Guanting Dong

Guanting Dong, Hongyi Yuan, Keming Lu et al. · tsinghua

Large language models (LLMs) with enormous pre-training tokens and parameters emerge diverse abilities, including math reasoning, code generation, and instruction following. These abilities are further enhanced by supervised fine-tuning (SFT). While the open-source community has explored ad-hoc SFT for enhancing individual capabilities, proprietary LLMs exhibit versatility across various skills. Therefore, understanding the facilitation of multiple abilities via SFT is paramount. In this study, we specifically focuses on the interplay of data composition between mathematical reasoning, code generation, and general human-aligning abilities during SFT. We propose four intriguing research questions to explore the association between model performance and various factors including data amount, composition ratio, model size and SFT strategies. Our experiments reveal that distinct capabilities scale differently and larger models generally show superior performance with same amount of data. Mathematical reasoning and code generation consistently improve with increasing data amount, whereas general abilities plateau after roughly a thousand samples. Moreover, we observe data composition appears to enhance various abilities under limited data conditions, yet can lead to performance conflicts when data is plentiful. Our findings also suggest the amount of composition data influences performance more than the composition ratio. In analysis of SFT strategies, we find that sequentially learning multiple skills risks catastrophic forgetting. Our proposed Dual-stage Mixed Fine-tuning (DMT) strategy offers a promising solution to learn multiple abilities with different scaling patterns.

Chengpeng Li, Zheng Yuan, Hongyi Yuan et al. · tsinghua

In math reasoning with large language models (LLMs), fine-tuning data augmentation by query evolution and diverse reasoning paths is empirically verified effective, profoundly narrowing the gap between open-sourced LLMs and cutting-edge proprietary LLMs. In this paper, we conduct an investigation for such data augmentation in math reasoning and are intended to answer: (1) What strategies of data augmentation are more effective; (2) What is the scaling relationship between the amount of augmented data and model performance; and (3) Can data augmentation incentivize generalization to out-of-domain mathematical reasoning tasks? To this end, we create two new dataset AugGSM8K and AugMATH, by complicating and diversifying the queries and sampling multiple reasoning paths from GSM8K and MATH. We obtained a series of LLMs called MuggleMath by fine-tuning LLaMA models on AugGSM8K and AugMATH. MuggleMath substantially achieves new state-of-the-art on GSM8K and MATH. A log-linear relationship and a segmented log-linear are presented between MuggleMath's performance and the amount of augmented data on GSM8K and MATH, respectively. We also find that it is weak in out-of-domain math reasoning generalization from AugGSM8K to MATH and from AugMATH to GSM8K, which suggests that augmenting queries that cover a broader range of subjects is more beneficial for generalization. We release our codes and augmented data in https://github.com/OFA-Sys/gsm8k-ScRel.

Zheng Yuan, Hongyi Yuan, Chengpeng Li et al. · tsinghua

Mathematical reasoning is a challenging task for large language models (LLMs), while the scaling relationship of it with respect to LLM capacity is under-explored. In this paper, we investigate how the pre-training loss, supervised data amount, and augmented data amount influence the reasoning performances of a supervised LLM. We find that pre-training loss is a better indicator of the model's performance than the model's parameter count. We apply supervised fine-tuning (SFT) with different amounts of supervised data and empirically find a log-linear relation between data amount and model performance, and we find better models improve less with enlarged supervised datasets. To augment more data samples for improving model performances without any human effort, we propose to apply Rejection sampling Fine-Tuning (RFT). RFT uses supervised models to generate and collect correct reasoning paths as augmented fine-tuning datasets. We find with augmented samples containing more distinct reasoning paths, RFT improves mathematical reasoning performance more for LLMs. We also find RFT brings more improvement for less performant LLMs. Furthermore, we combine rejection samples from multiple models which push LLaMA-7B to an accuracy of 49.3\% on GSM8K which outperforms the supervised fine-tuning (SFT) accuracy of 35.9\% significantly.

Runqi Qiao, Qiuna Tan, Guanting Dong et al.

Visual mathematical reasoning, as a fundamental visual reasoning ability, has received widespread attention from the Large Multimodal Models (LMMs) community. Existing benchmarks, such as MathVista and MathVerse, focus more on the result-oriented performance but neglect the underlying principles in knowledge acquisition and generalization. Inspired by human-like mathematical reasoning, we introduce WE-MATH, the first benchmark specifically designed to explore the problem-solving principles beyond end-to-end performance. We meticulously collect and categorize 6.5K visual math problems, spanning 67 hierarchical knowledge concepts and five layers of knowledge granularity. We decompose composite problems into sub-problems according to the required knowledge concepts and introduce a novel four-dimensional metric, namely Insufficient Knowledge (IK), Inadequate Generalization (IG), Complete Mastery (CM), and Rote Memorization (RM), to hierarchically assess inherent issues in LMMs' reasoning process. With WE-MATH, we conduct a thorough evaluation of existing LMMs in visual mathematical reasoning and reveal a negative correlation between solving steps and problem-specific performance. We confirm the IK issue of LMMs can be effectively improved via knowledge augmentation strategies. More notably, the primary challenge of GPT-4o has significantly transitioned from IK to IG, establishing it as the first LMM advancing towards the knowledge generalization stage. In contrast, other LMMs exhibit a marked inclination towards Rote Memorization - they correctly solve composite problems involving multiple knowledge concepts yet fail to answer sub-problems. We anticipate that WE-MATH will open new pathways for advancements in visual mathematical reasoning for LMMs. The WE-MATH data and evaluation code are available at https://github.com/We-Math/We-Math.

Chengpeng Li, Guanting Dong, Mingfeng Xue et al.

Large language models (LLMs) have made impressive progress in handling simple math problems, yet they still struggle with more challenging and complex mathematical tasks. In this paper, we introduce a series of LLMs that employs the Decomposition of thought with code assistance and self-correction for mathematical reasoning, dubbed as DotaMath. DotaMath models tackle complex mathematical tasks by decomposing them into simpler logical subtasks, leveraging code to solve these subtasks, obtaining fine-grained feedback from the code interpreter, and engaging in self-reflection and correction. By annotating diverse interactive tool-use trajectories and employing query evolution on GSM8K and MATH datasets, we generate an instruction fine-tuning dataset called DotaMathQA with 574K query-response pairs. We train a series of base LLMs using imitation learning on DotaMathQA, resulting in DotaMath models that achieve remarkable performance compared to open-source LLMs across various in-domain and out-of-domain benchmarks. Notably, DotaMath-deepseek-7B showcases an outstanding performance of 64.8% on the competitive MATH dataset and 86.7% on GSM8K. Besides, DotaMath-deepseek-7B maintains strong competitiveness on a series of in-domain and out-of-domain benchmarks (Avg. 80.1%). Looking forward, we anticipate that the DotaMath paradigm will open new pathways for addressing intricate mathematical problems. Our code is publicly available at https://github.com/ChengpengLi1003/DotaMath.

Guanting Dong, Jinxu Zhao, Tingfeng Hui et al.

With the increasing capabilities of large language models (LLMs), these high-performance models have achieved state-of-the-art results on a wide range of natural language processing (NLP) tasks. However, the models' performance on commonly-used benchmark datasets often fails to accurately reflect their reliability and robustness when applied to real-world noisy data. To address these challenges, we propose a unified robustness evaluation framework based on the slot-filling task to systematically evaluate the dialogue understanding capability of LLMs in diverse input perturbation scenarios. Specifically, we construct a input perturbation evaluation dataset, Noise-LLM, which contains five types of single perturbation and four types of mixed perturbation data. Furthermore, we utilize a multi-level data augmentation method (character, word, and sentence levels) to construct a candidate data pool, and carefully design two ways of automatic task demonstration construction strategies (instance-level and entity-level) with various prompt templates. Our aim is to assess how well various robustness methods of LLMs perform in real-world noisy scenarios. The experiments have demonstrated that the current open-source LLMs generally achieve limited perturbation robustness performance. Based on these experimental observations, we make some forward-looking suggestions to fuel the research in this direction.

Yejie Wang, Keqing He, Dayuan Fu et al.

Recently, there has been a growing interest in studying how to construct better code instruction tuning data. However, we observe Code models trained with these datasets exhibit high performance on HumanEval but perform worse on other benchmarks such as LiveCodeBench. Upon further investigation, we find that many datasets suffer from severe data leakage. After cleaning up most of the leaked data, some well-known high-quality datasets perform poorly. This discovery reveals a new challenge: identifying which dataset genuinely qualify as high-quality code instruction data. To address this, we propose an efficient code data pruning strategy for selecting good samples. Our approach is based on three dimensions: instruction complexity, response quality, and instruction diversity. Based on our selected data, we present XCoder, a family of models finetuned from LLaMA3. Our experiments show XCoder achieves new state-of-the-art performance using fewer training data, which verify the effectiveness of our data strategy. Moreover, we perform a comprehensive analysis on the data composition and find existing code datasets have different characteristics according to their construction methods, which provide new insights for future code LLMs. Our models and dataset are released in https://github.com/banksy23/XCoder

Jiachi Liu, Liwen Wang, Guanting Dong et al.

In real dialogue scenarios, as there are unknown input noises in the utterances, existing supervised slot filling models often perform poorly in practical applications. Even though there are some studies on noise-robust models, these works are only evaluated on rule-based synthetic datasets, which is limiting, making it difficult to promote the research of noise-robust methods. In this paper, we introduce a noise robustness evaluation dataset named Noise-SF for slot filling task. The proposed dataset contains five types of human-annotated noise, and all those noises are exactly existed in real extensive robust-training methods of slot filling into the proposed framework. By conducting exhaustive empirical evaluation experiments on Noise-SF, we find that baseline models have poor performance in robustness evaluation, and the proposed framework can effectively improve the robustness of models. Based on the empirical experimental results, we make some forward-looking suggestions to fuel the research in this direction. Our dataset Noise-SF will be released at https://github.com/dongguanting/Noise-SF.

Haoran Luo, Haihong E, Zichen Tang et al.

Knowledge Base Question Answering (KBQA) aims to answer natural language questions over large-scale knowledge bases (KBs), which can be summarized into two crucial steps: knowledge retrieval and semantic parsing. However, three core challenges remain: inefficient knowledge retrieval, mistakes of retrieval adversely impacting semantic parsing, and the complexity of previous KBQA methods. To tackle these challenges, we introduce ChatKBQA, a novel and simple generate-then-retrieve KBQA framework, which proposes first generating the logical form with fine-tuned LLMs, then retrieving and replacing entities and relations with an unsupervised retrieval method, to improve both generation and retrieval more directly. Experimental results show that ChatKBQA achieves new state-of-the-art performance on standard KBQA datasets, WebQSP, and CWQ. This work can also be regarded as a new paradigm for combining LLMs with knowledge graphs (KGs) for interpretable and knowledge-required question answering. Our code is publicly available.

Guanting Dong, Rumei Li, Sirui Wang et al.

Knowledge Base Question Answering (KBQA) aims to answer natural language questions with factual information such as entities and relations in KBs. However, traditional Pre-trained Language Models (PLMs) are directly pre-trained on large-scale natural language corpus, which poses challenges for them in understanding and representing complex subgraphs in structured KBs. To bridge the gap between texts and structured KBs, we propose a Structured Knowledge-aware Pre-training method (SKP). In the pre-training stage, we introduce two novel structured knowledge-aware tasks, guiding the model to effectively learn the implicit relationship and better representations of complex subgraphs. In downstream KBQA task, we further design an efficient linearization strategy and an interval attention mechanism, which assist the model to better encode complex subgraphs and shield the interference of irrelevant subgraphs during reasoning respectively. Detailed experiments and analyses on WebQSP verify the effectiveness of SKP, especially the significant improvement in subgraph retrieval (+4.08% H@10).

Guanting Dong, Zechen Wang, Jinxu Zhao et al.

The objective of few-shot named entity recognition is to identify named entities with limited labeled instances. Previous works have primarily focused on optimizing the traditional token-wise classification framework, while neglecting the exploration of information based on NER data characteristics. To address this issue, we propose a Multi-Task Semantic Decomposition Framework via Joint Task-specific Pre-training (MSDP) for few-shot NER. Drawing inspiration from demonstration-based and contrastive learning, we introduce two novel pre-training tasks: Demonstration-based Masked Language Modeling (MLM) and Class Contrastive Discrimination. These tasks effectively incorporate entity boundary information and enhance entity representation in Pre-trained Language Models (PLMs). In the downstream main task, we introduce a multi-task joint optimization framework with the semantic decomposing method, which facilitates the model to integrate two different semantic information for entity classification. Experimental results of two few-shot NER benchmarks demonstrate that MSDP consistently outperforms strong baselines by a large margin. Extensive analyses validate the effectiveness and generalization of MSDP.

Xuefeng Li, Liwen Wang, Guanting Dong et al.

Zero-shot cross-domain slot filling aims to transfer knowledge from the labeled source domain to the unlabeled target domain. Existing models either encode slot descriptions and examples or design handcrafted question templates using heuristic rules, suffering from poor generalization capability or robustness. In this paper, we propose a generative zero-shot prompt learning framework for cross-domain slot filling, both improving generalization and robustness than previous work. Besides, we introduce a novel inverse prompting strategy to distinguish different slot types to avoid the multiple prediction problem, and an efficient prompt-tuning strategy to boost higher performance by only training fewer prompt parameters. Experiments and analysis demonstrate the effectiveness of our proposed framework, especially huge improvements (+13.44% F1) on the unseen slots.

Weihao Zeng, Keqing He, Zechen Wang et al.

Recent advances in neural approaches greatly improve task-oriented dialogue (TOD) systems which assist users to accomplish their goals. However, such systems rely on costly manually labeled dialogs which are not available in practical scenarios. In this paper, we present our models for Track 2 of the SereTOD 2022 challenge, which is the first challenge of building semi-supervised and reinforced TOD systems on a large-scale real-world Chinese TOD dataset MobileCS. We build a knowledge-grounded dialog model to formulate dialog history and local KB as input and predict the system response. And we perform semi-supervised pre-training both on the labeled and unlabeled data. Our system achieves the first place both in the automatic evaluation and human interaction, especially with higher BLEU (+7.64) and Success (+13.6\%) than the second place.

Guanting Dong, Daichi Guo, Liwen Wang et al.

Most existing slot filling models tend to memorize inherent patterns of entities and corresponding contexts from training data. However, these models can lead to system failure or undesirable outputs when being exposed to spoken language perturbation or variation in practice. We propose a perturbed semantic structure awareness transferring method for training perturbation-robust slot filling models. Specifically, we introduce two MLM-based training strategies to respectively learn contextual semantic structure and word distribution from unsupervised language perturbation corpus. Then, we transfer semantic knowledge learned from upstream training procedure into the original samples and filter generated data by consistency processing. These procedures aim to enhance the robustness of slot filling models. Experimental results show that our method consistently outperforms the previous basic methods and gains strong generalization while preventing the model from memorizing inherent patterns of entities and contexts.

Xiaoshuai Song, Keqing He, Pei Wang et al.

The tasks of out-of-domain (OOD) intent discovery and generalized intent discovery (GID) aim to extend a closed intent classifier to open-world intent sets, which is crucial to task-oriented dialogue (TOD) systems. Previous methods address them by fine-tuning discriminative models. Recently, although some studies have been exploring the application of large language models (LLMs) represented by ChatGPT to various downstream tasks, it is still unclear for the ability of ChatGPT to discover and incrementally extent OOD intents. In this paper, we comprehensively evaluate ChatGPT on OOD intent discovery and GID, and then outline the strengths and weaknesses of ChatGPT. Overall, ChatGPT exhibits consistent advantages under zero-shot settings, but is still at a disadvantage compared to fine-tuned models. More deeply, through a series of analytical experiments, we summarize and discuss the challenges faced by LLMs including clustering, domain-specific understanding, and cross-domain in-context learning scenarios. Finally, we provide empirical guidance for future directions to address these challenges.

Daichi Guo, Guanting Dong, Dayuan Fu et al.

In real dialogue scenarios, the existing slot filling model, which tends to memorize entity patterns, has a significantly reduced generalization facing Out-of-Vocabulary (OOV) problems. To address this issue, we propose an OOV robust slot filling model based on multi-level data augmentations to solve the OOV problem from both word and slot perspectives. We present a unified contrastive learning framework, which pull representations of the origin sample and augmentation samples together, to make the model resistant to OOV problems. We evaluate the performance of the model from some specific slots and carefully design test data with OOV word perturbation to further demonstrate the effectiveness of OOV words. Experiments on two datasets show that our approach outperforms the previous sota methods in terms of both OOV slots and words.

Guanting Dong, Zechen Wang, Liwen Wang et al.

Few-shot named entity recognition (NER) aims at identifying named entities based on only few labeled instances. Most existing prototype-based sequence labeling models tend to memorize entity mentions which would be easily confused by close prototypes. In this paper, we proposed a Prototypical Semantic Decoupling method via joint Contrastive learning (PSDC) for few-shot NER. Specifically, we decouple class-specific prototypes and contextual semantic prototypes by two masking strategies to lead the model to focus on two different semantic information for inference. Besides, we further introduce joint contrastive learning objectives to better integrate two kinds of decoupling information and prevent semantic collapse. Experimental results on two few-shot NER benchmarks demonstrate that PSDC consistently outperforms the previous SOTA methods in terms of overall performance. Extensive analysis further validates the effectiveness and generalization of PSDC.

Xuefeng Li, Hao Lei, Liwen Wang et al.

Multi-domain text classification can automatically classify texts in various scenarios. Due to the diversity of human languages, texts with the same label in different domains may differ greatly, which brings challenges to the multi-domain text classification. Current advanced methods use the private-shared paradigm, capturing domain-shared features by a shared encoder, and training a private encoder for each domain to extract domain-specific features. However, in realistic scenarios, these methods suffer from inefficiency as new domains are constantly emerging. In this paper, we propose a robust contrastive alignment method to align text classification features of various domains in the same feature space by supervised contrastive learning. By this means, we only need two universal feature extractors to achieve multi-domain text classification. Extensive experimental results show that our method performs on par with or sometimes better than the state-of-the-art method, which uses the complex multi-classifier in a private-shared framework.

Gang Zhao, Xiaocheng Gong, Xinjie Yang et al.

Most current Event Extraction (EE) methods focus on the high-resource scenario, which requires a large amount of annotated data and can hardly be applied to low-resource domains. To address EE more effectively with limited resources, we propose the Demonstration-enhanced Schema-guided Generation (DemoSG) model, which benefits low-resource EE from two aspects: Firstly, we propose the demonstration-based learning paradigm for EE to fully use the annotated data, which transforms them into demonstrations to illustrate the extraction process and help the model learn effectively. Secondly, we formulate EE as a natural language generation task guided by schema-based prompts, thereby leveraging label semantics and promoting knowledge transfer in low-resource scenarios. We conduct extensive experiments under in-domain and domain adaptation low-resource settings on three datasets, and study the robustness of DemoSG. The results show that DemoSG significantly outperforms current methods in low-resource scenarios.

Yuxiang Wu, Guanting Dong, Weiran Xu

Zero-shot Dialogue State Tracking (DST) addresses the challenge of acquiring and annotating task-oriented dialogues, which can be time-consuming and costly. However, DST extends beyond simple slot-filling and requires effective updating strategies for tracking dialogue state as conversations progress. In this paper, we propose ParsingDST, a new In-Context Learning (ICL) method, to introduce additional intricate updating strategies in zero-shot DST. Our approach reformulates the DST task by leveraging powerful Large Language Models (LLMs) and translating the original dialogue text to JSON through semantic parsing as an intermediate state. We also design a novel framework that includes more modules to ensure the effectiveness of updating strategies in the text-to-JSON process. Experimental results demonstrate that our approach outperforms existing zero-shot DST methods on MultiWOZ, exhibiting significant improvements in Joint Goal Accuracy (JGA) and slot accuracy compared to existing ICL methods. Our code has been released.

Xiaoxi Li, Jiajie Jin, Guanting Dong et al.

Large reasoning models (LRMs), such as OpenAI-o1 and DeepSeek-R1, demonstrate impressive long-horizon reasoning capabilities. However, their reliance on static internal knowledge limits their performance on complex, knowledge-intensive tasks and hinders their ability to produce comprehensive research reports requiring synthesis of diverse web information. To address this, we propose WebThinker, a deep research agent that empowers LRMs to autonomously search the web, navigate among web pages, and draft reports during the reasoning process. WebThinker integrates a Deep Web Explorer module, enabling LRMs to dynamically search, navigate, and extract information from the web when encountering knowledge gaps. It also employs an Autonomous Think-Search-and-Draft strategy, allowing the model to seamlessly interleave reasoning, information gathering, and report writing in real time. To further enhance research tool utilization, we introduce an RL-based training strategy via iterative online Direct Preference Optimization (DPO). Extensive experiments on complex reasoning benchmarks (GPQA, GAIA, WebWalkerQA, HLE) and scientific report generation tasks (Glaive) demonstrate that WebThinker significantly outperforms existing methods and strong proprietary systems. Our approach enhances LRM reliability and applicability in complex scenarios, paving the way for more capable and versatile deep research systems. The code is available at https://github.com/RUC-NLPIR/WebThinker.

Dayuan Fu, Biqing Qi, Yihuai Gao et al.

Long-term memory is significant for agents, in which insights play a crucial role. However, the emergence of irrelevant insight and the lack of general insight can greatly undermine the effectiveness of insight. To solve this problem, in this paper, we introduce Multi-Scale Insight Agent (MSI-Agent), an embodied agent designed to improve LLMs' planning and decision-making ability by summarizing and utilizing insight effectively across different scales. MSI achieves this through the experience selector, insight generator, and insight selector. Leveraging a three-part pipeline, MSI can generate task-specific and high-level insight, store it in a database, and then use relevant insight from it to aid in decision-making. Our experiments show that MSI outperforms another insight strategy when planning by GPT3.5. Moreover, We delve into the strategies for selecting seed experience and insight, aiming to provide LLM with more useful and relevant insight for better decision-making. Our observations also indicate that MSI exhibits better robustness when facing domain-shifting scenarios.

Shanglin Lei, Xiaoping Wang, Guanting Dong et al.

Emotion Recognition in Conversation (ERC) has attracted widespread attention in the natural language processing field due to its enormous potential for practical applications. Existing ERC methods face challenges in achieving generalization to diverse scenarios due to insufficient modeling of context, ambiguous capture of dialogue relationships and overfitting in speaker modeling. In this work, we present a Hybrid Continuous Attributive Network (HCAN) to address these issues in the perspective of emotional continuation and emotional attribution. Specifically, HCAN adopts a hybrid recurrent and attention-based module to model global emotion continuity. Then a novel Emotional Attribution Encoding (EAE) is proposed to model intra- and inter-emotional attribution for each utterance. Moreover, aiming to enhance the robustness of the model in speaker modeling and improve its performance in different scenarios, A comprehensive loss function emotional cognitive loss $\mathcal{L}_{\rm EC}$ is proposed to alleviate emotional drift and overcome the overfitting of the model to speaker modeling. Our model achieves state-of-the-art performance on three datasets, demonstrating the superiority of our work. Another extensive comparative experiments and ablation studies on three benchmarks are conducted to provided evidence to support the efficacy of each module. Further exploration of generalization ability experiments shows the plug-and-play nature of the EAE module in our method.

Shanglin Lei, Guanting Dong, Xiaoping Wang et al.

The field of emotion recognition of conversation (ERC) has been focusing on separating sentence feature encoding and context modeling, lacking exploration in generative paradigms based on unified designs. In this study, we propose a novel approach, InstructERC, to reformulate the ERC task from a discriminative framework to a generative framework based on Large Language Models (LLMs). InstructERC makes three significant contributions: (1) it introduces a simple yet effective retrieval template module, which helps the model explicitly integrate multi-granularity dialogue supervision information. (2) We introduce two additional emotion alignment tasks, namely speaker identification and emotion prediction tasks, to implicitly model the dialogue role relationships and future emotional tendencies in conversations. (3) Pioneeringly, we unify emotion labels across benchmarks through the feeling wheel to fit real application scenarios. InstructERC still perform impressively on this unified dataset. Our LLM-based plugin framework significantly outperforms all previous models and achieves comprehensive SOTA on three commonly used ERC datasets. Extensive analysis of parameter-efficient and data-scaling experiments provides empirical guidance for applying it in practical scenarios.

Gang Zhao, Yidong Shi, Shudong Lu et al.

Document-level event extraction (DEE) faces two main challenges: arguments-scattering and multi-event. Although previous methods attempt to address these challenges, they overlook the interference of event-unrelated sentences during event detection and neglect the mutual interference of different event roles during argument extraction. Therefore, this paper proposes a novel Schema-based Explicitly Aggregating~(SEA) model to address these limitations. SEA aggregates event information into event type and role representations, enabling the decoding of event records based on specific type-aware representations. By detecting each event based on its event type representation, SEA mitigates the interference caused by event-unrelated information. Furthermore, SEA extracts arguments for each role based on its role-aware representations, reducing mutual interference between different roles. Experimental results on the ChFinAnn and DuEE-fin datasets show that SEA outperforms the SOTA methods.

Guanting Dong, Xiaoshuai Song, Yutao Zhu et al.

Following natural instructions is crucial for the effective application of Retrieval-Augmented Generation (RAG) systems. Despite recent advancements in Large Language Models (LLMs), research on assessing and improving instruction-following (IF) alignment within the RAG domain remains limited. To address this issue, we propose VIF-RAG, the first automated, scalable, and verifiable synthetic pipeline for instruction-following alignment in RAG systems. We start by manually crafting a minimal set of atomic instructions (<100) and developing combination rules to synthesize and verify complex instructions for a seed set. We then use supervised models for instruction rewriting while simultaneously generating code to automate the verification of instruction quality via a Python executor. Finally, we integrate these instructions with extensive RAG and general data samples, scaling up to a high-quality VIF-RAG-QA dataset (>100k) through automated processes. To further bridge the gap in instruction-following auto-evaluation for RAG systems, we introduce FollowRAG Benchmark, which includes approximately 3K test samples, covering 22 categories of general instruction constraints and four knowledge-intensive QA datasets. Due to its robust pipeline design, FollowRAG can seamlessly integrate with different RAG benchmarks. Using FollowRAG and eight widely-used IF and foundational abilities benchmarks for LLMs, we demonstrate that VIF-RAG markedly enhances LLM performance across a broad range of general instruction constraints while effectively leveraging its capabilities in RAG scenarios. Further analysis offers practical insights for achieving IF alignment in RAG systems. Our code and datasets are released at https://FollowRAG.github.io.

Jiajie Jin, Xiaoxi Li, Guanting Dong et al.

Complex information needs in real-world search scenarios demand deep reasoning and knowledge synthesis across diverse sources, which traditional retrieval-augmented generation (RAG) pipelines struggle to address effectively. Current reasoning-based approaches suffer from a fundamental limitation: they use a single model to handle both high-level planning and detailed execution, leading to inefficient reasoning and limited scalability. In this paper, we introduce HiRA, a hierarchical framework that separates strategic planning from specialized execution. Our approach decomposes complex search tasks into focused subtasks, assigns each subtask to domain-specific agents equipped with external tools and reasoning capabilities, and coordinates the results through a structured integration mechanism. This separation prevents execution details from disrupting high-level reasoning while enabling the system to leverage specialized expertise for different types of information processing. Experiments on four complex, cross-modal deep search benchmarks demonstrate that HiRA significantly outperforms state-of-the-art RAG and agent-based systems. Our results show improvements in both answer quality and system efficiency, highlighting the effectiveness of decoupled planning and execution for multi-step information seeking tasks. Our code is available at https://github.com/ignorejjj/HiRA.

Tingfeng Hui, Lulu Zhao, Guanting Dong et al.

Instruction tuning has been widely used to unleash the complete potential of large language models. Notably, complex and diverse instructions are of significant importance as they can effectively align models with various downstream tasks. However, current approaches to constructing large-scale instructions predominantly favour powerful models such as GPT-4 or those with over 70 billion parameters, under the empirical presumption that such larger language models (LLMs) inherently possess enhanced capabilities. In this study, we question this prevalent assumption and conduct an in-depth exploration into the potential of smaller language models (SLMs) in the context of instruction evolution. Extensive experiments across three scenarios of instruction evolution reveal that smaller language models (SLMs) can synthesize more effective instructions than LLMs. Further analysis demonstrates that SLMs possess a broader output space during instruction evolution, resulting in more complex and diverse variants. We also observe that the existing metrics fail to focus on the impact of the instructions. Thus, we propose Instruction Complex-Aware IFD (IC-IFD), which introduces instruction complexity in the original IFD score to evaluate the effectiveness of instruction data more accurately. Our source code is available at: \href{https://github.com/HypherX/Evolution-Analysis}{https://github.com/HypherX/Evolution-Analysis}

Jiajie Jin, Xiaoxi Li, Guanting Dong et al.

Real-world RAG applications often encounter long-context input scenarios, where redundant information and noise results in higher inference costs and reduced performance. To address these challenges, we propose LongRefiner, an efficient plug-and-play refiner that leverages the inherent structural characteristics of long documents. LongRefiner employs dual-level query analysis, hierarchical document structuring, and adaptive refinement through multi-task learning on a single foundation model. Experiments on seven QA datasets demonstrate that LongRefiner achieves competitive performance in various scenarios while using 10x fewer computational costs and latency compared to the best baseline. Further analysis validates that LongRefiner is scalable, efficient, and effective, providing practical insights for real-world long-text RAG applications. Our code is available at https://github.com/ignorejjj/LongRefiner.

Chenghao Zhang, Guanting Dong, Xinyu Yang et al.

Retrieval-Augmented Generation (RAG) has emerged as a powerful paradigm for enhancing large language models (LLMs) by retrieving relevant documents from an external corpus. However, existing RAG systems primarily focus on unimodal text documents, and often fall short in real-world scenarios where both queries and documents may contain mixed modalities (such as text and images). In this paper, we address the challenge of Universal Retrieval-Augmented Generation (URAG), which involves retrieving and reasoning over mixed-modal information to improve vision-language generation. To this end, we propose Nyx, a unified mixed-modal to mixed-modal retriever tailored for URAG scenarios. To mitigate the scarcity of realistic mixed-modal data, we introduce a four-stage automated pipeline for generation and filtering, leveraging web documents to construct NyxQA, a dataset comprising diverse mixed-modal question-answer pairs that better reflect real-world information needs. Building on this high-quality dataset, we adopt a two-stage training framework for Nyx: we first perform pre-training on NyxQA along with a variety of open-source retrieval datasets, followed by supervised fine-tuning using feedback from downstream vision-language models (VLMs) to align retrieval outputs with generative preferences. Experimental results demonstrate that Nyx not only performs competitively on standard text-only RAG benchmarks, but also excels in the more general and realistic URAG setting, significantly improving generation quality in vision-language tasks.

Guanting Dong, Yutao Zhu, Chenghao Zhang et al.

Retrieval-augmented generation (RAG) has demonstrated effectiveness in mitigating the hallucination problem of large language models (LLMs). However, the difficulty of aligning the retriever with the diverse LLMs' knowledge preferences inevitably poses an inevitable challenge in developing a reliable RAG system. To address this issue, we propose DPA-RAG, a universal framework designed to align diverse knowledge preferences within RAG systems. Specifically, we initially introduce a preference knowledge construction pipline and incorporate five novel query augmentation strategies to alleviate preference data scarcity. Based on preference data, DPA-RAG accomplishes both external and internal preference alignment: 1) It jointly integrate pair-wise, point-wise, and contrastive preference alignment abilities into the reranker, achieving external preference alignment among RAG components. 2) It further introduces a pre-aligned stage before vanilla Supervised Fine-tuning (SFT), enabling LLMs to implicitly capture knowledge aligned with their reasoning preferences, achieving LLMs' internal alignment. Experimental results across four knowledge-intensive QA datasets demonstrate that DPA-RAG outperforms all baselines and seamlessly integrates both black-box and open-sourced LLM readers. Further qualitative analysis and discussions also provide empirical guidance for achieving reliable RAG systems. Our code is publicly available at https://github.com/dongguanting/DPA-RAG.

Guanting Dong, Keming Lu, Chengpeng Li et al.

One core capability of large language models (LLMs) is to follow natural language instructions. However, the issue of automatically constructing high-quality training data to enhance the complex instruction-following abilities of LLMs without manual annotation remains unresolved. In this paper, we introduce AutoIF, the first scalable and reliable method for automatically generating instruction-following training data. AutoIF transforms the validation of instruction-following data quality into code verification, requiring LLMs to generate instructions, the corresponding code to check the correctness of the instruction responses, and unit test samples to verify the code's correctness. Then, execution feedback-based rejection sampling can generate data for Supervised Fine-Tuning (SFT) and Reinforcement Learning from Human Feedback (RLHF) training. AutoIF achieves significant improvements across three training algorithms, SFT, Offline DPO, and Online DPO, when applied to the top open-source LLMs, Qwen2 and LLaMA3, in self-alignment and strong-to-weak distillation settings. Our code is publicly available at https://github.com/QwenLM/AutoIF.

Xiaoshuai Song, Muxi Diao, Guanting Dong et al.

Large language models (LLMs) have demonstrated significant potential in advancing various fields of research and society. However, the current community of LLMs overly focuses on benchmarks for analyzing specific foundational skills (e.g. mathematics and code generation), neglecting an all-round evaluation of the computer science field. To bridge this gap, we introduce CS-Bench, the first multilingual (English, Chinese, French, German) benchmark dedicated to evaluating the performance of LLMs in computer science. CS-Bench comprises approximately 10K meticulously curated test samples, covering 26 subfields across 4 key areas of computer science, encompassing various task forms and divisions of knowledge and reasoning. Utilizing CS-Bench, we conduct a comprehensive evaluation of over 30 mainstream LLMs, revealing the relationship between CS performance and model scales. We also quantitatively analyze the reasons for failures in existing LLMs and highlight directions for improvements, including knowledge supplementation and CS-specific reasoning. Further cross-capability experiments show a high correlation between LLMs' capabilities in computer science and their abilities in mathematics and coding. Moreover, expert LLMs specialized in mathematics and coding also demonstrate strong performances in several CS subfields. Looking ahead, we envision CS-Bench serving as a cornerstone for LLM applications in the CS field and paving new avenues in assessing LLMs' diverse reasoning capabilities. The CS-Bench data and evaluation code are available at https://github.com/csbench/csbench.

Guanting Dong, Tingfeng Hui, Zhuoma GongQue et al.

Recently, prompt-based generative frameworks have shown impressive capabilities in sequence labeling tasks. However, in practical dialogue scenarios, relying solely on simplistic templates and traditional corpora presents a challenge for these methods in generalizing to unknown input perturbations. To address this gap, we propose a multi-task demonstration based generative framework for noisy slot filling, named DemoNSF. Specifically, we introduce three noisy auxiliary tasks, namely noisy recovery (NR), random mask (RM), and hybrid discrimination (HD), to implicitly capture semantic structural information of input perturbations at different granularities. In the downstream main task, we design a noisy demonstration construction strategy for the generative framework, which explicitly incorporates task-specific information and perturbed distribution during training and inference. Experiments on two benchmarks demonstrate that DemoNSF outperforms all baseline methods and achieves strong generalization. Further analysis provides empirical guidance for the practical application of generative frameworks. Our code is released at https://github.com/dongguanting/Demo-NSF.

Guanting Dong, Chenghao Zhang, Mengjie Deng et al.

Multi-step multimodal reasoning tasks pose significant challenges for multimodal large language models (MLLMs), and finding effective ways to enhance their performance in such scenarios remains an unresolved issue. In this paper, we propose AR-MCTS, a universal framework designed to progressively improve the reasoning capabilities of MLLMs through Active Retrieval (AR) and Monte Carlo Tree Search (MCTS). Our approach begins with the development of a unified retrieval module that retrieves key supporting insights for solving complex reasoning problems from a hybrid-modal retrieval corpus. To bridge the gap in automated multimodal reasoning verification, we employ the MCTS algorithm combined with an active retrieval mechanism, which enables the automatic generation of step-wise annotations. This strategy dynamically retrieves key insights for each reasoning step, moving beyond traditional beam search sampling to improve the diversity and reliability of the reasoning space. Additionally, we introduce a process reward model that aligns progressively to support the automatic verification of multimodal reasoning tasks. Experimental results across three complex multimodal reasoning benchmarks confirm the effectiveness of the AR-MCTS framework in enhancing the performance of various multimodal models. Further analysis demonstrates that AR-MCTS can optimize sampling diversity and accuracy, yielding reliable multimodal reasoning.

Yejie Wang, Keqing He, Guanting Dong et al.

Code Large Language Models (Code LLMs) have demonstrated outstanding performance in code-related tasks. Several instruction tuning approaches have been proposed to boost the code generation performance of pre-trained Code LLMs. In this paper, we introduce a diverse instruction model (DolphCoder) with self-evaluating for code generation. It learns diverse instruction targets and combines a code evaluation objective to enhance its code generation ability. Our model achieves superior performance on the HumanEval and MBPP benchmarks, demonstrating new insights for future code instruction tuning work. Our key findings are: (1) Augmenting more diverse responses with distinct reasoning paths increases the code capability of LLMs. (2) Improving one's ability to evaluate the correctness of code solutions also enhances their ability to create it.

Dayuan Fu, Jianzhao Huang, Siyuan Lu et al.

Addressing the disparity between forecasts and actual results can enable individuals to expand their thought processes and stimulate self-reflection, thus promoting accurate planning. In this research, we present **PreAct**, an agent framework that integrates **pre**diction, **rea**soning, and **act**ion. By utilizing the information derived from predictions, the large language model (LLM) agent can provide a wider range and more strategically focused reasoning. This leads to more efficient actions that aid the agent in accomplishing intricate tasks. Our experimental results show that PreAct surpasses the ReAct method in completing complex tasks and that PreAct's performance can be further improved when paired with other memory or selection strategy techniques. We presented the model with varying quantities of historical predictions and discovered that these predictions consistently enhance LLM planning.The variances in single-step reasoning between PreAct and ReAct indicate that PreAct indeed has benefits in terms of diversity and strategic orientation over ReAct.

Guanting Dong, Xiaoxi Li, Yuyao Zhang et al.

Real-world live retrieval-augmented generation (RAG) systems face significant challenges when processing user queries that are often noisy, ambiguous, and contain multiple intents. While RAG enhances large language models (LLMs) with external knowledge, current systems typically struggle with such complex inputs, as they are often trained or evaluated on cleaner data. This paper introduces Omni-RAG, a novel framework designed to improve the robustness and effectiveness of RAG systems in live, open-domain settings. Omni-RAG employs LLM-assisted query understanding to preprocess user inputs through three key modules: (1) Deep Query Understanding and Decomposition, which utilizes LLMs with tailored prompts to denoise queries (e.g., correcting spelling errors) and decompose multi-intent queries into structured sub-queries; (2) Intent-Aware Knowledge Retrieval, which performs retrieval for each sub-query from a corpus (i.e., FineWeb using OpenSearch) and aggregates the results; and (3) Reranking and Generation, where a reranker (i.e., BGE) refines document selection before a final response is generated by an LLM (i.e., Falcon-10B) using a chain-of-thought prompt. Omni-RAG aims to bridge the gap between current RAG capabilities and the demands of real-world applications, such as those highlighted by the SIGIR 2025 LiveRAG Challenge, by robustly handling complex and noisy queries.

Xiaoshuai Song, Zhengyang Wang, Keqing He et al.

Knowledge editing (KE) aims to efficiently and precisely modify the behavior of large language models (LLMs) to update specific knowledge without negatively influencing other knowledge. Current research primarily focuses on white-box LLMs editing, overlooking an important scenario: black-box LLMs editing, where LLMs are accessed through interfaces and only textual output is available. In this paper, we first officially introduce KE on black-box LLMs and then propose a comprehensive evaluation framework to overcome the limitations of existing evaluations that are not applicable to black-box LLMs editing and lack comprehensiveness. To tackle privacy leaks of editing data and style over-editing in current methods, we introduce a novel postEdit framework, resolving privacy concerns through downstream post-processing and maintaining textual style consistency via fine-grained editing to original responses. Experiments and analysis on two benchmarks demonstrate that postEdit outperforms all baselines and achieves strong generalization, especially with huge improvements on style retention (average $+20.82\%\uparrow$).

Jinxu Zhao, Guanting Dong, Yueyan Qiu et al.

In a realistic dialogue system, the input information from users is often subject to various types of input perturbations, which affects the slot-filling task. Although rule-based data augmentation methods have achieved satisfactory results, they fail to exhibit the desired generalization when faced with unknown noise disturbances. In this study, we address the challenges posed by input perturbations in slot filling by proposing Noise-BERT, a unified Perturbation-Robust Framework with Noise Alignment Pre-training. Our framework incorporates two Noise Alignment Pre-training tasks: Slot Masked Prediction and Sentence Noisiness Discrimination, aiming to guide the pre-trained language model in capturing accurate slot information and noise distribution. During fine-tuning, we employ a contrastive learning loss to enhance the semantic representation of entities and labels. Additionally, we introduce an adversarial attack training strategy to improve the model's robustness. Experimental results demonstrate the superiority of our proposed approach over state-of-the-art models, and further analysis confirms its effectiveness and generalization ability.

Tingfeng Hui, Pengyu Zhu, Bowen Ping et al.

Instruction-following has emerged as a crucial capability for large language models (LLMs). However, existing approaches often rely on pre-existing documents or external resources to synthesize instruction-following data, which limits their flexibility and generalizability. In this paper, we introduce DecIF, a fully autonomous, meta-decomposition guided framework that generates diverse and high-quality instruction-following data using only LLMs. DecIF is grounded in the principle of decomposition. For instruction generation, we guide LLMs to iteratively produce various types of meta-information, which are then combined with response constraints to form well-structured and semantically rich instructions. We further utilize LLMs to detect and resolve potential inconsistencies within the generated instructions. Regarding response generation, we decompose each instruction into atomic-level evaluation criteria, enabling rigorous validation and the elimination of inaccurate instruction-response pairs. Extensive experiments across a wide range of scenarios and settings demonstrate DecIF's superior performance on instruction-following tasks. Further analysis highlights its strong flexibility, scalability, and generalizability in automatically synthesizing high-quality instruction data.

YiFan Zhang, Shanglin Lei, Runqi Qiao et al.

The rapidly developing field of large multimodal models (LMMs) has led to the emergence of diverse models with remarkable capabilities. However, existing benchmarks fail to comprehensively, objectively and accurately evaluate whether LMMs align with the diverse needs of humans in real-world scenarios. To bridge this gap, we propose the Multi-Dimensional Insights (MDI) benchmark, which includes over 500 images covering six common scenarios of human life. Notably, the MDI-Benchmark offers two significant advantages over existing evaluations: (1) Each image is accompanied by two types of questions: simple questions to assess the model's understanding of the image, and complex questions to evaluate the model's ability to analyze and reason beyond basic content. (2) Recognizing that people of different age groups have varying needs and perspectives when faced with the same scenario, our benchmark stratifies questions into three age categories: young people, middle-aged people, and older people. This design allows for a detailed assessment of LMMs' capabilities in meeting the preferences and needs of different age groups. With MDI-Benchmark, the strong model like GPT-4o achieve 79% accuracy on age-related tasks, indicating that existing LMMs still have considerable room for improvement in addressing real-world applications. Looking ahead, we anticipate that the MDI-Benchmark will open new pathways for aligning real-world personalization in LMMs. The MDI-Benchmark data and evaluation code are available at https://mdi-benchmark.github.io/

Pei Wang, Yejie Wang, Muxi Diao et al.

In the deployment of large language models (LLMs), accurate confidence estimation is critical for assessing the credibility of model predictions. However, existing methods often fail to overcome the issue of overconfidence on incorrect answers. In this work, we focus on improving the confidence estimation of large language models. Considering the fragility of self-awareness in language models, we introduce a Multi-Perspective Consistency (MPC) method. We leverage complementary insights from different perspectives within models (MPC-Internal) and across different models (MPC-Across) to mitigate the issue of overconfidence arising from a singular viewpoint. The experimental results on eight publicly available datasets show that our MPC achieves state-of-the-art performance. Further analyses indicate that MPC can mitigate the problem of overconfidence and is effectively scalable to other models.