Underlying data distributions of natural language, programming code, and mathematical symbols vary vastly, presenting a complex challenge for large language models (LLMs) that strive to achieve high performance across all three domains simultaneously. Achieving a very high level of proficiency for an LLM within a specific domain often requires extensive training with relevant corpora, which is typically accompanied by a sacrifice in performance in other domains. In this paper, we aim to “play the dealt cards well” and propose to fuse models that are already highly-specialized directly. The proposed fusing framework, , consists of different distinct specialists that are already sufficiently trained on different domains (we mainly focus on language, coding, and mathematics in this paper). A token-level gating mechanism is introduced to blend the specialists’ outputs. A two-stage training strategy accompanied by balanced sampling is designed to ensure stability. To effectively train the fused model, we further construct a high-quality supervised instruction tuning dataset, , which includes text, code, and mathematical content. This dataset comprises approximately 300,000 instructions and covers a wide range of topics in each domain. Experiments show that our model could simultaneously achieve mastery of the three crucial domains.

This study investigates the structured generation capabilities of large language models (LLMs), focusing on producing valid JSON outputs against a given schema. Despite the widespread use of JSON in integrating language models with programs, there is a lack of comprehensive analysis and benchmarking of these capabilities. We explore various aspects of JSON generation, such as structure understanding, escaping, and natural language description, to determine how to assess and enable LLMs to generate valid responses. Building upon this, we propose SchemaBench features around 40K different JSON schemas to obtain and assess models’ abilities in generating valid JSON. We find that the latest LLMs are still struggling to generate a valid JSON string. Moreover, we demonstrate that incorporating reinforcement learning with a Fine-grained Schema Validator can further enhance models’ understanding of JSON schema, leading to improved performance. Our models demonstrate significant improvement in both generating JSON outputs and downstream tasks.

Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in chart understanding tasks. However, interpreting charts with textual descriptions often leads to information loss, as it fails to fully capture the dense information embedded in charts. In contrast, parsing charts into code provides lossless representations that can effectively contain all critical details. Although existing open-source MLLMs have achieved success in chart understanding tasks, they still face two major challenges when applied to chart-to-code tasks: (1) Low executability and poor restoration of chart details in the generated code and (2) Lack of large-scale and diverse training data. To address these challenges, we propose ChartCoder, the first dedicated chart-to-code MLLM, which leverages Code LLMs as the language backbone to enhance the executability of the generated code. Furthermore, we introduce Chart2Code-160k, the first large-scale and diverse dataset for chart-to-code generation, and propose the Snippet-of-Thought (SoT) method, which transforms direct chart-to-code generation data into step-by-step generation. Experiments demonstrate that ChartCoder, with only 7B parameters, surpasses existing open-source MLLMs on chart-to-code benchmarks, achieving superior chart restoration and code excitability. Our code is available at https://github.com/thunlp/ChartCoder.

Retrieval-Augmented Generation (RAG) is a powerful approach that enables large language models (LLMs) to incorporate external knowledge. However, evaluating the effectiveness of RAG systems in specialized scenarios remains challenging due to the high costs of data construction and the lack of suitable evaluation metrics. This paper introduces RAGEval, a framework designed to assess RAG systems across diverse scenarios by generating high-quality documents, questions, answers, and references through a schema-based pipeline. With a focus on factual accuracy, we propose three novel metrics—Completeness, Hallucination, and Irrelevance—to evaluate LLM-generated responses rigorously. Experimental results show that RAGEval outperforms zero-shot and one-shot methods in terms of clarity, safety, conformity, and richness of generated samples. Furthermore, the use of LLMs for scoring the proposed metrics demonstrates a high level of consistency with human evaluations. RAGEval establishes a new paradigm for evaluating RAG systems in real-world applications. The code and dataset are released at https://github.com/OpenBMB/RAGEval.

While long-context inference is crucial for advancing large language model (LLM) applications, its prefill speed remains a significant bottleneck. Current approaches, including sequence parallelism strategies and compute reduction through approximate attention mechanisms, still fall short of delivering optimal inference efficiency. This hinders scaling the inputs to longer sequences and processing long-context queries in a timely manner. To address this, we introduce APB, an efficient long-context inference framework that leverages multi-host approximate attention to enhance prefill speed by reducing compute and enhancing parallelism simultaneously. APB introduces a communication mechanism for essential key-value pairs within a sequence parallelism framework, enabling a faster inference speed while maintaining task performance. We implement APB by incorporating a tailored FlashAttn kernel alongside optimized distribution strategies, supporting diverse models and parallelism configurations. APB achieves speedups of up to 9.2×, 4.2×, and 1.6× compared with FlashAttn, RingAttn, and StarAttn, respectively, without any observable task performance degradation.

Large Language Models (LLMs) excel in traditional natural language processing tasks but struggle with problems that require complex domain-specific calculations or simulations. While equipping LLMs with external tools to build LLM-based agents can enhance their capabilities, existing approaches lack the flexibility to address diverse and ever-evolving user queries in open domains. Currently, there is also no existing dataset that evaluates LLMs on open-domain knowledge that requires tools to solve. To this end, we introduce OpenAct benchmark to evaluate the open-domain task-solving capability, which is built on human expert consultation and repositories in GitHub. It comprises 339 questions spanning 7 diverse domains that need to be solved with domain-specific methods. In our experiments, even state-of-the-art LLMs and LLM-based agents demonstrate unsatisfactory success rates, underscoring the need for a novel approach.Furthermore, we present OpenAgent, a novel LLM-based agent system that can tackle evolving queries in open domains through autonomously integrating specialized tools from GitHub. OpenAgent employs 1) a hierarchical framework where specialized agents handle specific tasks and can assign tasks to inferior agents, 2) a bi-level experience learning mechanism to learn from both humans’ and its own experiences to tackle tool flaws. Experiments demonstrate its superior effectiveness and efficiency, which significantly outperforms baselines. Our data and code are open-source at https://github.com/OpenBMB/OpenAct.

Existing LLM-based agents have achieved strong performance on held-in tasks, but their generalizability to unseen tasks remains poor. Hence, some recent work focus on fine-tuning the policy model with more diverse tasks to improve the generalizability. In this work, we find that finetuning a reward model to guide the policy model is more robust than directly finetuning the policy model.Based on this finding, we propose AgentRM, a 8B generalizable reward model, to guide the policy model for effective test-time search.We comprehensively investigate three approaches to construct the reward model, including explicit reward modeling, implicit reward modeling and LLM-as-a-judge.We then use AgentRM to guide the answer generation with Best-of-N sampling and beam search.We show that AgentRM is robust to paraphrasings of task instructions and can generalize to unseen tasks that require novel optimal behavior.Through extensive evaluation across nine tasks spanning four categories, AgentRM enhances the non-finetuned 8B policy model by 8.8 points on average, surpassing the top general agent by 4.0.Moreover, it demonstrates weak-to-strong generalization, yielding greater improvement on more powerful policy models.As for the specializability, AgentRM can also boost a finetuned policy model and outperform the top specialized agent by 11.4 on three held-in tasks.Further analysis verifies its effectiveness in test-time scaling.We release the code and data at https://github.com/thunlp/AgentRM.

Utilizing Graphic User Interfaces (GUIs) for human-computer interaction is essential for accessing various digital tools. Recent advancements in Vision Language Models (VLMs) reveal significant potential for developing versatile agents that assist humans in navigating GUIs. However, current VLMs face challenges related to fundamental abilities, such as OCR and grounding, as well as a lack of knowledge about GUI elements functionalities and control methods. These limitations hinder their effectiveness as practical GUI agents. To address these challenges, we introduce GUICourse, a series of datasets for training visual-based GUI agents using general VLMs. First, we enhance the OCR and grounding capabilities of VLMs using the GUIEnv dataset. Next, we enrich the GUI knowledge of VLMs using the GUIAct and GUIChat datasets. Our experiments demonstrate that even a small-sized GUI agent (with 3.1 billion parameters) performs effectively on both single-step and multi-step GUI tasks. We further finetune our GUI agents on other GUI tasks with different action spaces (AITW and Mind2Web), and the results show that our agents are better than their baseline VLMs. Additionally, we analyze the impact of OCR and grounding capabilities through an ablation study, revealing a positive correlation with GUI navigation ability.

We propose a training-free framework that enables large language models (LLMs) to effectively process long texts, using a divide-and-conquer strategy for comprehensive document understanding.The proposed LLM×MapReduce framework splits the entire document into several chunks for LLMs to read and then aggregates the intermediate outputs to produce the final response. The main challenge for divide-and-conquer long text processing frameworks lies in the risk of losing essential long-range information due to document splitting, which can lead the model to produce incomplete or incorrect answers based on the segmented texts.Disrupted long-range information can be classified into two categories: inter-chunk dependency and inter-chunk conflict.We design a structured information protocol to better cope with inter-chunk dependency and an in-context confidence calibration mechanism to resolve inter-chunk conflicts. Experiments demonstrate that LLM×MapReduce outperforms representative open-source and commercial long-context LLMs and is compatible with several models.Our framework can also function as a data synthesis engine, capable of generating high-quality long-alignment data using only short-context LLMs.

Grouped-Query Attention (GQA) is a widely adopted strategy for reducing the computational cost of attention layers in large language models (LLMs). However, current GQA configurations are often suboptimal because they overlook how context length influences inference cost. Since inference cost grows with context length, the most cost-efficient GQA configuration should vary accordingly. In this work, we analyze the relationship among context length, model size, GQA configuration, and model loss, and introduce two innovations: (1) we decouple the total head size from the hidden size, enabling more flexible control over attention FLOPs; and (2) we jointly optimize the model size and the GQA configuration to arrive at a better allocation of inference resources between attention layers and other components. Our analysis reveals that commonly used GQA configurations are highly suboptimal for long-context scenarios. Moreover, we propose a recipe for deriving cost-optimal GQA configurations. Our results show that for long-context scenarios, one should use fewer attention heads while scaling up the model size. Configurations selected by our recipe can reduce both memory usage and FLOPs by more than 50% compared to Llama-3’s GQA, with *no degradation in model capabilities*. Our findings offer valuable insights for designing efficient long-context LLMs.

As large language models (LLMs) increasingly exhibit human-like capabilities, a fundamental question emerges: How can we enable LLMs to learn the underlying patterns from limited examples in entirely novel environments and apply them effectively? This question is central to the ability of LLMs in inductive reasoning. Existing research on LLM-based inductive reasoning can be broadly categorized based on whether the underlying rules are expressible via explicit mathematical equations. However, many recent studies in the beyond-equations category have emphasized rule design without grounding them in specific scenarios. Inspired by the parallels between inductive reasoning and human scientific discovery, we propose the task of LLM-Based Scientific Inductive Reasoning Beyond Equations and introduce a new benchmark, SIRBench-V1, to evaluate the inductive reasoning abilities of LLMs in scientific settings. Our experimental results show that current LLMs still struggle with this task, underscoring its difficulty and the need for further advancement in this area.

Large language model agents have enabled GUI-based automation, particularly for mobile devices. However, deployment remains limited by noisy data, poor generalization, and lack of support for non-English GUIs. In this work, we present AgentCPM-GUI, an 8B-parameter GUI agent built for robust and efficient on-device GUI interaction. Our training pipeline includes grounding-aware pre-training to enhance perception, supervised fine-tuning on high-quality Chinese and English trajectories to imitate human-like actions, and reinforcement fine-tuning with GRPO to improve reasoning capability. AgentCPM-GUI achieves promising performance on five public benchmarks and our proposed Chinese benchmark CAGUI. To facilitate reproducibility and further research, we publicly release all code, model checkpoint, and evaluation data at: https://github.com/OpenBMB/AgentCPM-GUI

We introduce LLM×MapReduce-V3, a hierarchically modular agent system designed for long-form survey generation. Building on the prior work, LLM×MapReduce-V2, this version incorporates a multi-agent architecture where individual functional components, such as skeleton initialization, digest construction, and skeleton refinement, are implemented as independent model-context-protocol (MCP) servers. These atomic servers can be aggregated into higher-level servers, creating a hierarchically structured system. A high-level planner agent dynamically orchestrates the workflow by selecting appropriate modules based on their MCP tool descriptions and the execution history. This modular decomposition facilitates human-in-the-loop intervention, affording users greater control and customization over the research process. Through a multi-turn interaction, the system precisely captures the intended research perspectives to generate a comprehensive skeleton, which is then developed into an in-depth survey. Human evaluations demonstrate that our system surpasses representative baselines in both content depth and length, highlighting the strength of MCP-based modular planning. Demo, video and code are available at https://github.com/thunlp/LLMxMapReduce.

Code debugging is a vital stage of software development, essential for ensuring the reliability and performance of Large Language Models (LLMs) in the code generation task. Human debugging typically follows a multi-stage process, which includes Bug Localization, Bug Identification, Code Repair, and Code Recognition. However, existing code debugging benchmarks predominantly focus on the Code Repair stage, which offers only a limited perspective on evaluating the debugging capabilities of LLMs. In this paper, we introduce DEBUGEVAL, a comprehensive benchmark for evaluating the debugging abilities of LLMs by emulating the multi-stage human debugging process. Through evaluating on DEBUGEVAL, we observe that 7B-scale models consistently underperform compared to their larger counterparts, highlighting their limitations in comprehending code semantics. In this case, we propose the COmmunicative Agent-based data SynThesis (COAST) framework, which employs a multi-agent system to generate high-quality training data for supervised fine-tuning (SFT). Experimental results demonstrate that COAST-generated data outperform human-curated and GPT-4-generated data, enabling 7B-scale LLMs to achieve debugging performance comparable to GPT-3.5. All data and codes are available at https://github.com/NEUIR/COAST.

Parameter-Efficient Fine-Tuning (PEFT) methods such as LoRA have significantly improved the adaptation of LLMs to downstream tasksin a resource-efficient manner. However, in multi-task scenarios, challenges such as training imbalance and the seesaw effect frequently emerge. Mixture-of-LoRA (MoLoRA), which combines LoRA with sparse Mixture-of-Experts, mitigates some of these issues by promoting task-specific learning among experts. Despite this, MoLoRA remains inefficient in terms of training speed, parameter utilization, and overall multi-task performance. In this paper, we propose Mixture of Asymmetric Low-Rank Adaptaion (MALoRA), a flexible fine-tuning framework that leverages asymmetric optimization among LoRA experts. MALoRA reduces the number of trainable parameters by 30% to 48%, increases training speed by 1.2x, and matches the computational efficiency of single-task LoRA models. Additionally, MALoRA addresses overfitting issues commonly seen in high-rank configurations, enhancing performance stability. Extensive experiments across diverse multi-task learning scenarios demonstrate that MALoRA consistently outperforms all baseline methods in both inter-domain and intra-domain tasks.

Understanding alignment techniques begins with comprehending zero-shot generalization brought by instruction tuning, but little of the mechanism has been understood. Existing work has largely been confined to the task level, without considering that tasks are artificially defined and, to LLMs, merely consist of tokens and representations. To bridge this gap, we investigate zero-shot generalization from the perspective of the data itself. We first demonstrate that zero-shot generalization happens very early during instruction tuning, with loss serving as a stable indicator. Next, we investigate training data arrangement through similarity and granularity perspectives, confirming that the timing of exposure to certain training examples may greatly facilitate generalization on unseen tasks. Finally, we propose a more grounded training data arrangement framework, Test-centric Multi-turn Arrangement, and show its effectiveness in promoting continual learning and further loss reduction. For the first time, we show that zero-shot generalization during instruction tuning is a form of similarity-based generalization between training and test data at the instance level.

Retrieval-Augmented Generation (RAG) has proven its effectiveness in alleviating hallucinations for Large Language Models (LLMs). However, existing automated evaluation metrics cannot fairly evaluate the outputs generated by RAG models during training and evaluation. LLM-based judgment models provide the potential to produce high-quality judgments, but they are highly sensitive to evaluation prompts, leading to inconsistencies when judging the output of RAG models. This paper introduces the Judge-Consistency (ConsJudge) method, which aims to enhance LLMs to generate more accurate evaluations for RAG models. Specifically, ConsJudge prompts LLMs to generate different judgments based on various combinations of judgment dimensions, utilizes the judge-consistency to evaluate these judgments, and selects the chosen and rejected judgments for DPO training. Our experiments show that ConsJudge can effectively provide more accurate judgments for optimizing RAG models across various RAG models and datasets. Further analysis reveals that judgments generated by ConsJudge have a high agreement with the superior LLM. All codes are available at https://github.com/OpenBMB/ConsJudge.

The recent advancement of Multimodal Large Language Models (MLLMs) has significantly improved their fine-grained perception of single images and general comprehension across multiple images. However, existing MLLMs still face challenges in achieving precise grounding in complex multi-image scenarios. To address this, we first explore a Chain-of-Thought (CoT) framework that integrates single-image grounding with multi-image comprehension. While partially effective, it remains unstable and struggles to capture abstract visual information due to its non-end-to-end nature. Therefore, we introduce Migician, the first multi-image grounding model capable of performing free-form and accurate grounding across multiple images. To support this, we present the MGrounding-630k dataset, which comprises data for several multi-image grounding tasks derived from existing datasets, along with newly generated free-form grounding instruction-following data. Furthermore, we propose MIG-Bench, a comprehensive benchmark specifically designed for evaluating multi-image grounding capabilities. Experimental results demonstrate that our model achieves significantly superior multi-image grounding capabilities, outperforming the best existing MLLMs by 24.94% and even surpassing much larger 70B models. Our code, model, dataset, and benchmark are fully open-sourced at https://migician-vg.github.io/.

Large Language Model (LLM) based multi-agent systems (MAS) show remarkable potential in collaborative problem-solving, yet they still face critical challenges: low communication efficiency, poor scalability, and a lack of effective parameter-updating optimization methods. We present Optima, a novel framework that addresses these issues by significantly enhancing both communication efficiency and task effectiveness in LLM-based MAS through training. Optima employs an iterative generate, rank, select, and train paradigm with a reward function balancing task performance, token efficiency, and communication readability. We explore various algorithms, including Supervised Fine-Tuning, Direct Preference Optimization, and their hybrid approaches, providing insights into their effectiveness-efficiency trade-offs. We integrate Monte Carlo Tree Search-inspired techniques for DPO data generation, treating conversation turns as tree nodes to explore diverse interaction paths. Evaluated on common multi-agent tasks, including information-asymmetric question answering and complex reasoning, Optimashows consistent and substantial improvements over single-agent baselines and vanilla MAS based on Llama 3 8B / 3.2 3B, achieving up to 2.8x performance gain with less than 10% tokens on tasks requiring heavy information exchange. Moreover, Optima’s efficiency gains enable more effective compute utilization during inference, leading to improved inference-time scaling laws. By addressing fundamental challenges in LLM-based MAS, Optima shows the potential towards scalable, efficient, and effective MAS.

Web crawl is a main source of large language models’ (LLMs) pretraining data, but the majority of crawled web pages are discarded in pretraining due to low data quality. This paper presents Craw4LLM, an efficient web crawling method that explores the web graph based on the preference of LLM pretraining. Specifically, it leverages the influence of a webpage in LLM pretraining as the priority score of the web crawler’s scheduler, replacing the standard graph-connectivity-based priority. Our experiments on a web graph containing 900 million webpages from a commercial search engine’s index demonstrate the efficiency of Craw4LLM in obtaining high-quality pretraining data. With just 21% URLs crawled, LLMs pretrained on Craw4LLM data reach the same downstream performances of previous crawls, significantly reducing the crawling waste and alleviating the burdens on websites. Our code is publicly available at https://github.com/cxcscmu/Craw4LLM.

Triton, a high-level Python-like language designed for building efficient GPU kernels, is widely adopted in deep learning frameworks due to its portability, flexibility, and accessibility. However, programming and parallel optimization still require considerable trial and error from Triton developers. Despite advances in large language models (LLMs) for conventional code generation, these models struggle to generate accurate, performance-optimized Triton code, as they lack awareness of its specifications and the complexities of GPU programming. More critically, there is an urgent need for systematic evaluations tailored to Triton. In this work, we introduce TritonBench, the first comprehensive benchmark for Triton operator generation. TritonBench features two evaluation channels: a curated set of 184 real-world operators from GitHub and a collection of operators aligned with PyTorch interfaces. Unlike conventional code benchmarks prioritizing functional correctness, TritonBench also profiles efficiency performance on widely deployed GPUs aligned with industry applications. Our study reveals that current state-of-the-art code LLMs struggle to generate efficient Triton operators, highlighting a significant gap in high-performance code generation.

Large language models (LLMs) exhibit human-like intelligence, enabling them to simulate human behavior and support various applications that require both humanized communication and extensive knowledge reserves. Efforts are made to personify LLMs with special training data or hand-crafted prompts, while correspondingly faced with challenges such as insufficient data usage or rigid behavior patterns. Consequently, personified LLMs fail to capture personified knowledge or express persistent opinion. To fully unlock the potential of LLM personification, we propose PersLLM, a framework for better data construction and model tuning. For insufficient data usage, we incorporate strategies such as Chain-of-Thought prompting and anti-induction, improving the quality of data construction and capturing the personality experiences, knowledge, and thoughts more comprehensively. For rigid behavior patterns, we design the tuning process and introduce automated DPO to enhance the specificity and dynamism of the models’ personalities, which leads to a more natural opinion communication. Both automated metrics and expert human evaluations demonstrate the effectiveness of our approach. Case studies in human-machine interactions and multi-agent systems further suggest potential application scenarios and future directions for LLM personification.

Although retrieval-augmented generation (RAG) remains essential for knowledge-based question answering (KBQA), current paradigms face critical challenges under specific domains. Existing methods struggle with targeted adaptation on small-scale KBs: vanilla unsupervised training exhibits poor effectiveness, while fine-tuning incurs prohibitive costs of external signals. We present KBAlign, a self-supervised framework that enhances RAG systems through efficient model adaptation. Our key insight is to leverage the model’s intrinsic capabilities for knowledge alignment through two innovative mechanisms: multi-grained self-annotation that captures global knowledge for data construction, and iterative tuning that accelerates convergence through self verification. This framework enables cost-effective model adaptation to specific textual KBs, without human supervision or external model assistance. Experiments demonstrate that KBAlign can achieve 90% of the performance gain obtained through GPT-4-supervised adaptation, while relying entirely on self-annotation of much smaller models. KBAlign significantly improves downstream QA accuracy across multiple domains with tiny costs, particularly benefiting scenarios requiring deep knowledge integration from specialized corpora. We release our experimental data, models, and process analyses to the community for further exploration(https://anonymous.4open.science/r/KBAlign-D160).

Retrieval-Augmented Generation (RAG) augments Large Language Models (LLMs) with external knowledge to improve factuality. However, existing RAG systems frequently underutilize the retrieved documents, failing to extract and integrate the key clues needed to support faithful and interpretable reasoning, especially in cases where relevant evidence is implicit, scattered, or obscured by noise. To address this issue, we propose ClueAnchor, a novel framework for enhancing RAG via clue-anchored reasoning exploration and optimization. ClueAnchor extracts key clues from retrieved content and generates multiple reasoning paths based on different knowledge configurations, optimizing the model by selecting the most appropriate reasoning path for the given context through reward-based preference optimization. Experiments show that ClueAnchor significantly outperforms prior RAG baselines in the completeness and robustness of reasoning. Further analysis confirms its strong resilience to noisy or partially relevant retrieved content, as well as its capability to identify supporting evidence even in the absence of explicit clue supervision during inference. All codes are available at https://github.com/thunlp/ClueAnchor.

Training large language models (LLMs) heavily relies on distributed training strategies, among which pipeline parallelism (PP) plays a crucial role. As training sequences extend to 32k or even 128k tokens, current PP methods face severe bottlenecks, including substantial pipeline bubbles and high memory footprint, greatly hindering training throughput and model scalability. This paper introduces a sequence-level one-forward-one-backward (1F1B) PP method, named Seq1F1B, tailored for training LLMs on long sequences with high training throughput and memory efficiency. Unlike typical PP methods, which adopt batch-level pipeline schedule, Seq1F1B schedules the pipeline of training LLMs at the sequence level. It uses a computational strategy to partition sequences appropriately, significantly reducing pipeline bubbles and memory footprint. Compared to competitive PP baselines such as Megatron 1F1B PP, Seq1F1B achieves 1.14X training throughput with half memory footprint.Notably, Seq1F1B trains an LLM with 30B parameters on sequences up to 64k tokens using 64X NVIDIA A100 GPUs without using recomputation strategies, a feat unachievable with existing methods.We have released our code on GitHub to facilitate further research and development in LLM training on long sequences: https://github.com/thunlp/Seq1F1B.

Large language models (LLMs) have been applied across various intelligent educational tasks to assist teaching. While preliminary studies have focused on task-specific, independent LLM-empowered agents, the potential of LLMs within a multi-agent collaborative framework for classroom simulation with real user participation remains unexplored. In this work, we propose SimClass, a multi-agent classroom simulation teaching framework. We recognize representative class roles and introduce a novel class control mechanism for automatic classroom teaching, and conduct user experiments in two real-world courses. Using the Flanders Interactive Analysis System and Community of Inquiry theoretical frameworks from educational analysis, we demonstrate that LLMs can simulate a dynamic learning environment for users with active teacher-student and student-student interactions. We also observe group behaviors among agents in SimClass, where agents collaborate to create enlivening interactions in classrooms to improve user learning process. We hope this work pioneers the application of LLM-empowered multi-agent systems in virtual classroom teaching. Our implementation and service can be found at https://github.com/THU-MAIC/SimClass.

Evaluation is pivotal for honing Large Language Models (LLMs), pinpointing their capabilities and guiding enhancements. The rapid development of LLMs calls for a lightweight and easy-to-use framework for swift evaluation deployment. However, due to the various implementation details to consider, developing a comprehensive evaluation platform is never easy. Existing platforms are often complex and poorly modularized, hindering seamless incorporation into researcher’s workflows. This paper introduces UltraEval, a user-friendly evaluation framework characterized by lightweight, comprehensiveness, modularity, and efficiency. We identify and reimplement three core components of model evaluation (models, data, and metrics). The resulting composability allows for the free combination of different models, tasks, prompts, and metrics within a unified evaluation workflow. Additionally, UltraEval supports diverse models owing to a unified HTTP service and provides sufficient inference acceleration.

Despite advancements in Large Language Models (LLMs) and Large Multimodal Models (LMMs), their integration into language-grounded, human-like embodied agents remains incomplete, hindering complex real-life task performance in 3D environments. Existing integrations often feature limited open-sourcing, challenging collective progress in this field. We introduce LEGENT, an open, scalable platform for developing embodied agents using LLMs and LMMs. LEGENT offers a dual approach: a rich 3D environment with interactive, communicable, and actionable agents, paired with a user-friendly interface, and a sophisticated data generation pipeline utilizing advanced algorithms to exploit supervision from simulated worlds at scale. In our experiments, an embryonic vision-language-action model trained on LEGENT-generated data surpasses GPT-4V in embodied tasks, showcasing promising generalization capabilities. The demo video is available at the following link https://video.legent.ai.

The performance of Large Language Models (LLMs) is substantially influenced by the pretraining corpus, which consists of vast quantities of unsupervised data processed by the models. Despite its critical role in model performance, ensuring the quality of this data is challenging due to its sheer volume and the absence of sample-level quality annotations and enhancements. In this paper, we introduce DecorateLM, a data engineering method designed to refine the pretraining corpus through data rating, tagging and editing. Specifically, DecorateLM rates texts against quality criteria, tags texts with hierarchical labels, and edits texts into a more formalized format. Due to the massive size of the pretraining corpus, adopting an LLM for decorating the entire corpus is less efficient. Therefore, to balance performance with efficiency, we curate a meticulously annotated training corpus for DecorateLM using a large language model and distill data engineering expertise into a compact 1.2 billion parameter small language model (SLM). We then apply DecorateLM to enhance 100 billion tokens of the training corpus, selecting 45 billion tokens that exemplify high quality and diversity for the further training of another 1.2 billion parameter LLM. Our results demonstrate that employing such high-quality data can significantly boost model performance, showcasing a powerful approach to enhance the quality of the pretraining corpus.

Alignment in artificial intelligence pursues the consistency between model responses and human preferences as well as values. In practice, the multifaceted nature of human preferences inadvertently introduces what is known as the ”alignment tax”–a compromise where enhancements in alignment within one objective (e.g., harmlessness) can diminish performance in others (e.g., helpfulness). However, existing alignment techniques are mostly unidirectional, leading to suboptimal trade-offs and poor flexibility over various objectives. To navigate this challenge, we argue the prominence of grounding LLMs with evident preferences. We introduce controllable preference optimization (CPO), which explicitly specifies preference scores for different objectives, thereby guiding the model to generate responses that meet the requirements. Our experimental analysis reveals that the aligned models can provide responses that match various preferences among the ”3H” (helpfulness, honesty, harmlessness) desiderata. Furthermore, by introducing diverse data and alignment goals, we surpass baseline methods in aligning with single objectives, hence mitigating the impact of the alignment tax and achieving improvements in multi-objective alignment.

Legal case retrieval (LCR) aims to provide similar cases as references for a given fact description. This task is crucial for promoting consistent judgments in similar cases, effectively enhancing judicial fairness and improving work efficiency for judges. However, existing works face two main challenges for real-world applications: existing works mainly focus on case-to-case retrieval using lengthy queries, which does not match real-world scenarios; and the limited data scale, with current datasets containing only hundreds of queries, is insufficient to satisfy the training requirements of existing data-hungry neural models. To address these issues, we introduce an automated method to construct synthetic query-candidate pairs and build the largest LCR dataset to date, LEAD, which is hundreds of times larger than existing datasets. This data construction method can provide ample training signals for LCR models. Experimental results demonstrate that model training with our constructed data can achieve state-of-the-art results on two widely-used LCR benchmarks. Besides, the construction method can also be applied to civil cases and achieve promising results. The data and codes can be found in https://github.com/thunlp/LEAD.

As large language models (LLMs) increasingly permeate daily lives, there is a growing demand for real-time interactions that mirror human conversations. Traditional turn-based chat systems driven by LLMs prevent users from verbally interacting with the system while generating responses.To overcome these limitations, we adapt existing LLMs to duplex models so that they can listen to users while generating output and dynamically adjust themselves to provide instant feedback.Specifically, we divide the queries and responses of conversations into several time slices and then adopt a time-division-multiplexing (TDM) encoding-decoding strategy to process these slices pseudo-simultaneously.Furthermore, to make LLMs proficient enough to handle real-time conversations, we build a fine-tuning dataset consisting of alternating time slices of queries and responses and covering typical feedback types in instantaneous interactions.Our experiments show that although the queries and responses of conversations are segmented into incomplete slices for processing, LLMs can preserve their original performance on standard benchmarks with a few fine-tuning steps on our dataset. Automatic and human evaluation indicate that duplex models make user-AI interactions more natural and human-like, and greatly improve user satisfaction compared to vanilla LLMs. Our duplex model and dataset will be released soon.

Speculative decoding is a widely used method that accelerates the generation process of large language models (LLMs) with no compromise in model performance. It achieves this goal by using an existing smaller model for drafting and then employing the target LLM to verify the draft in a low-cost parallel manner. Under such a drafting-verification framework, drafting efficiency has become a bottleneck in the final speedup of speculative decoding. Therefore, generating longer drafts at less cost can lead to better decoding speedup. To achieve this, we introduce Ouroboros, which can generate draft phrases to parallelize the drafting process and meanwhile lengthen drafts in a training-free manner. The experimental results on various typical text generation tasks show that Ouroboros can achieve speedups of up to 2.4× over speculative decoding and 3.9× over vanilla decoding, without fine-tuning draft and target models. Code available at https://github.com/thunlp/Ouroboros.

Generative models have demonstrated considerable potential in software engineering, particularly in tasks such as code generation and debugging. However, their utilization in the domain of code documentation generation remains underexplored. To this end, we introduce RepoAgent, a large language model powered open-source framework aimed at proactively generating, maintaining, and updating code documentation. Through both qualitative and quantitative evaluations, we have validated the effectiveness of our approach, showing that RepoAgent excels in generating high-quality repository-level documentation. The code and results are publicly accessible at https://github.com/OpenBMB/RepoAgent.

Semi-structured interviews are a crucial method of data acquisition in qualitative research. Typically controlled by the interviewer, the process progresses through a question-and-answer format, aimed at eliciting information from the interviewee. However, interviews are highly time-consuming and demand considerable experience of the interviewers, which greatly limits the efficiency and feasibility of data collection. Therefore, we introduce LM-Interview, a novel system designed to automate the process of preparing, conducting and analyzing semi-structured interviews. Experimental results demonstrate that LM-interview achieves performance comparable to that of skilled human interviewers.

This paper introduces INTERVENOR (INTERactiVE chaiN Of Repair), a system designed to emulate the interactive code repair processes observed in humans, encompassing both code diagnosis and code repair. INTERVENOR prompts Large Language Models (LLMs) to play distinct roles during the code repair process, functioning as both a Code Learner and a Code Teacher. Specifically, the Code Learner is tasked with adhering to instructions to generate or repair code, while the Code Teacher is responsible for crafting a Chain-of-Repair (CoR) to serve as guidance for the Code Learner. During generating the CoR, the Code Teacher needs to check the generated codes from Code Learner and reassess how to address code bugs based on error feedback received from compilers. Experimental results demonstrate that INTERVENOR surpasses baseline models, exhibiting improvements of approximately 18% and 4.3% over GPT-3.5 in code generation and code translation tasks, respectively. Our further analyses show that CoR is effective to illuminate the reasons behind bugs and outline solution plans in natural language. With the feedback of code compilers, INTERVENOR can accurately identify syntax errors and assertion errors and provide precise instructions to repair codes. All data and codes are available at [https://github.com/NEUIR/INTERVENOR](https://github.com/NEUIR/INTERVENOR).

Large Language Models (LLMs) have demonstrated exceptional coding capability. However, as another critical component of programming proficiency, the debugging capability of LLMs remains relatively unexplored. Previous evaluations of LLMs’ debugging ability are significantly limited by the risk of data leakage, the scale of the dataset, and the variety of tested bugs. To overcome these deficiencies, we introduce ‘DebugBench’, an LLM debugging benchmark consisting of 4,253 instances. It covers four major bug categories and 18 minor types in C++, Java, and Python. To construct DebugBench, we collect code snippets from the LeetCode community, implant bugs into source data with GPT-4, and assure rigorous quality checks. We evaluate two commercial and four open-source models in a zero-shot scenario. We find that (1) while closed-source models exhibit inferior debugging performance compared to humans, open-source models relatively lower pass rate scores; (2) the complexity of debugging notably fluctuates depending on the bug category; (3) incorporating runtime feedback has a clear impact on debugging performance which is not always helpful. As an extension, we also compare LLM debugging and code generation, revealing a strong correlation between them for closed-source models. These findings will benefit the development of LLMs in debugging.

Large Language Models (LLMs) have witnessed remarkable advancements in recent years, prompting the exploration of tool learning, which integrates LLMs with external tools to address diverse real-world challenges. Assessing the capability of LLMs to utilise tools necessitates large-scale and stable benchmarks. However, previous works relied on either hand-crafted online tools with limited scale, or large-scale real online APIs suffering from instability of API status. To address this problem, we introduce StableToolBench, a benchmark evolving from ToolBench, proposing a virtual API server and stable evaluation system. The virtual API server contains a caching system and API simulators which are complementary to alleviate the change in API status. Meanwhile, the stable evaluation system designs solvable pass and win rates using GPT-4 as the automatic evaluator to eliminate the randomness during evaluation. Experimental results demonstrate the stability of StableToolBench, and further discuss the effectiveness of API simulators, the caching system, and the evaluator system.

Scientific data visualization plays a crucial role in research by enabling the direct display of complex information and assisting researchers in identifying implicit patterns. Despite its importance, the use of Large Language Models (LLMs) for scientific data visualization remains rather unexplored. In this study, we introduce MatPlotAgent, an efficient model-agnostic LLM agent framework designed to automate scientific data visualization tasks. Leveraging the capabilities of both code LLMs and multi-modal LLMs, MatPlotAgent consists of three core modules: query understanding, code generation with iterative debugging, and a visual feedback mechanism for error correction. To address the lack of benchmarks in this field, we present MatPlotBench, a high-quality benchmark consisting of 100 human-verified test cases. Additionally, we introduce a scoring approach that utilizes GPT-4V for automatic evaluation. Experimental results demonstrate that MatPlotAgent can improve the performance of various LLMs, including both commercial and open-source models. Furthermore, the proposed evaluation method shows a strong correlation with human-annotated scores.

In recent developments within the research community, the integration of Large Language Models (LLMs) in creating fully autonomous agents has garnered significant interest. Despite this, LLM-based agents frequently demonstrate notable shortcomings in adjusting to dynamic environments and fully grasping human needs. In this work, we introduce the problem of LLM-based human-agent collaboration for complex task-solving, exploring their synergistic potential. To tackle the problem, we propose a Reinforcement Learning-based Human-Agent Collaboration method, ReHAC, which trains a policy model designed to determine the most opportune stages for human intervention within the task-solving process. We conduct experiments under real and simulated human-agent collaboration scenarios. Experimental results demonstrate that the synergistic efforts of humans and LLM-based agents significantly improve performance in complex tasks, primarily through well-planned, limited human intervention. Datasets and code are available at: https://github.com/XueyangFeng/ReHAC/.

Natural language (NL) has long been the predominant format for human cognition and communication, and by extension, has been similarly pivotal in the development and application of Large Language Models (LLMs). Yet, besides NL, LLMs have seen various non-NL formats during pre-training, such as code and logical expression. NL’s status as the optimal format for LLMs, particularly in single-LLM reasoning and multi-agent communication, has not been thoroughly examined. In this work, we challenge the default use of NL by exploring the utility of non-NL formats in these contexts. We show that allowing LLMs to autonomously select the most suitable format before reasoning or communicating leads to a 3.3 to 5.7% improvement in reasoning efficiency for different LLMs, and up to a 72.7% reduction in token usage in multi-agent communication, all while maintaining communicative effectiveness. Our comprehensive analysis further reveals that LLMs can devise a format from limited task instructions and that the devised format is effectively transferable across different LLMs. Intriguingly, the structured communication format decided by LLMs exhibits notable parallels with established agent communication languages, suggesting a natural evolution towards efficient, structured communication in agent communication. Our code will be released to facilitate further exploration.

It is very common to use quotations (quotes) to make our writings more elegant or convincing. To help people find appropriate quotes efficiently, the task of quote recommendation is presented, aiming to recommend quotes that fit the current context of writing. There have been various quote recommendation approaches, but they are evaluated on different unpublished datasets. To facilitate the research on this task, we build a large and fully open quote recommendation dataset called QuoteR, which comprises three parts including English, standard Chinese and classical Chinese. Any part of it is larger than previous unpublished counterparts. We conduct an extensive evaluation of existing quote recommendation methods on QuoteR. Furthermore, we propose a new quote recommendation model that significantly outperforms previous methods on all three parts of QuoteR. All the code and data of this paper can be obtained at https://github.com/thunlp/QuoteR.

In recent years, researchers tend to pre-train ever-larger language models to explore the upper limit of deep models. However, large language model pre-training costs intensive computational resources, and most of the models are trained from scratch without reusing the existing pre-trained models, which is wasteful. In this paper, we propose bert2BERT, which can effectively transfer the knowledge of an existing smaller pre-trained model to a large model through parameter initialization and significantly improve the pre-training efficiency of the large model. Specifically, we extend the previous function-preserving method proposed in computer vision on the Transformer-based language model, and further improve it by proposing a novel method, advanced knowledge for large model’s initialization. In addition, a two-stage learning method is proposed to further accelerate the pre-training. We conduct extensive experiments on representative PLMs (e.g., BERT and GPT) and demonstrate that (1) our method can save a significant amount of training cost compared with baselines including learning from scratch, StackBERT and MSLT; (2) our method is generic and applicable to different types of pre-trained models. In particular, bert2BERT saves about 45% and 47% computational cost of pre-training BERT _BASE and GPT _BASE by reusing the models of almost their half sizes.

Tuning pre-trained language models (PLMs) with task-specific prompts has been a promising approach for text classification. Particularly, previous studies suggest that prompt-tuning has remarkable superiority in the low-data scenario over the generic fine-tuning methods with extra classifiers. The core idea of prompt-tuning is to insert text pieces, i.e., template, to the input and transform a classification problem into a masked language modeling problem, where a crucial step is to construct a projection, i.e., verbalizer, between a label space and a label word space. A verbalizer is usually handcrafted or searched by gradient descent, which may lack coverage and bring considerable bias and high variances to the results. In this work, we focus on incorporating external knowledge into the verbalizer, forming a knowledgeable prompttuning (KPT), to improve and stabilize prompttuning. Specifically, we expand the label word space of the verbalizer using external knowledge bases (KBs) and refine the expanded label word space with the PLM itself before predicting with the expanded label word space. Extensive experiments on zero and few-shot text classification tasks demonstrate the effectiveness of knowledgeable prompt-tuning.

Fine-grained entity typing (FGET) aims to classify named entity mentions into fine-grained entity types, which is meaningful for entity-related NLP tasks. For FGET, a key challenge is the low-resource problem — the complex entity type hierarchy makes it difficult to manually label data. Especially for those languages other than English, human-labeled data is extremely scarce. In this paper, we propose a cross-lingual contrastive learning framework to learn FGET models for low-resource languages. Specifically, we use multi-lingual pre-trained language models (PLMs) as the backbone to transfer the typing knowledge from high-resource languages (such as English) to low-resource languages (such as Chinese). Furthermore, we introduce entity-pair-oriented heuristic rules as well as machine translation to obtain cross-lingual distantly-supervised data, and apply cross-lingual contrastive learning on the distantly-supervised data to enhance the backbone PLMs. Experimental results show that by applying our framework, we can easily learn effective FGET models for low-resource languages, even without any language-specific human-labeled data. Our code is also available at https://github.com/thunlp/CrossET.

Selecting an appropriate pre-trained model (PTM) for a specific downstream task typically requires significant efforts of fine-tuning. To accelerate this process, researchers propose feature-based model selection (FMS) methods, which assess PTMs’ transferability to a specific task in a fast way without fine-tuning. In this work, we argue that current FMS methods are vulnerable, as the assessment mainly relies on the static features extracted from PTMs. However, such features are derived without training PTMs on downstream tasks, and are not necessarily reliable indicators for the PTM’s transferability. To validate our viewpoints, we design two methods to evaluate the robustness of FMS: (1) model disguise attack, which post-trains an inferior PTM with a contrastive objective, and (2) evaluation data selection, which selects a subset of the data points for FMS evaluation based on K-means clustering. Experimental results prove that both methods can successfully make FMS mistakenly judge the transferability of PTMs. Moreover, we find that these two methods can further be combined with the backdoor attack to misguide the FMS to select poisoned models. To the best of our knowledge, this is the first work to demonstrate the defects of current FMS algorithms and evaluate their potential security risks. By identifying previously unseen risks of FMS, our study indicates new directions for improving the robustness of FMS.

Hyperbolic neural networks have shown great potential for modeling complex data. However, existing hyperbolic networks are not completely hyperbolic, as they encode features in the hyperbolic space yet formalize most of their operations in the tangent space (a Euclidean subspace) at the origin of the hyperbolic model. This hybrid method greatly limits the modeling ability of networks. In this paper, we propose a fully hyperbolic framework to build hyperbolic networks based on the Lorentz model by adapting the Lorentz transformations (including boost and rotation) to formalize essential operations of neural networks. Moreover, we also prove that linear transformation in tangent spaces used by existing hyperbolic networks is a relaxation of the Lorentz rotation and does not include the boost, implicitly limiting the capabilities of existing hyperbolic networks. The experimental results on four NLP tasks show that our method has better performance for building both shallow and deep networks. Our code will be released to facilitate follow-up research.

Prompt-based tuning for pre-trained language models (PLMs) has shown its effectiveness in few-shot learning. Typically, prompt-based tuning wraps the input text into a cloze question. To make predictions, the model maps the output words to labels via a verbalizer, which is either manually designed or automatically built. However, manual verbalizers heavily depend on domain-specific prior knowledge and human efforts, while finding appropriate label words automatically still remains challenging. In this work, we propose the prototypical verbalizer (ProtoVerb) which is built directly from training data. Specifically, ProtoVerb learns prototype vectors as verbalizers by contrastive learning. In this way, the prototypes summarize training instances and are able to enclose rich class-level semantics. We conduct experiments on both topic classification and entity typing tasks, and the results demonstrate that ProtoVerb significantly outperforms current automatic verbalizers, especially when training data is extremely scarce. More surprisingly, ProtoVerb consistently boosts prompt-based tuning even on untuned PLMs, indicating an elegant non-tuning way to utilize PLMs. Our codes are avaliable at https://github.com/thunlp/OpenPrompt.

Program induction for answering complex questions over knowledge bases (KBs) aims to decompose a question into a multi-step program, whose execution against the KB produces the final answer. Learning to induce programs relies on a large number of parallel question-program pairs for the given KB. However, for most KBs, the gold program annotations are usually lacking, making learning difficult. In this paper, we propose the approach of program transfer, which aims to leverage the valuable program annotations on the rich-resourced KBs as external supervision signals to aid program induction for the low-resourced KBs that lack program annotations. For program transfer, we design a novel two-stage parsing framework with an efficient ontology-guided pruning strategy. First, a sketch parser translates the question into a high-level program sketch, which is the composition of functions. Second, given the question and sketch, an argument parser searches the detailed arguments from the KB for functions. During the searching, we incorporate the KB ontology to prune the search space. The experiments on ComplexWebQuestions and WebQuestionSP show that our method outperforms SOTA methods significantly, demonstrating the effectiveness of program transfer and our framework. Our codes and datasets can be obtained from https://github.com/THU-KEG/ProgramTransfer.

Prompts for pre-trained language models (PLMs) have shown remarkable performance by bridging the gap between pre-training tasks and various downstream tasks. Among these methods, prompt tuning, which freezes PLMs and only tunes soft prompts, provides an efficient and effective solution for adapting large-scale PLMs to downstream tasks. However, prompt tuning is yet to be fully explored. In our pilot experiments, we find that prompt tuning performs comparably with conventional full-model tuning when downstream data are sufficient, whereas it is much worse under few-shot learning settings, which may hinder the application of prompt tuning. We attribute this low performance to the manner of initializing soft prompts. Therefore, in this work, we propose to pre-train prompts by adding soft prompts into the pre-training stage to obtain a better initialization. We name this Pre-trained Prompt Tuning framework “PPT”. To ensure the generalization of PPT, we formulate similar classification tasks into a unified task form and pre-train soft prompts for this unified task. Extensive experiments show that tuning pre-trained prompts for downstream tasks can reach or even outperform full-model fine-tuning under both full-data and few-shot settings. Our approach is effective and efficient for using large-scale PLMs in practice.

Pre-trained language models (PLMs) cannot well recall rich factual knowledge of entities exhibited in large-scale corpora, especially those rare entities. In this paper, we propose to build a simple but effective Pluggable Entity Lookup Table (PELT) on demand by aggregating the entity’s output representations of multiple occurrences in the corpora. PELT can be compatibly plugged as inputs to infuse supplemental entity knowledge into PLMs. Compared to previous knowledge-enhanced PLMs, PELT only requires 0.2%-5% pre-computation with capability of acquiring knowledge from out-of-domain corpora for domain adaptation scenario. The experiments on knowledge-related tasks demonstrate that our method, PELT, can flexibly and effectively transfer entity knowledge from related corpora into PLMs with different architectures. Our code and models are publicly available at https://github.com/thunlp/PELT

Sememe knowledge bases (SKBs), which annotate words with the smallest semantic units (i.e., sememes), have proven beneficial to many NLP tasks. Building an SKB is very time-consuming and labor-intensive. Therefore, some studies have tried to automate the building process by predicting sememes for the unannotated words. However, all existing sememe prediction studies ignore the hierarchical structures of sememes, which are important in the sememe-based semantic description system. In this work, we tackle the structured sememe prediction problem for the first time, which is aimed at predicting a sememe tree with hierarchical structures rather than a set of sememes. We design a sememe tree generation model based on Transformer with adjusted attention mechanism, which shows its superiority over the baselines in experiments. We also conduct a series of quantitative and qualitative analyses of the effectiveness of our model. All the code and data of this paper are available at https://github.com/thunlp/STG.

In linguistics, a sememe is defined as the minimum semantic unit of languages. Sememe knowledge bases (KBs), which are built by manually annotating words with sememes, have been successfully applied to various NLP tasks. However, existing sememe KBs only cover a few languages, which hinders the wide utilization of sememes. To address this issue, the task of sememe prediction for BabelNet synsets (SPBS) is presented, aiming to build a multilingual sememe KB based on BabelNet, a multilingual encyclopedia dictionary. By automatically predicting sememes for a BabelNet synset, the words in many languages in the synset would obtain sememe annotations simultaneously. However, previous SPBS methods have not taken full advantage of the abundant information in BabelNet. In this paper, we utilize the multilingual synonyms, multilingual glosses and images in BabelNet for SPBS. We design a multimodal information fusion model to encode and combine this information for sememe prediction. Experimental results show the substantial outperformance of our model over previous methods (about 10 MAP and F1 scores). All the code and data of this paper can be obtained at https://github.com/thunlp/MSGI.

Recognizing facts is the most fundamental step in making judgments, hence detecting events in the legal documents is important to legal case analysis tasks. However, existing Legal Event Detection (LED) datasets only concern incomprehensive event types and have limited annotated data, which restricts the development of LED methods and their downstream applications. To alleviate these issues, we present LEVEN a large-scale Chinese LEgal eVENt detection dataset, with 8,116 legal documents and 150,977 human-annotated event mentions in 108 event types. Not only charge-related events, LEVEN also covers general events, which are critical for legal case understanding but neglected in existing LED datasets. To our knowledge, LEVEN is the largest LED dataset and has dozens of times the data scale of others, which shall significantly promote the training and evaluation of LED methods. The results of extensive experiments indicate that LED is challenging and needs further effort. Moreover, we simply utilize legal events as side information to promote downstream applications. The method achieves improvements of average 2.2 points precision in low-resource judgment prediction, and 1.5 points mean average precision in unsupervised case retrieval, which suggests the fundamentality of LED. The source code and dataset can be obtained from https://github.com/thunlp/LEVEN.

Recent work has shown that feed-forward networks (FFNs) in pre-trained Transformers are a key component, storing various linguistic and factual knowledge. However, the computational patterns of FFNs are still unclear. In this work, we study the computational patterns of FFNs and observe that most inputs only activate a tiny ratio of neurons of FFNs. This phenomenon is similar to the sparsity of the human brain, which drives research on functional partitions of the human brain. To verify whether functional partitions also emerge in FFNs, we propose to convert a model into its MoE version with the same parameters, namely MoEfication. Specifically, MoEfication consists of two phases: (1) splitting the parameters of FFNs into multiple functional partitions as experts, and (2) building expert routers to decide which experts will be used for each input. Experimental results show that MoEfication can conditionally use 10% to 30% of FFN parameters while maintaining over 95% original performance for different models on various downstream tasks. Besides, MoEfication brings two advantages: (1) it significantly reduces the FLOPS of inference, i.e., 2x speedup with 25% of FFN parameters, and (2) it provides a fine-grained perspective to study the inner mechanism of FFNs. The source code of this paper can be obtained from https://github.com/thunlp/MoEfication.

Current pre-trained language models (PLM) are typically trained with static data, ignoring that in real-world scenarios, streaming data of various sources may continuously grow. This requires PLMs to integrate the information from all the sources in a lifelong manner. Although this goal could be achieved by exhaustive pre-training on all the existing data, such a process is known to be computationally expensive. To this end, we propose ELLE, aiming at efficient lifelong pre-training for emerging data. Specifically, ELLE consists of (1) function preserved model expansion, which flexibly expands an existing PLM’s width and depth to improve the efficiency of knowledge acquisition; and (2) pre-trained domain prompts, which disentangle the versatile knowledge learned during pre-training and stimulate the proper knowledge for downstream tasks. We experiment ELLE with streaming data from 5 domains on BERT and GPT. The results show the superiority of ELLE over various lifelong learning baselines in both pre-training efficiency and downstream performances. The codes are publicly available at https://github.com/thunlp/ELLE.

Recent works have shown promising results of prompt tuning in stimulating pre-trained language models (PLMs) for natural language processing (NLP) tasks. However, to the best of our knowledge, existing works focus on prompt-tuning generative PLMs that are pre-trained to generate target tokens, such as BERT. It is still unknown whether and how discriminative PLMs, e.g., ELECTRA, can be effectively prompt-tuned. In this work, we present DPT, the first prompt tuning framework for discriminative PLMs, which reformulates NLP tasks into a discriminative language modeling problem. Comprehensive experiments on text classification and question answering show that, compared with vanilla fine-tuning, DPT achieves significantly higher performance, and also prevents the unstable problem in tuning large PLMs in both full-set and low-resource settings.

In recent years, pre-trained language models (PLMs) have been shown to capture factual knowledge from massive texts, which encourages the proposal of PLM-based knowledge graph completion (KGC) models. However, these models are still quite behind the SOTA KGC models in terms of performance. In this work, we find two main reasons for the weak performance: (1) Inaccurate evaluation setting. The evaluation setting under the closed-world assumption (CWA) may underestimate the PLM-based KGC models since they introduce more external knowledge; (2) Inappropriate utilization of PLMs. Most PLM-based KGC models simply splice the labels of entities and relations as inputs, leading to incoherent sentences that do not take full advantage of the implicit knowledge in PLMs. To alleviate these problems, we highlight a more accurate evaluation setting under the open-world assumption (OWA), which manual checks the correctness of knowledge that is not in KGs. Moreover, motivated by prompt tuning, we propose a novel PLM-based KGC model named PKGC. The basic idea is to convert each triple and its support information into natural prompt sentences, which is further fed into PLMs for classification. Experiment results on two KGC datasets demonstrate OWA is more reliable for evaluating KGC, especially on the link prediction, and the effectiveness of our PKCG model on both CWA and OWA settings.

Prompt-based learning paradigm bridges the gap between pre-training and fine-tuning, and works effectively under the few-shot setting. However, we find that this learning paradigm inherits the vulnerability from the pre-training stage, where model predictions can be misled by inserting certain triggers into the text. In this paper, we explore this universal vulnerability by either injecting backdoor triggers or searching for adversarial triggers on pre-trained language models using only plain text. In both scenarios, we demonstrate that our triggers can totally control or severely decrease the performance of prompt-based models fine-tuned on arbitrary downstream tasks, reflecting the universal vulnerability of the prompt-based learning paradigm. Further experiments show that adversarial triggers have good transferability among language models. We also find conventional fine-tuning models are not vulnerable to adversarial triggers constructed from pre-trained language models. We conclude by proposing a potential solution to mitigate our attack methods. Code and data are publicly available.

Recent explorations of large-scale pre-trained language models (PLMs) have revealed the power of PLMs with huge amounts of parameters, setting off a wave of training ever-larger PLMs. However, it requires tremendous computational resources to train a large-scale PLM, which may be practically unaffordable. In addition, existing large-scale PLMs are mainly trained from scratch individually, ignoring that many well-trained PLMs are available. To this end, we explore the question how could existing PLMs benefit training large-scale PLMs in future. Specifically, we introduce a pre-training framework named “knowledge inheritance” (KI) and explore how could knowledge distillation serve as auxiliary supervision during pre-training to efficiently learn larger PLMs. Experimental results demonstrate the superiority of KI in training efficiency. We also conduct empirical analyses to explore the effects of teacher PLMs’ pre-training settings, including model architecture, pre-training data, etc. Finally, we show that KI could be applied to domain adaptation and knowledge transfer.

Prompt tuning (PT) is a promising parameter-efficient method to utilize extremely large pre-trained language models (PLMs), which can achieve comparable performance to full-parameter fine-tuning by only tuning a few soft prompts. However, PT requires much more training time than fine-tuning. Intuitively, knowledge transfer can help to improve the efficiency. To explore whether we can improve PT via prompt transfer, we empirically investigate the transferability of soft prompts across different downstream tasks and PLMs in this work. We find that (1) in zero-shot setting, trained soft prompts can effectively transfer to similar tasks on the same PLM and also to other PLMs with a cross-model projector trained on similar tasks; (2) when used as initialization, trained soft prompts of similar tasks and projected prompts of other PLMs can significantly accelerate training and also improve the performance of PT. Moreover, to explore what decides prompt transferability, we investigate various transferability indicators and find that the overlapping rate of activated neurons strongly reflects the transferability, which suggests how the prompts stimulate PLMs is essential. Our findings show that prompt transfer is promising for improving PT, and further research shall focus more on prompts’ stimulation to PLMs. The source code can be obtained from https://github.com/thunlp/Prompt-Transferability.

Question Answering (QA) is a longstanding challenge in natural language processing. Existing QA works mostly focus on specific question types, knowledge domains, or reasoning skills. The specialty in QA research hinders systems from modeling commonalities between tasks and generalization for wider applications. To address this issue, we present ProQA, a unified QA paradigm that solves various tasks through a single model. ProQA takes a unified structural prompt as the bridge and improves the QA-centric ability by structural prompt-based pre-training. Through a structurally designed prompt-based input schema, ProQA concurrently models the knowledge generalization for all QA tasks while keeping the knowledge customization for every specific QA task. Furthermore, ProQA is pre-trained with structural prompt-formatted large-scale synthesized corpus, which empowers the model with the commonly-required QA ability. Experimental results on 11 QA benchmarks demonstrate that ProQA consistently boosts performance on both full data fine-tuning, few-shot learning, and zero-shot testing scenarios. Furthermore, ProQA exhibits strong ability in both continual learning and transfer learning by taking the advantages of the structural prompt.