Baichuan 2: Open Large-scale Language Models

Large language models (LLMs) have demonstrated remarkable performance on a variety of natural language tasks based on just a few examples of natural language instructions, reducing the need for extensive feature engineering. However, most powerful LLMs are closed-source or limited in their capability for languages other than English. In this technical report, we present Baichuan 2, a series of large-scale multilingual language models containing 7 billion and 13 billion parameters, trained from scratch, on 2.6 trillion tokens. Baichuan 2 matches or outperforms other open-source models of similar size on public benchmarks like MMLU, CMMLU, GSM8K, and HumanEval. Furthermore, Baichuan 2 excels in vertical domains such as medicine and law. We will release all pre-training model checkpoints to benefit the research community in better understanding the training dynamics of Baichuan 2.

Federated Learning in Big Model Era: Domain-Specific Multimodal Large Models

Multimodal data, which can comprehensively perceive and recognize the physical world, has become an essential path towards general artificial intelligence. However, multimodal large models trained on public datasets often underperform in specific industrial domains. This paper proposes a multimodal federated learning framework that enables multiple enterprises to utilize private domain data to collaboratively train large models for vertical domains, achieving intelligent services across scenarios. The authors discuss in-depth the strategic transformation of federated learning in terms of intelligence foundation and objectives in the era of big model, as well as the new challenges faced in heterogeneous data, model aggregation, performance and cost trade-off, data privacy, and incentive mechanism. The paper elaborates a case study of leading enterprises contributing multimodal data and expert knowledge to city safety operation management , including distributed deployment and efficient coordination of the federated learning platform, technical innovations on data quality improvement based on large model capabilities and efficient joint fine-tuning approaches. Preliminary experiments show that enterprises can enhance and accumulate intelligent capabilities through multimodal model federated learning, thereby jointly creating an smart city model that provides high-quality intelligent services covering energy infrastructure safety, residential community security, and urban operation management. The established federated learning cooperation ecosystem is expected to further aggregate industry, academia, and research resources, realize large models in multiple vertical domains, and promote the large-scale industrial application of artificial intelligence and cutting-edge research on multimodal federated learning.

Improving Text Semantic Similarity Modeling through a 3D Siamese Network

Siamese networks have gained popularity as a method for modeling text semantic similarity. Traditional methods rely on pooling operation to compress the semantic representations from Transformer blocks in encoding, resulting in two-dimensional semantic vectors and the loss of hierarchical semantic information from Transformer blocks. Moreover, this limited structure of semantic vectors is akin to a flattened landscape, which restricts the methods that can be applied in downstream modeling, as they can only navigate this flat terrain. To address this issue, we propose a novel 3D Siamese network for text semantic similarity modeling, which maps semantic information to a higher-dimensional space. The three-dimensional semantic tensors not only retains more precise spatial and feature domain information but also provides the necessary structural condition for comprehensive downstream modeling strategies to capture them. Leveraging this structural advantage, we introduce several modules to reinforce this 3D framework, focusing on three aspects: feature extraction, attention, and feature fusion. Our extensive experiments on four text semantic similarity benchmarks demonstrate the effectiveness and efficiency of our 3D Siamese Network.

TVPR: Text-to-Video Person Retrieval and a New Benchmark

Most existing methods for text-based person retrieval focus on text-to-image person retrieval. Nevertheless, due to the lack of dynamic information provided by isolated frames, the performance is hampered when the person is obscured in isolated frames or variable motion details are given in the textual description. In this paper, we propose a new task called Text-to-Video Person Retrieval(TVPR) which aims to effectively overcome the limitations of isolated frames. Since there is no dataset or benchmark that describes person videos with natural language, we construct a large-scale cross-modal person video dataset containing detailed natural language annotations, such as person's appearance, actions and interactions with environment, etc., termed as Text-to-Video Person Re-identification (TVPReid) dataset, which will be publicly available. To this end, a Text-to-Video Person Retrieval Network (TVPRN) is proposed. Specifically, TVPRN acquires video representations by fusing visual and motion representations of person videos, which can deal with temporal occlusion and the absence of variable motion details in isolated frames. Meanwhile, we employ the pre-trained BERT to obtain caption representations and the relationship between caption and video representations to reveal the most relevant person videos. To evaluate the effectiveness of the proposed TVPRN, extensive experiments have been conducted on TVPReid dataset. To the best of our knowledge, TVPRN is the first successful attempt to use video for text-based person retrieval task and has achieved state-of-the-art performance on TVPReid dataset. The TVPReid dataset will be publicly available to benefit future research.

Exploring the Potential of Large Language Models (LLMs) in Learning on Graphs

Learning on Graphs has attracted immense attention due to its wide real-world applications. The most popular pipeline for learning on graphs with textual node attributes primarily relies on Graph Neural Networks (GNNs), and utilizes shallow text embedding as initial node representations, which has limitations in general knowledge and profound semantic understanding. In recent years, Large Language Models (LLMs) have been proven to possess extensive common knowledge and powerful semantic comprehension abilities that have revolutionized existing workflows to handle text data. In this paper, we aim to explore the potential of LLMs in graph machine learning, especially the node classification task, and investigate two possible pipelines: LLMs-as-Enhancers and LLMs-as-Predictors. The former leverages LLMs to enhance nodes' text attributes with their massive knowledge and then generate predictions through GNNs. The latter attempts to directly employ LLMs as standalone predictors. We conduct comprehensive and systematical studies on these two pipelines under various settings. From comprehensive empirical results, we make original observations and find new insights that open new possibilities and suggest promising directions to leverage LLMs for learning on graphs.

Empowering Molecule Discovery for Molecule-Caption Translation with Large Language Models: A ChatGPT Perspective

Molecule discovery plays a crucial role in various scientific fields, advancing the design of tailored materials and drugs. Traditional methods for molecule discovery follow a trial-and-error process, which are both time-consuming and costly, while computational approaches such as artificial intelligence (AI) have emerged as revolutionary tools to expedite various tasks, like molecule-caption translation. Despite the importance of molecule-caption translation for molecule discovery, most of the existing methods heavily rely on domain experts, require excessive computational cost, and suffer from poor performance. On the other hand, Large Language Models (LLMs), like ChatGPT, have shown remarkable performance in various cross-modal tasks due to their great powerful capabilities in natural language understanding, generalization, and reasoning, which provides unprecedented opportunities to advance molecule discovery. To address the above limitations, in this work, we propose a novel LLMs-based framework (\textbf{MolReGPT}) for molecule-caption translation, where a retrieval-based prompt paradigm is introduced to empower molecule discovery with LLMs like ChatGPT without fine-tuning. More specifically, MolReGPT leverages the principle of molecular similarity to retrieve similar molecules and their text descriptions from a local database to ground the generation of LLMs through in-context few-shot molecule learning. We evaluate the effectiveness of MolReGPT via molecule-caption translation, which includes molecule understanding and text-based molecule generation. Experimental results show that MolReGPT outperforms fine-tuned models like MolT5-base without any additional training. To the best of our knowledge, MolReGPT is the first work to leverage LLMs in molecule-caption translation for advancing molecule discovery.

Do-GOOD: Towards Distribution Shift Evaluation for Pre-Trained Visual Document Understanding Models

Numerous pre-training techniques for visual document understanding (VDU) have recently shown substantial improvements in performance across a wide range of document tasks. However, these pre-trained VDU models cannot guarantee continued success when the distribution of test data differs from the distribution of training data. In this paper, to investigate how robust existing pre-trained VDU models are to various distribution shifts, we first develop an out-of-distribution (OOD) benchmark termed Do-GOOD for the fine-Grained analysis on Document image-related tasks specifically. The Do-GOOD benchmark defines the underlying mechanisms that result in different distribution shifts and contains 9 OOD datasets covering 3 VDU related tasks, e.g., document information extraction, classification and question answering. We then evaluate the robustness and perform a fine-grained analysis of 5 latest VDU pre-trained models and 2 typical OOD generalization algorithms on these OOD datasets. Results from the experiments demonstrate that there is a significant performance gap between the in-distribution (ID) and OOD settings for document images, and that fine-grained analysis of distribution shifts can reveal the brittle nature of existing pre-trained VDU models and OOD generalization algorithms. The code and datasets for our Do-GOOD benchmark can be found at https://github.com/MAEHCM/Do-GOOD.

Self-Explainable Graph Neural Networks for Link Prediction

Graph Neural Networks (GNNs) have achieved state-of-the-art performance for link prediction. However, GNNs suffer from poor interpretability, which limits their adoptions in critical scenarios that require knowing why certain links are predicted. Despite various methods proposed for the explainability of GNNs, most of them are post-hoc explainers developed for explaining node classification. Directly adopting existing post-hoc explainers for explaining link prediction is sub-optimal because: (i) post-hoc explainers usually adopt another strategy or model to explain a target model, which could misinterpret the target model; and (ii) GNN explainers for node classification identify crucial subgraphs around each node for the explanation; while for link prediction, one needs to explain the prediction for each pair of nodes based on graph structure and node attributes. Therefore, in this paper, we study a novel problem of self-explainable GNNs for link prediction, which can simultaneously give accurate predictions and explanations. Concretely, we propose a new framework and it can find various $K$ important neighbors of one node to learn pair-specific representations for links from this node to other nodes. These $K$ different neighbors represent important characteristics of the node and model various factors for links from it. Thus, $K$ neighbors can provide explanations for the existence of links. Experiments on both synthetic and real-world datasets verify the effectiveness of the proposed framework for link prediction and explanation.

Interpretable Multimodal Misinformation Detection with Logic Reasoning

Multimodal misinformation on online social platforms is becoming a critical concern due to increasing credibility and easier dissemination brought by multimedia content, compared to traditional text-only information. While existing multimodal detection approaches have achieved high performance, the lack of interpretability hinders these systems' reliability and practical deployment. Inspired by NeuralSymbolic AI which combines the learning ability of neural networks with the explainability of symbolic learning, we propose a novel logic-based neural model for multimodal misinformation detection which integrates interpretable logic clauses to express the reasoning process of the target task. To make learning effective, we parameterize symbolic logical elements using neural representations, which facilitate the automatic generation and evaluation of meaningful logic clauses. Additionally, to make our framework generalizable across diverse misinformation sources, we introduce five meta-predicates that can be instantiated with different correlations. Results on three public datasets (Twitter, Weibo, and Sarcasm) demonstrate the feasibility and versatility of our model.

T-SciQ: Teaching Multimodal Chain-of-Thought Reasoning via Large Language Model Signals for Science Question Answering

Large Language Models (LLMs) have recently demonstrated exceptional performance in various Natural Language Processing (NLP) tasks. They have also shown the ability to perform chain-of-thought (CoT) reasoning to solve complex problems. Recent studies have explored CoT reasoning in complex multimodal scenarios, such as the science question answering task, by fine-tuning multimodal models with high-quality human-annotated CoT rationales. However, collecting high-quality COT rationales is usually time-consuming and costly. Besides, the annotated rationales are hardly accurate due to the redundant information involved or the essential information missed. To address these issues, we propose a novel method termed \emph{T-SciQ} that aims at teaching science question answering with LLM signals. The T-SciQ approach generates high-quality CoT rationales as teaching signals and is advanced to train much smaller models to perform CoT reasoning in complex modalities. Additionally, we introduce a novel data mixing strategy to produce more effective teaching data samples for simple and complex science question answer problems. Extensive experimental results show that our T-SciQ method achieves a new state-of-the-art performance on the ScienceQA benchmark, with an accuracy of 96.18%. Moreover, our approach outperforms the most powerful fine-tuned baseline by 4.5%.