Large Language Models (LLMs) have demonstrated remarkable capabilities across various applications, fundamentally reshaping the landscape of natural language processing (NLP) research. However, recent evaluation frameworks often rely on the output probabilities of LLMs for predictions, primarily due to computational constraints, diverging from real-world LLM usage scenarios. While widely employed, the efficacy of these probability-based evaluation strategies remains an open research question. This study aims to scrutinize the validity of such probability-based evaluation methods within the context of using LLMs for Multiple Choice Questions (MCQs), highlighting their inherent limitations. Our empirical investigation reveals that the prevalent probability-based evaluation method inadequately aligns with generation-based prediction. Furthermore, current evaluation frameworks typically assess LLMs through predictive tasks based on output probabilities rather than directly generating responses, owing to computational limitations. We illustrate that these probability-based approaches do not effectively correspond with generative predictions. The outcomes of our study can enhance the understanding of LLM evaluation methodologies and provide insights for future research in this domain.

Homogeneity bias in Large Language Models (LLMs) refers to their tendency to homogenize the representations of some groups compared to others. Previous studies documenting this bias have predominantly used encoder models, which may have inadvertently introduced biases. To address this limitation, we prompted GPT-4 to generate single word/expression completions associated with 18 situation cues - specific, measurable elements of environments that influence how individuals perceive situations and compared the variability of these completions using probability of differentiation. This approach directly assessed homogeneity bias from the model's outputs, bypassing encoder models. Across five studies, we find that homogeneity bias is highly volatile across situation cues and writing prompts, suggesting that the bias observed in past work may reflect those within encoder models rather than LLMs. Furthermore, these results suggest that homogeneity bias in LLMs is brittle, as even minor and arbitrary changes in prompts can significantly alter the expression of biases. Future work should further explore how variations in syntactic features and topic choices in longer text generations influence homogeneity bias in LLMs.

The rapid evolution of Large Language Models (LLMs) has markedly expanded their application across diverse domains, transforming how complex problems are approached and solved. Initially conceived to predict subsequent words in texts, these models have transcended their original design to comprehend and respond to the underlying contexts of queries. Today, LLMs routinely perform tasks that once seemed formidable, such as writing essays, poems, stories, and even developing software code. As their capabilities continue to grow, so too do the expectations of their performance in even more sophisticated domains. Despite these advancements, LLMs still encounter significant challenges, particularly in scenarios requiring intricate decision-making, such as planning trips or choosing among multiple viable options. These tasks often demand a nuanced understanding of various outcomes and the ability to predict the consequences of different choices, which are currently outside the typical operational scope of LLMs. This paper proposes an innovative approach to bridge this capability gap. By enabling LLMs to request multiple potential options and their respective parameters from users, our system introduces a dynamic framework that integrates an optimization function within the decision-making process. This function is designed to analyze the provided options, simulate potential outcomes, and determine the most advantageous solution based on a set of predefined criteria. By harnessing this methodology, LLMs can offer tailored, optimal solutions to complex, multi-variable problems, significantly enhancing their utility and effectiveness in real-world applications. This approach not only expands the functional envelope of LLMs but also paves the way for more autonomous and intelligent systems capable of supporting sophisticated decision-making tasks.

In recent years, with the rapid advancements in large language models (LLMs), achieving excellent empathetic response capability has become a crucial prerequisite. Consequently, managing and understanding large-scale video datasets has gained increasing importance. However, empathetic data are typically trained without any quality selection, leading to inefficient data usage and wasted computational resources. Additionally, using raw data can result in low performance in empathetic dialogues. In this work, we present Efficient-Empathy, a sensibility and rationality score-based data selection algorithm that automatically selects sensibility and rationality data while discarding low-quality data. With only the sensibility data (59% of the full dataset), our trained sensibility model efficiently achieves state-of-the-art (SoTA) performance. Furthermore, with multiple data selection hyperparameters, the sensibility model demonstrates SoTA performance, showcasing the robustness of our method. By integrating sensibility and rationality data with a MoE structure, we achieve even higher performance, demonstrating the effectiveness of our Efficient-Empathy algorithm.

The widespread adoption of Large Language Models (LLMs), exemplified by OpenAI's ChatGPT, brings to the forefront the imperative to defend against adversarial threats on these models. These attacks, which manipulate an LLM's output by introducing malicious inputs, undermine the model's integrity and the trust users place in its outputs. In response to this challenge, our paper presents an innovative defensive strategy, given white box access to an LLM, that harnesses residual activation analysis between transformer layers of the LLM. We apply a novel methodology for analyzing distinctive activation patterns in the residual streams for attack prompt classification. We curate multiple datasets to demonstrate how this method of classification has high accuracy across multiple types of attack scenarios, including our newly-created attack dataset. Furthermore, we enhance the model's resilience by integrating safety fine-tuning techniques for LLMs in order to measure its effect on our capability to detect attacks. The results underscore the effectiveness of our approach in enhancing the detection and mitigation of adversarial inputs, advancing the security framework within which LLMs operate.

Emotions play important epistemological and cognitive roles in our lives, revealing our values and guiding our actions. Previous work has shown that LLMs display biases in emotion attribution along gender lines. However, unlike gender, which says little about our values, religion, as a socio-cultural system, prescribes a set of beliefs and values for its followers. Religions, therefore, cultivate certain emotions. Moreover, these rules are explicitly laid out and interpreted by religious leaders. Using emotion attribution, we explore how different religions are represented in LLMs. We find that: Major religions in the US and European countries are represented with more nuance, displaying a more shaded model of their beliefs. Eastern religions like Hinduism and Buddhism are strongly stereotyped. Judaism and Islam are stigmatized -- the models' refusal skyrocket. We ascribe these to cultural bias in LLMs and the scarcity of NLP literature on religion. In the rare instances where religion is discussed, it is often in the context of toxic language, perpetuating the perception of these religions as inherently toxic. This finding underscores the urgent need to address and rectify these biases. Our research underscores the crucial role emotions play in our lives and how our values influence them.

Global sustainable fund universe encompasses open-end funds and exchange-traded funds (ETF) that, by prospectus or other regulatory filings, claim to focus on Environment, Social and Governance (ESG). Challengingly, the claims can only be confirmed by examining the textual disclosures to check if there is presence of intentionality and ESG focus on its investment strategy. Currently, there is no regulation to enforce sustainability in ESG products space. This paper proposes a unique method and system to classify and score the fund prospectuses in the sustainable universe regarding specificity and transparency of language. We aim to employ few-shot learners to identify specific, ambiguous, and generic sustainable investment-related language. Additionally, we construct a ratio metric to determine language score and rating to rank products and quantify sustainability claims for US sustainable universe. As a by-product, we publish manually annotated quality training dataset on Hugging Face (ESG-Prospectus-Clarity-Category under cc-by-nc-sa-4.0) of more than 1K ESG textual statements. The performance of the few-shot finetuning approach is compared with zero-shot models e.g., Llama-13B, GPT 3.5 Turbo etc. We found that prompting large language models are not accurate for domain specific tasks due to misalignment issues. The few-shot finetuning techniques outperform zero-shot models by large margins of more than absolute ~30% in precision, recall and F1 metrics on completely unseen ESG languages (test set). Overall, the paper attempts to establish a systematic and scalable approach to measure and rate sustainability intention quantitatively for sustainable funds using texts in prospectus. Regulatory bodies, investors, and advisors may utilize the findings of this research to reduce cognitive load in investigating or screening of ESG funds which accurately reflects the ESG intention.

The task of Text-to-SQL enables anyone to retrieve information from SQL databases using natural language. Despite several challenges, recent models have made remarkable advancements in this task using large language models (LLMs). Interestingly, we find that LLM-based models without fine-tuning exhibit distinct natures compared to their fine-tuned counterparts, leading to inadequacies in current evaluation metrics to accurately convey their performance. Thus, we analyze the two primary metrics, Test Suite Execution Accuracy (EXE) and Exact Set Matching Accuracy (ESM), to examine their robustness for this task and address shortcomings. We compare the performance of 9 LLM-based models using EXE, the original ESM, and our improved ESM (called ESM+). Our results show that EXE and ESM have high false positive and negative rates of 11.3% and 13.9%, while ESM+ gives those of 0.1% and 2.6% respectively, providing a significantly more stable evaluation. We release the ESM+ script as open-source for the community to contribute, while enjoying a more reliable assessment of Text-to-SQL.

Mooncake is the serving platform for Kimi, a leading LLM service provided by Moonshot AI. It features a KVCache-centric disaggregated architecture that separates the prefill and decoding clusters. It also leverages the underutilized CPU, DRAM, and SSD resources of the GPU cluster to implement a disaggregated cache of KVCache. The core of Mooncake is its KVCache-centric scheduler, which balances maximizing overall effective throughput while meeting latency-related Service Level Objectives (SLOs). Unlike traditional studies that assume all requests will be processed, Mooncake faces challenges due to highly overloaded scenarios. To mitigate these, we developed a prediction-based early rejection policy. Experiments show that Mooncake excels in long-context scenarios. Compared to the baseline method, Mooncake can achieve up to a 525% increase in throughput in certain simulated scenarios while adhering to SLOs. Under real workloads, Mooncake's innovative architecture enables Kimi to handle 75% more requests.

The vocabulary size in temporal action localization (TAL) is constrained by the scarcity of large-scale annotated datasets. To address this, recent works incorporate powerful pre-trained vision-language models (VLMs), such as CLIP, to perform open-vocabulary TAL (OV-TAL). However, unlike VLMs trained on extensive image/video-text pairs, existing OV-TAL methods still rely on small, fully labeled TAL datasets for training an action localizer. In this paper, we explore the scalability of self-training with unlabeled YouTube videos for OV-TAL. Our self-training approach consists of two stages. First, a class-agnostic action localizer is trained on a human-labeled TAL dataset and used to generate pseudo-labels for unlabeled videos. Second, the large-scale pseudo-labeled dataset is combined with the human-labeled dataset to train the localizer. Extensive experiments demonstrate that leveraging web-scale videos in self-training significantly enhances the generalizability of an action localizer. Additionally, we highlighted issues with existing OV-TAL evaluation schemes and proposed a new evaluation protocol. Code is released at https://github.com/HYUNJS/STOV-TAL