The inference of transformer-based large language models consists of two sequential stages: 1) a prefilling stage to compute the KV cache of prompts and generate the first token, and 2) a decoding stage to generate subsequent tokens. For long prompts, the KV cache must be computed for all tokens during the prefilling stage, which can significantly increase the time needed to generate the first token. Consequently, the prefilling stage may become a bottleneck in the generation process. An open question remains whether all prompt tokens are essential for generating the first token. To answer this, we introduce a novel method, LazyLLM, that selectively computes the KV for tokens important for the next token prediction in both the prefilling and decoding stages. Contrary to static pruning approaches that prune the prompt at once, LazyLLM allows language models to dynamically select different subsets of tokens from the context in different generation steps, even though they might be pruned in previous steps. Extensive experiments on standard datasets across various tasks demonstrate that LazyLLM is a generic method that can be seamlessly integrated with existing language models to significantly accelerate the generation without fine-tuning. For instance, in the multi-document question-answering task, LazyLLM accelerates the prefilling stage of the LLama 2 7B model by 2.34x while maintaining accuracy.

Large language models (LLMs) as autonomous agents offer a novel avenue for tackling real-world challenges through a knowledge-driven manner. These LLM-enhanced methodologies excel in generalization and interpretability. However, the complexity of driving tasks often necessitates the collaboration of multiple, heterogeneous agents, underscoring the need for such LLM-driven agents to engage in cooperative knowledge sharing and cognitive synergy. Despite the promise of LLMs, current applications predominantly center around single agent scenarios. To broaden the horizons of knowledge-driven strategies and bolster the generalization capabilities of autonomous agents, we propose the KoMA framework consisting of multi-agent interaction, multi-step planning, shared-memory, and ranking-based reflection modules to enhance multi-agents' decision-making in complex driving scenarios. Based on the framework's generated text descriptions of driving scenarios, the multi-agent interaction module enables LLM agents to analyze and infer the intentions of surrounding vehicles, akin to human cognition. The multi-step planning module enables LLM agents to analyze and obtain final action decisions layer by layer to ensure consistent goals for short-term action decisions. The shared memory module can accumulate collective experience to make superior decisions, and the ranking-based reflection module can evaluate and improve agent behavior with the aim of enhancing driving safety and efficiency. The KoMA framework not only enhances the robustness and adaptability of autonomous driving agents but also significantly elevates their generalization capabilities across diverse scenarios. Empirical results demonstrate the superiority of our approach over traditional methods, particularly in its ability to handle complex, unpredictable driving environments without extensive retraining.

Large pre-trained Vision-Language Models (VLMs) have become ubiquitous foundational components of other models and downstream tasks. Although powerful, our empirical results reveal that such models might not be able to identify fine-grained concepts. Specifically, the explanations of VLMs with respect to fine-grained concepts are entangled and mislocalized. To address this issue, we propose to DisEntAngle and Localize (DEAL) the concept-level explanations for VLMs without human annotations. The key idea is encouraging the concept-level explanations to be distinct while maintaining consistency with category-level explanations. We conduct extensive experiments and ablation studies on a wide range of benchmark datasets and vision-language models. Our empirical results demonstrate that the proposed method significantly improves the concept-level explanations of the model in terms of disentanglability and localizability. Surprisingly, the improved explainability alleviates the model's reliance on spurious correlations, which further benefits the prediction accuracy.

We can evaluate features of problems and their potential solutions well before we can effectively solve them. When considering a game we have never played, for instance, we might infer whether it is likely to be challenging, fair, or fun simply from hearing the game rules, prior to deciding whether to invest time in learning the game or trying to play it well. Many studies of game play have focused on optimality and expertise, characterizing how people and computational models play based on moderate to extensive search and after playing a game dozens (if not thousands or millions) of times. Here, we study how people reason about a range of simple but novel connect-n style board games. We ask people to judge how fair and how fun the games are from very little experience: just thinking about the game for a minute or so, before they have ever actually played with anyone else, and we propose a resource-limited model that captures their judgments using only a small number of partial game simulations and almost no lookahead search.

Data-to-text (D2T) generation aims to generate human-readable text from semi-structured data, such as tables and graphs. The recent success of D2T is largely attributed to advancements in LLMs. Despite the success of LLMs, no research has been conducted to illustrate the impact of model size on the performance of fine-tuned LLMs for D2T tasks. D2T model performance is typically assessed based on three key qualities: \textit{readability} (indicates fluency and coherence), \textit{informativeness} (measures content similarity), and \textit{faithfulness} (assesses consistency of factual information). It is currently uncertain whether increasing the size of LLMs effectively improves performance in D2T tasks across these three qualities. The objective of this study is to investigate the performance of fine-tuned LLMs in D2T tasks in terms of model size. Through extensive comparative analysis, we aim to elucidate both the advantages and limitations of scaling model sizes across five widely used D2T datasets (E2E, ViGGo, WikiTableText, DART, and WebNLG) and twelve state-of-the-art LLMs with varying sizes from five different LLM families (T5, BART, OPT, BLOOM, and Llama 2). To comprehensively cover all the three essential qualities of D2T models, we incorporate six widely recognized automatic metrics -- \textsc{BLEU}, \textsc{METEOR}, \textsc{BERTScore}, \textsc{MoverScore}, \textsc{Parent}, and \textsc{BARTScore}. We also provide an in-depth analysis of LLM performance concerning model size in the presence of source-reference divergence, a critical aspect of D2T tasks. Our investigation reveals that increasing LLM size enhances \textit{readability} and \textit{informativeness} in D2T tasks, but larger (in terms of size) LLMs may sacrifice \textit{faithfulness}. Moreover, small-sized LLMs show more resilience than larger ones when source-reference divergence is present.

Many natural language processing (NLP) tasks in English or general domains are widely available and are often used to evaluate pre-trained language models. In contrast, there are fewer tasks available for languages other than English and for the financial domain. In particular, tasks in Japanese and the financial domain are limited. We construct two large datasets using materials published by a Japanese central government agency. The datasets provide three Japanese financial NLP tasks, which include a 3-class and 12-class classification for categorizing sentences, as well as a 5-class classification task for sentiment analysis. Our datasets are designed to be comprehensive and up-to-date, leveraging an automatic update framework that ensures the latest task datasets are publicly available anytime.

In recent years, code security has become increasingly important, especially with the rise of interconnected technologies. Detecting vulnerabilities early in the software development process has demonstrated numerous benefits. Consequently, the scientific community started using machine learning for automated detection of source code vulnerabilities. This work explores and refines the CVEFixes dataset, which is commonly used to train models for code-related tasks, specifically the C/C++ subset. To this purpose, the Source Code Processing Engine (SCoPE), a framework composed of strategized techniques that can be used to reduce the size and normalize C/C++ functions is presented. The output generated by SCoPE was used to create a new version of CVEFixes. This refined dataset was then employed in a feature representation analysis to assess the effectiveness of the tool's code processing techniques, consisting of fine-tuning three pre-trained LLMs for software vulnerability detection. The results show that SCoPE successfully helped to identify 905 duplicates within the evaluated subset. The LLM results corroborate with the literature regarding their suitability for software vulnerability detection, with the best model achieving 53% F1-score.

Although large language models (LLMs) have been largely successful in generating functionally correct programs, conditioning models to produce efficient solutions while ensuring correctness remains a challenge. Further, unreliability in benchmarking code efficiency is a hurdle across varying hardware specifications for popular interpreted languages such as Python. In this paper, we present ECCO, a reproducible benchmark for evaluating program efficiency via two paradigms: natural language (NL) based code generation and history-based code editing. On ECCO, we adapt and thoroughly investigate the three most promising existing LLM-based approaches: in-context learning, iterative refinement with execution or NL feedback, and fine-tuning conditioned on execution and editing history. While most methods degrade functional correctness and moderately increase program efficiency, we find that adding execution information often helps maintain functional correctness, and NL feedback enhances more on efficiency. We release our benchmark to support future work on LLM-based generation of efficient code.

The increasing proliferation of misinformation and its alarming impact have motivated both industry and academia to develop approaches for misinformation detection and fact checking. Recent advances on large language models (LLMs) have shown remarkable performance in various tasks, but whether and how LLMs could help with misinformation detection remains relatively underexplored. Most of existing state-of-the-art approaches either do not consider evidence and solely focus on claim related features or assume the evidence to be provided. Few approaches consider evidence retrieval as part of the misinformation detection but rely on fine-tuning models. In this paper, we investigate the potential of LLMs for misinformation detection in a zero-shot setting. We incorporate an evidence retrieval component into the process as it is crucial to gather pertinent information from various sources to detect the veracity of claims. To this end, we propose a novel re-ranking approach for multimodal evidence retrieval using both LLMs and large vision-language models (LVLM). The retrieved evidence samples (images and texts) serve as the input for an LVLM-based approach for multimodal fact verification (LVLM4FV). To enable a fair evaluation, we address the issue of incomplete ground truth for evidence samples in an existing evidence retrieval dataset by annotating a more complete set of evidence samples for both image and text retrieval. Our experimental results on two datasets demonstrate the superiority of the proposed approach in both evidence retrieval and fact verification tasks and also better generalization capability across dataset compared to the supervised baseline.

Cross-attention has become a fundamental module nowadays in many important artificial intelligence applications, e.g., retrieval-augmented generation (RAG), system prompt, guided stable diffusion, and many so on. Ensuring cross-attention privacy is crucial and urgently needed because its key and value matrices may contain sensitive information about companies and their users, many of which profit solely from their system prompts or RAG data. In this work, we design a novel differential privacy (DP) data structure to address the privacy security of cross-attention with a theoretical guarantee. In detail, let $n$ be the input token length of system prompt/RAG data, $d$ be the feature dimension, $0 < \alpha \le 1$ be the relative error parameter, $R$ be the maximum value of the query and key matrices, $R_w$ be the maximum value of the value matrix, and $r,s,\epsilon_s$ be parameters of polynomial kernel methods. Then, our data structure requires $\widetilde{O}(ndr^2)$ memory consumption with $\widetilde{O}(nr^2)$ initialization time complexity and $\widetilde{O}(\alpha^{-1} r^2)$ query time complexity for a single token query. In addition, our data structure can guarantee that the user query is $(\epsilon, \delta)$-DP with $\widetilde{O}(n^{-1} \epsilon^{-1} \alpha^{-1/2} R^{2s} R_w r^2)$ additive error and $n^{-1} (\alpha + \epsilon_s)$ relative error between our output and the true answer. Furthermore, our result is robust to adaptive queries in which users can intentionally attack the cross-attention system. To our knowledge, this is the first work to provide DP for cross-attention. We believe it can inspire more privacy algorithm design in large generative models (LGMs).