Abstract:Machine Reading Comprehension (MRC) is an essential task in evaluating natural language understanding. Existing MRC datasets primarily assess specific aspects of reading comprehension (RC), lacking a comprehensive MRC benchmark. To fill this gap, we first introduce a novel taxonomy that categorizes the key capabilities required for RC. Based on this taxonomy, we construct MRCEval, an MRC benchmark that leverages advanced Large Language Models (LLMs) as both sample generators and selection judges. MRCEval is a comprehensive, challenging and accessible benchmark designed to assess the RC capabilities of LLMs thoroughly, covering 13 distinct RC skills with a total of 2.1K high-quality multi-choice questions. We perform an extensive evaluation of 28 widely used open-source and proprietary models, highlighting that MRC continues to present significant challenges even in the era of LLMs.
Abstract:Large language models (LLMs) excel in tasks that require complex linguistic abilities, such as reference disambiguation and metaphor recognition/generation. Although LLMs possess impressive capabilities, their internal mechanisms for processing and representing linguistic knowledge remain largely opaque. Previous work on linguistic mechanisms has been limited by coarse granularity, insufficient causal analysis, and a narrow focus. In this study, we present a systematic and comprehensive causal investigation using sparse auto-encoders (SAEs). We extract a wide range of linguistic features from six dimensions: phonetics, phonology, morphology, syntax, semantics, and pragmatics. We extract, evaluate, and intervene on these features by constructing minimal contrast datasets and counterfactual sentence datasets. We introduce two indices-Feature Representation Confidence (FRC) and Feature Intervention Confidence (FIC)-to measure the ability of linguistic features to capture and control linguistic phenomena. Our results reveal inherent representations of linguistic knowledge in LLMs and demonstrate the potential for controlling model outputs. This work provides strong evidence that LLMs possess genuine linguistic knowledge and lays the foundation for more interpretable and controllable language modeling in future research.
Abstract:Reward models (RMs) are crucial for the training and inference-time scaling up of large language models (LLMs). However, existing reward models primarily focus on human preferences, neglecting verifiable correctness signals which have shown strong potential in training LLMs. In this paper, we propose agentic reward modeling, a reward system that combines reward models with verifiable correctness signals from different aspects to provide reliable rewards. We empirically implement a reward agent, named RewardAgent, that combines human preference rewards with two verifiable signals: factuality and instruction following, to provide more reliable rewards. We conduct comprehensive experiments on existing reward model benchmarks and inference time best-of-n searches on real-world downstream tasks. RewardAgent significantly outperforms vanilla reward models, demonstrating its effectiveness. We further construct training preference pairs using RewardAgent and train an LLM with the DPO objective, achieving superior performance on various NLP benchmarks compared to conventional reward models. Our codes are publicly released to facilitate further research (https://github.com/THU-KEG/Agentic-Reward-Modeling).
Abstract:Existing Large Vision-Language Models (LVLMs) can process inputs with context lengths up to 128k visual and text tokens, yet they struggle to generate coherent outputs beyond 1,000 words. We find that the primary limitation is the absence of long output examples during supervised fine-tuning (SFT). To tackle this issue, we introduce LongWriter-V-22k, a SFT dataset comprising 22,158 examples, each with multiple input images, an instruction, and corresponding outputs ranging from 0 to 10,000 words. Moreover, to achieve long outputs that maintain high-fidelity to the input images, we employ Direct Preference Optimization (DPO) to the SFT model. Given the high cost of collecting human feedback for lengthy outputs (e.g., 3,000 words), we propose IterDPO, which breaks long outputs into segments and uses iterative corrections to form preference pairs with the original outputs. Additionally, we develop MMLongBench-Write, a benchmark featuring six tasks to evaluate the long-generation capabilities of VLMs. Our 7B parameter model, trained with LongWriter-V-22k and IterDPO, achieves impressive performance on this benchmark, outperforming larger proprietary models like GPT-4o. Code and data: https://github.com/THU-KEG/LongWriter-V
Abstract:This paper introduces LongBench v2, a benchmark designed to assess the ability of LLMs to handle long-context problems requiring deep understanding and reasoning across real-world multitasks. LongBench v2 consists of 503 challenging multiple-choice questions, with contexts ranging from 8k to 2M words, across six major task categories: single-document QA, multi-document QA, long in-context learning, long-dialogue history understanding, code repository understanding, and long structured data understanding. To ensure the breadth and the practicality, we collect data from nearly 100 highly educated individuals with diverse professional backgrounds. We employ both automated and manual review processes to maintain high quality and difficulty, resulting in human experts achieving only 53.7% accuracy under a 15-minute time constraint. Our evaluation reveals that the best-performing model, when directly answers the questions, achieves only 50.1% accuracy. In contrast, the o1-preview model, which includes longer reasoning, achieves 57.7%, surpassing the human baseline by 4%. These results highlight the importance of enhanced reasoning ability and scaling inference-time compute to tackle the long-context challenges in LongBench v2. The project is available at https://longbench2.github.io.
Abstract:In real life, many dynamic events, such as major disasters and large-scale sports events, evolve continuously over time. Obtaining an overview of these events can help people quickly understand the situation and respond more effectively. This is challenging because the key information of the event is often scattered across multiple documents, involving complex event knowledge understanding and reasoning, which is under-explored in previous work. Therefore, we proposed the Event-Centric Multi-Document Summarization (ECS) task, which aims to generate concise and comprehensive summaries of a given event based on multiple related news documents. Based on this, we constructed the EventSum dataset, which was constructed using Baidu Baike entries and underwent extensive human annotation, to facilitate relevant research. It is the first large scale Chinese multi-document summarization dataset, containing 5,100 events and a total of 57,984 news documents, with an average of 11.4 input news documents and 13,471 characters per event. To ensure data quality and mitigate potential data leakage, we adopted a multi-stage annotation approach for manually labeling the test set. Given the complexity of event-related information, existing metrics struggle to comprehensively assess the quality of generated summaries. We designed specific metrics including Event Recall, Argument Recall, Causal Recall, and Temporal Recall along with corresponding calculation methods for evaluation. We conducted comprehensive experiments on EventSum to evaluate the performance of advanced long-context Large Language Models (LLMs) on this task. Our experimental results indicate that: 1) The event-centric multi-document summarization task remains challenging for existing long-context LLMs; 2) The recall metrics we designed are crucial for evaluating the comprehensiveness of the summary information.
Abstract:Recent advancements in large language models (LLMs) have led to significant improvements in various natural language processing tasks, but it is still challenging for LLMs to perform knowledge-intensive complex question answering due to LLMs' inefficacy in reasoning planning and the hallucination problem. A typical solution is to employ retrieval-augmented generation (RAG) coupled with chain-of-thought (CoT) reasoning, which decomposes complex questions into chain-like sub-questions and applies iterative RAG at each sub-question. However, prior works exhibit sub-optimal reasoning planning and overlook dynamic knowledge retrieval from heterogeneous sources. In this paper, we propose AtomR, a novel heterogeneous knowledge reasoning framework that conducts multi-source reasoning at the atomic level. Drawing inspiration from the graph modeling of knowledge, AtomR leverages large language models (LLMs) to decompose complex questions into combinations of three atomic knowledge operators, significantly enhancing the reasoning process at both the planning and execution stages. We also introduce BlendQA, a novel evaluation benchmark tailored to assess complex heterogeneous knowledge reasoning. Experiments show that AtomR significantly outperforms state-of-the-art baselines across three single-source and two multi-source reasoning benchmarks, with notable performance gains of 9.4% on 2WikiMultihop and 9.5% on BlendQA.
Abstract:Large language models (LLMs) struggle to follow instructions with complex constraints in format, length, etc. Following the conventional instruction-tuning practice, previous works conduct post-training on complex instruction-response pairs generated by feeding complex instructions to advanced LLMs. However, even advanced LLMs cannot follow complex instructions well, thus limiting the quality of generated data. In this work, we find that existing datasets inherently contain implicit complex constraints and propose a novel data generation technique, constraint back-translation. Specifically, we take the high-quality instruction-response pairs in existing datasets and only adopt advanced LLMs to add complex constraints already met by the responses to the instructions, which naturally reduces costs and data noise. In the experiments, we adopt Llama3-70B-Instruct to back-translate constraints and create a high-quality complex instruction-response dataset, named CRAB. We present that post-training on CRAB improves multiple backbone LLMs' complex instruction-following ability, evaluated on extensive instruction-following benchmarks. We further find that constraint back-translation also serves as a useful auxiliary training objective in post-training. Our code, data, and models will be released to facilitate future research.
Abstract:Though significant advancements have been achieved in developing long-context large language models (LLMs), the compromised quality of LLM-synthesized data for supervised fine-tuning (SFT) often affects the long-context performance of SFT models and leads to inherent limitations. In principle, reinforcement learning (RL) with appropriate reward signals can further enhance models' capacities. However, how to obtain reliable rewards in long-context scenarios remains unexplored. To this end, we propose LongReward, a novel method that utilizes an off-the-shelf LLM to provide rewards for long-context model responses from four human-valued dimensions: helpfulness, logicality, faithfulness, and completeness, each with a carefully designed assessment pipeline. By combining LongReward and offline RL algorithm DPO, we are able to effectively improve long-context SFT models. Our experiments indicate that LongReward not only significantly improves models' long-context performance but also enhances their ability to follow short instructions. We also find that long-context DPO with LongReward and conventional short-context DPO can be used together without hurting either one's performance.
Abstract:Reward models are critical in techniques like Reinforcement Learning from Human Feedback (RLHF) and Inference Scaling Laws, where they guide language model alignment and select optimal responses. Despite their importance, existing reward model benchmarks often evaluate models by asking them to distinguish between responses generated by models of varying power. However, this approach fails to assess reward models on subtle but critical content changes and variations in style, resulting in a low correlation with policy model performance. To this end, we introduce RM-Bench, a novel benchmark designed to evaluate reward models based on their sensitivity to subtle content differences and resistance to style biases. Extensive experiments demonstrate that RM-Bench strongly correlates with policy model performance, making it a reliable reference for selecting reward models to align language models effectively. We evaluate nearly 40 reward models on RM-Bench. Our results reveal that even state-of-the-art models achieve an average performance of only 46.6%, which falls short of random-level accuracy (50%) when faced with style bias interference. These findings highlight the significant room for improvement in current reward models. Related code and data are available at https://github.com/THU-KEG/RM-Bench.