Enabling multi-task adaptation in pre-trained Low-Rank Adaptation (LoRA) models is crucial for enhancing their generalization capabilities. Most existing pre-trained LoRA fusion methods decompose weight matrices, sharing similar parameters, while fusion divergent ones. However, this paradigm inevitably induces inter-weight conflicts and leads to catastrophic domain forgetting. While incremental learning enables adaptation to multiple tasks, it struggles to achieve generalization in few-shot scenarios. Consequently, when the weight data follows a long-tailed distribution, it can lead to forgetting in the fused weights. To address this issue, we propose In-Context Meta LoRA Fusion (ICM-Fusion), a novel framework that synergizes meta-learning with in-context adaptation. The key innovation lies in our task vector arithmetic, which dynamically balances conflicting optimization directions across domains through learned manifold projections. ICM-Fusion obtains the optimal task vector orientation for the fused model in the latent space by adjusting the orientation of the task vectors. Subsequently, the fused LoRA is reconstructed by a self-designed Fusion VAE (F-VAE) to realize multi-task LoRA generation. We have conducted extensive experiments on visual and linguistic tasks, and the experimental results demonstrate that ICM-Fusion can be adapted to a wide range of architectural models and applied to various tasks. Compared to the current pre-trained LoRA fusion method, ICM-Fusion fused LoRA can significantly reduce the multi-tasking loss and can even achieve task enhancement in few-shot scenarios.

Retrieval-augmented generation (RAG) has emerged to address the knowledge-intensive visual question answering (VQA) task. Current methods mainly employ separate retrieval and generation modules to acquire external knowledge and generate answers, respectively. We propose ReAuSE, an alternative to the previous RAG model for the knowledge-based VQA task, which seamlessly integrates knowledge retriever into the generative multi-modal large language model, serving as a built-in search engine. Specifically, our model functions both as a generative retriever and an accurate answer generator. It not only helps retrieve documents from the knowledge base by producing identifier for each document, but it also answers visual questions based on the retrieved documents. Furthermore, we also propose a reinforced retrieval calibration module from relevance feedback to improve retrieval performance and align with the preferences for accurate answer generation. Extensive experiments on two representative OKVQA and A-OKVQA datasets demonstrate significant improvements ranging from 2.9% to 9.6% across all evaluation metrics when compared to strong baselines.

Sihang Zeng, Kai Tian, Kaiyan Zhang, Yuru Wang, Junqi Gao, Runze Liu, Sa Yang, Jingxuan Li, Xinwei Long, Jiaheng Ma, Biqing Qi, Bowen Zhou. Proceedings of the 2025 Conference on Empirical Methods in Natural Language Processing. 2025.

Large language models (LLMs) have taken a great step towards AGI. Meanwhile, an increasing number of domain-specific problems such as math and programming boost these general-purpose models to continuously evolve via learning deeper expertise. Now is thus the time further to extend the diversity of specialized applications for knowledgeable LLMs, though collecting high quality data with unexpected and informative tasks is challenging. In this paper, we propose to use advertisement (ad) videos as a challenging test-bed to probe the ability of LLMs in perceiving beyond the objective physical content of common visual domain. Our motivation is to take full advantage of the clue-rich and information-dense ad videos' traits, e.g., marketing logic, persuasive strategies, and audience engagement. Our contribution is three-fold: (1) To our knowledge, this is the first attempt to use ad videos with well-designed tasks to evaluate LLMs. We contribute AdsQA, a challenging ad Video QA benchmark derived from 1,544 ad videos with 10,962 clips, totaling 22.7 hours, providing 5 challenging tasks. (2) We propose ReAd-R, a Deepseek-R1 styled RL model that reflects on questions, and generates answers via reward-driven optimization. (3) We benchmark 14 top-tier LLMs on AdsQA, and our \texttt{ReAd-R}~achieves the state-of-the-art outperforming strong competitors equipped with long-chain reasoning capabilities by a clear margin.

Low-rank Adaptation (LoRA) has demonstrated remarkable capabilities for task specific fine-tuning. However, in scenarios that involve multiple tasks, training a separate LoRA model for each one results in considerable inefficiency in terms of storage and inference. Moreover, existing parameter generation methods fail to capture the correlations among these tasks, making multi-task LoRA parameter generation challenging. To address these limitations, we propose In-Context Meta LoRA (ICM-LoRA), a novel approach that efficiently achieves task-specific customization of large language models (LLMs). Specifically, we use training data from all tasks to train a tailored generator, Conditional Variational Autoencoder (CVAE). CVAE takes task descriptions as inputs and produces task-aware LoRA weights as outputs. These LoRA weights are then merged with LLMs to create task-specialized models without the need for additional fine-tuning. Furthermore, we utilize in-context meta-learning for knowledge enhancement and task mapping, to capture the relationship between tasks and parameter distributions. As a result, our method achieves more accurate LoRA parameter generation for diverse tasks using CVAE. ICM-LoRA enables more accurate LoRA parameter reconstruction than current parameter reconstruction methods and is useful for implementing task-specific enhancements of LoRA parameters. At the same time, our method occupies 283MB, only 1% storage compared with the original LoRA. The code is available at https://github.com/YihuaJerry/ICM-LoRA.

In complex transportation systems, accurately sensing the surrounding environment and predicting the risk of potential accidents is crucial. Most existing accident prediction methods are based on temporal neural networks, such as RNN and LSTM. Recent multimodal fusion approaches improve vehicle localization through 3D target detection and assess potential risks by calculating inter-vehicle distances. However, these temporal networks and multimodal fusion methods suffer from limited detection robustness and high economic costs. To address these challenges, we propose AccidentBlip, a vision-only framework that employs our self-designed Motion Accident Transformer (MA-former) to process each frame of video. Unlike conventional self-attention mechanisms, MA-former replaces Q-former's self-attention with temporal attention, allowing the query corresponding to the previous frame to generate the query input for the next frame. Additionally, we introduce a residual module connection between queries of consecutive frames to enhance the model's temporal processing capabilities. For complex V2V and V2X scenarios, AccidentBlip adapts by concatenating queries from multiple cameras, effectively capturing spatial and temporal relationships. In particular, AccidentBlip achieves SOTA performance in both accident detection and prediction tasks on the DeepAccident dataset. It also outperforms current SOTA methods in V2V and V2X scenarios, demonstrating a superior capability to understand complex real-world environments.

Video Anomaly Detection (VAD) focuses on identifying anomalies within videos. Supervised methods require an amount of in-domain training data and often struggle to generalize to unseen anomalies. In contrast, training-free methods leverage the intrinsic world knowledge of large language models (LLMs) to detect anomalies but face challenges in localizing fine-grained visual transitions and diverse events. Therefore, we propose EventVAD, an event-aware video anomaly detection framework that combines tailored dynamic graph architectures and multimodal LLMs to perform fine-grained temporal-event reasoning. Specifically, EventVAD first employs dynamic spatiotemporal graph modeling with time-decay constraints to capture event-aware video features. Then, it performs adaptive noise filtering and uses signal ratio thresholding to detect event boundaries via unsupervised statistical features. Finally, it utilizes a hierarchical prompting strategy to guide MLLMs in performing reasoning and making final decisions. We conducted extensive experiments on the UCF-Crime and XD-Violence datasets. The results demonstrate that EventVAD with a 7B MLLM achieves state-of-the-art (SOTA) in training-free settings, outperforming strong baselines that use 7B or larger MLLMs. The code is available at https://github.com/YihuaJerry/EventVAD.

This paper investigates Reinforcement Learning (RL) on data without explicit labels for reasoning tasks in Large Language Models (LLMs). The core challenge of the problem is reward estimation during inference while not having access to ground-truth information. While this setting appears elusive, we find that common practices in Test-Time Scaling (TTS), such as majority voting, yield surprisingly effective rewards suitable for driving RL training. In this work, we introduce Test-Time Reinforcement Learning (TTRL), a novel method for training LLMs using RL on unlabeled data. TTRL enables self-evolution of LLMs by utilizing the priors in the pre-trained models. Our experiments demonstrate that TTRL consistently improves performance across a variety of tasks and models. Notably, TTRL boosts the pass@1 performance of Qwen-2.5-Math-7B by approximately 211% on the AIME 2024 with only unlabeled test data. Furthermore, although TTRL is only supervised by the Maj@N metric, TTRL has demonstrated performance to consistently surpass the upper limit of the initial model, and approach the performance of models trained directly on test data with ground-truth labels. Our experimental findings validate the general effectiveness of TTRL across various tasks and highlight TTRL's potential for broader tasks and domains.

Knowledge retrieval with multi-modal queries plays a crucial role in supporting knowledge-intensive multi-modal applications. However, existing methods face challenges in terms of their effectiveness and training efficiency, especially when it comes to training and integrating multiple retrievers to handle multi-modal queries. In this paper, we propose an innovative end-to-end generative framework for multi-modal knowledge retrieval. Our framework takes advantage of the fact that large language models (LLMs) can effectively serve as virtual knowledge bases, even when trained with limited data. We retrieve knowledge via a two-step process: 1) generating knowledge clues related to the queries, and 2) obtaining the relevant document by searching databases using the knowledge clue. In particular, we first introduce an object-aware prefix-tuning technique to guide multi-grained visual learning. Then, we align multi-grained visual features into the textual feature space of the LLM, employing the LLM to capture cross-modal interactions. Subsequently, we construct instruction data with a unified format for model training. Finally, we propose the knowledge-guided generation strategy to impose prior constraints in the decoding steps, thereby promoting the generation of distinctive knowledge clues. Through experiments conducted on three benchmarks, we demonstrate significant improvements ranging from 3.0% to 14.6% across all evaluation metrics when compared to strong baselines.

Xuekai Zhu, Biqing Qi, Kaiyan Zhang, Xinwei Long, Zhouhan Lin, Bowen Zhou. Proceedings of the 2024 Conference of the North American Chapter of the Association for Computational Linguistics: Human Language Technologies (Volume 1: Long Papers). 2024.

Instruction tuning has recently been recognized as an effective way of aligning Large Language Models (LLMs) to enhance their generalization ability across various tasks.However, when tuning publicly accessible, centralized LLMs with private instruction data, privacy concerns are inevitable.While direct transfer of parameterized modules between models is a plausible approach to address this, its implications and effectiveness need further exploration.This paper focuses on Offsite-Tuning (OFT), a representative technique that transfers transformer blocks between centralized LLMs and downstream emulators.Given the limited understanding of the underlying mechanism of OFT, we perform an empirical analysis on LLMs from the perspectives of representation and functional similarity.Interestingly, our findings reveal a unique modular structure within the layers of LLMs that appears to emerge as the model size expands.Simultaneously, we note subtle but potentially significant changes in representation and intermediate predictions across the layers.Inspired by these observations, we propose CRaSh, involving Clustering, Removing, and Sharing, a training-free strategy to derive improved emulators from LLMs.CRaSh significantly boosts performance of OFT with billions of parameters.Furthermore, we investigate the optimal solutions yielded by fine-tuning with and without full model through the lens of loss landscape.Our findings demonstrate a linear connectivity among these optima falling over the same basin, thereby highlighting the effectiveness of CRaSh and OFT.The source code is publicly available at https://github.com/TsinghuaC3I/CRaSh.

Relation Extraction is key to many downstream tasks. Dialogue relation extraction aims at discovering entity relations from multi-turn dialogue scenario. There exist utterance, topic and relation discrepancy mainly due to multi-speakers, utterances, and relations. In this paper, we propose a consistent learning and inference method to minimize possible contradictions from those distinctions. First, we design mask mechanisms to refine utterance-aware and speaker-aware representations respectively from the global dialogue representation for the utterance distinction. Then a gate mechanism is proposed to aggregate such bi-grained representations. Next, mutual attention mechanism is introduced to obtain the entity representation for various relation specific topic structures. Finally, the relational inference is performed through first order logic constraints over the labeled data to decrease logically contradictory predicted relations. Experimental results on two benchmark datasets show that the F1 performance improvement of the proposed method is at least 3.3% compared with SOTA.

Dialogue Relation Extraction (DRE) is a new kind of relation extraction task from multi-turn dialogues. Different from the previous tasks, speaker specific relations are implicitly mixed together in both a local utterance window and a speaker context. To tackle both local and speaker dependency challenges, we explicitly construct a unified mention co-occurrence graph within a local utterance window or all utterances of a speaker from different entities. For each dialogue, a position enhanced graph attention network over this graph is proposed to obtain position aware mention representations in terms of both contexts. A gate function is utilized to help obtain a discriminative representation enough for each relation from original and position aware mention representations. For each entity pair in this dialogue, a pairwise attention mechanism is deployed to aggregate those discriminative mention representations as pair representation, which is fed into a standard multi-label classifier for relation label prediction. Experimental results on two benchmarks show the performance improvement of the proposed method is at least 1.6% and 3.2% compared with SOTA.