LLMs. In this paper, we propose a novel framework, Induction through Deduction (ItD), to enable the LLMs to teach themselves induction through deduction. The ItD framework is composed of two main components: a Deductive Data Generation module to generate induction data and a Naive Bayesian Induction module to optimize the fine-tuning and decoding of LLMs. Our empirical results showcase the effectiveness of ItD on two induction benchmarks, achieving relative performance improvement of 36% and 10% compared with previous state-of-the-art, respectively. Our ablation study verifies the effectiveness of two key modules of ItD. We also verify the effectiveness of ItD across different LLMs and deductors. The data and code of this paper can be found at https://anonymous.4open.science/r/ItD-E844.

Photonic Crystal Surface Emitting Lasers (PCSEL)'s inverse design demands expert knowledge in physics, materials science, and quantum mechanics which is prohibitively labor-intensive. Advanced AI technologies, especially reinforcement learning (RL), have emerged as a powerful tool to augment and accelerate this inverse design process. By modeling the inverse design of PCSEL as a sequential decision-making problem, RL approaches can construct a satisfactory PCSEL structure from scratch. However, the data inefficiency resulting from online interactions with precise and expensive simulation environments impedes the broader applicability of RL approaches. Recently, sequential models, especially the Transformer architecture, have exhibited compelling performance in sequential decision-making problems due to their simplicity and scalability to large language models. In this paper, we introduce a novel framework named PCSEL Inverse Design Transformer (PiT) that abstracts the inverse design of PCSEL as a sequence modeling problem. The central part of our PiT is a Transformer-based structure that leverages the past trajectories and current states to predict the current actions. Compared with the traditional RL approaches, PiT can output the optimal actions and achieve target PCSEL designs by leveraging offline data and conditioning on the desired return. Results demonstrate that PiT achieves superior performance and data efficiency compared to baselines.

Current large language models have dangerous capabilities, which are likely to become more problematic in the future. Activation steering techniques can be used to reduce risks from these capabilities. In this paper, we investigate the efficacy of activation steering for broad skills and multiple behaviours. First, by comparing the effects of reducing performance on general coding ability and Python-specific ability, we find that steering broader skills is competitive to steering narrower skills. Second, we steer models to become more or less myopic and wealth-seeking, among other behaviours. In our experiments, combining steering vectors for multiple different behaviours into one steering vector is largely unsuccessful. On the other hand, injecting individual steering vectors at different places in a model simultaneously is promising.

Large Language Models (LLMs) stand at the forefront of a number of Natural Language Processing (NLP) tasks. Despite the widespread adoption of LLMs in NLP, much of their potential in broader fields remains largely unexplored, and significant limitations persist in their design and implementation. Notably, LLMs struggle with structured data, such as graphs, and often falter when tasked with answering domain-specific questions requiring deep expertise, such as those in biology and chemistry. In this paper, we explore a fundamental question: Can LLMs effectively handle molecule prediction tasks? Rather than pursuing top-tier performance, our goal is to assess how LLMs can contribute to diverse molecule tasks. We identify several classification and regression prediction tasks across six standard molecule datasets. Subsequently, we carefully design a set of prompts to query LLMs on these tasks and compare their performance with existing Machine Learning (ML) models, which include text-based models and those specifically designed for analysing the geometric structure of molecules. Our investigation reveals several key insights: Firstly, LLMs generally lag behind ML models in achieving competitive performance on molecule tasks, particularly when compared to models adept at capturing the geometric structure of molecules, highlighting the constrained ability of LLMs to comprehend graph data. Secondly, LLMs show promise in enhancing the performance of ML models when used collaboratively. Lastly, we engage in a discourse regarding the challenges and promising avenues to harness LLMs for molecule prediction tasks. The code and models are available at https://github.com/zhiqiangzhongddu/LLMaMol.