MLLM-Guided VLM Fine-Tuning with Joint Inference for Zero-Shot Composed Image Retrieval

Existing Zero-Shot Composed Image Retrieval (ZS-CIR) methods typically train adapters that convert reference images into pseudo-text tokens, which are concatenated with the modifying text and processed by frozen text encoders in pretrained VLMs or LLMs. While this design leverages the strengths of large pretrained models, it only supervises the adapter to produce encoder-compatible tokens that loosely preserve visual semantics. Crucially, it does not directly optimize the composed query representation to capture the full intent of the composition or to align with the target semantics, thereby limiting retrieval performance, particularly in cases involving fine-grained or complex visual transformations. To address this problem, we propose MLLM-Guided VLM Fine-Tuning with Joint Inference (MVFT-JI), a novel approach that leverages a pretrained multimodal large language model (MLLM) to construct two complementary training tasks using only unlabeled images: target text retrieval taskand text-to-image retrieval task. By jointly optimizing these tasks, our method enables the VLM to inherently acquire robust compositional retrieval capabilities, supported by the provided theoretical justifications and empirical validation. Furthermore, during inference, we further prompt the MLLM to generate target texts from composed queries and compute retrieval scores by integrating similarities between (i) the composed query and candidate images, and (ii) the MLLM-generated target text and candidate images. This strategy effectively combines the VLM's semantic alignment strengths with the MLLM's reasoning capabilities.

Multimodal Reasoning Agent for Zero-Shot Composed Image Retrieval

Zero-Shot Composed Image Retrieval (ZS-CIR) aims to retrieve target images given a compositional query, consisting of a reference image and a modifying text-without relying on annotated training data. Existing approaches often generate a synthetic target text using large language models (LLMs) to serve as an intermediate anchor between the compositional query and the target image. Models are then trained to align the compositional query with the generated text, and separately align images with their corresponding texts using contrastive learning. However, this reliance on intermediate text introduces error propagation, as inaccuracies in query-to-text and text-to-image mappings accumulate, ultimately degrading retrieval performance. To address these problems, we propose a novel framework by employing a Multimodal Reasoning Agent (MRA) for ZS-CIR. MRA eliminates the dependence on textual intermediaries by directly constructing triplets, <reference image, modification text, target image>, using only unlabeled image data. By training on these synthetic triplets, our model learns to capture the relationships between compositional queries and candidate images directly. Extensive experiments on three standard CIR benchmarks demonstrate the effectiveness of our approach. On the FashionIQ dataset, our method improves Average R@10 by at least 7.5\% over existing baselines; on CIRR, it boosts R@1 by 9.6\%; and on CIRCO, it increases mAP@5 by 9.5\%.

Can Visual Encoder Learn to See Arrows?

The diagram is a visual representation of a relationship illustrated with edges (lines or arrows), which is widely used in industrial and scientific communication. Although recognizing diagrams is essential for vision language models (VLMs) to comprehend domain-specific knowledge, recent studies reveal that many VLMs fail to identify edges in images. We hypothesize that these failures stem from an over-reliance on textual and positional biases, preventing VLMs from learning explicit edge features. Based on this idea, we empirically investigate whether the image encoder in VLMs can learn edge representation through training on a diagram dataset in which edges are biased neither by textual nor positional information. To this end, we conduct contrastive learning on an artificially generated diagram--caption dataset to train an image encoder and evaluate its diagram-related features on three tasks: probing, image retrieval, and captioning. Our results show that the finetuned model outperforms pretrained CLIP in all tasks and surpasses zero-shot GPT-4o and LLaVA-Mistral in the captioning task. These findings confirm that eliminating textual and positional biases fosters accurate edge recognition in VLMs, offering a promising path for advancing diagram understanding.

Hard Negative Contrastive Learning for Fine-Grained Geometric Understanding in Large Multimodal Models

Benefiting from contrastively trained visual encoders on large-scale natural scene images, Large Multimodal Models (LMMs) have achieved remarkable performance across various visual perception tasks. However, the inherent limitations of contrastive learning upon summarized descriptions fundamentally restrict the capabilities of models in meticulous reasoning, particularly in crucial scenarios of geometric problem-solving. To enhance geometric understanding, we propose a novel hard negative contrastive learning framework for the vision encoder, which combines image-based contrastive learning using generation-based hard negatives created by perturbing diagram generation code, and text-based contrastive learning using rule-based negatives derived from modified geometric descriptions and retrieval-based negatives selected based on caption similarity. We train CLIP using our strong negative learning method, namely MMCLIP (Multimodal Math CLIP), and subsequently train an LMM for geometric problem-solving. Experiments show that our trained model, MMGeoLM, significantly outperforms other open-source models on three geometric reasoning benchmarks. Even with a size of 7B, it can rival powerful closed-source models like GPT-4o. We further study the impact of different negative sample construction methods and the number of negative samples on the geometric reasoning performance of LMM, yielding fruitful conclusions. The code and dataset are available at https://github.com/THU-KEG/MMGeoLM.

Benchmarking Multimodal Knowledge Conflict for Large Multimodal Models

Large Multimodal Models(LMMs) face notable challenges when encountering multimodal knowledge conflicts, particularly under retrieval-augmented generation(RAG) frameworks where the contextual information from external sources may contradict the model's internal parametric knowledge, leading to unreliable outputs. However, existing benchmarks fail to reflect such realistic conflict scenarios. Most focus solely on intra-memory conflicts, while context-memory and inter-context conflicts remain largely investigated. Furthermore, commonly used factual knowledge-based evaluations are often overlooked, and existing datasets lack a thorough investigation into conflict detection capabilities. To bridge this gap, we propose MMKC-Bench, a benchmark designed to evaluate factual knowledge conflicts in both context-memory and inter-context scenarios. MMKC-Bench encompasses three types of multimodal knowledge conflicts and includes 1,573 knowledge instances and 3,381 images across 23 broad types, collected through automated pipelines with human verification. We evaluate three representative series of LMMs on both model behavior analysis and conflict detection tasks. Our findings show that while current LMMs are capable of recognizing knowledge conflicts, they tend to favor internal parametric knowledge over external evidence. We hope MMKC-Bench will foster further research in multimodal knowledge conflict and enhance the development of multimodal RAG systems. The source code is available at https://github.com/MLLMKCBENCH/MLLMKC.

BRIT: Bidirectional Retrieval over Unified Image-Text Graph

Retrieval-Augmented Generation (RAG) has emerged as a promising technique to enhance the quality and relevance of responses generated by large language models. While recent advancements have mainly focused on improving RAG for text-based queries, RAG on multi-modal documents containing both texts and images has not been fully explored. Especially when fine-tuning does not work. This paper proposes BRIT, a novel multi-modal RAG framework that effectively unifies various text-image connections in the document into a multi-modal graph and retrieves the texts and images as a query-specific sub-graph. By traversing both image-to-text and text-to-image paths in the graph, BRIT retrieve not only directly query-relevant images and texts but also further relevant contents to answering complex cross-modal multi-hop questions. To evaluate the effectiveness of BRIT, we introduce MM-RAG test set specifically designed for multi-modal question answering tasks that require to understand the text-image relations. Our comprehensive experiments demonstrate the superiority of BRIT, highlighting its ability to handle cross-modal questions on the multi-modal documents.

TNG-CLIP:Training-Time Negation Data Generation for Negation Awareness of CLIP

Vision-language models (VLMs), such as CLIP, have demonstrated strong performance across a range of downstream tasks. However, CLIP is still limited in negation understanding: the ability to recognize the absence or exclusion of a concept. Existing methods address the problem by using a large language model (LLM) to generate large-scale data of image captions containing negation for further fine-tuning CLIP. However, these methods are both time- and compute-intensive, and their evaluations are typically restricted to image-text matching tasks. To expand the horizon, we (1) introduce a training-time negation data generation pipeline such that negation captions are generated during the training stage, which only increases 2.5% extra training time, and (2) we propose the first benchmark, Neg-TtoI, for evaluating text-to-image generation models on prompts containing negation, assessing model's ability to produce semantically accurate images. We show that our proposed method, TNG-CLIP, achieves SOTA performance on diverse negation benchmarks of image-to-text matching, text-to-image retrieval, and image generation.

EvdCLIP: Improving Vision-Language Retrieval with Entity Visual Descriptions from Large Language Models

Vision-language retrieval (VLR) has attracted significant attention in both academia and industry, which involves using text (or images) as queries to retrieve corresponding images (or text). However, existing methods often neglect the rich visual semantics knowledge of entities, thus leading to incorrect retrieval results. To address this problem, we propose the Entity Visual Description enhanced CLIP (EvdCLIP), designed to leverage the visual knowledge of entities to enrich queries. Specifically, since humans recognize entities through visual cues, we employ a large language model (LLM) to generate Entity Visual Descriptions (EVDs) as alignment cues to complement textual data. These EVDs are then integrated into raw queries to create visually-rich, EVD-enhanced queries. Furthermore, recognizing that EVD-enhanced queries may introduce noise or low-quality expansions, we develop a novel, trainable EVD-aware Rewriter (EaRW) for vision-language retrieval tasks. EaRW utilizes EVD knowledge and the generative capabilities of the language model to effectively rewrite queries. With our specialized training strategy, EaRW can generate high-quality and low-noise EVD-enhanced queries. Extensive quantitative and qualitative experiments on image-text retrieval benchmarks validate the superiority of EvdCLIP on vision-language retrieval tasks.

Co-AttenDWG: Co-Attentive Dimension-Wise Gating and Expert Fusion for Multi-Modal Offensive Content Detection

Multi-modal learning has become a critical research area because integrating text and image data can significantly improve performance in tasks such as classification, retrieval, and scene understanding. However, despite progress with pre-trained models, current approaches are limited by inadequate cross-modal interactions and static fusion strategies that do not fully exploit the complementary nature of different modalities. To address these shortcomings, we introduce a novel multi-modal Co-AttenDWG architecture that leverages dual-path encoding, co-attention with dimension-wise gating, and advanced expert fusion. Our approach begins by projecting text and image features into a common embedding space, where a dedicated co-attention mechanism enables simultaneous, fine-grained interactions between modalities. This mechanism is further enhanced by a dimension-wise gating network that adaptively regulates the feature contributions at the channel level, ensuring that only the most relevant information is emphasized. In parallel, dual-path encoders refine the representations by processing cross-modal information separately before an additional cross-attention layer further aligns modalities. The refined features are then aggregated via an expert fusion module that combines learned gating and self-attention to produce a robust, unified representation. We validate our approach on the MIMIC and SemEval Memotion 1.0, where experimental results demonstrate significant improvements in cross-modal alignment and state-of-the-art performance, underscoring the potential of our model for a wide range of multi-modal applications.

Don't Look Only Once: Towards Multimodal Interactive Reasoning with Selective Visual Revisitation

We present v1, a lightweight extension to Multimodal Large Language Models (MLLMs) that enables selective visual revisitation during inference. While current MLLMs typically consume visual input only once and reason purely over internal memory, v1 introduces a simple point-and-copy mechanism that allows the model to dynamically retrieve relevant image regions throughout the reasoning process. This mechanism augments existing architectures with minimal modifications, enabling contextual access to visual tokens based on the model's evolving hypotheses. To train this capability, we construct v1g, a dataset of 300K multimodal reasoning traces with interleaved visual grounding annotations. Experiments on three multimodal mathematical reasoning benchmarks -- MathVista, MathVision, and MathVerse -- demonstrate that v1 consistently improves performance over comparable baselines, particularly on tasks requiring fine-grained visual reference and multi-step reasoning. Our results suggest that dynamic visual access is a promising direction for enhancing grounded multimodal reasoning. Code, models, and data will be released to support future research.