Abstract:Large Language Models (LLMs) are limited by their parametric knowledge, leading to hallucinations in knowledge-extensive tasks. To address this, Retrieval-Augmented Generation (RAG) incorporates external document chunks to expand LLM knowledge. Furthermore, compressing information from document chunks through extraction or summarization can improve LLM performance. Nonetheless, LLMs still struggle to notice and utilize scattered key information, a problem known as the "lost-in-the-middle" syndrome. Therefore, we typically need to restructure the content for LLM to recognize the key information. We propose $\textit{Refiner}$, an end-to-end extract-and-restructure paradigm that operates in the post-retrieval process of RAG. $\textit{Refiner}$ leverages a single decoder-only LLM to adaptively extract query-relevant contents verbatim along with the necessary context, and section them based on their interconnectedness, thereby highlights information distinction, and aligns downstream LLMs with the original context effectively. Experiments show that a trained $\textit{Refiner}$ (with 7B parameters) exhibits significant gain to downstream LLM in improving answer accuracy, and outperforms other state-of-the-art advanced RAG and concurrent compressing approaches in various single-hop and multi-hop QA tasks. Notably, $\textit{Refiner}$ achieves a 80.5% tokens reduction and a 1.6-7.0% improvement margin in multi-hop tasks compared to the next best solution. $\textit{Refiner}$ is a plug-and-play solution that can be seamlessly integrated with RAG systems, facilitating its application across diverse open-source frameworks.
Abstract:Projecting visual features into word embedding space has become a significant fusion strategy adopted by Multimodal Large Language Models (MLLMs). However, its internal mechanisms have yet to be explored. Inspired by multilingual research, we identify domain-specific neurons in multimodal large language models. Specifically, we investigate the distribution of domain-specific neurons and the mechanism of how MLLMs process features from diverse domains. Furthermore, we propose a three-stage framework for language model modules in MLLMs when handling projected image features, and verify this hypothesis using logit lens. Extensive experiments indicate that while current MLLMs exhibit Visual Question Answering (VQA) capability, they may not fully utilize domain-specific information. Manipulating domain-specific neurons properly will result in a 10\% change of accuracy at most, shedding light on the development of cross-domain, all-encompassing MLLMs in the future. Our code will be released upon paper notification.
Abstract:Driven by the demand for cross-sentence and large-scale relation extraction, document-level relation extraction (DocRE) has attracted increasing research interest. Despite the continuous improvement in performance, we find that existing DocRE models which initially perform well may make more mistakes when merely changing the entity names in the document, hindering the generalization to novel entity names. To this end, we systematically investigate the robustness of DocRE models to entity name variations in this work. We first propose a principled pipeline to generate entity-renamed documents by replacing the original entity names with names from Wikidata. By applying the pipeline to DocRED and Re-DocRED datasets, we construct two novel benchmarks named Env-DocRED and Env-Re-DocRED for robustness evaluation. Experimental results show that both three representative DocRE models and two in-context learned large language models consistently lack sufficient robustness to entity name variations, particularly on cross-sentence relation instances and documents with more entities. Finally, we propose an entity variation robust training method which not only improves the robustness of DocRE models but also enhances their understanding and reasoning capabilities. We further verify that the basic idea of this method can be effectively transferred to in-context learning for DocRE as well.
Abstract:LLM watermarking, which embeds imperceptible yet algorithmically detectable signals in model outputs to identify LLM-generated text, has become crucial in mitigating the potential misuse of large language models. However, the abundance of LLM watermarking algorithms, their intricate mechanisms, and the complex evaluation procedures and perspectives pose challenges for researchers and the community to easily experiment with, understand, and assess the latest advancements. To address these issues, we introduce MarkLLM, an open-source toolkit for LLM watermarking. MarkLLM offers a unified and extensible framework for implementing LLM watermarking algorithms, while providing user-friendly interfaces to ensure ease of access. Furthermore, it enhances understanding by supporting automatic visualization of the underlying mechanisms of these algorithms. For evaluation, MarkLLM offers a comprehensive suite of 12 tools spanning three perspectives, along with two types of automated evaluation pipelines. Through MarkLLM, we aim to support researchers while improving the comprehension and involvement of the general public in LLM watermarking technology, fostering consensus and driving further advancements in research and application. Our code is available at https://github.com/THU-BPM/MarkLLM.
Abstract:The perception of waterways based on human intent is significant for autonomous navigation and operations of Unmanned Surface Vehicles (USVs) in water environments. Inspired by visual grounding, we introduce WaterVG, the first visual grounding dataset designed for USV-based waterway perception based on human prompts. WaterVG encompasses prompts describing multiple targets, with annotations at the instance level including bounding boxes and masks. Notably, WaterVG includes 11,568 samples with 34,987 referred targets, whose prompts integrates both visual and radar characteristics. The pattern of text-guided two sensors equips a finer granularity of text prompts with visual and radar features of referred targets. Moreover, we propose a low-power visual grounding model, Potamoi, which is a multi-task model with a well-designed Phased Heterogeneous Modality Fusion (PHMF) mode, including Adaptive Radar Weighting (ARW) and Multi-Head Slim Cross Attention (MHSCA). Exactly, ARW extracts required radar features to fuse with vision for prompt alignment. MHSCA is an efficient fusion module with a remarkably small parameter count and FLOPs, elegantly fusing scenario context captured by two sensors with linguistic features, which performs expressively on visual grounding tasks. Comprehensive experiments and evaluations have been conducted on WaterVG, where our Potamoi archives state-of-the-art performances compared with counterparts.
Abstract:Entity Set Expansion (ESE) aims to identify new entities belonging to the same semantic class as a given set of seed entities. Traditional methods primarily relied on positive seed entities to represent a target semantic class, which poses challenge for the representation of ultra-fine-grained semantic classes. Ultra-fine-grained semantic classes are defined based on fine-grained semantic classes with more specific attribute constraints. Describing it with positive seed entities alone cause two issues: (i) Ambiguity among ultra-fine-grained semantic classes. (ii) Inability to define "unwanted" semantic. Due to these inherent shortcomings, previous methods struggle to address the ultra-fine-grained ESE (Ultra-ESE). To solve this issue, we first introduce negative seed entities in the inputs, which belong to the same fine-grained semantic class as the positive seed entities but differ in certain attributes. Negative seed entities eliminate the semantic ambiguity by contrast between positive and negative attributes. Meanwhile, it provide a straightforward way to express "unwanted". To assess model performance in Ultra-ESE, we constructed UltraWiki, the first large-scale dataset tailored for Ultra-ESE. UltraWiki encompasses 236 ultra-fine-grained semantic classes, where each query of them is represented with 3-5 positive and negative seed entities. A retrieval-based framework RetExpan and a generation-based framework GenExpan are proposed to comprehensively assess the efficacy of large language models from two different paradigms in Ultra-ESE. Moreover, we devised three strategies to enhance models' comprehension of ultra-fine-grained entities semantics: contrastive learning, retrieval augmentation, and chain-of-thought reasoning. Extensive experiments confirm the effectiveness of our proposed strategies and also reveal that there remains a large space for improvement in Ultra-ESE.
Abstract:Recent advancements in large language models (LLMs) have revealed their potential for achieving autonomous agents possessing human-level intelligence. However, existing benchmarks for evaluating LLM Agents either use static datasets, potentially leading to data leakage or focus only on single-agent scenarios, overlooking the complexities of multi-agent interactions. There is a lack of a benchmark that evaluates the diverse capabilities of LLM agents in multi-agent, dynamic environments. To this end, we introduce LLMArena, a novel and easily extensible framework for evaluating the diverse capabilities of LLM in multi-agent dynamic environments. LLMArena encompasses seven distinct gaming environments, employing Trueskill scoring to assess crucial abilities in LLM agents, including spatial reasoning, strategic planning, numerical reasoning, risk assessment, communication, opponent modeling, and team collaboration. We conduct an extensive experiment and human evaluation among different sizes and types of LLMs, showing that LLMs still have a significant journey ahead in their development towards becoming fully autonomous agents, especially in opponent modeling and team collaboration. We hope LLMArena could guide future research towards enhancing these capabilities in LLMs, ultimately leading to more sophisticated and practical applications in dynamic, multi-agent settings. The code and data will be available.
Abstract:Recently, large language models (LLMs) have achieved tremendous breakthroughs in the field of language processing, yet their mechanisms in processing multiple languages remain agnostic. Therefore, in this work we study the multilingual activation patterns of LLMs. By transforming the original Large Language Models (LLMs) into a Mixture of Experts (MoE) architecture, we analyze the expert activation patterns when processing various languages and demonstrate the connections of these activation patterns at the level of language families. We discover the existence of non-language-specific neurons as well as language-specific activation neurons. Further exploration even showcases that merely leveraging high-frequency activation neurons can accelerate inference while maintaining comparable performance. These findings shed light on the LLMs' multilingual processing mechanism, and are of significant importance in guiding the multilingual training and model pruning of LLMs.
Abstract:Recently, Large Language Models (LLMs) have been widely studied by researchers for their roles in various downstream NLP tasks. As a fundamental task in the NLP field, Chinese Grammatical Error Correction (CGEC) aims to correct all potential grammatical errors in the input sentences. Previous studies have shown that LLMs' performance as correctors on CGEC remains unsatisfactory due to its challenging task focus. To promote the CGEC field to better adapt to the era of LLMs, we rethink the roles of LLMs in the CGEC task so that they can be better utilized and explored in CGEC. Considering the rich grammatical knowledge stored in LLMs and their powerful semantic understanding capabilities, we utilize LLMs as explainers to provide explanation information for the CGEC small models during error correction to enhance performance. We also use LLMs as evaluators to bring more reasonable CGEC evaluations, thus alleviating the troubles caused by the subjectivity of the CGEC task. In particular, our work is also an active exploration of how LLMs and small models better collaborate in downstream tasks. Extensive experiments and detailed analyses on widely used datasets verify the effectiveness of our thinking intuition and the proposed methods.
Abstract:Recently, Large Language Models (LLMs) have made remarkable evolutions in language understanding and generation. Following this, various benchmarks for measuring all kinds of capabilities of LLMs have sprung up. In this paper, we challenge the reasoning and understanding abilities of LLMs by proposing a FaLlacy Understanding Benchmark (FLUB) containing cunning questions that are easy for humans to understand but difficult for models to grasp. Specifically, the cunning questions that FLUB focuses on mainly consist of the tricky, humorous, and misleading questions collected from the real internet environment. And we design three tasks with increasing difficulty in the FLUB benchmark to evaluate the fallacy understanding ability of LLMs. Based on FLUB, we investigate the performance of multiple representative and advanced LLMs, reflecting our FLUB is challenging and worthy of more future study. Interesting discoveries and valuable insights are achieved in our extensive experiments and detailed analyses. We hope that our benchmark can encourage the community to improve LLMs' ability to understand fallacies.