Yichi Zhang

Yichi Zhang, Ankush Garg, Yuan Cao et al. · deepmind

The rapid scaling of language models is motivating research using low-bitwidth quantization. In this work, we propose a novel binarization technique for Transformers applied to machine translation (BMT), the first of its kind. We identify and address the problem of inflated dot-product variance when using one-bit weights and activations. Specifically, BMT leverages additional LayerNorms and residual connections to improve binarization quality. Experiments on the WMT dataset show that a one-bit weight-only Transformer can achieve the same quality as a float one, while being 16x smaller in size. One-bit activations incur varying degrees of quality drop, but mitigated by the proposed architectural changes. We further conduct a scaling law study using production-scale translation datasets, which shows that one-bit weight Transformers scale and generalize well in both in-domain and out-of-domain settings. Implementation in JAX/Flax will be open sourced.

Yichi Zhang, Jianing Yang, Jiayi Pan et al. · berkeley

Recent years have seen an increasing amount of work on embodied AI agents that can perform tasks by following human language instructions. However, most of these agents are reactive, meaning that they simply learn and imitate behaviors encountered in the training data. These reactive agents are insufficient for long-horizon complex tasks. To address this limitation, we propose a neuro-symbolic deliberative agent that, while following language instructions, proactively applies reasoning and planning based on its neural and symbolic representations acquired from past experience (e.g., natural language and egocentric vision). We show that our deliberative agent achieves greater than 70% improvement over reactive baselines on the challenging TEACh benchmark. Moreover, the underlying reasoning and planning processes, together with our modular framework, offer impressive transparency and explainability to the behaviors of the agent. This enables an in-depth understanding of the agent's capabilities, which shed light on challenges and opportunities for future embodied agents for instruction following. The code is available at https://github.com/sled-group/DANLI.

Xiangru Tang, Yuliang Liu, Zefan Cai et al. · pku

Despite Large Language Models (LLMs) like GPT-4 achieving impressive results in function-level code generation, they struggle with repository-scale code understanding (e.g., coming up with the right arguments for calling routines), requiring a deeper comprehension of complex file interactions. Also, recently, people have developed LLM agents that attempt to interact with repository code (e.g., compiling and evaluating its execution), prompting the need to evaluate their performance. These gaps have motivated our development of ML-Bench, a benchmark rooted in real-world programming applications that leverage existing code repositories to perform tasks. Addressing the need for LLMs to interpret long code contexts and translate instructions into precise, executable scripts, ML-Bench encompasses annotated 9,641 examples across 18 GitHub repositories, challenging LLMs to accommodate user-specified arguments and documentation intricacies effectively. To evaluate both LLMs and AI agents, two setups are employed: ML-LLM-Bench for assessing LLMs' text-to-code conversion within a predefined deployment environment, and ML-Agent-Bench for testing autonomous agents in an end-to-end task execution within a Linux sandbox environment. Our findings indicate that while GPT-4o leads with a Pass@5 rate surpassing 50%, there remains significant scope for improvement, highlighted by issues such as hallucinated outputs and difficulties with bash script generation. Notably, in the more demanding ML-Agent-Bench, GPT-4o achieves a 76.47% success rate, reflecting the efficacy of iterative action and feedback in complex task resolution. Our code, dataset, and models are available at https://github.com/gersteinlab/ML-bench.

Ziyi Lin, Chris Liu, Renrui Zhang et al.

We present SPHINX, a versatile multi-modal large language model (MLLM) with a joint mixing of model weights, tuning tasks, and visual embeddings. First, for stronger vision-language alignment, we unfreeze the large language model (LLM) during pre-training, and introduce a weight mix strategy between LLMs trained by real-world and synthetic data. By directly integrating the weights from two domains, the mixed LLM can efficiently incorporate diverse semantics with favorable robustness. Then, to enable multi-purpose capabilities, we mix a variety of tasks for joint visual instruction tuning, and design task-specific instructions to avoid inter-task conflict. In addition to the basic visual question answering, we include more challenging tasks such as region-level understanding, caption grounding, document layout detection, and human pose estimation, contributing to mutual enhancement over different scenarios. Additionally, we propose to extract comprehensive visual embeddings from various network architectures, pre-training paradigms, and information granularity, providing language models with more robust image representations. Based on our proposed joint mixing, SPHINX exhibits superior multi-modal understanding capabilities on a wide range of applications. On top of this, we further propose an efficient strategy aiming to better capture fine-grained appearances of high-resolution images. With a mixing of different scales and high-resolution sub-images, SPHINX attains exceptional visual parsing and reasoning performance on existing evaluation benchmarks. We hope our work may cast a light on the exploration of joint mixing in future MLLM research. Code is released at https://github.com/Alpha-VLLM/LLaMA2-Accessory.

Yichi Zhang, Jiayi Pan, Yuchen Zhou et al. · berkeley

Vision-Language Models (VLMs) are trained on vast amounts of data captured by humans emulating our understanding of the world. However, known as visual illusions, human's perception of reality isn't always faithful to the physical world. This raises a key question: do VLMs have the similar kind of illusions as humans do, or do they faithfully learn to represent reality? To investigate this question, we build a dataset containing five types of visual illusions and formulate four tasks to examine visual illusions in state-of-the-art VLMs. Our findings have shown that although the overall alignment is low, larger models are closer to human perception and more susceptible to visual illusions. Our dataset and initial findings will promote a better understanding of visual illusions in humans and machines and provide a stepping stone for future computational models that can better align humans and machines in perceiving and communicating about the shared visual world. The code and data are available at https://github.com/vl-illusion/dataset.

Liang Chen, Yichi Zhang, Shuhuai Ren et al. · pku

In this study, we explore the potential of Multimodal Large Language Models (MLLMs) in improving embodied decision-making processes for agents. While Large Language Models (LLMs) have been widely used due to their advanced reasoning skills and vast world knowledge, MLLMs like GPT4-Vision offer enhanced visual understanding and reasoning capabilities. We investigate whether state-of-the-art MLLMs can handle embodied decision-making in an end-to-end manner and whether collaborations between LLMs and MLLMs can enhance decision-making. To address these questions, we introduce a new benchmark called PCA-EVAL, which evaluates embodied decision-making from the perspectives of Perception, Cognition, and Action. Additionally, we propose HOLMES, a multi-agent cooperation framework that allows LLMs to leverage MLLMs and APIs to gather multimodal information for informed decision-making. We compare end-to-end embodied decision-making and HOLMES on our benchmark and find that the GPT4-Vision model demonstrates strong end-to-end embodied decision-making abilities, outperforming GPT4-HOLMES in terms of average decision accuracy (+3%). However, this performance is exclusive to the latest GPT4-Vision model, surpassing the open-source state-of-the-art MLLM by 26%. Our results indicate that powerful MLLMs like GPT4-Vision hold promise for decision-making in embodied agents, offering new avenues for MLLM research. Code and data are open at https://github.com/pkunlp-icler/PCA-EVAL/.

Yichi Zhang, Zhuo Chen, Yin Fang et al.

Deploying large language models (LLMs) to real scenarios for domain-specific question answering (QA) is a key thrust for LLM applications, which poses numerous challenges, especially in ensuring that responses are both accommodating to user requirements and appropriately leveraging domain-specific knowledge bases. They are the two major difficulties for LLM application as vanilla fine-tuning falls short of addressing. Combining these requirements, we conceive of them as the requirement for the model's preference to be harmoniously aligned with humans'. Thus, we introduce Knowledgeable Preference AlignmenT (KnowPAT), which constructs two kinds of preference sets to tackle the two issues. Besides, we design a new alignment objective to align the LLM preference with different human preferences uniformly, aiming to optimize LLM performance in real-world, domain-specific QA settings. Adequate experiments and comprehensive comparisons with 15 baseline methods illustrate that our KnowPAT is a superior pipeline for real-scenario domain-specific QA with LLMs.

Yichi Zhang, Mingyang Chen, Wen Zhang

Negative sampling (NS) is widely used in knowledge graph embedding (KGE), which aims to generate negative triples to make a positive-negative contrast during training. However, existing NS methods are unsuitable when multi-modal information is considered in KGE models. They are also inefficient due to their complex design. In this paper, we propose Modality-Aware Negative Sampling (MANS) for multi-modal knowledge graph embedding (MMKGE) to address the mentioned problems. MANS could align structural and visual embeddings for entities in KGs and learn meaningful embeddings to perform better in multi-modal KGE while keeping lightweight and efficient. Empirical results on two benchmarks demonstrate that MANS outperforms existing NS methods. Meanwhile, we make further explorations about MANS to confirm its effectiveness.

Yuwei Bao, Keunwoo Peter Yu, Yichi Zhang et al.

Despite tremendous advances in AI, it remains a significant challenge to develop interactive task guidance systems that can offer situated, personalized guidance and assist humans in various tasks. These systems need to have a sophisticated understanding of the user as well as the environment, and make timely accurate decisions on when and what to say. To address this issue, we created a new multimodal benchmark dataset, Watch, Talk and Guide (WTaG) based on natural interaction between a human user and a human instructor. We further proposed two tasks: User and Environment Understanding, and Instructor Decision Making. We leveraged several foundation models to study to what extent these models can be quickly adapted to perceptually enabled task guidance. Our quantitative, qualitative, and human evaluation results show that these models can demonstrate fair performances in some cases with no task-specific training, but a fast and reliable adaptation remains a significant challenge. Our benchmark and baselines will provide a stepping stone for future work on situated task guidance.

Yichi Zhang, Zihao Zeng, Dongbai Li et al.

Large Reasoning Models (LRMs), such as OpenAI o1 and DeepSeek-R1, have been rapidly progressing and achieving breakthrough performance on complex reasoning tasks such as mathematics and coding. However, the open-source R1 models have raised safety concerns in wide applications, such as the tendency to comply with malicious queries, which greatly impacts the utility of these powerful models in their applications. In this paper, we introduce RealSafe-R1 as safety-aligned versions of DeepSeek-R1 distilled models. To train these models, we construct a dataset of 15k safety-aware reasoning trajectories generated by DeepSeek-R1, under explicit instructions for expected refusal behavior. Both quantitative experiments and qualitative case studies demonstrate the models' improvements, which are shown in their safety guardrails against both harmful queries and jailbreak attacks. Importantly, unlike prior safety alignment efforts that often compromise reasoning performance, our method preserves the models' reasoning capabilities by maintaining the training data within the original distribution of generation. Model weights of RealSafe-R1 are open-source at https://huggingface.co/RealSafe.

Liang Chen, Yichi Zhang, Shuhuai Ren et al. · pku

We present PCA-Bench, a multimodal decision-making benchmark for evaluating the integrated capabilities of Multimodal Large Language Models (MLLMs). Departing from previous benchmarks focusing on simplistic tasks and individual model capability, PCA-Bench introduces three complex scenarios: autonomous driving, domestic robotics, and open-world games. Given task instructions and diverse contexts, the model is required to seamlessly integrate multiple capabilities of Perception, Cognition, and Action in a reasoning chain to make accurate decisions. Moreover, PCA-Bench features error localization capabilities, scrutinizing model inaccuracies in areas such as perception, knowledge, or reasoning. This enhances the reliability of deploying MLLMs. To balance accuracy and efficiency in evaluation, we propose PCA-Eval, an automatic evaluation protocol, and assess 10 prevalent MLLMs. The results reveal significant performance disparities between open-source models and powerful proprietary models like GPT-4 Vision. To address this, we introduce Embodied-Instruction-Evolution (EIE), an automatic framework for synthesizing instruction tuning examples in multimodal embodied environments. EIE generates 7,510 training examples in PCA-Bench and enhances the performance of open-source MLLMs, occasionally surpassing GPT-4 Vision (+3\% in decision accuracy), thereby validating the effectiveness of EIE. Our findings suggest that robust MLLMs like GPT4-Vision show promise for decision-making in embodied agents, opening new avenues for MLLM research.

Xiaoze Liu, Feijie Wu, Tianyang Xu et al.

The advent of Large Language Models (LLMs) has significantly transformed the AI landscape, enhancing machine learning and AI capabilities. Factuality issue is a critical concern for LLMs, as they may generate factually incorrect responses. In this paper, we propose GraphEval to evaluate an LLM's performance using a substantially large test dataset. Specifically, the test dataset is retrieved from a large knowledge graph with more than 10 million facts without expensive human efforts. Unlike conventional methods that evaluate LLMs based on generated responses, GraphEval streamlines the evaluation process by creating a judge model to estimate the correctness of the answers given by the LLM. Our experiments demonstrate that the judge model's factuality assessment aligns closely with the correctness of the LLM's generated outputs, while also substantially reducing evaluation costs. Besides, our findings offer valuable insights into LLM performance across different metrics and highlight the potential for future improvements in ensuring the factual integrity of LLM outputs. The code is publicly available at https://github.com/xz-liu/GraphEval.

Yichi Zhang, Binbin Hu, Zhuo Chen et al.

Knowledge graphs (KGs) provide reliable external knowledge for a wide variety of AI tasks in the form of structured triples. Knowledge graph pre-training (KGP) aims to pre-train neural networks on large-scale KGs and provide unified interfaces to enhance different downstream tasks, which is a key direction for KG management, maintenance, and applications. Existing works often focus on purely research questions in open domains, or they are not open source due to data security and privacy in real scenarios. Meanwhile, existing studies have not explored the training efficiency and transferability of KGP models in depth. To address these problems, We propose a framework MuDoK to achieve multi-domain collaborative pre-training and efficient prefix prompt tuning to serve diverse downstream tasks like recommendation and text understanding. Our design is a plug-and-play prompt learning approach that can be flexibly adapted to different downstream task backbones. In response to the lack of open-source benchmarks, we constructed a new multi-domain KGP benchmark called KPI with two large-scale KGs and six different sub-domain tasks to evaluate our method and open-sourced it for subsequent research. We evaluated our approach based on constructed KPI benchmarks using diverse backbone models in heterogeneous downstream tasks. The experimental results show that our framework brings significant performance gains, along with its generality, efficiency, and transferability.

Zhuo Chen, Yin Fang, Yichi Zhang et al.

The rise of Multi-modal Pre-training highlights the necessity for a unified Multi-Modal Knowledge Graph (MMKG) representation learning framework. Such a framework is essential for embedding structured knowledge into multi-modal Large Language Models effectively, alleviating issues like knowledge misconceptions and multi-modal hallucinations. In this work, we explore the efficacy of models in accurately embedding entities within MMKGs through two pivotal tasks: Multi-modal Knowledge Graph Completion (MKGC) and Multi-modal Entity Alignment (MMEA). Building on this foundation, we propose a novel SNAG method that utilizes a Transformer-based architecture equipped with modality-level noise masking to robustly integrate multi-modal entity features in KGs. By incorporating specific training objectives for both MKGC and MMEA, our approach achieves SOTA performance across a total of ten datasets, demonstrating its versatility. Moreover, SNAG can not only function as a standalone model but also enhance other existing methods, providing stable performance improvements. Code and data are available at https://github.com/zjukg/SNAG.

Yichi Zhang, Yao Huang, Yifan Wang et al.

The trustworthiness of Multimodal Large Language Models (MLLMs) remains an intense concern despite the significant progress in their capabilities. Existing evaluation and mitigation approaches often focus on narrow aspects and overlook risks introduced by the multimodality. To tackle these challenges, we propose MultiTrust-X, a comprehensive benchmark for evaluating, analyzing, and mitigating the trustworthiness issues of MLLMs. We define a three-dimensional framework, encompassing five trustworthiness aspects which include truthfulness, robustness, safety, fairness, and privacy; two novel risk types covering multimodal risks and cross-modal impacts; and various mitigation strategies from the perspectives of data, model architecture, training, and inference algorithms. Based on the taxonomy, MultiTrust-X includes 32 tasks and 28 curated datasets, enabling holistic evaluations over 30 open-source and proprietary MLLMs and in-depth analysis with 8 representative mitigation methods. Our extensive experiments reveal significant vulnerabilities in current models, including a gap between trustworthiness and general capabilities, as well as the amplification of potential risks in base LLMs by both multimodal training and inference. Moreover, our controlled analysis uncovers key limitations in existing mitigation strategies that, while some methods yield improvements in specific aspects, few effectively address overall trustworthiness, and many introduce unexpected trade-offs that compromise model utility. These findings also provide practical insights for future improvements, such as the benefits of reasoning to better balance safety and performance. Based on these insights, we introduce a Reasoning-Enhanced Safety Alignment (RESA) approach that equips the model with chain-of-thought reasoning ability to discover the underlying risks, achieving state-of-the-art results.

Yao Huang, Yitong Sun, Yichi Zhang et al.

Despite the remarkable advances of Large Language Models (LLMs) across diverse cognitive tasks, the rapid enhancement of these capabilities also introduces emergent deceptive behaviors that may induce severe risks in high-stakes deployments. More critically, the characterization of deception across realistic real-world scenarios remains underexplored. To bridge this gap, we establish DeceptionBench, the first benchmark that systematically evaluates how deceptive tendencies manifest across different societal domains, what their intrinsic behavioral patterns are, and how extrinsic factors affect them. Specifically, on the static count, the benchmark encompasses 150 meticulously designed scenarios in five domains, i.e., Economy, Healthcare, Education, Social Interaction, and Entertainment, with over 1,000 samples, providing sufficient empirical foundations for deception analysis. On the intrinsic dimension, we explore whether models exhibit self-interested egoistic tendencies or sycophantic behaviors that prioritize user appeasement. On the extrinsic dimension, we investigate how contextual factors modulate deceptive outputs under neutral conditions, reward-based incentivization, and coercive pressures. Moreover, we incorporate sustained multi-turn interaction loops to construct a more realistic simulation of real-world feedback dynamics. Extensive experiments across LLMs and Large Reasoning Models (LRMs) reveal critical vulnerabilities, particularly amplified deception under reinforcement dynamics, demonstrating that current models lack robust resistance to manipulative contextual cues and the urgent need for advanced safeguards against various deception behaviors. Code and resources are publicly available at https://github.com/Aries-iai/DeceptionBench.

Xunlin Zhan, Yangxin Wu, Xiao Dong et al.

Nowadays, customer's demands for E-commerce are more diversified, which introduces more complications to the product retrieval industry. Previous methods are either subject to single-modal input or perform supervised image-level product retrieval, thus fail to accommodate real-life scenarios where enormous weakly annotated multi-modal data are present. In this paper, we investigate a more realistic setting that aims to perform weakly-supervised multi-modal instance-level product retrieval among fine-grained product categories. To promote the study of this challenging task, we contribute Product1M, one of the largest multi-modal cosmetic datasets for real-world instance-level retrieval. Notably, Product1M contains over 1 million image-caption pairs and consists of two sample types, i.e., single-product and multi-product samples, which encompass a wide variety of cosmetics brands. In addition to the great diversity, Product1M enjoys several appealing characteristics including fine-grained categories, complex combinations, and fuzzy correspondence that well mimic the real-world scenes. Moreover, we propose a novel model named Cross-modal contrAstive Product Transformer for instance-level prodUct REtrieval (CAPTURE), that excels in capturing the potential synergy between multi-modal inputs via a hybrid-stream transformer in a self-supervised manner.CAPTURE generates discriminative instance features via masked multi-modal learning as well as cross-modal contrastive pretraining and it outperforms several SOTA cross-modal baselines. Extensive ablation studies well demonstrate the effectiveness and the generalization capacity of our model. Dataset and codes are available at https: //github.com/zhanxlin/Product1M.

Zijun Yao, Chengjiang Li, Tiansi Dong et al.

Entity Matching (EM) aims at recognizing entity records that denote the same real-world object. Neural EM models learn vector representation of entity descriptions and match entities end-to-end. Though robust, these methods require many resources for training, and lack of interpretability. In this paper, we propose a novel EM framework that consists of Heterogeneous Information Fusion (HIF) and Key Attribute Tree (KAT) Induction to decouple feature representation from matching decision. Using self-supervised learning and mask mechanism in pre-trained language modeling, HIF learns the embeddings of noisy attribute values by inter-attribute attention with unlabeled data. Using a set of comparison features and a limited amount of annotated data, KAT Induction learns an efficient decision tree that can be interpreted by generating entity matching rules whose structure is advocated by domain experts. Experiments on 6 public datasets and 3 industrial datasets show that our method is highly efficient and outperforms SOTA EM models in most cases. Our codes and datasets can be obtained from https://github.com/THU-KEG/HIF-KAT.

Yichi Zhang, Ziqiao Ma, Xiaofeng Gao et al.

Most multimodal large language models (MLLMs) learn language-to-object grounding through causal language modeling where grounded objects are captured by bounding boxes as sequences of location tokens. This paradigm lacks pixel-level representations that are important for fine-grained visual understanding and diagnosis. In this work, we introduce GROUNDHOG, an MLLM developed by grounding Large Language Models to holistic segmentation. GROUNDHOG incorporates a masked feature extractor and converts extracted features into visual entity tokens for the MLLM backbone, which then connects groundable phrases to unified grounding masks by retrieving and merging the entity masks. To train GROUNDHOG, we carefully curated M3G2, a grounded visual instruction tuning dataset with Multi-Modal Multi-Grained Grounding, by harvesting a collection of segmentation-grounded datasets with rich annotations. Our experimental results show that GROUNDHOG achieves superior performance on various language grounding tasks without task-specific fine-tuning, and significantly reduces object hallucination. GROUNDHOG also demonstrates better grounding towards complex forms of visual input and provides easy-to-understand diagnosis in failure cases.

Yichi Zhang, Yinpeng Dong, Siyuan Zhang et al.

Although Multimodal Large Language Models (MLLMs) have demonstrated promising versatile capabilities, their performance is still inferior to specialized models on downstream tasks, which makes adaptation necessary to enhance their utility. However, fine-tuning methods require independent training for every model, leading to huge computation and memory overheads. In this paper, we propose a novel setting where we aim to improve the performance of diverse MLLMs with a group of shared parameters optimized for a downstream task. To achieve this, we propose Transferable Visual Prompting (TVP), a simple and effective approach to generate visual prompts that can transfer to different models and improve their performance on downstream tasks after trained on only one model. We introduce two strategies to address the issue of cross-model feature corruption of existing visual prompting methods and enhance the transferability of the learned prompts, including 1) Feature Consistency Alignment: which imposes constraints to the prompted feature changes to maintain task-agnostic knowledge; 2) Task Semantics Enrichment: which encourages the prompted images to contain richer task-specific semantics with language guidance. We validate the effectiveness of TVP through extensive experiments with 6 modern MLLMs on a wide variety of tasks ranging from object recognition and counting to multimodal reasoning and hallucination correction.

Haozhe Zhao, Shuzheng Si, Liang Chen et al.

Large vision-language models (LVLMs) have achieved impressive results in various vision-language tasks. However, despite showing promising performance, LVLMs suffer from hallucinations caused by language bias, leading to diminished focus on images and ineffective visual comprehension. We identify two primary reasons for this bias: 1. Different scales of training data between the pretraining stage of LLM and multimodal alignment stage. 2. The learned inference bias due to short-term dependency of text data. Therefore, we propose LACING, a systemic framework designed to address the language bias of LVLMs with muLtimodal duAl-attention meChanIsm (MDA) aNd soft-image Guidance (IFG). Specifically, MDA introduces a parallel dual-attention mechanism that enhances the integration of visual inputs across the model. IFG introduces a learnable soft visual prompt during training and inference to replace visual inputs, designed to compel LVLMs to prioritize text inputs. Then, IFG further proposes a novel decoding strategy using the soft visual prompt to mitigate the model's over-reliance on adjacent text inputs. Comprehensive experiments demonstrate that our method effectively debiases LVLMs from their language bias, enhancing visual comprehension and reducing hallucinations without requiring additional training resources or data. The code and model are available at [lacing-lvlm.github.io](https://lacing-lvlm.github.io).

Yichi Zhang, Zhuo Chen, Lingbing Guo et al.

Large language models (LLMs) have demonstrated exceptional performance in text generation within current NLP research. However, the lack of factual accuracy is still a dark cloud hanging over the LLM skyscraper. Structural knowledge prompting (SKP) is a prominent paradigm to integrate external knowledge into LLMs by incorporating structural representations, achieving state-of-the-art results in many knowledge-intensive tasks. However, existing methods often focus on specific problems, lacking a comprehensive exploration of the generalization and capability boundaries of SKP. This paper aims to evaluate and rethink the generalization capability of the SKP paradigm from four perspectives including Granularity, Transferability, Scalability, and Universality. To provide a thorough evaluation, we introduce a novel multi-granular, multi-level benchmark called SUBARU, consisting of 9 different tasks with varying levels of granularity and difficulty.

Lingbing Guo, Yichi Zhang, Zhongpu Bo et al.

Recent advancements in large language models (LLMs) have significantly improved various natural language processing (NLP) tasks. Typically, LLMs are trained to predict the next token, aligning well with many NLP tasks. However, in knowledge graph (KG) scenarios, entities are the fundamental units and identifying an entity requires at least several tokens. This leads to a granularity mismatch between KGs and natural languages. To address this issue, we propose K-ON, which integrates KG knowledge into the LLM by employing multiple head layers for next k-step prediction. K-ON can not only generate entity-level results in one step, but also enables contrastive loss against entities, which is the most powerful tool in KG representation learning. Experimental results show that K-ON outperforms state-of-the-art methods that incorporate text and even the other modalities.

Yichi Zhang, Alexander Belloni, Ethan X. Fang et al.

Motivated by the need for rigorous and scalable evaluation of large language models, we study contextual preference inference for pairwise comparison functionals of context-dependent preference score functions across domains. Focusing on the contextual Bradley-Terry-Luce model, we develop a semiparametric efficient estimator that automates the debiased estimation through aggregating weighted residual balancing terms across the comparison graph. We show that the efficiency is achieved when the weights are derived from a novel strategy called Fisher random walk. We also propose a computationally feasible method to compute the weights by a potential representation of nuisance weight functions. We show our inference procedure is valid for general score function estimators accommodating the practitioners' need to implement flexible deep learning methods. We extend the procedure to multiple hypothesis testing using a Gaussian multiplier bootstrap that controls familywise error and to distributional shift via a cross-fitted importance-sampling adjustment for target-domain inference. Numerical studies, including language model evaluations under diverse contexts, corroborate the accuracy, efficiency, and practical utility of our method.

Yichi Zhang, Yao Huang, Yitong Sun et al.

Despite the superior capabilities of Multimodal Large Language Models (MLLMs) across diverse tasks, they still face significant trustworthiness challenges. Yet, current literature on the assessment of trustworthy MLLMs remains limited, lacking a holistic evaluation to offer thorough insights into future improvements. In this work, we establish MultiTrust, the first comprehensive and unified benchmark on the trustworthiness of MLLMs across five primary aspects: truthfulness, safety, robustness, fairness, and privacy. Our benchmark employs a rigorous evaluation strategy that addresses both multimodal risks and cross-modal impacts, encompassing 32 diverse tasks with self-curated datasets. Extensive experiments with 21 modern MLLMs reveal some previously unexplored trustworthiness issues and risks, highlighting the complexities introduced by the multimodality and underscoring the necessity for advanced methodologies to enhance their reliability. For instance, typical proprietary models still struggle with the perception of visually confusing images and are vulnerable to multimodal jailbreaking and adversarial attacks; MLLMs are more inclined to disclose privacy in text and reveal ideological and cultural biases even when paired with irrelevant images in inference, indicating that the multimodality amplifies the internal risks from base LLMs. Additionally, we release a scalable toolbox for standardized trustworthiness research, aiming to facilitate future advancements in this important field. Code and resources are publicly available at: https://multi-trust.github.io/.

Shane Storks, Qiaozi Gao, Yichi Zhang et al.

Large-scale, pre-trained language models (LMs) have achieved human-level performance on a breadth of language understanding tasks. However, evaluations only based on end task performance shed little light on machines' true ability in language understanding and reasoning. In this paper, we highlight the importance of evaluating the underlying reasoning process in addition to end performance. Toward this goal, we introduce Tiered Reasoning for Intuitive Physics (TRIP), a novel commonsense reasoning dataset with dense annotations that enable multi-tiered evaluation of machines' reasoning process. Our empirical results show that while large LMs can achieve high end performance, they struggle to support their predictions with valid supporting evidence. The TRIP dataset and our baseline results will motivate verifiable evaluation of commonsense reasoning and facilitate future research toward developing better language understanding and reasoning models.

Yichi Zhang, Joyce Chai

Despite recent progress, learning new tasks through language instructions remains an extremely challenging problem. On the ALFRED benchmark for task learning, the published state-of-the-art system only achieves a task success rate of less than 10% in an unseen environment, compared to the human performance of over 90%. To address this issue, this paper takes a closer look at task learning. In a departure from a widely applied end-to-end architecture, we decomposed task learning into three sub-problems: sub-goal planning, scene navigation, and object manipulation; and developed a model HiTUT (stands for Hierarchical Tasks via Unified Transformers) that addresses each sub-problem in a unified manner to learn a hierarchical task structure. On the ALFRED benchmark, HiTUT has achieved the best performance with a remarkably higher generalization ability. In the unseen environment, HiTUT achieves over 160% performance gain in success rate compared to the previous state of the art. The explicit representation of task structures also enables an in-depth understanding of the nature of the problem and the ability of the agent, which provides insight for future benchmark development and evaluation.

Yichi Zhang, Zhijian Ou, Huixin Wang et al.

Structured belief states are crucial for user goal tracking and database query in task-oriented dialog systems. However, training belief trackers often requires expensive turn-level annotations of every user utterance. In this paper we aim at alleviating the reliance on belief state labels in building end-to-end dialog systems, by leveraging unlabeled dialog data towards semi-supervised learning. We propose a probabilistic dialog model, called the LAtent BElief State (LABES) model, where belief states are represented as discrete latent variables and jointly modeled with system responses given user inputs. Such latent variable modeling enables us to develop semi-supervised learning under the principled variational learning framework. Furthermore, we introduce LABES-S2S, which is a copy-augmented Seq2Seq model instantiation of LABES. In supervised experiments, LABES-S2S obtains strong results on three benchmark datasets of different scales. In utilizing unlabeled dialog data, semi-supervised LABES-S2S significantly outperforms both supervised-only and semi-supervised baselines. Remarkably, we can reduce the annotation demands to 50% without performance loss on MultiWOZ.

Silin Gao, Yichi Zhang, Zhijian Ou et al.

Neural generative models have achieved promising performance on dialog generation tasks if given a huge data set. However, the lack of high-quality dialog data and the expensive data annotation process greatly limit their application in real-world settings. We propose a paraphrase augmented response generation (PARG) framework that jointly trains a paraphrase model and a response generation model to improve the dialog generation performance. We also design a method to automatically construct paraphrase training data set based on dialog state and dialog act labels. PARG is applicable to various dialog generation models, such as TSCP (Lei et al., 2018) and DAMD (Zhang et al., 2019). Experimental results show that the proposed framework improves these state-of-the-art dialog models further on CamRest676 and MultiWOZ. PARG also significantly outperforms other data augmentation methods in dialog generation tasks, especially under low resource settings.

Yichi Zhang, Zhijian Ou, Zhou Yu

Conversations have an intrinsic one-to-many property, which means that multiple responses can be appropriate for the same dialog context. In task-oriented dialogs, this property leads to different valid dialog policies towards task completion. However, none of the existing task-oriented dialog generation approaches takes this property into account. We propose a Multi-Action Data Augmentation (MADA) framework to utilize the one-to-many property to generate diverse appropriate dialog responses. Specifically, we first use dialog states to summarize the dialog history, and then discover all possible mappings from every dialog state to its different valid system actions. During dialog system training, we enable the current dialog state to map to all valid system actions discovered in the previous process to create additional state-action pairs. By incorporating these additional pairs, the dialog policy learns a balanced action distribution, which further guides the dialog model to generate diverse responses. Experimental results show that the proposed framework consistently improves dialog policy diversity, and results in improved response diversity and appropriateness. Our model obtains state-of-the-art results on MultiWOZ.

Qingyang Wu, Yichi Zhang, Yu Li et al.

Existing dialog system models require extensive human annotations and are difficult to generalize to different tasks. The recent success of large pre-trained language models such as BERT and GPT-2 (Devlin et al., 2019; Radford et al., 2019) have suggested the effectiveness of incorporating language priors in down-stream NLP tasks. However, how much pre-trained language models can help dialog response generation is still under exploration. In this paper, we propose a simple, general, and effective framework: Alternating Roles Dialog Model (ARDM). ARDM models each speaker separately and takes advantage of the large pre-trained language model. It requires no supervision from human annotations such as belief states or dialog acts to achieve effective conversations. ARDM outperforms or is on par with state-of-the-art methods on two popular task-oriented dialog datasets: CamRest676 and MultiWOZ. Moreover, we can generalize ARDM to more challenging, non-collaborative tasks such as persuasion. In persuasion tasks, ARDM is capable of generating human-like responses to persuade people to donate to a charity.

Yinpei Dai, Yichi Zhang, Hong Liu et al.

Slot filling is a crucial component in task-oriented dialog systems that is used to parse (user) utterances into semantic concepts called slots. An ontology is defined by the collection of slots and the values that each slot can take. The most widely used practice of treating slot filling as a sequence labeling task suffers from two main drawbacks. First, the ontology is usually pre-defined and fixed and therefore is not able to detect new labels for unseen slots. Second, the one-hot encoding of slot labels ignores the correlations between slots with similar semantics, which makes it difficult to share knowledge learned across different domains. To address these problems, we propose a new model called elastic conditional random field (eCRF), where each slot is represented by the embedding of its natural language description and modeled by a CRF layer. New slot values can be detected by eCRF whenever a language description is available for the slot. In our experiment, we show that eCRFs outperform existing models in both in-domain and cross-domain tasks, especially in predicting unseen slots and values.