Multimodal large language models (MLLMs) have made significant strides by integrating visual and textual modalities. A critical factor in training MLLMs is the quality of image-text pairs within multimodal pretraining datasets. However, in the process of high-quality data curation, filter-based paradigms often discard a substantial portion of high-quality images due to inadequate semantic alignment between images and texts, leading to inefficiency in data utilization and scalability. In this paper, we propose the Adaptive Image-Text Quality Enhancer (AITQE), a model that dynamically assesses and enhances the quality of image-text pairs. AITQE employs a text rewriting mechanism for low-quality pairs and incorporates a negative sample learning strategy to improve evaluative capabilities by integrating deliberately generated low-quality samples during training. Unlike prior approaches that significantly alter text distributions, our method minimally adjusts text to preserve data volume while enhancing quality. Experimental results demonstrate that AITQE surpasses existing methods on various benchmarks, effectively leveraging raw data and scaling with increasing data volumes. Codes and model are available at https://github.com/hanhuang22/AITQE.

Zican Dong, Junyi Li, Jinhao Jiang, Mingyu Xu, Xin Zhao, Bingning Wang, Weipeng Chen. Proceedings of the 63rd Annual Meeting of the Association for Computational Linguistics (Volume 1: Long Papers). 2025.

Deep Neural Networks (DNNs) have achieved remarkable success in a variety of tasks, particularly in terms of prediction accuracy. However, in real-world scenarios, especially in safety-critical applications, accuracy alone is insufficient; reliable uncertainty estimates are essential. Modern DNNs, often trained with cross-entropy loss, tend to exhibit overconfidence, especially on ambiguous samples. Many techniques aim to improve uncertainty calibration, yet they often come at the cost of reduced accuracy or increased computational demands. To address this challenge, we propose Differentiated Deep Mutual Learning (Diff-DML), an efficient ensemble approach that simultaneously enhances accuracy and uncertainty calibration. Diff-DML draws inspiration from Deep Mutual Learning (DML) while introducing two strategies to maintain prediction diversity: (1) Differentiated Training Strategy (DTS) and (2) Diversity-Preserving Learning Objective (DPLO). Our theoretical analysis shows that Diff-DML’s diversified learning framework not only leverages ensemble benefits but also avoids the loss of prediction diversity observed in traditional DML setups, which is crucial for improved calibration. Extensive evaluations on various benchmarks confirm the effectiveness of Diff-DML. For instance, on the CIFAR-100 dataset, Diff-DML on ResNet34/50 models achieved substantial improvements over the previous state-of-the-art method, MDCA, with absolute accuracy gains of 1.3%/3.1%, relative ECE reductions of 49.6%/43.8%, and relative classwise-ECE reductions of 7.7%/13.0%.

Most current video MLLMs rely on uniform frame sampling and image-level encoders, resulting in inefficient data processing and limited motion awareness. To address these challenges, we introduce EMA, an Efficient Motion-Aware video MLLM that utilizes compressed video structures as inputs. We propose a motion-aware GOP (Group of Pictures) encoder that fuses spatial and motion information within a GOP unit in the compressed video stream, generating compact, informative visual tokens. By integrating fewer but denser RGB frames with more but sparser motion vectors in this native slow-fast input architecture, our approach reduces redundancy and enhances motion representation. Additionally, we introduce MotionBench, a benchmark for evaluating motion understanding across four motion types: linear, curved, rotational, and contact-based. Experimental results show that EMA achieves state-of-the-art performance on both MotionBench and popular video question answering benchmarks, while reducing inference costs. Moreover, EMA demonstrates strong scalability, as evidenced by its competitive performance on long video understanding benchmarks.

Multi-lingual ability transfer has become increasingly important for the broad application of large language models (LLMs).Existing work highly relies on training with the multi-lingual ability-related data, which may not be available for low-resource languages.To solve it, we propose a Multilingual Abilities Extraction and Combination approach (MAEC), which decomposes and extracts language-agnostic ability-related weights from LLMs, and combines them across different languages by simple addition and subtraction operations without training.Specifically, our MAEC consists of the extraction and combination stages.In the extraction stage, we firstly locate key neurons that are highly related to specific abilities, and then employ them to extract the transferable ability-related weights.In the combination stage, we further select the ability-related tensors that mitigate the linguistic effects, and design a combining strategy based on them and the languagespecific weights, to build the multi-lingual ability-enhanced LLM.To assess the effectiveness of our approach, we conduct extensive experiments on LLaMA-3 8B on mathematical and scientific tasks in both high-resource and low-resource lingual scenarios.Empirical results have shown that MAEC can effectively and efficiently extract and combine the advanced abilities, achieving comparable performance with PaLM.Resources are available at https://github.com/RUCAIBox/MAET.

Video Multimodal Large Language Models~(MLLMs) have shown remarkable capability of understanding the video semantics on various downstream tasks. Despite the advancements, there is still a lack of systematic research on visual context representation, which refers to the scheme to select frames from a video and further select the tokens from a frame. In this paper, we explore the design space for visual context representation, and aim to improve the performance of video MLLMs by finding more effective representation schemes. Firstly, we formulate the task of visual context representation as a constrained optimization problem, and model the language modeling loss as a function of the number of frames and the number of embeddings (or tokens) per frame, given the maximum visual context window size. Then, we explore the scaling effects in frame selection and token selection respectively, and fit the corresponding function curve by conducting extensive empirical experiments. We examine the effectiveness of typical selection strategies and present empirical findings to determine the two factors. Furthermore, we study the joint effect of frame selection and token selection, and derive the optimal formula for determining the two factors. We demonstrate that the derived optimal settings show alignment with the best-performed results of empirical experiments. The data and code are available at: https://github.com/RUCAIBox/Opt-Visor.

Video understanding is a crucial next step for multimodal large language models (MLLMs). Various benchmarks are introduced for better evaluating the MLLMs. Nevertheless, current video benchmarks are still inefficient for evaluating video models during iterative development due to the high cost of constructing datasets and the difficulty in isolating specific skills. In this paper, we propose VideoNIAH (Video Needle in A Haystack), a benchmark construction framework through synthetic video generation. VideoNIAH decouples video content from their query-responses by inserting unrelated visual 'needles' into original videos. The framework automates the generation of query-response pairs using predefined rules, minimizing manual labor. The queries focus on specific aspects of video understanding, enabling more skill-specific evaluations. The separation between video content and the queries also allow for increased video variety and evaluations across different lengths. Utilizing VideoNIAH, we compile a video benchmark, VNBench, which includes tasks such as retrieval, ordering, and counting to evaluate three key aspects of video understanding: temporal perception, chronological ordering, and spatio-temporal coherence. We conduct a comprehensive evaluation of both proprietary and open-source models, uncovering significant differences in their video understanding capabilities across various tasks. Additionally, we perform an in-depth analysis of the test results and model configurations. Based on these findings, we provide some advice for improving video MLLM training, offering valuable insights to guide future research and model development.

The rapid proliferation of large language models (LLMs), such as GPT-4 and Gemini, underscores the intense demand for resources during their training processes, posing significant challenges due to substantial computational and environmental costs. In this paper, we introduce a novel checkpoint merging strategy aimed at making efficient use of intermediate checkpoints during LLM pretraining. This method utilizes intermediate checkpoints with shared training trajectories, and is rooted in an extensive search space exploration for the best merging weight via Bayesian optimization. Through various experiments, we demonstrate that: (1) Our proposed methodology exhibits the capacity to augment pretraining, presenting an opportunity akin to obtaining substantial benefits at minimal cost; (2) Our proposed methodology, despite requiring a given held-out dataset, still demonstrates robust generalization capabilities across diverse domains, a pivotal aspect in pretraining.

Current large language models (LLMs) predominantly rely on decode-only transformer architectures, which exhibit exceptional in-context learning (ICL) capabilities. It is widely acknowledged that the cornerstone of their ICL ability lies in the induction heads mechanism, which necessitates at least two layers of attention. To more effectively harness the model's induction capabilities, we revisit the induction heads mechanism and provide theoretical proof that KV shifting attention reduces the model's dependency on the depth and width of the induction heads mechanism. Our experimental results confirm that KV shifting attention enhances the learning of induction heads and improves language modeling performance. This leads to superior performance or accelerated convergence, spanning from toy models to pre-trained models with over 10 billion parameters.

Yao Xu, Shizhu He, Jiabei Chen, ZengXiangrong ZengXiangrong, Bingning Wang, Guang Liu, Jun Zhao, Kang Liu. Proceedings of the 2025 Conference of the Nations of the Americas Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2025.

The advent of large language models (LLMs) like GPT-4 has catalyzed the exploration of multi-task learning (MTL), in which a single model demonstrates proficiency across diverse tasks.Task arithmetic has emerged as a costeffective approach for MTL.It enables performance enhancement across multiple tasks by adding their corresponding task vectors to a pre-trained model.However, the current lack of a method that can achieve optimal performance with low computational cost and protecting the data privacy, which limits their application to LLMs.In this paper, we propose Model Exclusive Task Arithmetic for merging GPT-scale models (MetaGPT), which formalizes the objective of model merging into a multi-task learning framework, aiming to minimize the average loss difference between the merged model and each individual task model.Since data privacy limits the use of multi-task training data, we leverage LLMs' local linearity and task vectors' orthogonality to separate the data term and scaling coefficients term and derive a model-exclusive task arithmetic method.Our proposed MetaGPT is dataagnostic and bypasses the heavy search process, making it cost-effective and easy to implement for LLMs.Extensive experiments demonstrate that MetaGPT leads to improvements in task arithmetic and achieves state-of-the-art performance on multiple tasks.

Transformer-based large language models (LLMs) typically have a limited context window, resulting in significant performance degradation when processing text beyond the length of the context window. Extensive studies have been proposed to extend the context window and achieve length extrapolation of LLMs, but there is still a lack of in-depth interpretation of these approaches. In this study, we explore the positional information within and beyond the context window for deciphering the underlying mechanism of LLMs. By using a mean-based decomposition method, we disentangle positional vectors from hidden states of LLMs and analyze their formation and effect on attention. Furthermore, when texts exceed the context window, we analyze the change of positional vectors in two settings, i.e., direct extrapolation and context window extension. Based on our findings, we design two training-free context window extension methods, positional vector replacement and attention window extension. Experimental results show that our methods can effectively extend the context window length.

Position embedding is a core component of current Large Language Models (LLMs). Rotary position embedding (RoPE), a technique that encodes the position information with a rotation matrix, has been the de facto choice for position embedding in many LLMs, such as the Llama series. RoPE has been further utilized to extend long context capability, which is roughly based on adjusting the \textit{base} parameter of RoPE to mitigate out-of-distribution (OOD) problems in position embedding. However, in this paper, we find that LLMs may obtain a superficial long-context ability based on the OOD theory. We revisit the role of RoPE in LLMs and propose a novel property of long-term decay, we derive that the \textit{base of RoPE bounds context length}: there is an absolute lower bound for the base value to obtain certain context length capability. Our work reveals the relationship between context length and RoPE base both theoretically and empirically, which may shed light on future long context training.

Passage ranking involves two stages: passage retrieval and passage re-ranking, which are important and challenging topics for both academics and industries in the area of Information Retrieval (IR). However, the commonly-used datasets for passage ranking usually focus on the English language. For non-English scenarios, such as Chinese, the existing datasets are limited in terms of data scale, fine-grained relevance annotation and false negative issues. To address this problem, we introduce T2Ranking, a large-scale Chinese benchmark for passage ranking. T2Ranking comprises more than 300K queries and over 2M unique passages from real-world search engines. Expert annotators are recruited to provide 4-level graded relevance scores (fine-grained) for query-passage pairs instead of binary relevance judgments (coarse-grained). To ease the false negative issues, more passages with higher diversities are considered when performing relevance annotations, especially in the test set, to ensure a more accurate evaluation. Apart from the textual query and passage data, other auxiliary resources are also provided, such as query types and XML files of documents which passages are generated from, to facilitate further studies. To evaluate the dataset, commonly used ranking models are implemented and tested on T2Ranking as baselines. The experimental results show that T2Ranking is challenging and there is still scope for improvement. The full data and all codes are available at https://github.com/THUIR/T2Ranking/.

Visual question answering is an important task in both natural language and vision understanding. However, in most of the public visual question answering datasets such as VQA, CLEVR, the questions are human generated that specific to the given image, such as 'What color are her eyes?'. The human generated crowdsourcing questions are relatively simple and sometimes have the bias toward certain entities or attributes [1, 55].

Knowledge graph (KG) embedding aims to encode entities and relations into low-dimensional vector spaces, in turn, can support various machine learning models on KG related tasks with good performance. However, existing methods for knowledge graph embedding fail to consider the influence of the embedding space, which makes them still unsatisfactory in practical applications. In this study, we try to improve the expressiveness of the embedding space from the perspective of the metric. Specifically, we first point out the implications of Minkowski metric used in KG embedding and then make a quantitative analysis. To solve the limitations, we introduce a new metric, named Cycle metric, based on the oscillation property of the periodic function. Furthermore, we find that the function period has a significant influence on the expressiveness of the embedding space. Given a fully trained model, the smaller the period, the better the expressive ability. Finally, to validate the findings, we propose a new model, named CyclE by combining Cycle Metric and the popular KG embeddings models. Comprehensive experimental results show that Cycle is more appropriate than Minkowski for KG embedding.

With the development of deep learning techniques and large scale datasets, the question answering (QA) systems have been quickly improved, providing more accurate and satisfying answers. However, current QA systems either focus on the sentence-level answer, i.e., answer selection, or phrase-level answer, i.e., machine reading comprehension. How to produce compositional answers has not been throughout investigated. In compositional question answering, the systems should assemble several supporting evidence from the document to generate the final answer, which is more difficult than sentence-level or phrase-level QA. In this paper, we present a large-scale compositional question answering dataset containing more than 120k human-labeled questions. The answer in this dataset is composed of discontiguous sentences in the corresponding document. To tackle the ComQA problem, we proposed a hierarchical graph neural networks, which represent the document from the low-level word to the high-level sentence. We also devise a question selection and node selection task for pre-training. Our proposed model achieves a significant improvement over previous machine reading comprehension methods and pre-training methods. Codes, dataset can be found at https://github.com/benywon/ComQA.

Neural question generation (NQG) is the task of generating questions from the given context with deep neural networks. Previous answer-aware NQG methods suffer from the problem that the generated answers are focusing on entity and most of the questions are trivial to be answered. The answer-agnostic NQG methods reduce the bias towards named entities and increasing the model's degrees of freedom, but sometimes result in generating unanswerable questions which are not valuable for the subsequent machine reading comprehension system. In this paper, we treat the answers as the hidden pivot for question generation and combine the question generation and answer selection process in a joint model. We achieve the state-of-the-art result on the SQuAD dataset according to automatic metric and human evaluation.

This paper presents the ReCO, a human-curated Chinese Reading Comprehension dataset on Opinion. The questions in ReCO are opinion based queries issued to commercial search engine. The passages are provided by the crowdworkers who extract the support snippet from the retrieved documents. Finally, an abstractive yes/no/uncertain answer was given by the crowdworkers. The release of ReCO consists of 300k questions that to our knowledge is the largest in Chinese reading comprehension. A prominent characteristic of ReCO is that in addition to the original context paragraph, we also provided the support evidence that could be directly used to answer the question. Quality analysis demonstrates the challenge of ReCO that it requires various types of reasoning skills such as causal inference, logical reasoning, etc. Current QA models that perform very well on many question answering problems, such as BERT (Devlin et al. 2018), only achieves 77% accuracy on this dataset, a large margin behind humans nearly 92% performance, indicating ReCO present a good challenge for machine reading comprehension. The codes, dataset and leaderboard will be freely available at https://github.com/benywon/ReCO.

Aspect-based sentiment analysis (ABSA) aims to predict the sentiment towards a specific aspect in the text.However, existing ABSA test sets cannot be used to probe whether a model can distinguish the sentiment of the target aspect from the non-target aspects.To solve this problem, we develop a simple but effective approach to enrich ABSA test sets.Specifically, we generate new examples to disentangle the confounding sentiments of the non-target aspects from the target aspect's sentiment.Based on the SemEval 2014 dataset, we construct the Aspect Robustness Test Set (ARTS) as a comprehensive probe of the aspect robustness of ABSA models.Over 92% data of ARTS show high fluency and desired sentiment on all aspects by human evaluation.Using ARTS, we analyze the robustness of nine ABSA models, and observe, surprisingly, that their accuracy drops by up to 69.73%.We explore several ways to improve aspect robustness, and find that adversarial training can improve models' performance on ARTS by up to 32.85%. 1

Open-domain question answering focuses on using diverse information resources to answer any types of question. Recent years, with the development of large-scale data set and various deep neural networks models, some recent advances in open domain question answering system first utilize the distantly supervised dataset as the knowledge resource, then apply deep learning based machine comprehension techniques to generate the right answers, which achieves impressive results compared with traditional feature-based pipeline methods.

With the amount of data has been rapidly growing over recent decades, binary hashing has become an attractive approach for fast search over large databases, in which the high-dimensional data such as image, video or text is mapped into a low-dimensional binary code. Searching in this hamming space is extremely efficient which is independent of the data size. A lot of methods have been proposed to learn this binary mapping. However, to make the binary codes conserves the input information, previous works mostly resort to mean squared error, which is prone to lose a lot of input information [11]. On the other hand, most of the previous works adopt the norm constraint or approximation on the hidden representation to make it as close as possible to binary, but the norm constraint is too strict that harms the expressiveness and flexibility of the code.

Recently proposed Story Cloze Test [Mostafazadeh et al., 2016] is a commonsense machine comprehension application to deal with natural language understanding problem. This dataset contains a lot of story tests which require commonsense inference ability. Unfortunately, the training data is almost unsupervised where each context document followed with only one positive sentence that can be inferred from the context. However, in the testing period, we must make inference from two candidate sentences. To tackle this problem, we employ the generative adversarial networks (GANs) to generate fake sentence. We proposed a Conditional GANs in which the generator is conditioned by the context. Our experiments show the advantage of the CGANs in discriminating sentence and achieve state-of-the-art results in commonsense story reading comprehension task compared with previous feature engineering and deep learning methods.

Attention based recurrent neural networks have shown advantages in representing natural language sentences (Hermann et al., 2015;Rocktäschel et al., 2015;Tan et al., 2015).Based on recurrent neural networks (RNN), external attention information was added to hidden representations to get an attentive sentence representation.Despite the improvement over nonattentive models, the attention mechanism under RNN is not well studied.In this work, we analyze the deficiency of traditional attention based RNN models quantitatively and qualitatively.Then we present three new RNN models that add attention information before RNN hidden representation, which shows advantage in representing sentence and achieves new stateof-art results in answer selection task.