Large language models (LLMs) have achieved unprecedented success in the field of natural language processing. However, the black-box nature of their internal mechanisms has brought many concerns about their trustworthiness and interpretability. Recent research has discovered a class of abnormal tokens in the model's vocabulary space and named them "glitch tokens". Those tokens, once included in the input, may induce the model to produce incorrect, irrelevant, or even harmful results, drastically undermining the reliability and practicality of LLMs. In this work, we aim to enhance the understanding of glitch tokens and propose techniques for their detection and mitigation. We first reveal the characteristic features induced by glitch tokens on LLMs, which are evidenced by significant deviations in the distributions of attention patterns and dynamic information from intermediate model layers. Based on the insights, we develop GlitchProber, a tool for efficient glitch token detection and mitigation. GlitchProber utilizes small-scale sampling, principal component analysis for accelerated feature extraction, and a simple classifier for efficient vocabulary screening. Taking one step further, GlitchProber rectifies abnormal model intermediate layer values to mitigate the destructive effects of glitch tokens. Evaluated on five mainstream open-source LLMs, GlitchProber demonstrates higher efficiency, precision, and recall compared to existing approaches, with an average F1 score of 0.86 and an average repair rate of 50.06%. GlitchProber unveils a novel path to address the challenges posed by glitch tokens and inspires future research toward more robust and interpretable LLMs.

Our work presents a novel angle for evaluating language models' (LMs) mathematical abilities, by investigating whether they can discern skills and concepts enabled by math content. We contribute two datasets: one consisting of 385 fine-grained descriptions of K-12 math skills and concepts, or standards, from Achieve the Core (ATC), and another of 9.9K problems labeled with these standards (MathFish). Working with experienced teachers, we find that LMs struggle to tag and verify standards linked to problems, and instead predict labels that are close to ground truth, but differ in subtle ways. We also show that LMs often generate problems that do not fully align with standards described in prompts. Finally, we categorize problems in GSM8k using math standards, allowing us to better understand why some problems are more difficult to solve for models than others.

Large language models (LLMs) have demonstrated impressive performance in mathematical and commonsense reasoning tasks using chain-of-thought (CoT) prompting techniques. But can they perform emotional reasoning by concatenating `Let's think step-by-step' to the input prompt? In this paper we investigate this question along with introducing a novel approach to zero-shot emotion detection and emotional reasoning using LLMs. Existing state of the art zero-shot approaches rely on textual entailment models to choose the most appropriate emotion label for an input text. We argue that this strongly restricts the model to a fixed set of labels which may not be suitable or sufficient for many applications where emotion analysis is required. Instead, we propose framing the problem of emotion analysis as a generative question-answering (QA) task. Our approach uses a two step methodology of generating relevant context or background knowledge to answer the emotion detection question step-by-step. Our paper is the first work on using a generative approach to jointly address the tasks of emotion detection and emotional reasoning for texts. We evaluate our approach on two popular emotion detection datasets and also release the fine-grained emotion labels and explanations for further training and fine-tuning of emotional reasoning systems.

This study explores the potential of Large Language Models (LLMs), specifically GPT-4, to enhance objectivity in organizational task performance evaluations. Through comparative analyses across two studies, including various task performance outputs, we demonstrate that LLMs can serve as a reliable and even superior alternative to human raters in evaluating knowledge-based performance outputs, which are a key contribution of knowledge workers. Our results suggest that GPT ratings are comparable to human ratings but exhibit higher consistency and reliability. Additionally, combined multiple GPT ratings on the same performance output show strong correlations with aggregated human performance ratings, akin to the consensus principle observed in performance evaluation literature. However, we also find that LLMs are prone to contextual biases, such as the halo effect, mirroring human evaluative biases. Our research suggests that while LLMs are capable of extracting meaningful constructs from text-based data, their scope is currently limited to specific forms of performance evaluation. By highlighting both the potential and limitations of LLMs, our study contributes to the discourse on AI role in management studies and sets a foundation for future research to refine AI theoretical and practical applications in management.

Although LLMs have substantial potential in financial applications, there are notable challenges in using pre-trained models to Extraction and interpretation of intricate information from unstructured extract information from financial documents outside their training text data arising in financial applications, such as earnings data while also reducing hallucination [7, 8]. Financial documents call transcripts, present substantial challenges to large language typically contain domain-specific language, multiple data formats, models (LLMs) even using the current best practices to use Retrieval and unique contextual relationships that general purpose-trained Augmented Generation (RAG) (referred to as VectorRAG LLMs do not handle well. In addition, extracting consistent and techniques which utilize vector databases for information retrieval) coherent information from multiple financial documents can be due to challenges such as domain specific terminology and complex challenging due to variations in terminology, format, and context formats of the documents. We introduce a novel approach based across different textual sources. The specialized terminology and on a combination, called HybridRAG, of the Knowledge Graphs complex data formats in financial documents make it difficult for (KGs) based RAG techniques (called GraphRAG) and VectorRAG models to extract meaningful insights, in turn, causing inaccurate techniques to enhance question-answer (Q&A) systems for information predictions, overlooked insights, and unreliable analysis, which extraction from financial documents that is shown to be ultimately hinder the ability to make well-informed decisions.

Metaphor Components Identification (MCI) contributes to enhancing machine understanding of metaphors, thereby advancing downstream natural language processing tasks. However, the complexity, diversity, and dependency on context and background knowledge pose significant challenges for MCI. Large language models (LLMs) offer new avenues for accurate comprehension of complex natural language texts due to their strong semantic analysis and extensive commonsense knowledge. In this research, a new LLM-based framework is proposed, named Linguistics-aware In-context Learning with Data Augmentation (LaiDA). Specifically, ChatGPT and supervised fine-tuning are utilized to tailor a high-quality dataset. LaiDA incorporates a simile dataset for pre-training. A graph attention network encoder generates linguistically rich feature representations to retrieve similar examples. Subsequently, LLM is fine-tuned with prompts that integrate linguistically similar examples. LaiDA ranked 2nd in Subtask 2 of NLPCC2024 Shared Task 9, demonstrating its effectiveness. Code and data are available at https://github.com/WXLJZ/LaiDA.

Much work on the cultural awareness of large language models (LLMs) focuses on the models' sensitivity to geo-cultural diversity. However, in addition to cross-cultural differences, there also exists common ground across cultures. For instance, a bridal veil in the United States plays a similar cultural-relevant role as a honggaitou in China. In this study, we introduce a benchmark dataset CUNIT for evaluating decoder-only LLMs in understanding the cultural unity of concepts. Specifically, CUNIT consists of 1,425 evaluation examples building upon 285 traditional cultural-specific concepts across 10 countries. Based on a systematic manual annotation of cultural-relevant features per concept, we calculate the cultural association between any pair of cross-cultural concepts. Built upon this dataset, we design a contrastive matching task to evaluate the LLMs' capability to identify highly associated cross-cultural concept pairs. We evaluate 3 strong LLMs, using 3 popular prompting strategies, under the settings of either giving all extracted concept features or no features at all on CUNIT Interestingly, we find that cultural associations across countries regarding clothing concepts largely differ from food. Our analysis shows that LLMs are still limited to capturing cross-cultural associations between concepts compared to humans. Moreover, geo-cultural proximity shows a weak influence on model performance in capturing cross-cultural associations.

Debugging is a vital aspect of software development, yet the debugging capabilities of Large Language Models (LLMs) remain largely unexplored. This paper first introduces DEBUGEVAL, a comprehensive benchmark designed to evaluate the debugging capabilities of LLMs. DEBUGEVAL collects data from existing high-quality datasets and designs four different tasks to evaluate the debugging effectiveness, including BUG Localization, BUG Identification, Code Review, and Code Repair. Additionally, to enhance the code debugging ability of LLMs, this paper proposes a CoMmunicative Agent BaSed DaTa REfinement FRamework (MASTER), which generates the refined code debugging data for supervised finetuning. Specifically, MASTER employs the Code Quizzer to generate refined data according to the defined tasks of DEBUGEVAL. Then the Code Learner acts as a critic and reserves the generated problems that it can not solve. Finally, the Code Teacher provides a detailed Chain-of-Thought based solution to deal with the generated problem. We collect the synthesized data and finetune the Code Learner to enhance the debugging ability and conduct the NeuDebugger model. Our experiments evaluate various LLMs and NeuDebugger in the zero-shot setting on DEBUGEVAL. Experimental results demonstrate that these 7B-scale LLMs have weaker debugging capabilities, even these code-oriented LLMs. On the contrary, these larger models (over 70B) show convincing debugging ability. Our further analyses illustrate that MASTER is an effective method to enhance the code debugging ability by synthesizing data for Supervised Fine-Tuning (SFT) LLMs.

With the advent of multi-modal large language models (MLLMs), datasets used for visual question answering (VQA) and referring expression comprehension have seen a resurgence. However, the most popular datasets used to evaluate MLLMs are some of the earliest ones created, and they have many known problems, including extreme bias, spurious correlations, and an inability to permit fine-grained analysis. In this paper, we pioneer evaluating recent MLLMs (LLaVA 1.5, LLaVA-NeXT, BLIP2, InstructBLIP, GPT-4V, and GPT-4o) on datasets designed to address weaknesses in earlier ones. We assess three VQA datasets: 1) TDIUC, which permits fine-grained analysis on 12 question types; 2) TallyQA, which has simple and complex counting questions; and 3) DVQA, which requires optical character recognition for chart understanding. We also study VQDv1, a dataset that requires identifying all image regions that satisfy a given query. Our experiments reveal the weaknesses of many MLLMs that have not previously been reported. Our code is integrated into the widely used LAVIS framework for MLLM evaluation, enabling the rapid assessment of future MLLMs.

Despite their impressive capabilities, large language models (LLMs) can be prone to generating hallucinations -- undesirable outputs that are incorrect, unfaithful, or inconsistent with respect to the inputs (or the output itself) [1]. These unreliable behaviors pose significant risks for adopting LLMs in real-world applications. Challenges in detecting hallucinations lie, among other things, in hallucinations taking different forms, being context-dependent and sometimes being in conflict with other desirable properties of generated text [2, 3]. Hallucinations may be harmless in some contexts, but can be undesired or potentially dangerous in other applications (e.g., erroneous medical advice). Detecting and quantifying hallucination risk is thus a critical capability to enable safe applications of LLMs and improve generated outputs. Prior work has proposed various approaches for detecting and mitigating hallucinations in LLM-generated outputs, including verifying faithfulness to inputs [4], assessing internal coherence [5], consulting external knowledge sources [6], and quantifying model uncertainty [2, 3, 7, 8]. However, deploying these methods in production settings is far from trivial due to several challenges: First, there is limited comparative evaluation illuminating how different detection methods perform. Second, existing approaches for detecting hallucinations differ greatly in their computational demands, and guidelines are lacking on cost-effectiveness trade-offs to inform method selection for real-world applications with constraints. Third, hallucination detection in the real world often requires careful consideration of risks and false positive/negative trade-offs, requiring methods to provide well-calibrated probability scores.