CAER introduces a span-grounded evidence router that transforms claim representations into soft textual queries and retrieves corresponding evidence from frozen visual tokens, enabling fine-grained conflict estimation and design a dual-prefix expert routing mechanism that learns separate experts for visually supported and contradicted inputs, enabling conflict-aware generation through explicit expert selection.
Abstract
Multimodal Large Language Models (MLLMs) have demonstrated remarkable capabilities in multimodal understanding and generation. However, when textual inputs conflict with visual evidence, they still suffer from hallucinations and produce responses inconsistent with visual content. Existing approaches mainly rely on decoding strategies, additional training, verification methods, or prompting techniques, but often lack fine-grained conflict localization and conflict-aware generation. In this work, we propose CAER, a backbone-agnostic framework for visual-language conflict detection and conflict-aware generation. CAER introduces a span-grounded evidence router that transforms claim representations into soft textual queries and retrieves corresponding evidence from frozen visual tokens, enabling fine-grained conflict estimation. Furthermore, we design a dual-prefix expert routing mechanism that learns separate experts for visually supported and contradicted inputs, enabling conflict-aware generation through explicit expert selection. Experiments on the public MMMC benchmark and our newly curated AgriConflict dataset demonstrate that CAER effectively detects visual-language conflicts and improves the reliability of open-source MLLMs without updating their backbone parameters.
Multimodal Large Language Models (MLLMs) are increasingly deployed as multi-step agents, where explicit reasoning supports task decomposition and tool coordination but also accumulates self-generated text. Over long trajectories, this text can dominate the context and suppress visual evidence, creating textual debt. We observe that reasoning becomes redundant once task-relevant visual evidence is grounded, while stale hypotheses can misguide later inference when grounding remains uncertain. Pruning must therefore remove redundant text without discarding visual evidence. We propose SPARE, a Kullback-Leibler (KL)-guided framework for pruning accumulated reasoning in multimodal tool-use agents. SPARE uses a compact task-state summary as privileged diagnostic context. For each candidate segment, it replays the same model under the original and summary-conditioned contexts. Reverse-KL divergence from on-policy self-distillation (OPSD) then tests whether the summary sufficiently covers the segment without disrupting future reasoning. We further fine-tune the summarizer with supervised fine-tuning (SFT), enabling more compact summaries, broader coverage, and more aggressive pruning. Across multi-step visual tool-use benchmarks, SPARE achieves the highest average accuracy among pruning methods while removing 37.89-64.58\% of reasoning tokens. This favorable accuracy-context trade-off shows that reducing textual dominance restores reliance on visual evidence and mitigates over-conditioning on self-generated language.
Yuchen Huang, Sijia Li, Jun Zhang et al.· 0 citations
An evidence-driven multimodal reasoning framework that utilizes a Bloom-inspired taxonomy as a hierarchical reasoning protocol and quantitatively analyzes the trace to make evidence usage and reasoning progression explicit is proposed.
While Multimodal Retrieval-Augmented Generation (MM-RAG) has shown promising results, it still struggles with complex multi-hop reasoning tasks. Existing methods primarily focus on independent instance-level matching, which often fails to capture explicit relationships across modalities and documents. Although Graph-enhanced methods introduce structural modeling, they face a fundamental challenge in multimodal scenarios: incorporating fine-grained visual features leads to rapid graph expansion and retrieval noise, whereas coarse-grained representations cause the discarding of critical local evidence. To address this dilemma, we propose DualG-MRAG, a Dual-tier framework that introduces a decoupled architecture comprising Macro-reasoning and Micro-matching Graphs for Multimodal RAG. Specifically, to suppress retrieval noise by isolating global structural reasoning from fine-grained evidence matching, we construct a Macro Graph for global topological routing and a Micro Graph for precise local verification. Subsequently, to enable dynamic relevance propagation across heterogeneous evidence sources, we formulate retrieval as a query-driven message passing process via a GNN Retriever. Furthermore, to provide the generative model with coherent structural guidance, we introduce a dynamic programming decoding mechanism that extracts explicit reasoning paths directly from the GNN's forward pass, replacing the standard input of isolated document chunks. Extensive experiments demonstrate that DualG-MRAG outperforms baselines in both evidence recall and complex QA accuracy.
Jiachen Tao, Qingyun Sun, Haonan Yuan et al.· 0 citations
Multimodal fake news is difficult to detect when text and image remain topically aligned but diverge in local factual details such as entities, locations, time, or event attributes. Existing methods often rely on global cross-modal fusion and may therefore overlook these localized contradictions. We present LCFND, a staged and fully specified bounded-evidence verification pipeline for localized conflict discovery, structured evidence generation, and graph-based verification. Optimal transport identifies high-conflict token-region pairs and constructs localized evidence packets. A parameter-efficient LLM trained with OT-grounded structured targets and training-split label-informed weak relation calibration then converts these packets into structured evidence, including claims, visual evidence, contradiction types, and grounded rationales. A heterogeneous evidence graph integrates OT conflict priors and LLM-generated evidence for veracity prediction, and all graph-construction rules, training objectives, and evaluation protocols are stated explicitly for reproducibility. Under a controlled matched-backbone protocol on Weibo, Twitter, and GossipCop, LCFND improves F1 over the top matched-backbone baseline MGCA by 1.47, 1.73, and 1.67 points, respectively. The gain remains positive when the visual backbone is strengthened to ViT-B/16 or Swin-T. Manual evidence evaluation on 500 samples per dataset reports 85.9–87.3% claim grounding and 78.1–80.2% contradiction-type correctness. Under the P3b matched-input protocol, LCFND remains above FND-LLM-M and MMRGV-lite, and improves average cross-dataset transfer F1 from 67.10 to 69.98. Full online inference takes 1.47 s per sample in our setting, so the framework is better suited to offline or high-risk verification than to large-scale real-time screening.
Yuechuan Zhang, Hongyu Jin, Yaxuan Wang et al.· Journal of King Saud Univers...· 0 citations
Advancing multimodal retrieval-augmented generation (RAG) for complex document understanding presents a formidable dual dilemma of accuracy and efficiency, particularly in graph RAG. Processing structurally sparse yet visually dense layouts, such as extracting a tiny data marker from a financial chart, often incurs computationally prohibitive token overhead while still triggering catastrophic hallucination. However, multimodal Graph RAG pipelines rely on graph-construction stages that assume Vision-Language Models (VLMs) can resolve sparse semantics within high-density layouts. We challenge this assumption, revealing that forcing VLMs to localize visual evidence, interpret semantics, and extract relations triggers a"Visual Attention Sink,"a mechanism driving catastrophic semantic loss, while full-page processing incurs massive computational overhead. Controlled interventions verify that this failure is boundary-driven rather than content-specific and that semantic anchoring mitigates it. To fundamentally correct this flawed paradigm, we introduce DeCoRAG, a multimodal Graph RAG pipeline that shifts knowledge processing from coupled visual-semantic reasoning to"Cognitive Decoupling."Rather than passively processing raw pixels, its graph-construction stage establishes a macroscopic Semantic Anchor to neutralize the attention sink. This anchor subsequently drives our Region-Aware Pruning and Cropping (RAP-Crop) mechanism, shifting the reasoning space from dense, noisy backgrounds to purified, intent-driven semantic clusters. The resulting graph supports hybrid retrieval and answer generation. Across complex document benchmarks, DeCoRAG improves the semantic pass rate by up to 12.5 percentage points over the strongest baseline and generalizes to DocVQA. RAP-Crop reduces offline graph-construction prompt tokens by 40.8% without sacrificing end-to-end accuracy.
Shuoshuo Wang, Kai Zhang, Wenyuan Huang et al.· 0 citations
Multimodal representation learning is a cornerstone of modern AI. By encoding multimodal queries and targets into vectors, it powers industrial search and recommendation and underpins modern agents. Real-world platforms with complex modalities and massive-scale content, such as Douyin, Xiaohongshu, and YouTube, demand both efficiency under billion-scale indexing and fine-grained discrimination for hard matching. Existing MLLM embedding models rarely satisfy both. Contrastive models are efficient but rely on pair-level supervision too coarse for fine-grained distinctions, while CoT-based models improve discrimination through explicit generation impractical to serve online. We present Douyin Multimodal Embedding (DME), a model trained in two stages to combine both strengths. Stage 1 performs large-scale contrastive pre-training that establishes a unified multimodal embedding space with broad modality and task coverage. Stage 2 supplements semantic sufficiency, the property that an embedding is grounded in retrieval-relevant evidence and preserves fine-grained counterpart-side semantics, via two mechanisms. Evidence-Grounded Typed Latent Reasoning organizes retrieval evidence through hidden-space latent reasoning, and Cross-Conditional Reconstruction enforces counterpart-side semantics through cross-directional autoregressive reconstruction. Both act only during training and add only marginal query-side overhead, so DME serves as efficiently as a standard contrastive encoder. On MMEB-v2, DME reaches state-of-the-art results at comparable scales for its 2B and 9B variants (74.8 and 78.4), with especially strong video and visual-document tasks. In production, DME delivers a 2.92% relative gain on Douyin's in-house offline evaluation set, is deployed across Douyin scenarios such as generative, image, and AI search, and yields a 0.1% Lifetime (LT) gain in online A/B testing on Douyin search.
Haonan Chen, Chu Li, Zhi-Cheng Wang et al.· 0 citations