Yuanxin Liu

Yuanxin Liu, Lei Li, Shuhuai Ren et al. · pku

Recently, open-domain text-to-video (T2V) generation models have made remarkable progress. However, the promising results are mainly shown by the qualitative cases of generated videos, while the quantitative evaluation of T2V models still faces two critical problems. Firstly, existing studies lack fine-grained evaluation of T2V models on different categories of text prompts. Although some benchmarks have categorized the prompts, their categorization either only focuses on a single aspect or fails to consider the temporal information in video generation. Secondly, it is unclear whether the automatic evaluation metrics are consistent with human standards. To address these problems, we propose FETV, a benchmark for Fine-grained Evaluation of Text-to-Video generation. FETV is multi-aspect, categorizing the prompts based on three orthogonal aspects: the major content, the attributes to control and the prompt complexity. FETV is also temporal-aware, which introduces several temporal categories tailored for video generation. Based on FETV, we conduct comprehensive manual evaluations of four representative T2V models, revealing their pros and cons on different categories of prompts from different aspects. We also extend FETV as a testbed to evaluate the reliability of automatic T2V metrics. The multi-aspect categorization of FETV enables fine-grained analysis of the metrics' reliability in different scenarios. We find that existing automatic metrics (e.g., CLIPScore and FVD) correlate poorly with human evaluation. To address this problem, we explore several solutions to improve CLIPScore and FVD, and develop two automatic metrics that exhibit significant higher correlation with humans than existing metrics. Benchmark page: https://github.com/llyx97/FETV.

Yuanxin Liu, Fandong Meng, Zheng Lin et al. · pku, tsinghua

Recent studies on the lottery ticket hypothesis (LTH) show that pre-trained language models (PLMs) like BERT contain matching subnetworks that have similar transfer learning performance as the original PLM. These subnetworks are found using magnitude-based pruning. In this paper, we find that the BERT subnetworks have even more potential than these studies have shown. Firstly, we discover that the success of magnitude pruning can be attributed to the preserved pre-training performance, which correlates with the downstream transferability. Inspired by this, we propose to directly optimize the subnetwork structure towards the pre-training objectives, which can better preserve the pre-training performance. Specifically, we train binary masks over model weights on the pre-training tasks, with the aim of preserving the universal transferability of the subnetwork, which is agnostic to any specific downstream tasks. We then fine-tune the subnetworks on the GLUE benchmark and the SQuAD dataset. The results show that, compared with magnitude pruning, mask training can effectively find BERT subnetworks with improved overall performance on downstream tasks. Moreover, our method is also more efficient in searching subnetworks and more advantageous when fine-tuning within a certain range of data scarcity. Our code is available at https://github.com/llyx97/TAMT.

Yuanxin Liu, Fandong Meng, Zheng Lin et al. · pku, tsinghua

Despite the remarkable success of pre-trained language models (PLMs), they still face two challenges: First, large-scale PLMs are inefficient in terms of memory footprint and computation. Second, on the downstream tasks, PLMs tend to rely on the dataset bias and struggle to generalize to out-of-distribution (OOD) data. In response to the efficiency problem, recent studies show that dense PLMs can be replaced with sparse subnetworks without hurting the performance. Such subnetworks can be found in three scenarios: 1) the fine-tuned PLMs, 2) the raw PLMs and then fine-tuned in isolation, and even inside 3) PLMs without any parameter fine-tuning. However, these results are only obtained in the in-distribution (ID) setting. In this paper, we extend the study on PLMs subnetworks to the OOD setting, investigating whether sparsity and robustness to dataset bias can be achieved simultaneously. To this end, we conduct extensive experiments with the pre-trained BERT model on three natural language understanding (NLU) tasks. Our results demonstrate that \textbf{sparse and robust subnetworks (SRNets) can consistently be found in BERT}, across the aforementioned three scenarios, using different training and compression methods. Furthermore, we explore the upper bound of SRNets using the OOD information and show that \textbf{there exist sparse and almost unbiased BERT subnetworks}. Finally, we present 1) an analytical study that provides insights on how to promote the efficiency of SRNets searching process and 2) a solution to improve subnetworks' performance at high sparsity. The code is available at https://github.com/llyx97/sparse-and-robust-PLM.

Qingyi Si, Yuanxin Liu, Zheng Lin et al. · pku

Despite the excellent performance of vision-language pre-trained models (VLPs) on conventional VQA task, they still suffer from two problems: First, VLPs tend to rely on language biases in datasets and fail to generalize to out-of-distribution (OOD) data. Second, they are inefficient in terms of memory footprint and computation. Although promising progress has been made in both problems, most existing works tackle them independently. To facilitate the application of VLP to VQA tasks, it is imperative to jointly study VLP compression and OOD robustness, which, however, has not yet been explored. This paper investigates whether a VLP can be compressed and debiased simultaneously by searching sparse and robust subnetworks. To this end, we systematically study the design of a training and compression pipeline to search the subnetworks, as well as the assignment of sparsity to different modality-specific modules. Our experiments involve 3 VLPs, 2 compression methods, 4 training methods, 2 datasets and a range of sparsity levels and random seeds. Our results show that there indeed exist sparse and robust subnetworks, which are competitive with the debiased full VLP and clearly outperform the debiasing SoTAs with fewer parameters on OOD datasets VQA-CP v2 and VQA-VS. The codes can be found at https://github.com/PhoebusSi/Compress-Robust-VQA.

Qingyi Si, Fandong Meng, Mingyu Zheng et al. · pku, tsinghua

Visual Question Answering (VQA) models are prone to learn the shortcut solution formed by dataset biases rather than the intended solution. To evaluate the VQA models' reasoning ability beyond shortcut learning, the VQA-CP v2 dataset introduces a distribution shift between the training and test set given a question type. In this way, the model cannot use the training set shortcut (from question type to answer) to perform well on the test set. However, VQA-CP v2 only considers one type of shortcut and thus still cannot guarantee that the model relies on the intended solution rather than a solution specific to this shortcut. To overcome this limitation, we propose a new dataset that considers varying types of shortcuts by constructing different distribution shifts in multiple OOD test sets. In addition, we overcome the three troubling practices in the use of VQA-CP v2, e.g., selecting models using OOD test sets, and further standardize OOD evaluation procedure. Our benchmark provides a more rigorous and comprehensive testbed for shortcut learning in VQA. We benchmark recent methods and find that methods specifically designed for particular shortcuts fail to simultaneously generalize to our varying OOD test sets. We also systematically study the varying shortcuts and provide several valuable findings, which may promote the exploration of shortcut learning in VQA.

Qingyi Si, Yuanxin Liu, Fandong Meng et al. · pku, tsinghua

Models for Visual Question Answering (VQA) often rely on the spurious correlations, i.e., the language priors, that appear in the biased samples of training set, which make them brittle against the out-of-distribution (OOD) test data. Recent methods have achieved promising progress in overcoming this problem by reducing the impact of biased samples on model training. However, these models reveal a trade-off that the improvements on OOD data severely sacrifice the performance on the in-distribution (ID) data (which is dominated by the biased samples). Therefore, we propose a novel contrastive learning approach, MMBS, for building robust VQA models by Making the Most of Biased Samples. Specifically, we construct positive samples for contrastive learning by eliminating the information related to spurious correlation from the original training samples and explore several strategies to use the constructed positive samples for training. Instead of undermining the importance of biased samples in model training, our approach precisely exploits the biased samples for unbiased information that contributes to reasoning. The proposed method is compatible with various VQA backbones. We validate our contributions by achieving competitive performance on the OOD dataset VQA-CP v2 while preserving robust performance on the ID dataset VQA v2.

Shicheng Li, Lei Li, Shuhuai Ren et al. · pku

The ability to perceive how objects change over time is a crucial ingredient in human intelligence. However, current benchmarks cannot faithfully reflect the temporal understanding abilities of video-language models (VidLMs) due to the existence of static visual shortcuts. To remedy this issue, we present VITATECS, a diagnostic VIdeo-Text dAtaset for the evaluation of TEmporal Concept underStanding. Specifically, we first introduce a fine-grained taxonomy of temporal concepts in natural language in order to diagnose the capability of VidLMs to comprehend different temporal aspects. Furthermore, to disentangle the correlation between static and temporal information, we generate counterfactual video descriptions that differ from the original one only in the specified temporal aspect. We employ a semi-automatic data collection framework using large language models and human-in-the-loop annotation to obtain high-quality counterfactual descriptions efficiently. Evaluation of representative video-language understanding models confirms their deficiency in temporal understanding, revealing the need for greater emphasis on the temporal elements in video-language research.

Bowen Shen, Zheng Lin, Yuanxin Liu et al. · pku

Transformer-based pre-trained language models (PLMs) mostly suffer from excessive overhead despite their advanced capacity. For resource-constrained devices, there is an urgent need for a spatially and temporally efficient model which retains the major capacity of PLMs. However, existing statically compressed models are unaware of the diverse complexities between input instances, potentially resulting in redundancy and inadequacy for simple and complex inputs. Also, miniature models with early exiting encounter challenges in the trade-off between making predictions and serving the deeper layers. Motivated by such considerations, we propose a collaborative optimization for PLMs that integrates static model compression and dynamic inference acceleration. Specifically, the PLM is slenderized in width while the depth remains intact, complementing layer-wise early exiting to speed up inference dynamically. To address the trade-off of early exiting, we propose a joint training approach that calibrates slenderization and preserves contributive structures to each exit instead of only the final layer. Experiments are conducted on GLUE benchmark and the results verify the Pareto optimality of our approach at high compression and acceleration rate with 1/8 parameters and 1/19 FLOPs of BERT.

Kimi Team, Angang Du, Bohong Yin et al. · pku, tsinghua

We present Kimi-VL, an efficient open-source Mixture-of-Experts (MoE) vision-language model (VLM) that offers advanced multimodal reasoning, long-context understanding, and strong agent capabilities - all while activating only 2.8B parameters in its language decoder (Kimi-VL-A3B). Kimi-VL demonstrates strong performance across challenging domains: as a general-purpose VLM, Kimi-VL excels in multi-turn agent tasks (e.g., OSWorld), matching flagship models. Furthermore, it exhibits remarkable capabilities across diverse challenging vision language tasks, including college-level image and video comprehension, OCR, mathematical reasoning, and multi-image understanding. In comparative evaluations, it effectively competes with cutting-edge efficient VLMs such as GPT-4o-mini, Qwen2.5-VL-7B, and Gemma-3-12B-IT, while surpassing GPT-4o in several key domains. Kimi-VL also advances in processing long contexts and perceiving clearly. With a 128K extended context window, Kimi-VL can process diverse long inputs, achieving impressive scores of 64.5 on LongVideoBench and 35.1 on MMLongBench-Doc. Its native-resolution vision encoder, MoonViT, further allows it to see and understand ultra-high-resolution visual inputs, achieving 83.2 on InfoVQA and 34.5 on ScreenSpot-Pro, while maintaining lower computational cost for common tasks. Building upon Kimi-VL, we introduce an advanced long-thinking variant: Kimi-VL-Thinking-2506. Developed through long chain-of-thought (CoT) supervised fine-tuning (SFT) and reinforcement learning (RL), the latest model exhibits strong long-horizon reasoning capabilities (64.0 on MMMU, 46.3 on MMMU-Pro, 56.9 on MathVision, 80.1 on MathVista, 65.2 on VideoMMMU) while obtaining robust general abilities. Code and models are publicly accessible at https://github.com/MoonshotAI/Kimi-VL.

Kun Ouyang, Yuanxin Liu, Haoning Wu et al.

Video spatial reasoning, which involves inferring the underlying spatial structure from observed video frames, poses a significant challenge for existing Multimodal Large Language Models (MLLMs). This limitation stems primarily from 1) the absence of high-quality datasets for this task, and 2) the lack of effective training strategies to develop spatial reasoning capabilities. Motivated by the success of Reinforcement Learning with Verifiable Reward (RLVR) in unlocking LLM reasoning abilities, this work aims to improve MLLMs in video spatial reasoning through the RLVR paradigm. To this end, we introduce the $\textbf{SpaceR}$ framework. First, we present $\textbf{SpaceR-151k}$, a dataset with 91k questions spanning diverse spatial reasoning scenarios with verifiable answers, and 60k samples for maintaining general multimodal understanding. Second, we propose $\textbf{Spatially-Guided RLVR (SG-RLVR)}$, a novel reinforcement learning approach that extends Group Relative Policy Optimization (GRPO) with a novel map imagination mechanism, which encourages the model to infer spatial layouts in the thinking process, thereby facilitating more effective spatial reasoning. Extensive experiments demonstrate that SpaceR achieves state-of-the-art performance on spatial reasoning benchmarks (e.g., VSI-Bench, STI-Bench, and SPAR-Bench), while maintaining competitive results on video understanding benchmarks (e.g., Video-MME, TempCompass, and LongVideoBench). Remarkably, SpaceR surpasses the advanced GPT-4o by 11.6\% accuracy on VSI-Bench and is on par with the leading proprietary model Gemini-2.0-Flash, highlighting the effectiveness of our SpaceR-151k dataset and SG-RLVR in reinforcing spatial reasoning ability of MLLMs. Code, model, and dataset are available at https://github.com/OuyangKun10/SpaceR.

Yuchi Wang, Yishuo Cai, Shuhuai Ren et al. · pku

Image recaptioning is widely used to generate training datasets with enhanced quality for various multimodal tasks. Existing recaptioning methods typically rely on powerful multimodal large language models (MLLMs) to enhance textual descriptions, but often suffer from inaccuracies due to hallucinations and incompleteness caused by missing fine-grained details. To address these limitations, we propose RICO, a novel framework that refines captions through visual reconstruction. Specifically, we leverage a text-to-image model to reconstruct a caption into a reference image, and prompt an MLLM to identify discrepancies between the original and reconstructed images to refine the caption. This process is performed iteratively, further progressively promoting the generation of more faithful and comprehensive descriptions. To mitigate the additional computational cost induced by the iterative process, we introduce RICO-Flash, which learns to generate captions like RICO using DPO. Extensive experiments demonstrate that our approach significantly improves caption accuracy and completeness, outperforms most baselines by approximately 10% on both CapsBench and CompreCap. Code released at https://github.com/wangyuchi369/RICO.

Yuanxin Liu, Zheng Lin, Fengcheng Yuan

Pre-trained language models of the BERT family have defined the state-of-the-arts in a wide range of NLP tasks. However, the performance of BERT-based models is mainly driven by the enormous amount of parameters, which hinders their application to resource-limited scenarios. Faced with this problem, recent studies have been attempting to compress BERT into a small-scale model. However, most previous work primarily focuses on a single kind of compression technique, and few attention has been paid to the combination of different methods. When BERT is compressed with integrated techniques, a critical question is how to design the entire compression framework to obtain the optimal performance. In response to this question, we integrate three kinds of compression methods (weight pruning, low-rank factorization and knowledge distillation (KD)) and explore a range of designs concerning model architecture, KD strategy, pruning frequency and learning rate schedule. We find that a careful choice of the designs is crucial to the performance of the compressed model. Based on the empirical findings, our best compressed model, dubbed Refined BERT cOmpreSsion with InTegrAted techniques (ROSITA), is $7.5 \times$ smaller than BERT while maintains $98.5\%$ of the performance on five tasks of the GLUE benchmark, outperforming the previous BERT compression methods with similar parameter budget. The code is available at https://github.com/llyx97/Rosita.

Fenglin Liu, Yuanxin Liu, Xuancheng Ren et al.

In vision-and-language grounding problems, fine-grained representations of the image are considered to be of paramount importance. Most of the current systems incorporate visual features and textual concepts as a sketch of an image. However, plainly inferred representations are usually undesirable in that they are composed of separate components, the relations of which are elusive. In this work, we aim at representing an image with a set of integrated visual regions and corresponding textual concepts, reflecting certain semantics. To this end, we build the Mutual Iterative Attention (MIA) module, which integrates correlated visual features and textual concepts, respectively, by aligning the two modalities. We evaluate the proposed approach on two representative vision-and-language grounding tasks, i.e., image captioning and visual question answering. In both tasks, the semantic-grounded image representations consistently boost the performance of the baseline models under all metrics across the board. The results demonstrate that our approach is effective and generalizes well to a wide range of models for image-related applications. (The code is available at https://github.com/fenglinliu98/MIA)

Fenglin Liu, Xuancheng Ren, Yuanxin Liu et al.

The encode-decoder framework has shown recent success in image captioning. Visual attention, which is good at detailedness, and semantic attention, which is good at comprehensiveness, have been separately proposed to ground the caption on the image. In this paper, we propose the Stepwise Image-Topic Merging Network (simNet) that makes use of the two kinds of attention at the same time. At each time step when generating the caption, the decoder adaptively merges the attentive information in the extracted topics and the image according to the generated context, so that the visual information and the semantic information can be effectively combined. The proposed approach is evaluated on two benchmark datasets and reaches the state-of-the-art performances.(The code is available at https://github.com/lancopku/simNet)

Linli Yao, Yicheng Li, Yuancheng Wei et al. · pku

The rapid growth of online video platforms, particularly live streaming services, has created an urgent need for real-time video understanding systems. These systems must process continuous video streams and respond to user queries instantaneously, presenting unique challenges for current Video Large Language Models (VideoLLMs). While existing VideoLLMs excel at processing complete videos, they face significant limitations in streaming scenarios due to their inability to handle dense, redundant frames efficiently. We introduce TimeChat-Online, a novel online VideoLLM that revolutionizes real-time video interaction. At its core lies our innovative Differential Token Drop (DTD) module, which addresses the fundamental challenge of visual redundancy in streaming videos. Drawing inspiration from human visual perception's Change Blindness phenomenon, DTD preserves meaningful temporal changes while filtering out static, redundant content between frames. Remarkably, our experiments demonstrate that DTD achieves an 82.8% reduction in video tokens while maintaining 98% performance on StreamingBench, revealing that over 80% of visual content in streaming videos is naturally redundant without requiring language guidance. To enable seamless real-time interaction, we present TimeChat-Online-139K, a comprehensive streaming video dataset featuring diverse interaction patterns including backward-tracing, current-perception, and future-responding scenarios. TimeChat-Online's unique Proactive Response capability, naturally achieved through continuous monitoring of video scene transitions via DTD, sets it apart from conventional approaches. Our extensive evaluation demonstrates TimeChat-Online's superior performance on streaming benchmarks (StreamingBench and OvOBench) and maintaining competitive results on long-form video tasks such as Video-MME and MLVU.

Han Zhong, Yutong Yin, Shenao Zhang et al.

Large Language Models (LLMs) have demonstrated remarkable capabilities in complex reasoning tasks, yet generating reliable reasoning processes remains a significant challenge. We present a unified probabilistic framework that formalizes LLM reasoning through a novel graphical model incorporating latent thinking processes and evaluation signals. Within this framework, we introduce the Bootstrapping Reinforced Thinking Process (BRiTE) algorithm, which works in two steps. First, it generates high-quality rationales by approximating the optimal thinking process through reinforcement learning, using a novel reward shaping mechanism. Second, it enhances the base LLM by maximizing the joint probability of rationale generation with respect to the model's parameters. Theoretically, we demonstrate BRiTE's convergence at a rate of $1/T$ with $T$ representing the number of iterations. Empirical evaluations on math and coding benchmarks demonstrate that our approach consistently improves performance across different base models without requiring human-annotated thinking processes. In addition, BRiTE demonstrates superior performance compared to existing algorithms that bootstrap thinking processes use alternative methods such as rejection sampling, and can even match or exceed the results achieved through supervised fine-tuning with human-annotated data.

Yuanxin Liu, Kun Ouyang, Haoning Wu et al.

Recent studies have shown that long chain-of-thought (CoT) reasoning can significantly enhance the performance of large language models (LLMs) on complex tasks. However, this benefit is yet to be demonstrated in the domain of video understanding, since most existing benchmarks lack the reasoning depth required to demonstrate the advantages of extended CoT chains. While recent efforts have proposed benchmarks aimed at video reasoning, the tasks are often knowledge-driven and do not rely heavily on visual content. To bridge this gap, we introduce VideoReasonBench, a benchmark designed to evaluate vision-centric, complex video reasoning. To ensure visual richness and high reasoning complexity, each video in VideoReasonBench depicts a sequence of fine-grained operations on a latent state that is only visible in part of the video. The questions evaluate three escalating levels of video reasoning skills: recalling observed visual information, inferring the content of latent states, and predicting information beyond the video. Under such task setting, models have to precisely recall multiple operations in the video, and perform step-by-step reasoning to get correct final answers for these questions. Using VideoReasonBench, we comprehensively evaluate 18 state-of-the-art multimodal LLMs (MLLMs), finding that most perform poorly on complex video reasoning, e.g., GPT-4o achieves only 6.9% accuracy, while the thinking-enhanced Gemini-2.5-Pro significantly outperforms others with 56.0% accuracy. Our investigations on "test-time scaling" further reveal that extended thinking budget, while offering none or minimal benefits on existing video benchmarks, is essential for improving the performance on VideoReasonBench.

Sishuo Chen, Lei Li, Shuhuai Ren et al. · pku

Video paragraph captioning (VPC) involves generating detailed narratives for long videos, utilizing supportive modalities such as speech and event boundaries. However, the existing models are constrained by the assumption of constant availability of a single auxiliary modality, which is impractical given the diversity and unpredictable nature of real-world scenarios. To this end, we propose a Missing-Resistant framework MR-VPC that effectively harnesses all available auxiliary inputs and maintains resilience even in the absence of certain modalities. Under this framework, we propose the Multimodal VPC (MVPC) architecture integrating video, speech, and event boundary inputs in a unified manner to process various auxiliary inputs. Moreover, to fortify the model against incomplete data, we introduce DropAM, a data augmentation strategy that randomly omits auxiliary inputs, paired with DistillAM, a regularization target that distills knowledge from teacher models trained on modality-complete data, enabling efficient learning in modality-deficient environments. Through exhaustive experimentation on YouCook2 and ActivityNet Captions, MR-VPC has proven to deliver superior performance on modality-complete and modality-missing test data. This work highlights the significance of developing resilient VPC models and paves the way for more adaptive, robust multimodal video understanding.

Shicheng Li, Lei Li, Kun Ouyang et al. · pku

Video Large Language Models (Video LLMs) have achieved significant success by adopting the paradigm of large-scale pre-training followed by supervised fine-tuning (SFT). However, existing approaches struggle with temporal reasoning due to weak temporal correspondence in the data and over-reliance on the next-token prediction paradigm}, which collectively result in the absence temporal supervision. To address these limitations, we propose TEMPLE (TEMporal Preference LEarning), a systematic framework that enhances temporal reasoning capabilities through Direct Preference Optimization (DPO). To address temporal information scarcity in data, we introduce an automated pipeline for systematically constructing temporality-intensive preference pairs comprising three steps: selecting temporally rich videos, designing video-specific perturbation strategies, and evaluating model responses on clean and perturbed inputs. Complementing this data pipeline, we provide additional supervision signals via preference learning and propose a novel Progressive Pre-SFT Alignment strategy featuring two key innovations: a curriculum learning strategy which progressively increases perturbation difficulty to maximize data efficiency; and applying preference optimization before instruction tuning to incentivize fundamental temporal alignment. Extensive experiments demonstrate that our approach consistently improves Video LLM performance across multiple benchmarks with a relatively small set of self-generated DPO data. Our findings highlight TEMPLE as a scalable and efficient complement to SFT-based methods, paving the way for developing reliable Video LLMs.

Yuanxin Liu, Rui Zhu, Shuhuai Ren et al.

With the rapid growth of video generative models (VGMs), it is essential to develop reliable and comprehensive automatic metrics for AI-generated videos (AIGVs). Existing methods either use off-the-shelf models optimized for other tasks or rely on human assessment data to train specialized evaluators. These approaches are constrained to specific evaluation aspects and are difficult to scale with the increasing demands for finer-grained and more comprehensive evaluations. To address this issue, this work investigates the feasibility of using multimodal large language models (MLLMs) as a unified evaluator for AIGVs, leveraging their strong visual perception and language understanding capabilities. To evaluate the performance of automatic metrics in unified AIGV evaluation, we introduce a benchmark called UVE-Bench. UVE-Bench collects videos generated by state-of-the-art VGMs and provides pairwise human preference annotations across 15 evaluation aspects. Using UVE-Bench, we extensively evaluate 18 MLLMs. Our empirical results suggest that while advanced MLLMs (e.g., Qwen2VL-72B and InternVL2.5-78B) still lag behind human evaluators, they demonstrate promising ability in unified AIGV evaluation, significantly surpassing existing specialized evaluation methods. Additionally, we conduct an in-depth analysis of key design choices that impact the performance of MLLM-driven evaluators, offering valuable insights for future research on AIGV evaluation.

Kun Ouyang, Yuanxin Liu, Shicheng Li et al.

Multimodal punchlines, which involve humor or sarcasm conveyed in image-caption pairs, are a popular way of communication on online multimedia platforms. With the rapid development of multimodal large language models (MLLMs), it is essential to assess their ability to effectively comprehend these punchlines. However, existing benchmarks on punchline comprehension suffer from three major limitations: 1) language shortcuts that allow models to solely rely on text, 2) lack of question diversity, and 3) narrow focus on a specific domain of multimodal content (e.g., cartoon). To address these limitations, we introduce a multimodal \textbf{Punch}line comprehension \textbf{Bench}mark, named \textbf{PunchBench}, which is tailored for accurate and comprehensive evaluation of punchline comprehension. To enhance the evaluation accuracy, we generate synonymous and antonymous captions by modifying original captions, which mitigates the impact of shortcuts in the captions. To provide a comprehensive evaluation, PunchBench incorporates diverse question formats and image-captions from various domains. On this basis, we conduct extensive evaluations and reveal a significant gap between state-of-the-art MLLMs and humans in punchline comprehension. To improve punchline comprehension, we propose Simple-to-Complex Chain-of-Question (SC-CoQ) strategy, enabling the models to incrementally address complicated questions by first mastering simple ones. SC-CoQ effectively enhances the performance of various MLLMs on PunchBench, surpassing in-context learning and chain-of-thought.

Kun Ouyang, Yuanxin Liu, Linli Yao et al.

Video reasoning, which requires multi-step deduction across frames, remains a major challenge for multimodal large language models (MLLMs). While reinforcement learning (RL)-based methods enhance reasoning capabilities, they often rely on text-only chains that yield ungrounded or hallucinated conclusions. Conversely, frame-retrieval approaches introduce visual grounding, yet still struggle with inaccurate evidence localization. To address these limitations, we present Conan, a framework for evidence-grounded multi-step video reasoning. Conan identifies context and evidence frames, reasons over cross-frame clues, and adaptively decides when to conclude or explore further. To achieve this, we 1) construct Conan-91K, a large-scale dataset of automatically generated reasoning traces that include frame identification, evidence reasoning, and action decision, and 2) design a multi-stage progressive cold-start strategy combined with an Identification-Reasoning-Action (AIR) RLVR training framework to progressively incentivize multi-step visual reasoning. Extensive experiments on six multi-step reasoning benchmarks demonstrate that Conan surpasses the baseline Qwen2.5-VL-7B-Instruct by an average of over 10% in accuracy, achieving state-of-the-art performance. Furthermore, Conan generalizes effectively to long video understanding tasks, validating its strong scalability and robustness.

Yuanxin Liu, Fandong Meng, Zheng Lin et al.

Recently, knowledge distillation (KD) has shown great success in BERT compression. Instead of only learning from the teacher's soft label as in conventional KD, researchers find that the rich information contained in the hidden layers of BERT is conducive to the student's performance. To better exploit the hidden knowledge, a common practice is to force the student to deeply mimic the teacher's hidden states of all the tokens in a layer-wise manner. In this paper, however, we observe that although distilling the teacher's hidden state knowledge (HSK) is helpful, the performance gain (marginal utility) diminishes quickly as more HSK is distilled. To understand this effect, we conduct a series of analysis. Specifically, we divide the HSK of BERT into three dimensions, namely depth, length and width. We first investigate a variety of strategies to extract crucial knowledge for each single dimension and then jointly compress the three dimensions. In this way, we show that 1) the student's performance can be improved by extracting and distilling the crucial HSK, and 2) using a tiny fraction of HSK can achieve the same performance as extensive HSK distillation. Based on the second finding, we further propose an efficient KD paradigm to compress BERT, which does not require loading the teacher during the training of student. For two kinds of student models and computing devices, the proposed KD paradigm gives rise to training speedup of 2.7x ~ 3.4x.

Qingyi Si, Yuanxin Liu, Peng Fu et al.

Zero-shot intent detection (ZSID) aims to deal with the continuously emerging intents without annotated training data. However, existing ZSID systems suffer from two limitations: 1) They are not good at modeling the relationship between seen and unseen intents. 2) They cannot effectively recognize unseen intents under the generalized intent detection (GZSID) setting. A critical problem behind these limitations is that the representations of unseen intents cannot be learned in the training stage. To address this problem, we propose a novel framework that utilizes unseen class labels to learn Class-Transductive Intent Representations (CTIR). Specifically, we allow the model to predict unseen intents during training, with the corresponding label names serving as input utterances. On this basis, we introduce a multi-task learning objective, which encourages the model to learn the distinctions among intents, and a similarity scorer, which estimates the connections among intents more accurately. CTIR is easy to implement and can be integrated with existing methods. Experiments on two real-world datasets show that CTIR brings considerable improvement to the baseline systems.

Fenglin Liu, Xuancheng Ren, Yuanxin Liu et al.

Recently, attention-based encoder-decoder models have been used extensively in image captioning. Yet there is still great difficulty for the current methods to achieve deep image understanding. In this work, we argue that such understanding requires visual attention to correlated image regions and semantic attention to coherent attributes of interest. Based on the Transformer, to perform effective attention, we explore image captioning from a cross-modal perspective and propose the Global-and-Local Information Exploring-and-Distilling approach that explores and distills the source information in vision and language. It globally provides the aspect vector, a spatial and relational representation of images based on caption contexts, through the extraction of salient region groupings and attribute collocations, and locally extracts the fine-grained regions and attributes in reference to the aspect vector for word selection. Our Transformer-based model achieves a CIDEr score of 129.3 in offline COCO evaluation on the COCO testing set with remarkable efficiency in terms of accuracy, speed, and parameter budget.

Yuanxin Liu, Zheng Lin

Recently, unsupervised pre-training is gaining increasing popularity in the realm of computational linguistics, thanks to its surprising success in advancing natural language understanding (NLU) and the potential to effectively exploit large-scale unlabelled corpus. However, regardless of the success in NLU, the power of unsupervised pre-training is only partially excavated when it comes to natural language generation (NLG). The major obstacle stems from an idiosyncratic nature of NLG: Texts are usually generated based on certain context, which may vary with the target applications. As a result, it is intractable to design a universal architecture for pre-training as in NLU scenarios. Moreover, retaining the knowledge learned from pre-training when learning on the target task is also a non-trivial problem. This review summarizes the recent efforts to enhance NLG systems with unsupervised pre-training, with a special focus on the methods to catalyse the integration of pre-trained models into downstream tasks. They are classified into architecture-based methods and strategy-based methods, based on their way of handling the above obstacle. Discussions are also provided to give further insights into the relationship between these two lines of work, some informative empirical phenomenons, as well as some possible directions where future work can be devoted to.