Large language models (LLMs) are still struggling in aligning with human preference in complex tasks and scenarios. They are prone to overfit into the unexpected patterns or superficial styles in the training data. We conduct an empirical study that only selects the top-10% most updated parameters in LLMs for alignment training, and see improvements in the convergence process and final performance. It indicates the existence of redundant neurons in LLMs for alignment training. To reduce its influence, we propose a low-redundant alignment method named **ALLO**, focusing on optimizing the most related neurons with the most useful supervised signals. Concretely, we first identify the neurons that are related to the human preference data by a gradient-based strategy, then identify the alignment-related key tokens by reward models for computing loss. Besides, we also decompose the alignment process into the forgetting and learning stages, where we first forget the tokens with unaligned knowledge and then learn aligned knowledge, by updating different ratios of neurons, respectively. Experimental results on 10 datasets have shown the effectiveness of ALLO. Our code and data will be publicly released.

Reinforcement learning (RL) has been widely used in training large language models (LLMs) for preventing unexpected outputs, e.g., reducing harmfulness and errors. However, existing RL methods mainly adopt instance-level reward, which cannot provide fine-grained supervision for complex reasoning tasks. As a result, the RL training cannot be fully aware of the specific part or step that actually leads to the incorrectness in model response. To address it, we propose a new RL method named RLMEC that incorporates a generative model as the reward model, which is trained by the erroneous solution rewriting task under the minimum editing constraint, which can produce token-level supervision for RL training. Based 0on the generative reward model, we design the token-level RL objective for training and an imitation-based regularization for stabilizing RL process. And these two objectives focus on the revision of the key tokens for the erroneous solution, reducing the effect of other unimportant tokens. Experiment results on 8 tasks have demonstrated the effectiveness of our approach. Our code and data will be publicly released.

Recently, multi-task instruction tuning has been utilized to improve sentence representation learning (SRL). It enables SRL models to generate task-specific representations with the guidance of task instruction, thus exhibiting strong generalization ability on unseen tasks. However, these methods mostly neglect the potential interference problems across different tasks and instances, which may affect the training of the model.To address this issue, we propose a data curriculum method, namely **Data-CUBE**, that arranges the order of all the multi-task data for training, to minimize the interference risks from two aspects.At the task level, we aim to find the optimal task order to minimize the total cross-task interference risk and formulate this problem as the traveling salesman problem, which is further solved by a specially designed simulated annealing algorithm. At the instance level, we propose a measurement method to quantify the difficulty of all instances per task, and then arrange instances in an easy-to-difficult order for training.Experimental results show that our approach can boost the performance of state-of-the-art methods. Our code and data will be publicly released.

Recently, continuous diffusion models (CDM) have been introduced into non-autoregressive (NAR) text-to-text generation. However, the discrete nature of text increases the difficulty of CDM to generate coherent and fluent texts, and also causes the incompatibility problem between CDM and advanced NLP techniques, especially the popular pre-trained language models (PLMs).To solve it, we propose Diffusion-NAT, which introduces discrete diffusion models (DDM) into NAR text-to-text generation and integrates BART to improve the performance.By revising the decoding process of BART and the typical settings of DDM, we unify the inference process of BART and the denoising process of DDM into the same NAR masked tokens recovering task.In this way, DDM can rely on BART to perform denoising, which can benefit from both the rich pre-learned knowledge of BART and the iterative refining paradigm of DDM.Besides, we also propose the iterative self-prompting strategy to further improve the generation quality.Experimental results on 7 datasets show that our approach can outperform competitive NAR methods, and even surpass autoregressive methods.Our code and data are released at https://github.com/RUCAIBox/DiffusionNAT.

Humans often interact with large language models (LLMs) in multi-turn interaction to obtain desired answers or more information. However, most existing studies overlook the multi-turn instruction following ability of LLMs, in terms of training dataset, training method, and evaluation benchmark. In this paper, we introduce Parrot, a solution aiming to enhance multi-turn instruction following for LLMs. First, we introduce an efficient but effective method for collecting multi-turn instructions that feature human-like queries, such as anaphora and ellipsis. Second, we propose a context-aware preference optimization strategy to further enhance LLMs for complex queries in multi-turn interaction. Moreover, to quantitatively evaluate LLMs in multi-turn instruction following, we manually build a multi-turn benchmark derived from existing ones. Extensive experiments show that Parrot improves current LLMs by up to 7.2% in multi-turn instruction following. Our dataset and codes will be open-sourced to facilitate future research.

Large Language Models (LLMs) have shown remarkable capabilities in tasks such as summarization, arithmetic reasoning, and question answering. However, they encounter significant challenges in the domain of moral reasoning and ethical decision-making, especially in complex scenarios with multiple stakeholders. This paper introduces the Skin-in-the-Game (SKIG) framework, aimed at enhancing moral reasoning in LLMs by exploring decisions’ consequences from multiple stakeholder perspectives. The core components of the framework consist of simulating accountability for decisions, conducting empathy exercises on different stakeholders, and evaluating the risks associated with the impacts of potential actions. We study SKIG’s performance across various moral reasoning benchmarks with proprietary and open-source LLMs, and investigate its crucial components through extensive ablation analyses. Our framework exhibits marked improvements in performance compared to baselines across different language models and benchmarks.

To facilitate the research on large language models (LLMs), this paper presents a comprehensive and unified library, LLMBox, to ease the development, use, and evaluation of LLMs. This library is featured with three main merits: (1) a unified data interface that supports the flexible implementation of various training strategies, (2) a comprehensive evaluation that covers extensive tasks, datasets, and models, and (3) more practical consideration, especially on user-friendliness and efficiency. With our library, users can easily reproduce existing methods, train new models, and conduct comprehensive performance comparisons. To rigorously test LLMBox, we conduct extensive experiments in a diverse coverage of evaluation settings, and experimental results demonstrate the effectiveness and efficiency of our library in supporting various implementations related to LLMs. The detailed introduction and usage guidance can be found at https://github.com/RUCAIBox/LLMBox.

By scaling the model size, large pre-trained language models (PLMs) have shown remarkable performance in various natural language processing tasks, mostly outperforming small PLMs by a large margin. However, due to the high computational cost, the huge number of parameters also restricts the applicability of large PLMs in real-world systems. In this paper, we focus on scaling up the parameters of PLMs only during fine-tuning, to benefit from the over-parameterization, while without increasing the inference latency. Given a relatively small PLM, we over-parameterize it by employing a matrix product operator, an efficient and almost lossless decomposition method to factorize its contained parameter matrices into a set of higher-dimensional tensors.Considering the efficiency, we further propose both static and dynamic strategies to select the most important parameter matrices for over-parameterization.Extensive experiments have demonstrated that our approach can significantly boost the fine-tuning performance of small PLMs and even help small PLMs outperform 3× parameterized larger ones.Our code is publicly available at https://github.com/zfgao66/OPF.

Although pre-trained language models (PLMs) have shown impressive performance by text-only self-supervised training, they are found lack of visual semantics or commonsense. Existing solutions often rely on explicit images for visual knowledge augmentation (requiring time-consuming retrieval or generation), and they also conduct the augmentation for the whole input text, without considering whether it is actually needed in specific inputs or tasks. To address these issues, we propose a novel **V**isually-**A**ugmented fine-tuning approach that can be generally applied to various PLMs or NLP tasks, **W**ithout using any retrieved or generated **I**mages, namely **VAWI**. Experimental results show that our approach can consistently improve the performance of BERT, RoBERTa, BART, and T5 at different scales, and outperform several competitive baselines on ten tasks. Our codes and data are publicly available at https://github.com/RUCAIBox/VAWI.

Although large language models (LLMs) have achieved excellent performance in a variety of evaluation benchmarks, they still struggle in complex reasoning tasks which require specific knowledge and multi-hop reasoning. To improve the reasoning abilities, we propose ChatCoT, a tool-augmented chain-of-thought reasoning framework for chat-based LLMs (e.g., ChatGPT). In ChatCoT, we model the chain-of-thought (CoT) reasoning as multi-turn conversations, to utilize tools in a more natural way through chatting. At each turn, LLMs can either interact with tools or perform the reasoning. Our approach can effectively leverage the multi-turn conversation ability of chat-based LLMs, and integrate the thought chain following and tools manipulation in a unified way. Specially, we initialize the early turns of the conversation by the knowledge about tools, tasks, and reasoning format, and propose an iterative tool-augmented reasoning step to perform step-by-step tool-augmented reasoning. The experiment results on two complex reasoning datasets (MATH and HotpotQA) have shown the effectiveness of ChatCoT on complex reasoning tasks, achieving a 7.9% relative improvement over the state-of-the-art baseline.

Inspired by the superior language abilities of large language models (LLM), large vision-language models (LVLM) have been recently proposed by integrating powerful LLMs for improving the performance on complex multimodal tasks. Despite the promising progress on LVLMs, we find that they suffer from object hallucinations, i.e., they tend to generate objects inconsistent with the target images in the descriptions. To investigate it, this work presents the first systematic study on object hallucination of LVLMs. We conduct the evaluation experiments on several representative LVLMs, and show that they mostly suffer from severe object hallucination issues. We further discuss that the visual instructions may influence the hallucination, and find that: objects that frequently appear in the visual instructions or co-occur with the image objects are obviously prone to be hallucinated by LVLMs. Besides, we further design a polling-based query method called POPE for better evaluation of object hallucination. Experiment results show that our POPE can evaluate object hallucination in a more stable and flexible way.

Question Answering over Knowledge Graph (KGQA) aims to seek answer entities for the natural language question from a large-scale Knowledge Graph (KG). To better perform reasoning on KG, recent work typically adopts a pre-trained language model (PLM) to model the question, and a graph neural network (GNN) based module to perform multi-hop reasoning on the KG. Despite the effectiveness, due to the divergence in model architecture, the PLM and GNN are not closely integrated, limiting the knowledge sharing and fine-grained feature interactions. To solve it, we aim to simplify the above two-module approach, and develop a more capable PLM that can directly support subgraph reasoning for KGQA, namely ReasoningLM. In our approach, we propose a subgraph-aware self-attention mechanism to imitate the GNN for performing structured reasoning, and also adopt an adaptation tuning strategy to adapt the model parameters with 20,000 subgraphs with synthesized questions. After adaptation, the PLM can be parameter-efficient fine-tuned on downstream tasks. Experiments show that ReasoningLM surpasses state-of-the-art models by a large margin, even with fewer updated parameters and less training data. Our codes and data are publicly available at https://github.com/RUCAIBox/ReasoningLM.

In this paper, we aim to improve the reasoning ability of large language models (LLMs) over structured data in a unified way. Inspired by the studies on tool augmentation for LLMs, we develop an Iterative Reading-then-Reasoning (IRR) framework to solve question answering tasks based on structured data, called StructGPT. In this framework, we construct the specialized interfaces to collect relevant evidence from structured data (i.e., reading), and let LLMs concentrate on the reasoning task based on the collected information (i.e., reasoning). Specially, we propose an invoking-linearization-generation procedure to support LLMs in reasoning on the structured data with the help of the interfaces. By iterating this procedure with provided interfaces, our approach can gradually approach the target answers to a given query. Experiments conducted on three types of structured data show that StructGPT greatly improves the performance of LLMs, under the few-shot and zero-shot settings.

We investigate and refine denoising methods for NER task on data that potentially contains extremely noisy labels from multi-sources. In this paper, we first summarized all possible noise types and noise generation schemes, based on which we built a thorough evaluation system. We then pinpoint the bottleneck of current state-of-art denoising methods using our evaluation system. Correspondingly, we propose several refinements, including using a two-stage framework to avoid error accumulation; a novel confidence score utilizing minimal clean supervision to increase predictive power; an automatic cutoff fitting to save extensive hyper-parameter tuning; a warm started weighted partial CRF to better learn on the noisy tokens. Additionally, we propose to use adaptive sampling to further boost the performance in long-tailed entity settings. Our method improves F1 score by on average at least 5 10% over current state-of-art across extensive experiments.

In this paper, we study how to continually pre-train language models for improving the understanding of math problems. Specifically, we focus on solving a fundamental challenge in modeling math problems, how to fuse the semantics of textual description and formulas, which are highly different in essence. To address this issue, we propose a new approach called COMUS to continually pre-train language models for math problem understanding with syntax-aware memory network. In this approach, we first construct the math syntax graph to model the structural semantic information, by combining the parsing trees of the text and formulas, and then design the syntax-aware memory networks to deeply fuse the features from the graph and text. With the help of syntax relations, we can model the interaction between the token from the text and its semantic-related nodes within the formulas, which is helpful to capture fine-grained semantic correlations between texts and formulas. Besides, we devise three continual pre-training tasks to further align and fuse the representations of the text and math syntax graph. Experimental results on four tasks in the math domain demonstrate the effectiveness of our approach. Our code and data are publicly available at the link: bluehttps://github.com/RUCAIBox/COMUS.

Recently, contrastive learning has been shown to be effective in improving pre-trained language models (PLM) to derive high-quality sentence representations. It aims to pull close positive examples to enhance the alignment while push apart irrelevant negatives for the uniformity of the whole representation space. However, previous works mostly adopt in-batch negatives or sample from training data at random. Such a way may cause the sampling bias that improper negatives (false negatives and anisotropy representations) are used to learn sentence representations, which will hurt the uniformity of the representation space. To address it, we present a new framework DCLR (Debiased Contrastive Learning of unsupervised sentence Representations) to alleviate the influence of these improper negatives.In DCLR, we design an instance weighting method to punish false negatives and generate noise-based negatives to guarantee the uniformity of the representation space.Experiments on seven semantic textual similarity tasks show that our approach is more effective than competitive baselines. Our code and data are publicly available at the link: bluehttps://github.com/RUCAIBox/DCLR.

Sampling proper negatives from a large document pool is vital to effectively train a dense retrieval model. However, existing negative sampling strategies suffer from the uninformative or false negative problem. In this work, we empirically show that according to the measured relevance scores, the negatives ranked around the positives are generally more informative and less likely to be false negatives. Intuitively, these negatives are not too hard (may be false negatives) or too easy (uninformative). They are the ambiguous negatives and need more attention during training.Thus, we propose a simple ambiguous negatives sampling method, SimANS, which incorporates a new sampling probability distribution to sample more ambiguous negatives.Extensive experiments on four public and one industry datasets show the effectiveness of our approach.We made the code and models publicly available in https://github.com/microsoft/SimXNS.

Commonsense reasoning in natural language is a desired ability of artificial intelligent systems. For solving complex commonsense reasoning tasks, a typical solution is to enhance pre-trained language models (PTMs) with a knowledge-aware graph neural network (GNN) encoder that models a commonsense knowledge graph (CSKG).Despite the effectiveness, these approaches are built on heavy architectures, and can’t clearly explain how external knowledge resources improve the reasoning capacity of PTMs. Considering this issue, we conduct a deep empirical analysis, and find that it is indeed relation features from CSKGs (but not node features) that mainly contribute to the performance improvement of PTMs. Based on this finding, we design a simple MLP-based knowledge encoder that utilizes statistical relation paths as features. Extensive experiments conducted on five benchmarks demonstrate the effectiveness of our approach, which also largely reduces the parameters for encoding CSKGs.Our codes and data are publicly available at https://github.com/RUCAIBox/SAFE.

In recent years, conversational recommender systems (CRSs) have drawn a wide attention in the research community, which focus on providing high-quality recommendations to users via natural language conversations. However, due to diverse scenarios and data formats, existing studies on CRSs lack unified and standardized implementation or comparison. To tackle this challenge, we release an open-source toolkit CRSLab, which provides a unified and extensible framework with highly-decoupled modules to develop CRSs. Based on this framework, we collect 6 commonly used human-annotated CRS datasets and implement 19 models that include advanced techniques such as graph neural networks and pre-training models. Besides, our toolkit provides a series of automatic evaluation protocols and a human-machine interaction interface to evaluate and compare different CRS methods. The project and documents are released at https://github.com/RUCAIBox/CRSLab.

Recent works have shown that powerful pre-trained language models (PLM) can be fooled by small perturbations or intentional attacks. To solve this issue, various data augmentation techniques are proposed to improve the robustness of PLMs. However, it is still challenging to augment semantically relevant examples with sufficient diversity. In this work, we present Virtual Data Augmentation (VDA), a general framework for robustly fine-tuning PLMs. Based on the original token embeddings, we construct a multinomial mixture for augmenting virtual data embeddings, where a masked language model guarantees the semantic relevance and the Gaussian noise provides the augmentation diversity. Furthermore, a regularized training strategy is proposed to balance the two aspects. Extensive experiments on six datasets show that our approach is able to improve the robustness of PLMs and alleviate the performance degradation under adversarial attacks. Our codes and data are publicly available at bluehttps://github.com/RUCAIBox/VDA.

Weakly supervised machine reading comprehension (MRC) task is practical and promising for its easily available and massive training data, but inevitablely introduces noise. Existing related methods usually incorporate extra submodels to help filter noise before the noisy data is input to main models. However, these multistage methods often make training difficult, and the qualities of submodels are hard to be controlled. In this paper, we first explore and analyze the essential characteristics of noise from the perspective of loss distribution, and find that in the early stage of training, noisy samples usually lead to significantly larger loss values than clean ones. Based on the observation, we propose a hierarchical loss correction strategy to avoid fitting noise and enhance clean supervision signals, including using an unsupervisedly fitted Gaussian mixture model to calculate the weight factors for all losses to correct the loss distribution, and employ a hard bootstrapping loss to modify loss function. Experimental results on different weakly supervised MRC datasets show that the proposed methods can help improve models significantly.

Conversational recommender systems (CRS) aim to recommend high-quality items to users through interactive conversations. To develop an effective CRS, the support of high-quality datasets is essential. Existing CRS datasets mainly focus on immediate requests from users, while lack proactive guidance to the recommendation scenario. In this paper, we contribute a new CRS dataset named TG-ReDial (Recommendation through Topic-Guided Dialog). Our dataset has two major features. First, it incorporates topic threads to enforce natural semantic transitions towards the recommendation scenario. Second, it is created in a semi-automatic way, hence human annotation is more reasonable and controllable. Based on TG-ReDial, we present the task of topic-guided conversational recommendation, and propose an effective approach to this task. Extensive experiments have demonstrated the effectiveness of our approach on three sub-tasks, namely topic prediction, item recommendation and response generation. TG-ReDial is available at bluehttps://github.com/RUCAIBox/TG-ReDial.

Context modeling has a pivotal role in open domain conversation. Existing works either use heuristic methods or jointly learn context modeling and response generation with an encoder-decoder framework. This paper proposes an explicit context rewriting method, which rewrites the last utterance by considering context history. We leverage pseudo-parallel data and elaborate a context rewriting network, which is built upon the CopyNet with the reinforcement learning method. The rewritten utterance is beneficial to candidate retrieval, explainable context modeling, as well as enabling to employ a single-turn framework to the multi-turn scenario. The empirical results show that our model outperforms baselines in terms of the rewriting quality, the multi-turn response generation, and the end-to-end retrieval-based chatbots.