Yifan Xu

Xiao Liu, Hao Yu, Hanchen Zhang et al. · berkeley, microsoft-research

The potential of Large Language Model (LLM) as agents has been widely acknowledged recently. Thus, there is an urgent need to quantitatively \textit{evaluate LLMs as agents} on challenging tasks in interactive environments. We present AgentBench, a multi-dimensional benchmark that consists of 8 distinct environments to assess LLM-as-Agent's reasoning and decision-making abilities. Our extensive test over \num API-based and open-sourced (OSS) LLMs shows that, while top commercial LLMs present a strong ability of acting as agents in complex environments, there is a significant disparity in performance between them and many OSS competitors that are no larger than 70B. We identify the typical reasons of failures in environments and LLMs, showing that poor long-term reasoning, decision-making, and instruction following abilities are the main obstacles for developing usable LLM agents. Improving instruction following and training on high quality multi-round alignment data could improve agent performance. And different from existing assumptions, training on code present ambivalent impacts on different agent tasks. Datasets, environments, and an integrated evaluation package for AgentBench are released at https://github.com/THUDM/AgentBench.

Xiao Liu, Xuanyu Lei, Shengyuan Wang et al. · tsinghua

Alignment has become a critical step for instruction-tuned Large Language Models (LLMs) to become helpful assistants. However, the effective evaluation of alignment for emerging Chinese LLMs is still largely unexplored. To fill in this gap, we introduce AlignBench, a comprehensive multi-dimensional benchmark for evaluating LLMs' alignment in Chinese. We design a human-in-the-loop data curation pipeline, containing eight main categories, 683 real-scenario rooted queries and corresponding human verified references. To ensure the correctness of references, each knowledge-intensive query is accompanied with evidences collected from reliable web sources (including URLs and quotations) by our annotators. For automatic evaluation, our benchmark employs a rule-calibrated multi-dimensional LLM-as-Judge~\cite{zheng2023judging} approach with Chain-of-Thought to generate explanations and final ratings, ensuring high reliability and interpretability. All evaluation code, data, and LLM generations are available at \url{https://github.com/THUDM/AlignBench}. Since its release, AlignBench has been adopted by top (Chinese) LLMs for evaluating their alignment capabilities in Chinese, including ChatGLM, Qwen, DeepSeek, Yi, Baichuan, and Abab.

Kai Sun, Yifan Ethan Xu, Hanwen Zha et al.

Since the recent prosperity of Large Language Models (LLMs), there have been interleaved discussions regarding how to reduce hallucinations from LLM responses, how to increase the factuality of LLMs, and whether Knowledge Graphs (KGs), which store the world knowledge in a symbolic form, will be replaced with LLMs. In this paper, we try to answer these questions from a new angle: How knowledgeable are LLMs? To answer this question, we constructed Head-to-Tail, a benchmark that consists of 18K question-answer (QA) pairs regarding head, torso, and tail facts in terms of popularity. We designed an automated evaluation method and a set of metrics that closely approximate the knowledge an LLM confidently internalizes. Through a comprehensive evaluation of 16 publicly available LLMs, we show that existing LLMs are still far from being perfect in terms of their grasp of factual knowledge, especially for facts of torso-to-tail entities.

Yifan Xu, Xiao Liu, Xinghan Liu et al. · tsinghua

Large language models (LLMs) have shown excellent mastering of human language, but still struggle in real-world applications that require mathematical problem-solving. While many strategies and datasets to enhance LLMs' mathematics are developed, it remains a challenge to simultaneously maintain and improve both language and mathematical capabilities in deployed LLM systems.In this work, we tailor the Self-Critique pipeline, which addresses the challenge in the feedback learning stage of LLM alignment. We first train a general Math-Critique model from the LLM itself to provide feedback signals. Then, we sequentially employ rejective fine-tuning and direct preference optimization over the LLM's own generations for data collection. Based on ChatGLM3-32B, we conduct a series of experiments on both academic and our newly created challenging dataset, MathUserEval. Results show that our pipeline significantly enhances the LLM's mathematical problem-solving while still improving its language ability, outperforming LLMs that could be two times larger. Related techniques have been deployed to ChatGLM\footnote{\url{https://chatglm.cn}}, an online serving LLM. Related evaluation dataset and scripts are released at \url{https://github.com/THUDM/ChatGLM-Math}.

Xue Jiang, Lubin Meng, Xinru Chen et al.

Electroencephalogram-based motor imagery (MI) classification is an important paradigm of non-invasive brain-computer interfaces. Common spatial pattern (CSP), which exploits different energy distributions on the scalp while performing different MI tasks, is very popular in MI classification. Convolutional neural networks (CNNs) have also achieved great success, due to their powerful learning capabilities. This paper proposes two CSP-empowered neural networks (CSP-Nets), which integrate knowledge-driven CSP filters with data-driven CNNs to enhance the performance in MI classification. CSP-Net-1 directly adds a CSP layer before a CNN to improve the input discriminability. CSP-Net-2 replaces a convolutional layer in CNN with a CSP layer. The CSP layer parameters in both CSP-Nets are initialized with CSP filters designed from the training data. During training, they can either be kept fixed or optimized using gradient descent. Experiments on four public MI datasets demonstrated that the two CSP-Nets consistently improved over their CNN backbones, in both within-subject and cross-subject classifications. They are particularly useful when the number of training samples is very small. Our work demonstrates the advantage of integrating knowledge-driven traditional machine learning with data-driven deep learning in EEG-based brain-computer interfaces.

Yifan Xu, Xue Jiang, Dongrui Wu

Emotion recognition is a critical component of affective computing. Training accurate machine learning models for emotion recognition typically requires a large amount of labeled data. Due to the subtleness and complexity of emotions, multiple evaluators are usually needed for each affective sample to obtain its ground-truth label, which is expensive. To save the labeling cost, this paper proposes an inconsistency-based active learning approach for cross-task transfer between emotion classification and estimation. Affective norms are utilized as prior knowledge to connect the label spaces of categorical and dimensional emotions. Then, the prediction inconsistency on the two tasks for the unlabeled samples is used to guide sample selection in active learning for the target task. Experiments on within-corpus and cross-corpus transfers demonstrated that cross-task inconsistency could be a very valuable metric in active learning. To our knowledge, this is the first work that utilizes prior knowledge on affective norms and data in a different task to facilitate active learning for a new task, even the two tasks are from different datasets.

Jifan Yu, Xiaohan Zhang, Yifan Xu et al. · tsinghua

Empowered by the large-scale pretrained language models, existing dialogue systems have demonstrated impressive performance conducting fluent and natural-sounding conversations. However, they are still plagued by the hallucination problem, causing unpredictable factual errors in the generated responses. Recently, knowledge-grounded dialogue generation models, that intentionally invoke external knowledge resources to more informative responses, are also proven to be effective in reducing hallucination. Following the idea of getting high-quality knowledge, a few efforts have achieved pretty good performance on this issue. As some inevitable knowledge noises may also lead to hallucinations, it is emergent to investigate the reason and future directions for building noise-tolerant methods in KGD tasks. In this paper, we analyze the causal story behind this problem with counterfactual reasoning methods. Based on the causal effect analysis, we propose a possible solution for alleviating the hallucination in KGD by exploiting the dialogue-knowledge interaction. Experimental results of our example implementation show that this method can reduce hallucination without disrupting other dialogue performance, while keeping adaptive to different generation models. We hope our efforts can support and call for more attention to developing lightweight techniques towards robust and trusty dialogue systems.

Yushi Sun, Hao Xin, Kai Sun et al.

Large language models (LLMs) demonstrate an impressive ability to internalize knowledge and answer natural language questions. Although previous studies validate that LLMs perform well on general knowledge while presenting poor performance on long-tail nuanced knowledge, the community is still doubtful about whether the traditional knowledge graphs should be replaced by LLMs. In this paper, we ask if the schema of knowledge graph (i.e., taxonomy) is made obsolete by LLMs. Intuitively, LLMs should perform well on common taxonomies and at taxonomy levels that are common to people. Unfortunately, there lacks a comprehensive benchmark that evaluates the LLMs over a wide range of taxonomies from common to specialized domains and at levels from root to leaf so that we can draw a confident conclusion. To narrow the research gap, we constructed a novel taxonomy hierarchical structure discovery benchmark named TaxoGlimpse to evaluate the performance of LLMs over taxonomies. TaxoGlimpse covers ten representative taxonomies from common to specialized domains with in-depth experiments of different levels of entities in this taxonomy from root to leaf. Our comprehensive experiments of eighteen state-of-the-art LLMs under three prompting settings validate that LLMs can still not well capture the knowledge of specialized taxonomies and leaf-level entities.

Tyler A. Chang, Yifan Xu, Weijian Xu et al.

In this paper, we detail the relationship between convolutions and self-attention in natural language tasks. We show that relative position embeddings in self-attention layers are equivalent to recently-proposed dynamic lightweight convolutions, and we consider multiple new ways of integrating convolutions into Transformer self-attention. Specifically, we propose composite attention, which unites previous relative position embedding methods under a convolutional framework. We conduct experiments by training BERT with composite attention, finding that convolutions consistently improve performance on multiple downstream tasks, replacing absolute position embeddings. To inform future work, we present results comparing lightweight convolutions, dynamic convolutions, and depthwise-separable convolutions in language model pre-training, considering multiple injection points for convolutions in self-attention layers.

Yuqi Cui, Dongrui Wu, Yifan Xu

Takagi-Sugeno-Kang (TSK) fuzzy system with Gaussian membership functions (MFs) is one of the most widely used fuzzy systems in machine learning. However, it usually has difficulty handling high-dimensional datasets. This paper explores why TSK fuzzy systems with Gaussian MFs may fail on high-dimensional inputs. After transforming defuzzification to an equivalent form of softmax function, we find that the poor performance is due to the saturation of softmax. We show that two defuzzification operations, LogTSK and HTSK, the latter of which is first proposed in this paper, can avoid the saturation. Experimental results on datasets with various dimensionalities validated our analysis and demonstrated the effectiveness of LogTSK and HTSK.

Yifan Xu, Tianqi Fan, Yi Yuan et al.

Deep implicit field regression methods are effective for 3D reconstruction from single-view images. However, the impact of different sampling patterns on the reconstruction quality is not well-understood. In this work, we first study the effect of point set discrepancy on the network training. Based on Farthest Point Sampling algorithm, we propose a sampling scheme that theoretically encourages better generalization performance, and results in fast convergence for SGD-based optimization algorithms. Secondly, based on the reflective symmetry of an object, we propose a feature fusion method that alleviates issues due to self-occlusions which makes it difficult to utilize local image features. Our proposed system Ladybird is able to create high quality 3D object reconstructions from a single input image. We evaluate Ladybird on a large scale 3D dataset (ShapeNet) demonstrating highly competitive results in terms of Chamfer distance, Earth Mover's distance and Intersection Over Union (IoU).

Xin Luna Dong, Xiang He, Andrey Kan et al.

Can one build a knowledge graph (KG) for all products in the world? Knowledge graphs have firmly established themselves as valuable sources of information for search and question answering, and it is natural to wonder if a KG can contain information about products offered at online retail sites. There have been several successful examples of generic KGs, but organizing information about products poses many additional challenges, including sparsity and noise of structured data for products, complexity of the domain with millions of product types and thousands of attributes, heterogeneity across large number of categories, as well as large and constantly growing number of products. We describe AutoKnow, our automatic (self-driving) system that addresses these challenges. The system includes a suite of novel techniques for taxonomy construction, product property identification, knowledge extraction, anomaly detection, and synonym discovery. AutoKnow is (a) automatic, requiring little human intervention, (b) multi-scalable, scalable in multiple dimensions (many domains, many products, and many attributes), and (c) integrative, exploiting rich customer behavior logs. AutoKnow has been operational in collecting product knowledge for over 11K product types.

Yaqing Wang, Yifan Ethan Xu, Xian Li et al.

Product catalogs are valuable resources for eCommerce website. In the catalog, a product is associated with multiple attributes whose values are short texts, such as product name, brand, functionality and flavor. Usually individual retailers self-report these key values, and thus the catalog information unavoidably contains noisy facts. Although existing deep neural network models have shown success in conducting cross-checking between two pieces of texts, their success has to be dependent upon a large set of quality labeled data, which are hard to obtain in this validation task: products span a variety of categories. To address the aforementioned challenges, we propose a novel meta-learning latent variable approach, called MetaBridge, which can learn transferable knowledge from a subset of categories with limited labeled data and capture the uncertainty of never-seen categories with unlabeled data. More specifically, we make the following contributions. (1) We formalize the problem of validating the textual attribute values of products from a variety of categories as a natural language inference task in the few-shot learning setting, and propose a meta-learning latent variable model to jointly process the signals obtained from product profiles and textual attribute values. (2) We propose to integrate meta learning and latent variable in a unified model to effectively capture the uncertainty of various categories. (3) We propose a novel objective function based on latent variable model in the few-shot learning setting, which ensures distribution consistency between unlabeled and labeled data and prevents overfitting by sampling from the learned distribution. Extensive experiments on real eCommerce datasets from hundreds of categories demonstrate the effectiveness of MetaBridge on textual attribute validation and its outstanding performance compared with state-of-the-art approaches.

Dongrui Wu, Yifan Xu, Bao-Liang Lu

A brain-computer interface (BCI) enables a user to communicate with a computer directly using brain signals. The most common non-invasive BCI modality, electroencephalogram (EEG), is sensitive to noise/artifact and suffers between-subject/within-subject non-stationarity. Therefore, it is difficult to build a generic pattern recognition model in an EEG-based BCI system that is optimal for different subjects, during different sessions, for different devices and tasks. Usually, a calibration session is needed to collect some training data for a new subject, which is time-consuming and user unfriendly. Transfer learning (TL), which utilizes data or knowledge from similar or relevant subjects/sessions/devices/tasks to facilitate learning for a new subject/session/device/task, is frequently used to reduce the amount of calibration effort. This paper reviews journal publications on TL approaches in EEG-based BCIs in the last few years, i.e., since 2016. Six paradigms and applications -- motor imagery, event-related potentials, steady-state visual evoked potentials, affective BCIs, regression problems, and adversarial attacks -- are considered. For each paradigm/application, we group the TL approaches into cross-subject/session, cross-device, and cross-task settings and review them separately. Observations and conclusions are made at the end of the paper, which may point to future research directions.

Yifan Xu, Lu Dai, Udaikaran Singh et al.

Neural inductive program synthesis is a task generating instructions that can produce desired outputs from given inputs. In this paper, we focus on the generation of a chunk of assembly code that can be executed to match a state change inside the CPU and RAM. We develop a neural program synthesis algorithm, AutoAssemblet, learned via self-learning reinforcement learning that explores the large code space efficiently. Policy networks and value networks are learned to reduce the breadth and depth of the Monte Carlo Tree Search, resulting in better synthesis performance. We also propose an effective multi-entropy policy sampling technique to alleviate online update correlations. We apply AutoAssemblet to basic programming tasks and show significant higher success rates compared to several competing baselines.

Yifan Xu, Kening Zhang, Haoyu Dong et al.

Exposure bias describes the phenomenon that a language model trained under the teacher forcing schema may perform poorly at the inference stage when its predictions are conditioned on its previous predictions unseen from the training corpus. Recently, several generative adversarial networks (GANs) and reinforcement learning (RL) methods have been introduced to alleviate this problem. Nonetheless, a common issue in RL and GANs training is the sparsity of reward signals. In this paper, we adopt two simple strategies, multi-range reinforcing, and multi-entropy sampling, to amplify and denoise the reward signal. Our model produces an improvement over competing models with regards to BLEU scores and road exam, a new metric we designed to measure the robustness against exposure bias in language models.

Yuqi Cuui, Yifan Xu, Dongrui Wu

Drowsy driving is pervasive, and also a major cause of traffic accidents. Estimating a driver's drowsiness level by monitoring the electroencephalogram (EEG) signal and taking preventative actions accordingly may improve driving safety. However, individual differences among different drivers make this task very challenging. A calibration session is usually required to collect some subject-specific data and tune the model parameters before applying it to a new subject, which is very inconvenient and not user-friendly. Many approaches have been proposed to reduce the calibration effort, but few can completely eliminate it. This paper proposes a novel approach, feature weighted episodic training (FWET), to completely eliminate the calibration requirement. It integrates two techniques: feature weighting to learn the importance of different features, and episodic training for domain generalization. Experiments on EEG-based driver drowsiness estimation demonstrated that both feature weighting and episodic training are effective, and their integration can further improve the generalization performance. FWET does not need any labelled or unlabelled calibration data from the new subject, and hence could be very useful in plug-and-play brain-computer interfaces.

Zihan Liu, Bo Huang, Yuqi Cui et al.

Time-lapse is a technology used to record the development of embryos during in-vitro fertilization (IVF). Accurate classification of embryo early development stages can provide embryologists valuable information for assessing the embryo quality, and hence is critical to the success of IVF. This paper proposes a multi-task deep learning with dynamic programming (MTDL-DP) approach for this purpose. It first uses MTDL to pre-classify each frame in the time-lapse video to an embryo development stage, and then DP to optimize the stage sequence so that the stage number is monotonically non-decreasing, which usually holds in practice. Different MTDL frameworks, e.g., one-to-many, many-to-one, and many-to-many, are investigated. It is shown that the one-to-many MTDL framework achieved the best compromise between performance and computational cost. To our knowledge, this is the first study that applies MTDL to embryo early development stage classification from time-lapse videos.

Jang Ik Cho, Xiaofeng Wang, Yifan Xu et al.

Large sequences of images (or movies) can now be obtained on an unprecedented scale, which poses fundamental challenges to the existing image analysis techniques. The challenges include heterogeneity, (automatic) alignment, multiple comparisons, potential artifacts, and hidden noises. This paper introduces our MATLAB package, Longitudinal Image Sequence Analysis (LISA), as a one-stop ensemble of image processing and analysis tool for comparing a general class of images from either different times, sessions, or subjects. Given two contrasting sequences of images, the image processing in LISA starts with selecting a region of interest in two representative images, followed by automatic or manual segmentation and registration. Automatic segmentation de-noises an image using a mixture of Gaussian distributions of the pixel intensity values, while manual segmentation applies a user-chosen intensity cut-off value to filter out noises. Automatic registration aligns the contrasting images based on a mid-line regression whereas manual registration lines up the images along a reference line formed by two user-selected points. The processed images are then rendered for simultaneous statistical comparisons to generate D, S, T, and P-maps. The D map represents a curated difference of contrasting images, the S map is the non-parametrically smoothed differences, the T map presents the variance-adjusted, smoothed differences, and the P-map provides multiplicity-controlled p-values. These maps reveal the regions with significant differences due to either longitudinal, subject-specific, or treatment changes. A user can skip the image processing step to dive directly into the statistical analysis step if the images have already been processed. Hence, LISA offers flexibility in applying other image pre-processing tools. LISA also has a parallel computing option for high definition images.