Large language models (LLMs) have achieved impressive success in text-formatted learning problems, and most popular LLMs have been deployed in a black-box fashion. Meanwhile, fine-tuning is usually necessary for a specific downstream task to obtain better performance, and this functionality is provided by the owners of the black-box LLMs. To fine-tune a black-box LLM, labeled data are always required to adjust the model parameters. However, in many real-world applications, LLMs can label textual datasets with even better quality than skilled human annotators, motivating us to explore the possibility of fine-tuning black-box LLMs with unlabeled data. In this paper, we propose unsupervised prompt learning for classification with black-box LLMs, where the learning parameters are the prompt itself and the pseudo labels of unlabeled data. Specifically, the prompt is modeled as a sequence of discrete tokens, and every token has its own to-be-learned categorical distribution. On the other hand, for learning the pseudo labels, we are the first to consider the in-context learning (ICL) capabilities of LLMs: we first identify reliable pseudo-labeled data using the LLM, and then assign pseudo labels to other unlabeled data based on the prompt, allowing the pseudo-labeled data to serve as in-context demonstrations alongside the prompt. Those in-context demonstrations matter: previously, they are involved when the prompt is used for prediction while they are not involved when the prompt is trained; thus, taking them into account during training makes the prompt-learning and prompt-using stages more consistent. Experiments on benchmark datasets show the effectiveness of our proposed algorithm. After unsupervised prompt learning, we can use the pseudo-labeled dataset for further fine-tuning by the owners of the black-box LLMs.

While language models have demonstrated impressive capabilities across a range of tasks, they still struggle with tasks that require complex planning and reasoning. Recent studies have proposed training language models on search processes rather than optimal solutions, resulting in better generalization performance even though search processes are noisy and even suboptimal. However, these studies overlook the value of optimal solutions, which can serve as step-by-step landmarks to guide more effective search. In this work, we explore how to leverage optimal solutions to enhance the search and planning abilities of language models. To this end, we propose guided stream of search (GSoS), which seamlessly incorporates optimal solutions into the self-generation process in a progressive manner, producing high-quality search trajectories. These trajectories are then distilled into the pre-trained model via supervised fine-tuning. Our approach significantly enhances the search and planning abilities of language models on Countdown, a simple yet challenging mathematical reasoning task. Notably, combining our method with RL fine-tuning yields further improvements, whereas previous supervised fine-tuning methods do not benefit from RL. Furthermore, our approach exhibits greater effectiveness than leveraging optimal solutions in the form of subgoal rewards.

In this paper, we present a novel approach to coal mining question answering (QA) using large language models (LLMs) combined with tailored prompt engineering techniques. Coal mining is a complex, high-risk industry where accurate, context-aware information is critical for safe and efficient operations. Current QA systems struggle to handle the technical and dynamic nature of mining-related queries. To address these challenges, we propose a multi-turn prompt engineering framework designed to guide LLMs, such as GPT-4, in answering coal mining questions with higher precision and relevance. By breaking down complex queries into structured components, our approach allows LLMs to process nuanced technical information more effectively. We manually curated a dataset of 500 questions from real-world mining scenarios and evaluated the system's performance using both accuracy (ACC) and GPT-4-based scoring metrics. Experiments comparing ChatGPT, Claude2, and GPT-4 across baseline, chain-of-thought (CoT), and multi-turn prompting methods demonstrate that our method significantly improves both accuracy and contextual relevance, with an average accuracy improvement of 15-18\% and a notable increase in GPT-4 scores. The results show that our prompt-engineering approach provides a robust, adaptable solution for domain-specific question answering in high-stakes environments like coal mining.

Despite large language models (LLMs) increasingly becoming important components of news recommender systems, employing LLMs in such systems introduces new risks, such as the influence of cognitive biases in LLMs. Cognitive biases refer to systematic patterns of deviation from norms or rationality in the judgment process, which can result in inaccurate outputs from LLMs, thus threatening the reliability of news recommender systems. Specifically, LLM-based news recommender systems affected by cognitive biases could lead to the propagation of misinformation, reinforcement of stereotypes, and the formation of echo chambers. In this paper, we explore the potential impact of multiple cognitive biases on LLM-based news recommender systems, including anchoring bias, framing bias, status quo bias and group attribution bias. Furthermore, to facilitate future research at improving the reliability of LLM-based news recommender systems, we discuss strategies to mitigate these biases through data augmentation, prompt engineering and learning algorithms aspects.

How can we automatically select an out-of-distribution (OOD) detection model for various underlying tasks? This is crucial for maintaining the reliability of open-world applications by identifying data distribution shifts, particularly in critical domains such as online transactions, autonomous driving, and real-time patient diagnosis. Despite the availability of numerous OOD detection methods, the challenge of selecting an optimal model for diverse tasks remains largely underexplored, especially in scenarios lacking ground truth labels. In this work, we introduce MetaOOD, the first zero-shot, unsupervised framework that utilizes meta-learning to automatically select an OOD detection model. As a meta-learning approach, MetaOOD leverages historical performance data of existing methods across various benchmark OOD datasets, enabling the effective selection of a suitable model for new datasets without the need for labeled data at the test time. To quantify task similarities more accurately, we introduce language model-based embeddings that capture the distinctive OOD characteristics of both datasets and detection models. Through extensive experimentation with 24 unique test dataset pairs to choose from among 11 OOD detection models, we demonstrate that MetaOOD significantly outperforms existing methods and only brings marginal time overhead. Our results, validated by Wilcoxon statistical tests, show that MetaOOD surpasses a diverse group of 11 baselines, including established OOD detectors and advanced unsupervised selection methods.

Large Language Models (LLMs) are rapidly becoming commodity components of larger software systems. This poses natural security and privacy problems: poisoned data retrieved from one component can change the model's behavior and compromise the entire system, including coercing the model to spread confidential data to untrusted components. One promising approach is to tackle this problem at the system level via dynamic information flow (aka taint) tracking. Unfortunately, the traditional approach of propagating the most restrictive input label to the output is too conservative for applications where LLMs operate on inputs retrieved from diverse sources. In this paper, we propose a novel, more permissive approach to propagate information flow labels through LLM queries. The key idea behind our approach is to propagate only the labels of the samples that were influential in generating the model output and to eliminate the labels of unnecessary input. We implement and investigate the effectiveness of two variations of this approach, based on (i) prompt-based retrieval augmentation, and (ii) a $k$-nearest-neighbors language model. We compare these with the baseline of an introspection-based influence estimator that directly asks the language model to predict the output label. The results obtained highlight the superiority of our prompt-based label propagator, which improves the label in more than 85% of the cases in an LLM agent setting. These findings underscore the practicality of permissive label propagation for retrieval augmentation.

LLM-as-a-Judge has been widely utilized as an evaluation method in various benchmarks and served as supervised rewards in model training. However, despite their excellence in many domains, potential issues are under-explored, undermining their reliability and the scope of their utility. Therefore, we identify 12 key potential biases and propose a new automated bias quantification framework-CALM-which systematically quantifies and analyzes each type of bias in LLM-as-a-Judge by using automated and principle-guided modification. Our experiments cover multiple popular language models, and the results indicate that while advanced models have achieved commendable overall performance, significant biases persist in certain specific tasks. Empirical results suggest that there remains room for improvement in the reliability of LLM-as-a-Judge. Moreover, we also discuss the explicit and implicit influence of these biases and give some suggestions for the reliable application of LLM-as-a-Judge. Our work highlights the need for stakeholders to address these issues and remind users to exercise caution in LLM-as-a-Judge applications.

Information retrieval (IR) methods, like retrieval augmented generation, are fundamental to modern applications but often lack statistical guarantees. Conformal prediction addresses this by retrieving sets guaranteed to include relevant information, yet existing approaches produce large-sized sets, incurring high computational costs and slow response times. In this work, we introduce a score refinement method that applies a simple monotone transformation to retrieval scores, leading to significantly smaller conformal sets while maintaining their statistical guarantees. Experiments on various BEIR benchmarks validate the effectiveness of our approach in producing compact sets containing relevant information.

Artistic inspiration remains one of the least understood aspects of the creative process. It plays a crucial role in producing works that resonate deeply with audiences, but the complexity and unpredictability of aesthetic stimuli that evoke inspiration have eluded systematic study. This work proposes a novel framework for computationally modeling artistic preferences in different individuals through key linguistic and stylistic properties, with a focus on lyrical content. In addition to the framework, we introduce \textit{EvocativeLines}, a dataset of annotated lyric lines, categorized as either "inspiring" or "not inspiring," to facilitate the evaluation of our framework across diverse preference profiles. Our computational model leverages the proposed linguistic and poetic features and applies a calibration network on top of it to accurately forecast artistic preferences among different creative individuals. Our experiments demonstrate that our framework outperforms an out-of-the-box LLaMA-3-70b, a state-of-the-art open-source language model, by nearly 18 points. Overall, this work contributes an interpretable and flexible framework that can be adapted to analyze any type of artistic preferences that are inherently subjective across a wide spectrum of skill levels.

Scaling laws in language modeling traditionally quantify training loss as a function of dataset size and model parameters, providing compute-optimal estimates but often neglecting the impact of data quality on model generalization. In this paper, we extend the conventional understanding of scaling law by offering a microscopic view of data quality within the original formulation -- effective training tokens -- which we posit to be a critical determinant of performance for parameter-constrained language models. Specifically, we formulate the proposed term of effective training tokens to be a combination of two readily-computed indicators of text: (i) text diversity and (ii) syntheticity as measured by a teacher model. We pretrained over $200$ models of 25M to 1.5B parameters on a diverse set of sampled, synthetic data, and estimated the constants that relate text quality, model size, training tokens, and eight reasoning task accuracy scores. We demonstrated the estimated constants yield +0.83 Pearson correlation with true accuracies, and analyzed it in scenarios involving widely-used data techniques such as data sampling and synthesis which aim to improve data quality.