Yansong Feng

Quzhe Huang, Shibo Hao, Yuan Ye et al. · pku

DocRED is a widely used dataset for document-level relation extraction. In the large-scale annotation, a \textit{recommend-revise} scheme is adopted to reduce the workload. Within this scheme, annotators are provided with candidate relation instances from distant supervision, and they then manually supplement and remove relational facts based on the recommendations. However, when comparing DocRED with a subset relabeled from scratch, we find that this scheme results in a considerable amount of false negative samples and an obvious bias towards popular entities and relations. Furthermore, we observe that the models trained on DocRED have low recall on our relabeled dataset and inherit the same bias in the training data. Through the analysis of annotators' behaviors, we figure out the underlying reason for the problems above: the scheme actually discourages annotators from supplementing adequate instances in the revision phase. We appeal to future research to take into consideration the issues with the recommend-revise scheme when designing new models and annotation schemes. The relabeled dataset is released at \url{https://github.com/AndrewZhe/Revisit-DocRED}, to serve as a more reliable test set of document RE models.

Xiao Liu, Yansong Feng, Jizhi Tang et al. · pku

People can acquire knowledge in an unsupervised manner by reading, and compose the knowledge to make novel combinations. In this paper, we investigate whether pretrained language models can perform compositional generalization in a realistic setting: recipe generation. We design the counterfactual recipe generation task, which asks models to modify a base recipe according to the change of an ingredient. This task requires compositional generalization at two levels: the surface level of incorporating the new ingredient into the base recipe, and the deeper level of adjusting actions related to the changing ingredient. We collect a large-scale recipe dataset in Chinese for models to learn culinary knowledge, and a subset of action-level fine-grained annotations for evaluation. We finetune pretrained language models on the recipe corpus, and use unsupervised counterfactual generation methods to generate modified recipes. Results show that existing models have difficulties in modifying the ingredients while preserving the original text style, and often miss actions that need to be adjusted. Although pretrained language models can generate fluent recipe texts, they fail to truly learn and use the culinary knowledge in a compositional way. Code and data are available at https://github.com/xxxiaol/counterfactual-recipe-generation.

Zhenwei An, Quzhe Huang, Cong Jiang et al. · pku

The charge prediction task aims to predict the charge for a case given its fact description. Recent models have already achieved impressive accuracy in this task, however, little is understood about the mechanisms they use to perform the judgment.For practical applications, a charge prediction model should conform to the certain legal theory in civil law countries, as under the framework of civil law, all cases are judged according to certain local legal theories. In China, for example, nearly all criminal judges make decisions based on the Four Elements Theory (FET).In this paper, we argue that trustworthy charge prediction models should take legal theories into consideration, and standing on prior studies in model interpretation, we propose three principles for trustworthy models should follow in this task, which are sensitive, selective, and presumption of innocence.We further design a new framework to evaluate whether existing charge prediction models learn legal theories. Our findings indicate that, while existing charge prediction models meet the selective principle on a benchmark dataset, most of them are still not sensitive enough and do not satisfy the presumption of innocence. Our code and dataset are released at https://github.com/ZhenweiAn/EXP_LJP.

Xiao Liu, Da Yin, Yansong Feng et al. · pku

Spatial commonsense, the knowledge about spatial position and relationship between objects (like the relative size of a lion and a girl, and the position of a boy relative to a bicycle when cycling), is an important part of commonsense knowledge. Although pretrained language models (PLMs) succeed in many NLP tasks, they are shown to be ineffective in spatial commonsense reasoning. Starting from the observation that images are more likely to exhibit spatial commonsense than texts, we explore whether models with visual signals learn more spatial commonsense than text-based PLMs. We propose a spatial commonsense benchmark that focuses on the relative scales of objects, and the positional relationship between people and objects under different actions. We probe PLMs and models with visual signals, including vision-language pretrained models and image synthesis models, on this benchmark, and find that image synthesis models are more capable of learning accurate and consistent spatial knowledge than other models. The spatial knowledge from image synthesis models also helps in natural language understanding tasks that require spatial commonsense.

Chen Zhang, Mingxu Tao, Quzhe Huang et al. · pku

Current large language models demonstrate deficiencies in understanding low-resource languages, particularly the minority languages in China. This limitation stems from the scarcity of available pre-training data. To address this accessibility challenge, we present MC$^2$, a Multilingual Corpus of Minority Languages in China, which is the largest open-source corpus of its kind so far. MC$^2$ includes four underrepresented languages: Tibetan, Uyghur, Kazakh, and Mongolian. Notably, we focus on the less common writing systems of Kazakh and Mongolian, i.e., Kazakh Arabic script and traditional Mongolian script, respectively, which have been long neglected in previous corpus construction efforts. Recognizing the prevalence of language contamination within existing corpora, we adopt a quality-centric solution for collecting MC$^2$, prioritizing accuracy while enhancing diversity. Furthermore, we underscore the importance of attending to the multiplicity of writing systems, which is closely related to the cultural awareness of the resulting models. The MC$^2$ corpus and related models are made public to the community.

Chen Zhang, Jiuheng Lin, Xiao Liu et al. · pku

The multi-answer phenomenon, where a question may have multiple answers scattered in the document, can be well handled by humans but is challenging enough for machine reading comprehension (MRC) systems. Despite recent progress in multi-answer MRC, there lacks a systematic analysis of how this phenomenon arises and how to better address it. In this work, we design a taxonomy to categorize commonly-seen multi-answer MRC instances, with which we inspect three multi-answer datasets and analyze where the multi-answer challenge comes from. We further analyze how well different paradigms of current multi-answer MRC models deal with different types of multi-answer instances. We find that some paradigms capture well the key information in the questions while others better model the relationship between questions and contexts. We thus explore strategies to make the best of the strengths of different paradigms. Experiments show that generation models can be a promising platform to incorporate different paradigms. Our annotations and code are released for further research.

Chen Zhang, Yuxuan Lai, Yansong Feng et al. · pku

Although many large-scale knowledge bases (KBs) claim to contain multilingual information, their support for many non-English languages is often incomplete. This incompleteness gives birth to the task of cross-lingual question answering over knowledge base (xKBQA), which aims to answer questions in languages different from that of the provided KB. One of the major challenges facing xKBQA is the high cost of data annotation, leading to limited resources available for further exploration. Another challenge is mapping KB schemas and natural language expressions in the questions under cross-lingual settings. In this paper, we propose a novel approach for xKBQA in a reading comprehension paradigm. We convert KB subgraphs into passages to narrow the gap between KB schemas and questions, which enables our model to benefit from recent advances in multilingual pre-trained language models (MPLMs) and cross-lingual machine reading comprehension (xMRC). Specifically, we use MPLMs, with considerable knowledge of cross-lingual mappings, for cross-lingual reading comprehension. Existing high-quality xMRC datasets can be further utilized to finetune our model, greatly alleviating the data scarcity issue in xKBQA. Extensive experiments on two xKBQA datasets in 12 languages show that our approach outperforms various baselines and achieves strong few-shot and zero-shot performance. Our dataset and code are released for further research.

Zhibin Chen, Yansong Feng, Dongyan Zhao

Typed entailment graphs try to learn the entailment relations between predicates from text and model them as edges between predicate nodes. The construction of entailment graphs usually suffers from severe sparsity and unreliability of distributional similarity. We propose a two-stage method, Entailment Graph with Textual Entailment and Transitivity (EGT2). EGT2 learns local entailment relations by recognizing possible textual entailment between template sentences formed by typed CCG-parsed predicates. Based on the generated local graph, EGT2 then uses three novel soft transitivity constraints to consider the logical transitivity in entailment structures. Experiments on benchmark datasets show that EGT2 can well model the transitivity in entailment graph to alleviate the sparsity issue, and lead to significant improvement over current state-of-the-art methods.

Mingxu Tao, Yansong Feng, Dongyan Zhao

Large pre-trained language models help to achieve state of the art on a variety of natural language processing (NLP) tasks, nevertheless, they still suffer from forgetting when incrementally learning a sequence of tasks. To alleviate this problem, recent works enhance existing models by sparse experience replay and local adaption, which yield satisfactory performance. However, in this paper we find that pre-trained language models like BERT have a potential ability to learn sequentially, even without any sparse memory replay. To verify the ability of BERT to maintain old knowledge, we adopt and re-finetune single-layer probe networks with the parameters of BERT fixed. We investigate the models on two types of NLP tasks, text classification and extractive question answering. Our experiments reveal that BERT can actually generate high quality representations for previously learned tasks in a long term, under extremely sparse replay or even no replay. We further introduce a series of novel methods to interpret the mechanism of forgetting and how memory rehearsal plays a significant role in task incremental learning, which bridges the gap between our new discovery and previous studies about catastrophic forgetting.

Mingxu Tao, Chen Zhang, Quzhe Huang et al. · pku

Adapting large language models (LLMs) to new languages typically involves continual pre-training (CT) followed by supervised fine-tuning (SFT). However, this CT-then-SFT approach struggles with limited data in the context of low-resource languages, failing to balance language modeling and task-solving capabilities. We thus propose model merging as an alternative for low-resource languages, combining models with distinct capabilities into a single model without additional training. We use model merging to develop task-solving LLMs for low-resource languages without SFT data in the target languages. Our experiments based on Llama-2-7B demonstrate that model merging effectively endows LLMs for low-resource languages with task-solving abilities, outperforming CT-then-SFT in scenarios with extremely scarce data. Observing performance saturation in model merging with more training tokens, we further analyze the merging process and introduce a slack variable to the model merging algorithm to mitigate the loss of important parameters, thereby enhancing performance. We hope that model merging can benefit more human languages suffering from data scarcity with its higher data efficiency.

Chen Zhang, Jiuheng Lin, Zhiyuan Liao et al. · pku

Adapting large language models (LLMs) to low-resource languages (LRLs) is constrained by the scarcity of task data and computational resources. Although Proxy Tuning offers a logit-level strategy for introducing scaling effects, it often fails in LRL settings because the large model's weak LRL competence might overwhelm the knowledge of specialized smaller models. We thus propose TriMix, a test-time logit fusion framework that dynamically balances capabilities from three different sources: LRL competence from a continually pretrained small model, task competence from high-resource language instruction tuning, and the scaling benefits of large models. It is data- and compute-efficient, requiring no LRL task annotations, and only continual pretraining on a small model. Experiments across four model families and eight LRLs show that TriMix consistently outperforms single-model baselines and Proxy Tuning. Our analysis reveals that prioritizing the small LRL-specialized model's logits is crucial for success, challenging the prevalent large-model-dominant assumption.

Zhibin Chen, Yansong Feng, Dongyan Zhao

Entailment Graphs (EGs) have been constructed based on extracted corpora as a strong and explainable form to indicate context-independent entailment relations in natural languages. However, EGs built by previous methods often suffer from the severe sparsity issues, due to limited corpora available and the long-tail phenomenon of predicate distributions. In this paper, we propose a multi-stage method, Typed Predicate-Entailment Graph Generator (TP-EGG), to tackle this problem. Given several seed predicates, TP-EGG builds the graphs by generating new predicates and detecting entailment relations among them. The generative nature of TP-EGG helps us leverage the recent advances from large pretrained language models (PLMs), while avoiding the reliance on carefully prepared corpora. Experiments on benchmark datasets show that TP-EGG can generate high-quality and scale-controllable entailment graphs, achieving significant in-domain improvement over state-of-the-art EGs and boosting the performance of down-stream inference tasks.

Yifei Zhou, Yansong Feng

Recent works show that discourse analysis benefits from modeling intra- and inter-sentential levels separately, where proper representations for text units of different granularities are desired to capture both the meaning of text units and their relations to the context. In this paper, we propose to take advantage of transformers to encode contextualized representations of units of different levels to dynamically capture the information required for discourse dependency analysis on intra- and inter-sentential levels. Motivated by the observation of writing patterns commonly shared across articles, we propose a novel method that treats discourse relation identification as a sequence labelling task, which takes advantage of structural information from the context of extracted discourse trees, and substantially outperforms traditional direct-classification methods. Experiments show that our model achieves state-of-the-art results on both English and Chinese datasets.

Yutong Hu, Quzhe Huang, Yansong Feng · pku

Event Temporal Relation Extraction (ETRE) aims to identify the temporal relationship between two events, which plays an important role in natural language understanding. Most previous works follow a single-label classification style, classifying an event pair into either a specific temporal relation (e.g., \textit{Before}, \textit{After}), or a special label \textit{Vague} when there may be multiple possible temporal relations between the pair. In our work, instead of directly making predictions on \textit{Vague}, we propose a multi-label classification solution for ETRE (METRE) to infer the possibility of each temporal relation independently, where we treat \textit{Vague} as the cases when there is more than one possible relation between two events. We design a speculation mechanism to explore the possible relations hidden behind \textit{Vague}, which enables the latent information to be used efficiently. Experiments on TB-Dense, MATRES and UDS-T show that our method can effectively utilize the \textit{Vague} instances to improve the recognition for specific temporal relations and outperforms most state-of-the-art methods.

Xiao Liu, Zirui Wu, Xueqing Wu et al. · stanford

Quantitative reasoning is a critical skill to analyze data, yet the assessment of such ability remains limited. To address this gap, we introduce the Quantitative Reasoning with Data (QRData) benchmark, aiming to evaluate Large Language Models' capability in statistical and causal reasoning with real-world data. The benchmark comprises a carefully constructed dataset of 411 questions accompanied by data sheets from textbooks, online learning materials, and academic papers. To compare models' quantitative reasoning abilities on data and text, we enrich the benchmark with an auxiliary set of 290 text-only questions, namely QRText. We evaluate natural language reasoning, program-based reasoning, and agent reasoning methods including Chain-of-Thought, Program-of-Thoughts, ReAct, and code interpreter assistants on diverse models. The strongest model GPT-4 achieves an accuracy of 58%, which has much room for improvement. Among open-source models, Deepseek-coder-instruct, a code LLM pretrained on 2T tokens, gets the highest accuracy of 37%. Analysis reveals that models encounter difficulties in data analysis and causal reasoning, and struggle in using causal knowledge and provided data simultaneously. Code and data are in https://github.com/xxxiaol/QRData.

Quzhe Huang, Zhenwei An, Nan Zhuang et al. · pku

In this paper, we introduce a novel dynamic expert selection framework for Mixture of Experts (MoE) models, aiming to enhance computational efficiency and model performance by adjusting the number of activated experts based on input difficulty. Unlike traditional MoE approaches that rely on fixed Top-K routing, which activates a predetermined number of experts regardless of the input's complexity, our method dynamically selects experts based on the confidence level in expert selection for each input. This allows for a more efficient utilization of computational resources, activating more experts for complex tasks requiring advanced reasoning and fewer for simpler tasks. Through extensive evaluations, our dynamic routing method demonstrates substantial improvements over conventional Top-2 routing across various benchmarks, achieving an average improvement of 0.7% with less than 90% activated parameters. Further analysis shows our model dispatches more experts to tasks requiring complex reasoning skills, like BBH, confirming its ability to dynamically allocate computational resources in alignment with the input's complexity. Our findings also highlight a variation in the number of experts needed across different layers of the transformer model, offering insights into the potential for designing heterogeneous MoE frameworks. The code and models are available at https://github.com/ZhenweiAn/Dynamic_MoE.

Zirui Wu, Yansong Feng

Tables play a crucial role in conveying information in various domains. We propose a Plan-then-Reason framework to answer different types of user queries over tables with sentence context. The framework first plans the reasoning paths over the context, then assigns each step to program-based or textual reasoning to reach the final answer. This framework enhances the table reasoning abilities for both in-context learning and fine-tuning methods. GPT-3.5-Turbo following Plan-then-Reason framework surpasses other prompting baselines without self-consistency while using less API calls and in-context demonstrations. We also construct an instruction tuning set TrixInstruct to evaluate the effectiveness of fine-tuning with this framework. We present ProTrix model family by finetuning models on TrixInstruct. Our experiments show that ProTrix family generalizes to diverse unseen tabular tasks with only 6k training instances. We further demonstrate that ProTrix can generate accurate and faithful explanations to answer complex free-form questions. Our work underscores the importance of the planning and reasoning abilities towards a model over tabular tasks with generalizability and interpretability. We open-source our dataset and models at https://github.com/WilliamZR/ProTrix.

Jiuheng Lin, Yuxuan Lai, Yansong Feng

Conditional question answering (CQA) is an important task that aims to find probable answers and identify missing conditions. Existing approaches struggle with CQA due to two challenges: (1) precisely identifying necessary conditions and the logical relationship, and (2) verifying conditions to detect any that are missing. In this paper, we propose a novel prompting approach, Chain of condition, by first identifying all conditions and constructing their logical relationships explicitly according to the document, then verifying whether these conditions are satisfied, finally solving the logical expression to indicate any missing conditions and generating the answer accordingly. Experiments on two CQA benchmark datasets show our chain of condition outperforms existing prompting baselines, establishing a new state of the art. Furthermore, with only a few examples, our method can facilitate GPT-3.5-Turbo or GPT-4 to outperform all existing supervised models.

Mingxu Tao, Quzhe Huang, Kun Xu et al. · pku

The advancement of Multimodal Large Language Models (MLLMs) has greatly accelerated the development of applications in understanding integrated texts and images. Recent works leverage image-caption datasets to train MLLMs, achieving state-of-the-art performance on image-to-text tasks. However, there are few studies exploring which layers of MLLMs make the most effort to the global image information, which plays vital roles in multimodal comprehension and generation. In this study, we find that the intermediate layers of models can encode more global semantic information, whose representation vectors perform better on visual-language entailment tasks, rather than the topmost layers. We further probe models regarding local semantic representations through object recognition tasks. We find that the topmost layers may excessively focus on local information, leading to a diminished ability to encode global information. Our code and data are released via https://github.com/kobayashikanna01/probing_MLLM_rep.

Kunhang Li, Yansong Feng

The task of text2motion is to generate human motion sequences from given textual descriptions, where the model explores diverse mappings from natural language instructions to human body movements. While most existing works are confined to coarse-grained motion descriptions, e.g., "A man squats.", fine-grained descriptions specifying movements of relevant body parts are barely explored. Models trained with coarse-grained texts may not be able to learn mappings from fine-grained motion-related words to motion primitives, resulting in the failure to generate motions from unseen descriptions. In this paper, we build a large-scale language-motion dataset specializing in fine-grained textual descriptions, FineHumanML3D, by feeding GPT-3.5-turbo with step-by-step instructions with pseudo-code compulsory checks. Accordingly, we design a new text2motion model, FineMotionDiffuse, making full use of fine-grained textual information. Our quantitative evaluation shows that FineMotionDiffuse trained on FineHumanML3D improves FID by a large margin of 0.38, compared with competitive baselines. According to the qualitative evaluation and case study, our model outperforms MotionDiffuse in generating spatially or chronologically composite motions, by learning the implicit mappings from fine-grained descriptions to the corresponding basic motions. We release our data at https://github.com/KunhangL/finemotiondiffuse.

Mingxu Tao, Dongyan Zhao, Yansong Feng

Open-ended question answering requires models to find appropriate evidence to form wellreasoned, comprehensive and helpful answers. In practical applications, models also need to engage in extended discussions on potential scenarios closely relevant to the question. With augmentation of retrieval module, open-source Large Language Models (LLMs) can produce coherent answers often with different focuses, but are still sub-optimal in terms of reliable evidence selection and in-depth question analysis. In this paper, we propose a novel Chain-ofDiscussion framework to leverage the synergy among multiple open-source LLMs aiming to provide more correct and more comprehensive answers for open-ended QA, although they are not strong enough individually. Our experiments show that discussions among multiple LLMs play a vital role in enhancing the quality of answers.

Zirui Wu, Lin Zheng, Zhihui Xie et al.

Diffusion Language Models (DLMs) present a compelling alternative to autoregressive models, offering flexible, any-order infilling without specialized prompting design. However, their practical utility is blocked by a critical limitation: the requirement of a fixed-length masked sequence for generation. This constraint severely degrades code infilling performance when the predefined mask size mismatches the ideal completion length. To address this, we propose DreamOn, a novel diffusion framework that enables dynamic, variable-length generation. DreamOn augments the diffusion process with two length control states, allowing the model to autonomously expand or contract the output length based solely on its own predictions. We integrate this mechanism into existing DLMs with minimal modifications to the training objective and no architectural changes. Built upon Dream-Coder-7B and DiffuCoder-7B, DreamOn achieves infilling performance on par with state-of-the-art autoregressive models on HumanEval-Infilling and SantaCoder-FIM and matches oracle performance achieved with ground-truth length. Our work removes a fundamental barrier to the practical deployment of DLMs, significantly advancing their flexibility and applicability for variable-length generation. Our code is available at https://github.com/DreamLM/DreamOn.

Zirui Wu, Xiao Liu, Jiayi Li et al.

While Large Language Model-based agents have demonstrated substantial progress in task completion, existing evaluation benchmarks tend to overemphasize single-task performance, with insufficient attention given to the crucial aspects of multitask planning and execution efficiency required in real-world scenarios. To bridge this gap, we present Recipe2Plan, a novel benchmark framework based on real-world cooking scenarios. Unlike conventional benchmarks, Recipe2Plan challenges agents to optimize cooking time through parallel task execution while respecting temporal constraints i.e. specific actions need to be performed within a particular time intervals following the preceding steps. Overly aggressive local parallelization may disrupt this constraint, potentially compromising the entire cooking process. This strict time constraint between actions raises a unique challenge for agents to balance between maximizing concurrent operations and adhering to critical timing constraints. Extensive experiments with state-of-the-art models reveal challenges in maintaining this balance between efficiency and feasibility. The results highlight the need for improved temporal awareness and global multitasking capabilities in large language models. We open-source our benchmark and code at https://github.com/WilliamZR/Recipe2Plan.

Huanghai Liu, Quzhe Huang, Qingjing Chen et al. · pku

In recent years, Large Language Models (LLMs) have been widely applied to legal tasks. To enhance their understanding of legal texts and improve reasoning accuracy, a promising approach is to incorporate legal theories. One of the most widely adopted theories is the Four-Element Theory (FET), which defines the crime constitution through four elements: Subject, Object, Subjective Aspect, and Objective Aspect. While recent work has explored prompting LLMs to follow FET, our evaluation demonstrates that LLM-generated four-elements are often incomplete and less representative, limiting their effectiveness in legal reasoning. To address these issues, we present JUREX-4E, an expert-annotated four-element knowledge base covering 155 criminal charges. The annotations follow a progressive hierarchical framework grounded in legal source validity and incorporate diverse interpretive methods to ensure precision and authority. We evaluate JUREX-4E on the Similar Charge Disambiguation task and apply it to Legal Case Retrieval. Experimental results validate the high quality of JUREX-4E and its substantial impact on downstream legal tasks, underscoring its potential for advancing legal AI applications. The dataset and code are available at: https://github.com/THUlawtech/JUREX

Chen Zhang, Mingxu Tao, Zhiyuan Liao et al. · pku

Large language models (LLMs) excel in high-resource languages but struggle with low-resource languages (LRLs), particularly those spoken by minority communities in China, such as Tibetan, Uyghur, Kazakh, and Mongolian. To systematically track the progress in these languages, we introduce MiLiC-Eval, a benchmark designed for minority languages in China, featuring 24K instances across 9 tasks. MiLiC-Eval focuses on underrepresented writing systems. Its parallelism between tasks and languages can provide a faithful and fine-grained assessment of linguistic and problem-solving skills. Our evaluation reveals that open-source LLMs perform poorly on syntax-intensive tasks and multi-script languages. We further demonstrate how MiLiC-Eval can help advance LRL research in handling diverse writing systems and understanding the process of language adaptation.

Xiao Liu, Yansong Feng, Kai-Wei Chang · pku

The argument sufficiency assessment task aims to determine if the premises of a given argument support its conclusion. To tackle this task, existing works often train a classifier on data annotated by humans. However, annotating data is laborious, and annotations are often inconsistent due to subjective criteria. Motivated by the definition of probability of sufficiency (PS) in the causal literature, we proposeCASA, a zero-shot causality-driven argument sufficiency assessment framework. PS measures how likely introducing the premise event would lead to the conclusion when both the premise and conclusion events are absent. To estimate this probability, we propose to use large language models (LLMs) to generate contexts that are inconsistent with the premise and conclusion and revise them by injecting the premise event. Experiments on two logical fallacy detection datasets demonstrate that CASA accurately identifies insufficient arguments. We further deploy CASA in a writing assistance application, and find that suggestions generated by CASA enhance the sufficiency of student-written arguments. Code and data are available at https://github.com/xxxiaol/CASA.

Yutong Hu, Kangcheng Luo, Yansong Feng

Despite remarkable performance in legal consultation exhibited by legal Large Language Models(LLMs) combined with legal article retrieval components, there are still cases when the advice given is incorrect or baseless. To alleviate these problems, we propose {\bf ELLA}, a tool for {\bf E}mpowering {\bf L}LMs for interpretable, accurate, and informative {\bf L}egal {\bf A}dvice. ELLA visually presents the correlation between legal articles and LLM's response by calculating their similarities, providing users with an intuitive legal basis for the responses. Besides, based on the users' queries, ELLA retrieves relevant legal articles and displays them to users. Users can interactively select legal articles for LLM to generate more accurate responses. ELLA also retrieves relevant legal cases for user reference. Our user study shows that presenting the legal basis for the response helps users understand better. The accuracy of LLM's responses also improves when users intervene in selecting legal articles for LLM. Providing relevant legal cases also aids individuals in obtaining comprehensive information.

Jiuheng Lin, Cong Jiang, Zirui Wu et al.

Training expert LLMs in domains with scarce data is difficult, often relying on multiple-choice questions (MCQs). However, standard outcome-based reinforcement learning (RL) on MCQs is risky. While it may improve accuracy, we observe it often degrades reasoning quality such as logical consistency. Existing solutions to supervise reasoning, such as large-scale Process Reward Models (PRMs), are prohibitively expensive. To address this, we propose CLARity, a cost-effective RL framework that enhances reasoning quality using only a small, general-purpose LLM. CLARity integrates a consistency-aware reward mechanism with a 2-stage refine-then-monitor training pipeline to enhance reasoning consistency, and a dynamic data reformulation strategy to to better exploit limited data. Experiments demonstrate that CLARity improves response consistency by 16.5% and accuracy by 7.5% over baselines. Human evaluations further confirm holistic improvements in coherence and professionalism. Thus, CLARity offers a generalizable solution that enables smaller models to effectively guide expert models by reasoning consistency.Our code is open sourced at: https://github.com/Infinite-set/CLARity

Qin Wang, Youhuan Li, Yansong Feng et al.

People without a database background usually rely on file systems or tools such as Excel for data management, which often lead to redundancy and data inconsistency. Relational databases possess strong data management capabilities, but require a high level of professional expertise from users. Although there are already many works on Text2SQL to automate the translation of natural language into SQL queries for data manipulation, all of them presuppose that the database schema is pre-designed. In practice, schema design itself demands domain expertise, and research on directly generating schemas from textual requirements remains unexplored. In this paper, we systematically define a new problem, called Text2Schema, to convert a natural language text requirement into a relational database schema. With an effective Text2Schema technique, users can effortlessly create database table structures using natural language, and subsequently leverage existing Text2SQL techniques to perform data manipulations, which significantly narrows the gap between non-technical personnel and highly efficient, versatile relational database systems. We propose SchemaAgent, an LLM-based multi-agent framework for Text2Schema. We emulate the workflow of manual schema design by assigning specialized roles to agents and enabling effective collaboration to refine their respective subtasks. We also incorporate dedicated roles for reflection and inspection, along with an innovative error detection and correction mechanism to identify and rectify issues across various phases. Moreover, we build and open source a benchmark containing 381 pairs of requirement description and schema. Experimental results demonstrate the superiority of our approach over comparative work.

Chengang Hu, Xiao Liu, Yansong Feng

Most of the existing compositional generalization datasets are synthetically-generated, resulting in a lack of natural language variation. While there have been recent attempts to introduce non-synthetic datasets for compositional generalization, they suffer from either limited data scale or a lack of diversity in the forms of combinations. To better investigate compositional generalization with more linguistic phenomena and compositional diversity, we propose the DIsh NamE Recognition (DiNeR) task and create a large realistic Chinese dataset. Given a recipe instruction, models are required to recognize the dish name composed of diverse combinations of food, actions, and flavors. Our dataset consists of 3,811 dishes and 228,114 recipes, and involves plenty of linguistic phenomena such as anaphora, omission and ambiguity. We provide two strong baselines based on T5 and large language models (LLMs). This work contributes a challenging task, baseline methods to tackle the task, and insights into compositional generalization in the context of dish name recognition. Code and data are available at https://github.com/Jumpy-pku/DiNeR.

Quzhe Huang, Yutong Hu, Shengqi Zhu et al.

Event temporal relation extraction~(ETRE) is usually formulated as a multi-label classification task, where each type of relation is simply treated as a one-hot label. This formulation ignores the meaning of relations and wipes out their intrinsic dependency. After examining the relation definitions in various ETRE tasks, we observe that all relations can be interpreted using the start and end time points of events. For example, relation \textit{Includes} could be interpreted as event 1 starting no later than event 2 and ending no earlier than event 2. In this paper, we propose a unified event temporal relation extraction framework, which transforms temporal relations into logical expressions of time points and completes the ETRE by predicting the relations between certain time point pairs. Experiments on TB-Dense and MATRES show significant improvements over a strong baseline and outperform the state-of-the-art model by 0.3\% on both datasets. By representing all relations in a unified framework, we can leverage the relations with sufficient data to assist the learning of other relations, thus achieving stable improvement in low-data scenarios. When the relation definitions are changed, our method can quickly adapt to the new ones by simply modifying the logic expressions that map time points to new event relations. The code is released at \url{https://github.com/AndrewZhe/A-Unified-Framework-for-ETRE}.

Quzhe Huang, Shengqi Zhu, Yansong Feng et al.

Document-level Relation Extraction (RE) is a more challenging task than sentence RE as it often requires reasoning over multiple sentences. Yet, human annotators usually use a small number of sentences to identify the relationship between a given entity pair. In this paper, we present an embarrassingly simple but effective method to heuristically select evidence sentences for document-level RE, which can be easily combined with BiLSTM to achieve good performance on benchmark datasets, even better than fancy graph neural network based methods. We have released our code at https://github.com/AndrewZhe/Three-Sentences-Are-All-You-Need.

Yuxuan Lai, Yijia Liu, Yansong Feng et al.

Chinese pre-trained language models usually process text as a sequence of characters, while ignoring more coarse granularity, e.g., words. In this work, we propose a novel pre-training paradigm for Chinese -- Lattice-BERT, which explicitly incorporates word representations along with characters, thus can model a sentence in a multi-granularity manner. Specifically, we construct a lattice graph from the characters and words in a sentence and feed all these text units into transformers. We design a lattice position attention mechanism to exploit the lattice structures in self-attention layers. We further propose a masked segment prediction task to push the model to learn from rich but redundant information inherent in lattices, while avoiding learning unexpected tricks. Experiments on 11 Chinese natural language understanding tasks show that our model can bring an average increase of 1.5% under the 12-layer setting, which achieves new state-of-the-art among base-size models on the CLUE benchmarks. Further analysis shows that Lattice-BERT can harness the lattice structures, and the improvement comes from the exploration of redundant information and multi-granularity representations. Our code will be available at https://github.com/alibaba/pretrained-language-models/LatticeBERT.

Yuxi Xie, Liangming Pan, Dongzhe Wang et al.

Recent question generation (QG) approaches often utilize the sequence-to-sequence framework (Seq2Seq) to optimize the log-likelihood of ground-truth questions using teacher forcing. However, this training objective is inconsistent with actual question quality, which is often reflected by certain global properties such as whether the question can be answered by the document. As such, we directly optimize for QG-specific objectives via reinforcement learning to improve question quality. We design three different rewards that target to improve the fluency, relevance, and answerability of generated questions. We conduct both automatic and human evaluations in addition to a thorough analysis to explore the effect of each QG-specific reward. We find that optimizing question-specific rewards generally leads to better performance in automatic evaluation metrics. However, only the rewards that correlate well with human judgement (e.g., relevance) lead to real improvement in question quality. Optimizing for the others, especially answerability, introduces incorrect bias to the model, resulting in poor question quality. Our code is publicly available at https://github.com/YuxiXie/RL-for-Question-Generation.

Liangming Pan, Yuxi Xie, Yansong Feng et al.

This paper proposes the problem of Deep Question Generation (DQG), which aims to generate complex questions that require reasoning over multiple pieces of information of the input passage. In order to capture the global structure of the document and facilitate reasoning, we propose a novel framework which first constructs a semantic-level graph for the input document and then encodes the semantic graph by introducing an attention-based GGNN (Att-GGNN). Afterwards, we fuse the document-level and graph-level representations to perform joint training of content selection and question decoding. On the HotpotQA deep-question centric dataset, our model greatly improves performance over questions requiring reasoning over multiple facts, leading to state-of-the-art performance. The code is publicly available at https://github.com/WING-NUS/SG-Deep-Question-Generation.

Yao Fu, Yansong Feng, John P. Cunningham

Paraphrase generation is a longstanding important problem in natural language processing. In addition, recent progress in deep generative models has shown promising results on discrete latent variables for text generation. Inspired by variational autoencoders with discrete latent structures, in this work, we propose a latent bag of words (BOW) model for paraphrase generation. We ground the semantics of a discrete latent variable by the BOW from the target sentences. We use this latent variable to build a fully differentiable content planning and surface realization model. Specifically, we use source words to predict their neighbors and model the target BOW with a mixture of softmax. We use Gumbel top-k reparameterization to perform differentiable subset sampling from the predicted BOW distribution. We retrieve the sampled word embeddings and use them to augment the decoder and guide its generation search space. Our latent BOW model not only enhances the decoder, but also exhibits clear interpretability. We show the model interpretability with regard to \emph{(i)} unsupervised learning of word neighbors \emph{(ii)} the step-by-step generation procedure. Extensive experiments demonstrate the transparent and effective generation process of this model.\footnote{Our code can be found at \url{https://github.com/FranxYao/dgm_latent_bow}}

Yuan Ye, Yansong Feng, Bingfeng Luo et al.

Recent years have seen rapid progress in identifying predefined relationship between entity pairs using neural networks NNs. However, such models often make predictions for each entity pair individually, thus often fail to solve the inconsistency among different predictions, which can be characterized by discrete relation constraints. These constraints are often defined over combinations of entity-relation-entity triples, since there often lack of explicitly well-defined type and cardinality requirements for the relations. In this paper, we propose a unified framework to integrate relation constraints with NNs by introducing a new loss term, ConstraintLoss. Particularly, we develop two efficient methods to capture how well the local predictions from multiple instance pairs satisfy the relation constraints. Experiments on both English and Chinese datasets show that our approach can help NNs learn from discrete relation constraints to reduce inconsistency among local predictions, and outperform popular neural relation extraction NRE models even enhanced with extra post-processing. Our source code and datasets will be released at https://github.com/PKUYeYuan/Constraint-Loss-AAAI-2020.

Zechang Li, Yuxuan Lai, Yuxi Xie et al.

This paper presents our semantic parsing system for the evaluation task of open domain semantic parsing in NLPCC 2019. Many previous works formulate semantic parsing as a sequence-to-sequence(seq2seq) problem. Instead, we treat the task as a sketch-based problem in a coarse-to-fine(coarse2fine) fashion. The sketch is a high-level structure of the logical form exclusive of low-level details such as entities and predicates. In this way, we are able to optimize each part individually. Specifically, we decompose the process into three stages: the sketch classification determines the high-level structure while the entity labeling and the matching network fill in missing details. Moreover, we adopt the seq2seq method to evaluate logical form candidates from an overall perspective. The co-occurrence relationship between predicates and entities contribute to the reranking as well. Our submitted system achieves the exactly matching accuracy of 82.53% on full test set and 47.83% on hard test subset, which is the 3rd place in NLPCC 2019 Shared Task 2. After optimizations for parameters, network structure and sampling, the accuracy reaches 84.47% on full test set and 63.08% on hard test subset(Our code and data are available at https://github.com/zechagl/NLPCC2019-Semantic-Parsing).

Jiayi Li, Xiao Liu, Yansong Feng

Large Language Model (LLM)-based multi-agent systems are increasingly used to simulate human interactions and solve collaborative tasks. A common practice is to assign agents with personas to encourage behavioral diversity. However, this raises a critical yet underexplored question: do personas introduce biases into multi-agent interactions? This paper presents a systematic investigation into persona-induced biases in multi-agent interactions, with a focus on social traits like trustworthiness (how an agent's opinion is received by others) and insistence (how strongly an agent advocates for its opinion). Through a series of controlled experiments in collaborative problem-solving and persuasion tasks, we reveal that (1) LLM-based agents exhibit biases in both trustworthiness and insistence, with personas from historically advantaged groups (e.g., men and White individuals) perceived as less trustworthy and demonstrating less insistence; and (2) agents exhibit significant in-group favoritism, showing a higher tendency to conform to others who share the same persona. These biases persist across various LLMs, group sizes, and numbers of interaction rounds, highlighting an urgent need for awareness and mitigation to ensure the fairness and reliability of multi-agent systems.

Kangcheng Luo, Tinglang Wu, Yansong Feng

Recent search-augmented LLMs trained with reinforcement learning (RL) can interleave searching and reasoning for multi-hop reasoning tasks. However, they face two critical failure modes as the accumulating context becomes flooded with both crucial evidence and irrelevant information: (1) ineffective search chain construction that produces incorrect queries or omits retrieval of critical information, and (2) reasoning hijacking by peripheral evidence that causes models to misidentify distractors as valid evidence. To address these challenges, we propose **D$^2$Plan**, a **D**ual-agent **D**ynamic global **Plan**ning paradigm for complex retrieval-augmented reasoning. **D$^2$Plan** operates through the collaboration of a *Reasoner* and a *Purifier*: the *Reasoner* constructs explicit global plans during reasoning and dynamically adapts them based on retrieval feedback; the *Purifier* assesses retrieval relevance and condenses key information for the *Reasoner*. We further introduce a two-stage training framework consisting of supervised fine-tuning (SFT) cold-start on synthesized trajectories and RL with plan-oriented rewards to teach LLMs to master the **D$^2$Plan** paradigm. Extensive experiments demonstrate that **D$^2$Plan** enables more coherent multi-step reasoning and stronger resilience to irrelevant information, thereby achieving superior performance on challenging QA benchmarks.

Chen Zhang, Xiao Liu, Jiuheng Lin et al. · pku

Existing large language models struggle to support numerous low-resource languages, particularly the extremely low-resource ones, for which there is minimal training data available for effective parameter updating. We thus investigate whether LLMs can learn a new language on the fly solely through prompting. To study this question, we collect a research suite for Zhuang, a language supported by no LLMs currently. We introduce DiPMT++, a framework for adapting LLMs to unseen languages by in-context learning. Using a dictionary and 5K parallel sentences only, DiPMT++ significantly enhances the performance of GPT-4 from 0 to 16 BLEU for Chinese-to-Zhuang translation and achieves 32 BLEU for Zhuang-to-Chinese translation. We also validate the effectiveness of our framework on Kalamang, another unseen language. Furthermore, we demonstrate the practical utility of DiPMT++ in aiding humans in translating completely unseen languages, which could contribute to the preservation of linguistic diversity.

Chengjun Pan, Shichun Liu, Jiahang Lin et al.

Reinforcement learning (RL) for LLM post-training faces a fundamental design choice: whether to use a learned critic as a baseline for policy optimization. Classical theory favors critic-based methods such as PPO for variance reduction, yet critic-free alternatives like GRPO have gained widespread adoption due to their simplicity and competitive performance. We show that in sparse-reward settings, a learned critic can inject estimation noise that exceeds the state signal it captures, increasing rather than reducing advantage variance. By casting baseline selection as a Kalman filtering problem, we unify PPO and GRPO as two extremes of the Kalman gain and prove that explained variance (EV), computable from a single training batch, identifies the exact boundary: positive EV indicates the critic reduces variance, while zero or negative EV signals that it inflates variance. Building on this insight, we propose Explained Variance Policy Optimization (EVPO), which monitors batch-level EV at each training step and adaptively switches between critic-based and batch-mean advantage estimation, provably achieving no greater variance than the better of the two at every step. Across four tasks spanning classical control, agentic interaction, and mathematical reasoning, EVPO consistently outperforms both PPO and GRPO regardless of which fixed baseline is stronger on a given task. Further analysis confirms that the adaptive gating tracks critic maturation over training and that the theoretically derived zero threshold is empirically optimal.

Kangcheng Luo, Quzhe Huang, Cong Jiang et al. · pku

Interpreting the law is always essential for the law to adapt to the ever-changing society. It is a critical and challenging task even for legal practitioners, as it requires meticulous and professional annotations and summarizations by legal experts, which are admittedly time-consuming and expensive to collect at scale. To alleviate the burden on legal experts, we propose a method for automated legal interpretation. Specifically, by emulating doctrinal legal research, we introduce a novel framework, ATRIE, to address Legal Concept Interpretation, a typical task in legal interpretation. ATRIE utilizes large language models (LLMs) to AuTomatically Retrieve concept-related information, Interpret legal concepts, and Evaluate generated interpretations, eliminating dependence on legal experts. ATRIE comprises a legal concept interpreter and a legal concept interpretation evaluator. The interpreter uses LLMs to retrieve relevant information from previous cases and interpret legal concepts. The evaluator uses performance changes on Legal Concept Entailment, a downstream task we propose, as a proxy of interpretation quality. Automated and multifaceted human evaluations indicate that the quality of our interpretations is comparable to those written by legal experts, with superior comprehensiveness and readability. Although there remains a slight gap in accuracy, it can already assist legal practitioners in improving the efficiency of legal interpretation.

Chen Zhang, Zhiyuan Liao, Yansong Feng · pku

Despite substantial research efforts evaluating how well large language models~(LLMs) handle global cultural diversity, the mechanisms behind their cultural knowledge acquisition, particularly in multilingual settings, remain unclear. We study this question by investigating how cultural knowledge transfers across languages during language adaptation of LLMs. We introduce an interpretable framework for studying this transfer, ensuring training data transparency and controlling transfer effects. Through a study of four non-Anglophonic cultures, we observe bidirectional cultural transfer between English and other high-resource languages, while low-resource languages primarily transfer knowledge to English with limited reverse flow. To explain this asymmetric phenomenon, we propose a frequency-based hypothesis: cultural knowledge appearing more frequently in the pretraining data transfers more easily, which is supported by empirical analysis of the training corpora.

Yiming Gao, Yansong Feng, Honggang Hu et al.

We study the \emph{Subset Balancing} problem: given $\mathbf{x} \in \mathbb{Z}^n$ and a coefficient set $C \subseteq \mathbb{Z}$, find a nonzero vector $\mathbf{c} \in C^n$ such that $\mathbf{c}\cdot\mathbf{x} = 0$. The standard meet-in-the-middle algorithm runs in time $\tilde{O}(|C|^{n/2})=\tilde{O}(2^{n\log |C|/2})$, and recent improvements (SODA~2022, Chen, Jin, Randolph, and Servedio; STOC~2026, Randolph and Węgrzycki) beyond this barrier apply mainly when $d$ is constant. We give a reduction from Subset Balancing with $C = \{-d, \dots, d\}$ to a single instance of $\mathrm{SVP}_{\infty}$ in dimension $n+1$, which yields a deterministic algorithm with running time $\tilde{O}((6\sqrt{2πe})^n) \approx \tilde{O}(2^{4.632n})$, and a randomized algorithm with running time $\tilde{O}(2^{2.443n})$ (here $\tilde{O}$ suppresses $\operatorname{poly}(n)$ factors). We also show that for sufficiently large $d$, Subset Balancing is solvable in polynomial time. More generally, we extend the box constraint $[-d,d]^n$ to an arbitrary centrally symmetric convex body $K \subseteq \mathbb{R}^n$ with a deterministic $\tilde{O}(2^{c_K n})$-time algorithm, where $c_K$ depends only on the shape of $K$. We further study the \emph{Generalized Subset Sum} problem of finding $\mathbf{c} \in C^n$ such that $\mathbf{c} \cdot \mathbf{x} = τ$. For $C = \{-d, \dots, d\}$, we reduce the worst-case problem to a single instance of $\mathrm{CVP}_{\infty}$. Although no general single exponential time algorithm is known for exact $\mathrm{CVP}_{\infty}$, we show that in the average-case setting, for both $C = \{-d, \dots, d\}$ and $C = \{-d, \dots, d\} \setminus \{0\}$, the embedded instance satisfies a bounded-distance promise with high probability. This yields a deterministic algorithm running in time $\tilde{O}((18\sqrt{2πe})^n) \approx \tilde{O}(2^{6.217n})$.

Chen Zhang, Jiuheng Lin, Xiao Liu et al. · pku

While large language models (LLMs) have shown promise in translating extremely low-resource languages using resources like dictionaries, the effectiveness of grammar books remains debated. This paper investigates the role of grammar books in translating extremely low-resource languages by decomposing it into two key steps: grammar rule retrieval and application. To facilitate the study, we introduce ZhuangRules, a modularized dataset of grammar rules and their corresponding test sentences. Our analysis reveals that rule retrieval constitutes a primary bottleneck in grammar-based translation. Moreover, although LLMs can apply simple rules for translation when explicitly provided, they encounter difficulties in handling more complex rules. To address these challenges, we propose representing grammar rules as code functions, considering their similarities in structure and the benefit of code in facilitating LLM reasoning. Our experiments show that using code rules significantly boosts both rule retrieval and application, ultimately resulting in a 13.1% BLEU improvement in translation.

Xiao Liu, Xinyi Dong, Xinyang Gao et al.

Recent advancements in large language models (LLMs) have shown promise in generating novel research ideas. However, these ideas often face challenges related to feasibility and expected effectiveness. This paper explores how augmenting LLMs with relevant data during the idea generation process can enhance the quality of generated ideas. We introduce two ways of incorporating data: (1) providing metadata during the idea generation stage to guide LLMs toward feasible directions, and (2) adding automatic validation during the idea selection stage to assess the empirical plausibility of hypotheses within ideas. We conduct experiments in the social science domain, specifically with climate negotiation topics, and find that metadata improves the feasibility of generated ideas by 20%, while automatic validation improves the overall quality of selected ideas by 7%. A human study shows that LLM-generated ideas, along with their related data and validation processes, inspire researchers to propose research ideas with higher quality. Our work highlights the potential of data-driven research idea generation, and underscores the practical utility of LLM-assisted ideation in real-world academic settings.

Mingxu Tao, Jie Hu, Mingchuan Yang et al.

The remarkable performance of Large language models (LLMs) relies heavily on the availability of abundant high-quality training data. However, the high cost of acquiring annotated data often prevents models from obtaining capabilities to tackle downstream tasks. In this paper, we introduce a novel method, EpiCoDe that boosts model performance in data-scarcity scenarios without extra training. We first employ model extrapolation to enhance a finetuned model with its inferior version, and then adopt contrastive decoding to further reduce predicted errors, by comparing the logit scores given by the extrapolated and the vanilla finetuned model. Experiments across three tasks over four different LLMs show that EpiCoDe consistently outperforms existing methods with significant and robust improvement. We also propose a new theoretical framework to reveal the mechanism behind contrastive decoding in data-scarcity scenarios, which further helps us better understand the effectiveness of EpiCoDe.

Xiao Liu, Da Yin, Zirui Wu et al.

Tools enhance the reasoning capabilities of large language models (LLMs) in complex problem-solving tasks, but not all tasks have available tools. In the absence of predefined tools, prior works have explored instructing LLMs to generate tools on their own. However, such approaches rely heavily on the models' internal knowledge and would fail in domains beyond the LLMs' knowledge scope. To address this limitation, we propose RefTool, a reference-guided framework for automatic tool creation that leverages structured external materials such as textbooks. RefTool consists of two modules: (1) tool creation, where LLMs generate executable tools from reference content, validate them using illustrative examples, and organize them hierarchically into a toolbox; and (2) tool utilization, where LLMs navigate the toolbox structure to select and apply the appropriate tools to solve problems. Experiments on causality, physics, and chemistry benchmarks demonstrate that RefTool outperforms existing tool-creation and domain-specific reasoning methods by 11.3% on average accuracy, while being cost-efficient and broadly generalizable. Analyses reveal that grounding tool creation in references produces accurate and faithful tools, and that the hierarchical structure facilitates effective tool selection. RefTool enables LLMs to overcome knowledge limitations, demonstrating the value of grounding tool creation in external references for enhanced and generalizable reasoning.

Kunhang Li, Jason Naradowsky, Yansong Feng et al.

We explore the human motion knowledge of Large Language Models (LLMs) through 3D avatar control. Given a motion instruction, we prompt LLMs to first generate a high-level movement plan with consecutive steps (High-level Planning), then specify body part positions in each step (Low-level Planning), which we linearly interpolate into avatar animations. Using 20 representative motion instructions that cover fundamental movements and balance body part usage, we conduct comprehensive evaluations, including human and automatic scoring of both high-level movement plans and generated animations, as well as automatic comparison with oracle positions in low-level planning. Our findings show that LLMs are strong at interpreting high-level body movements but struggle with precise body part positioning. While decomposing motion queries into atomic components improves planning, LLMs face challenges in multi-step movements involving high-degree-of-freedom body parts. Furthermore, LLMs provide reasonable approximations for general spatial descriptions, but fall short in handling precise spatial specifications. Notably, LLMs demonstrate promise in conceptualizing creative motions and distinguishing culturally specific motion patterns.