In this work, we investigate the controllability of large language models (LLMs) on scientific summarization tasks. We identify key stylistic and content coverage factors that characterize different types of summaries such as paper reviews, abstracts, and lay summaries. By controlling stylistic features, we find that non-fine-tuned LLMs outperform humans in the MuP review generation task, both in terms of similarity to reference summaries and human preferences. Also, we show that we can improve the controllability of LLMs with keyword-based classifier-free guidance (CFG) while achieving lexical overlap comparable to strong fine-tuned baselines on arXiv and PubMed. However, our results also indicate that LLMs cannot consistently generate long summaries with more than 8 sentences. Furthermore, these models exhibit limited capacity to produce highly abstractive lay summaries. Although LLMs demonstrate strong generic summarization competency, sophisticated content control without costly fine-tuning remains an open problem for domain-specific applications.
相關內容
Learning meaningful word embeddings is key to training a robust language model. The recent rise of Large Language Models (LLMs) has provided us with many new word/sentence/document embedding models. Although LLMs have shown remarkable advancement in various NLP tasks, it is still unclear whether the performance improvement is merely because of scale or whether underlying embeddings they produce significantly differ from classical encoding models like Sentence-BERT (SBERT) or Universal Sentence Encoder (USE). This paper systematically investigates this issue by comparing classical word embedding techniques against LLM-based word embeddings in terms of their latent vector semantics. Our results show that LLMs tend to cluster semantically related words more tightly than classical models. LLMs also yield higher average accuracy on the Bigger Analogy Test Set (BATS) over classical methods. Finally, some LLMs tend to produce word embeddings similar to SBERT, a relatively lighter classical model.
Out-of-distribution (OOD) detection plays a vital role in enhancing the reliability of machine learning (ML) models. The emergence of large language models (LLMs) has catalyzed a paradigm shift within the ML community, showcasing their exceptional capabilities across diverse natural language processing tasks. While existing research has probed OOD detection with relative small-scale Transformers like BERT, RoBERTa and GPT-2, the stark differences in scales, pre-training objectives, and inference paradigms call into question the applicability of these findings to LLMs. This paper embarks on a pioneering empirical investigation of OOD detection in the domain of LLMs, focusing on LLaMA series ranging from 7B to 65B in size. We thoroughly evaluate commonly-used OOD detectors, scrutinizing their performance in both zero-grad and fine-tuning scenarios. Notably, we alter previous discriminative in-distribution fine-tuning into generative fine-tuning, aligning the pre-training objective of LLMs with downstream tasks. Our findings unveil that a simple cosine distance OOD detector demonstrates superior efficacy, outperforming other OOD detectors. We provide an intriguing explanation for this phenomenon by highlighting the isotropic nature of the embedding spaces of LLMs, which distinctly contrasts with the anisotropic property observed in smaller BERT family models. The new insight enhances our understanding of how LLMs detect OOD data, thereby enhancing their adaptability and reliability in dynamic environments.
In recent years, large language models (LLMs) have demonstrated substantial potential in addressing automatic program repair (APR) tasks. However, the current evaluation of these models for APR tasks focuses solely on the limited context of the single function or file where the bug is located, overlooking the valuable information in the repository-level context. This paper investigates the performance of popular LLMs in handling repository-level repair tasks. We introduce RepoBugs, a new benchmark comprising 124 typical repository-level bugs from open-source repositories. Preliminary experiments using GPT3.5 based on the function where the error is located, reveal that the repair rate on RepoBugs is only 22.58%, significantly diverging from the performance of GPT3.5 on function-level bugs in related studies. This underscores the importance of providing repository-level context when addressing bugs at this level. However, the repository-level context offered by the preliminary method often proves redundant and imprecise and easily exceeds the prompt length limit of LLMs. To solve the problem, we propose a simple and universal repository-level context extraction method (RLCE) designed to provide more precise context for repository-level code repair tasks. Evaluations of three mainstream LLMs show that RLCE significantly enhances the ability to repair repository-level bugs. The improvement reaches a maximum of 160% compared to the preliminary method. Additionally, we conduct a comprehensive analysis of the effectiveness and limitations of RLCE, along with the capacity of LLMs to address repository-level bugs, offering valuable insights for future research.
Direct alignment from preferences (DAP) methods, such as DPO, have recently emerged as efficient alternatives to reinforcement learning from human feedback (RLHF), that do not require a separate reward model. However, the preference datasets used in DAP methods are usually collected ahead of training and never updated, thus the feedback is purely offline. Moreover, responses in these datasets are often sampled from a language model distinct from the one being aligned, and since the model evolves over training, the alignment phase is inevitably off-policy. In this study, we posit that online feedback is key and improves DAP methods. Our method, online AI feedback (OAIF), uses an LLM as annotator: on each training iteration, we sample two responses from the current model and prompt the LLM annotator to choose which one is preferred, thus providing online feedback. Despite its simplicity, we demonstrate via human evaluation in several tasks that OAIF outperforms both offline DAP and RLHF methods. We further show that the feedback leveraged in OAIF is easily controllable, via instruction prompts to the LLM annotator.
We explore the viability of Large Language Models (LLMs), specifically OpenAI's GPT-3.5 and GPT-4, in emulating human survey respondents and eliciting preferences, with a focus on intertemporal choices. Leveraging the extensive literature on intertemporal discounting for benchmarking, we examine responses from LLMs across various languages and compare them to human responses, exploring preferences between smaller, sooner, and larger, later rewards. Our findings reveal that both GPT models demonstrate less patience than humans, with GPT-3.5 exhibiting a lexicographic preference for earlier rewards, unlike human decision-makers. Though GPT-4 does not display lexicographic preferences, its measured discount rates are still considerably larger than those found in humans. Interestingly, GPT models show greater patience in languages with weak future tense references, such as German and Mandarin, aligning with existing literature that suggests a correlation between language structure and intertemporal preferences. We demonstrate how prompting GPT to explain its decisions, a procedure we term "chain-of-thought conjoint," can mitigate, but does not eliminate, discrepancies between LLM and human responses. While directly eliciting preferences using LLMs may yield misleading results, combining chain-of-thought conjoint with topic modeling aids in hypothesis generation, enabling researchers to explore the underpinnings of preferences. Chain-of-thought conjoint provides a structured framework for marketers to use LLMs to identify potential attributes or factors that can explain preference heterogeneity across different customers and contexts.
In this paper, we apply the variational information bottleneck approach to end-to-end neural diarization with encoder-decoder attractors (EEND-EDA). This allows us to investigate what information is essential for the model. EEND-EDA utilizes vector representations of the speakers in a conversation - attractors. Our analysis shows that, attractors do not necessarily have to contain speaker characteristic information. On the other hand, giving the attractors more freedom allowing them to encode some extra (possibly speaker-specific) information leads to small but consistent diarization performance improvements. Despite architectural differences in EEND systems, the notion of attractors and frame embeddings is common to most of them and not specific to EEND-EDA. We believe that the main conclusions of this work can apply to other variants of EEND. Thus, we hope this paper will be a valuable contribution to guide the community to make more informed decisions when designing new systems.
Knowledge editing aims to inject knowledge updates into language models to keep them correct and up-to-date. However, its current evaluation strategies are notably impractical: they solely update with well-curated structured facts (triplets with subjects, relations, and objects), whereas real-world knowledge updates commonly emerge in unstructured texts like news articles. In this paper, we propose a new benchmark, Unstructured Knowledge Editing (UKE). It evaluates editing performance directly using unstructured texts as knowledge updates, termed unstructured facts. Hence UKE avoids the laborious construction of structured facts and enables efficient and responsive knowledge editing, becoming a more practical benchmark. We conduct extensive experiments on newly built datasets and demonstrate that UKE poses a significant challenge to state-of-the-art knowledge editing methods, resulting in their critical performance declines. We further show that this challenge persists even if we extract triplets as structured facts. Our analysis discloses key insights to motivate future research in UKE for more practical knowledge editing.
Large language models (LLMs) demonstrate remarkable performance across a spectrum of languages. In this work, we delve into the question: How do LLMs handle multilingualism? We introduce a framework that depicts LLMs' processing of multilingual inputs: In the first several layers, LLMs understand the question, converting multilingual inputs into English to facilitate the task-solving phase. In the intermediate layers, LLMs engage in problem-solving by thinking in English and incorporating multilingual knowledge to obtain factual content, leveraging the self-attention and feed-forward structures, respectively. In the last several layers, LLMs generate responses that align with the original language of the query. In addition, we investigate the existence of language-specific neurons when processing a certain language. To detect neurons activated by the input language, even without labels, we innovatively design a Parallel Language specific Neuron Detection ($\texttt{PLND}$) method that effectively measures the significance of neurons when handling multilingual inputs. By comprehensive ablation analysis through deactivating neurons of different layers and structures, we verify the framework that we propose. Additionally, we demonstrate that we can utilize such a framework to effectively enhance the multilingual ability with much less training effort.
How can we estimate the importance of nodes in a knowledge graph (KG)? A KG is a multi-relational graph that has proven valuable for many tasks including question answering and semantic search. In this paper, we present GENI, a method for tackling the problem of estimating node importance in KGs, which enables several downstream applications such as item recommendation and resource allocation. While a number of approaches have been developed to address this problem for general graphs, they do not fully utilize information available in KGs, or lack flexibility needed to model complex relationship between entities and their importance. To address these limitations, we explore supervised machine learning algorithms. In particular, building upon recent advancement of graph neural networks (GNNs), we develop GENI, a GNN-based method designed to deal with distinctive challenges involved with predicting node importance in KGs. Our method performs an aggregation of importance scores instead of aggregating node embeddings via predicate-aware attention mechanism and flexible centrality adjustment. In our evaluation of GENI and existing methods on predicting node importance in real-world KGs with different characteristics, GENI achieves 5-17% higher NDCG@100 than the state of the art.
Graph Neural Networks (GNNs) for representation learning of graphs broadly follow a neighborhood aggregation framework, where the representation vector of a node is computed by recursively aggregating and transforming feature vectors of its neighboring nodes. Many GNN variants have been proposed and have achieved state-of-the-art results on both node and graph classification tasks. However, despite GNNs revolutionizing graph representation learning, there is limited understanding of their representational properties and limitations. Here, we present a theoretical framework for analyzing the expressive power of GNNs in capturing different graph structures. Our results characterize the discriminative power of popular GNN variants, such as Graph Convolutional Networks and GraphSAGE, and show that they cannot learn to distinguish certain simple graph structures. We then develop a simple architecture that is provably the most expressive among the class of GNNs and is as powerful as the Weisfeiler-Lehman graph isomorphism test. We empirically validate our theoretical findings on a number of graph classification benchmarks, and demonstrate that our model achieves state-of-the-art performance.
小貼士
登錄享
注冊地址： 北京市海淀區羊坊店路18號2幢3層301-191