state tracking

This work addresses the prevalent issue of factually inconsistent hallucinations in multimodal generative models. To mitigate this, the authors propose TIGER, a framework that, during inference, constructs an input observation graph and an output claim graph, then computes a risk score for each generated claim based on graph-structural alignment. Only high-risk claims undergo localized correction while the backbone model remains frozen. TIGER is the first method to enable fact-level traceable hallucination rectification, preventing erroneous content from corrupting input interpretation and supporting fine-grained, claim-wise feedback ranking and scheduling. Experiments across four cross-modal tasks—spanning images, text, audio, and video—demonstrate that TIGER substantially reduces unsupported content without compromising original task performance and is compatible with diverse backbone architectures and multi-source information settings.

Existing approaches to video misinformation detection typically model entire videos holistically, overlooking the fact that deceptive content often relies on sparse, critical cues. This leads to computational redundancy and dilutes discriminative evidence. To address this, this work proposes SIEVE, a novel framework that introduces, for the first time, an agent-based paradigm for actively searching multimodal sparse evidence. SIEVE decouples evidence acquisition from verification: an agent employs evidence-aware reinforcement learning to efficiently extract a minimal set of key cues, forming a compact evidence package, which a verifier then uses to make interpretable judgments. The method significantly outperforms current state-of-the-art techniques across multiple benchmarks, achieving high detection accuracy with highly distilled evidence while providing transparent and traceable decision rationales.

Existing interactive long-video generation methods suffer from insufficient historical context modeling, leading to poor scene consistency. To address this, we propose a novel “context-as-memory” paradigm: historical frames are directly treated as retrievable memory, enabling conditional modeling via frame-wise concatenation—eliminating the need for auxiliary control modules. We further design a lightweight memory retrieval mechanism based on camera field-of-view (FOV) overlap, ensuring information integrity while substantially reducing computational overhead. Our approach is embodied in an end-to-end trainable video diffusion architecture. On interactive long-video generation benchmarks, our method significantly outperforms state-of-the-art approaches, demonstrating strong generalization to unseen open-domain scenes and reducing redundant computation by over 40%.

To address the inefficiency and poor temporal consistency in long-term Gaussian scene reconstruction caused by frequent object mutations and subtle changes in everyday environments, this paper proposes a templated temporal Gaussian modeling framework. Methodologically, it leverages structured Gaussian object templates as transferable priors, integrated with sparse-view observations and few-shot self-adaptive optimization, enabling lightweight, incremental scene evolution modeling via spatio-temporal transformations; initialization employs Gaussian splatting for rapid online updates. Experiments on a newly constructed real-world dynamic scene dataset demonstrate significant improvements in reconstruction fidelity and inter-frame geometric consistency, alongside a 3.2× speedup in inference latency and a 57% reduction in memory overhead. To our knowledge, this is the first approach achieving high-fidelity, long-duration, low-latency Gaussian scene chronology construction.

This work addresses the limitations of existing medical imaging systems—typically black-box models performing single-pass inference—by introducing R⁴, a novel self-improving multi-agent framework for medical vision-language tasks. R⁴ integrates four collaborative agents responsible for routing, retrieval, reflection, and repair, enabling dynamic prompt configuration, joint image-text generation, clinical error detection, and constraint-driven iterative refinement. The framework supports explainable inference, self-diagnosis of errors, and joint spatial-linguistic optimization without requiring model fine-tuning. Evaluated on chest X-ray data, R⁴ achieves a 1.7–2.5 point improvement in LLM-as-a-Judge report generation scores and a 2.5–3.5 percentage point gain in weakly supervised detection mAP50, significantly outperforming single vision-language model baselines.

This work addresses the challenge of complex temporal logical reasoning in video question answering—such as event existence, ordering, duration, and boundary overlap—by proposing a visual evidence routing framework that decouples perception from symbolic temporal reasoning. The approach first parses the question structure and dynamically routes it to an appropriate processing strategy based on video length and operator complexity. It leverages multimodal large language models to generate structured visual evidence, which is then subjected to precise logical operations via a programmatic verifier and a deterministic temporal reducer. Innovatively integrating structured evidence with formal temporal rules, the method employs a conservative answer fusion strategy to enhance consistency. Evaluated on the TimeLogicQA benchmark, it achieves an average accuracy of 81.8%.

This work addresses the unreliability of RGB reconstruction evidence across diverse product categories and the absence of category-adaptive fusion strategies during testing in multimodal industrial anomaly detection. To overcome these limitations, the authors propose TC-MAF, a unified anchor-based multimodal evidence fusion framework that integrates multimodal detectors, complementary evidence from Dinomaly, and cross-modal consistency cues. Notably, TC-MAF introduces a lightweight Training-time Discrete Confidence (TDC) mechanism—relying solely on normal training statistics—to dynamically adjust the weights of auxiliary evidence without requiring category information at test time, thereby enabling robust and reliable fusion. The method achieves state-of-the-art average performance among multimodal approaches, with an image-level AUROC of 0.979 and pixel-level AUPRO of 0.990 on MVTec-3D, and demonstrates consistent robustness across various ablations and cross-dataset evaluations.

This work addresses the challenges of identity drift and cross-shot inconsistency in multi-shot long-form video generation by proposing a triadic collaborative closed-loop generation paradigm. The approach introduces a vision–language–memory agent framework that jointly models narrative intent, persistent memory, and generated visuals. Central to the method are an entity-centric memory evolution mechanism and a dual-path (intra-shot and inter-shot) dynamic prompt optimization strategy. These components are integrated with a vision–language model analyzer, a mutable visual state memory structure, image-to-video prompt fine-tuning, and evidence-driven prompt rewriting grounded in generation outputs. Evaluated on the StoryBench benchmark, the proposed method significantly outperforms existing state-of-the-art approaches in cross-shot consistency, prompt adherence, and cinematic continuity.

Existing multimodal reasoning approaches often suffer from the loss of fine-grained visual details or compromised visual faithfulness due to suboptimal timing and manner of visual evidence integration. To address this limitation, this work proposes the CSMR framework, which introduces a novel cognitive scheduling mechanism: a language model dynamically controls an independent visual perception module, invoking task-relevant visual evidence on demand during reasoning. This design overcomes the constraints of static fusion and end-to-end joint optimization paradigms. Evaluated under zero-shot settings, the proposed method achieves substantial performance gains over state-of-the-art baselines across multiple multimodal benchmarks, demonstrating that dynamic scheduling effectively enhances both reasoning accuracy and visual faithfulness.

This work addresses the ambiguity regarding whether existing single-stage video object detectors genuinely leverage temporal context, as standard evaluation metrics often fail to reveal their actual reliance on temporal information. To this end, we propose TemporalLens, a diagnostic framework that quantifies a model’s temporal dependency through controlled perturbations—including temporal shuffling, structured occlusion, and redundancy injection. Furthermore, we design YOLO-3D based on YOLOv8, explicitly preserving the temporal dimension within the backbone to enhance genuine temporal reasoning. Experiments demonstrate that TemporalLens effectively distinguishes between stacked 2D models and true temporal architectures, while YOLO-3D achieves an average mAP@50 improvement of 3.7 percentage points with 32-frame inputs, underscoring the critical role of temporal depth in performance gains.