Visual Commonsense Reasoning (VCR) aims to enable models to perform human-like cognitive reasoning based on images. Although existing methods primarily rely on large-scale cross-modal pre-trained models for implicit cross-modal modeling, they still fall short in terms of external knowledge incorporation, contextual modeling, and the consistency between answers and rationales. To address these limitations, the paper propose a knowledge-augmented collaborative reasoning framework for VCR. The proposed framework introduces a dual-layer knowledge injection mechanism comprising scene commonsense graphs and regional factual descriptions. By explicitly integrating scene-level commonsense with entity-level visual facts, this mechanism mitigates knowledge deficiency and semantic ambiguity. Meanwhile, a dual-task collaborative learning framework based on large language models is designed to jointly optimize question answering and rationale inference. A cognitive alignment loss is further employed to enhance the logical consistency between the two tasks in the latent space. Experimental results on the VCR benchmark demonstrate that the proposed framework effectively improves model reasoning performance. Notably, it outperforms state-of-the-art baselines by 4.2 percentage points on the joint Q→AR task, thereby validating the effectiveness of the knowledge-augmented collaborative reasoning strategy.
自VCR任务提出以来,相关研究主要围绕推理结构与多模态表示学习展开,形成两条技术路线:一类为结构化推理模型,通过注意力机制或图神经网络对图像实体及关系进行显式建模,并与文本信息交互融合;另一类基于视觉-语言预训练模型(Vision-Language Pretraining, VLP)及多模态大语言模型(Multimodal Large Language Models, MLLM),在大规模图文数据上学习统一语义表示并迁移至 VCR 任务。随着Transformer[3]、BERT[4]等模型的发展,视觉-文本联合建模方法[5-6]不断涌现,并在多项任务上取得良好效果。近年来,大语言模型及其多模态扩展[7-8]进一步提升了复杂视觉语境下的推理能力。
此外,VCR 还定义了联合任务Q→AR(Question to Answer and Rationale, Q→AR),要求模型同时预测正确答案及其支撑理由。该任务采用严格的联合评估标准,只有当Q→A与QA→R均预测正确时,预测结果才被视为正确。因此,模型通常在训练阶段分别学习两个子任务,并在测试阶段综合评估其整体推理能力。
通过此方法,对于图像中的 person_1,模块能够生成如“person_1 is a man with short, light-colored hair, wearing a dark blue uniform with a round patch on the chest.”这样的高质量事实描述,而非“person_1 looks happy and is greeting someone.”这样的推理性描述。
为了监督模型在每个子任务上的性能,我们为每个任务引入了独立的线性分类头(Separate Linear Classification Heads)。这种设计允许模型从一个被对齐的、深度的认知状态出发,学习到将该状态映射到具体任务决策空间(答案选项或理由选项)的特定线性变换,从而为每个任务提供了更高的模型灵活性。
Q→A任务损失():我们定义一个答案分类头,其中d是隐藏层维度,是候选选项的数量。该分类头将问答任务的认知状态向量映射为4个候选答案的原始预测得分(logits)。随后,我们通过Softmax函数计算概率分布,并与真实答案标签 yA 计算标准的交叉熵损失:
从表3可以看出,在当前实验配置下,KACR 的在线推理延迟保持在0.29 s /样本的量级,能够满足常见视觉常识推理任务对响应速度的需求。该延迟主要来自于模型在推理阶段对更丰富多模态知识信息的联合建模过程。在此基础上,KACR在Q→AR联合指标上提升了3.8个百分点的性能,表明引入结构化场景常识与区域事实信息能够显著提升模型的推理质量。综合来看,KACR 在保持可接受推理效率的同时,实现了对视觉常识推理性能的有效增强,适用于对推理准确性具有较高要求的应用场景。
为了直观展示模型在复杂场景下的推理优势,图6对比了本模型(KACR)与基线模型(Qwen2-VL-2B)在不同案例上的表现。如图6左上角所示,当问题询问“为何Person 0不住在该处”时,基线模型忽略了人物衣着细节,给出了无法验证的猜测。相比之下,我们的模型敏锐捕捉到了 Person 0“身穿外套(wearing a coat)”这一细粒度视觉事实,并将其与“室内穿着厚外套通常意味着外来访客”的常识建立关联,从而实现了准确推理。这证明了引入区域事实描述能有效帮助模型捕捉被忽视的关键线索。同样的差异也体现在对抗性较强的场景中。在图6g的案例中,面对“为何点燃火柴”的提问,基线模型受昏暗的环境背景误导,错误判断其意图为“照亮房间”。而我们的模型通过场景常识图(SCG)准确定位了画面边缘的“木堆(pile of wood)”实体,成功排除了背景干扰,推断出“生火”这一符合物理逻辑的行为。此外,在左下角涉及细粒度状态理解的案例中,基线模型仅能给出专注这类宽泛且模糊的情感描述。而我们的模型结合人物面部神态与上下文语境,精准识别出“悲伤与困惑”的复合情绪,选择了比基线模型更为深刻且具体的理由。然而,模型在处理深层隐喻或特定叙事套路时仍面临挑战。如图6h案例所示,正确答案依赖于“刺客准备行动”的戏剧性隐喻。由于VGFD模块为抑制幻觉被限制于生成客观描述,且SCG检索的常识偏向日常逻辑,导致模型缺乏支撑高阶意图推理的关键线索,倾向于选择保守的错误选项。这一现象深刻揭示了在追求“视觉事实准确性”与保留“叙事想象力”之间存在的内在权衡,为未来提升模型的高阶语义理解能力指明了方向。
ZellersR, BiskY, FarhadiA, et al.From recognition to cognition: Visual commonsense reasoning[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2019: 6720-6731.
[2]
JainA, VatsaM, SinghR.Words over pixels? Rethinking vision in multimodal large language models[C]//Proceedings of the 34th International Joint Conference on Artificial Intelligence (IJCAI-25), 2025: 10481-10489.
[3]
VaswaniA, ShazeerN, ParmarN, et al.Attention is all you need[C]//Advances in Neural Information Processing Systems 30, 2017: 5998-6008.
[4]
KoroteevM V.BERT: A review of applications in natural language processing and understanding[PP/OL].[2025-11-20].
[5]
ChengH, YeH, ZhouX, et al.Vision-language pre-training via modal interaction[J].Pattern Recogn, 2024, 156: 110809.
[6]
YuanY, LiZ, ZhaoB.A survey of multimodal learning: Methods, applications, and future[J].ACM Comput Surv, 2025, 57(7): 1-34.
[7]
YuY, ShiC, TangJ, et al.Qwen-VL2 model with NEFTune technique for medical report generation[C]//Proceedings of the 4th International Symposium on Computer Applications and Information Technology, 2025: 165-168.
[8]
LiJ, LiD, SavareseS, et al.BLIP-2: Bootstrapping language-image pre-training with frozen image encoders and large language models[C]//Proceedings of the International Conference on Machine Learning, 2023: 19730-19742.
[9]
JabriA, JoulinA, Van der MaatenL.Revisiting visual question answering baselines[C]//Proceedings of the European Conference on Computer Vision, 2016: 727-739.
[10]
AndersonP, HeX, BuehlerC, et al.Bottom-up and top-down attention for image captioning and visual question answering[C]//Proceedings of the IEEE Conference on Computer Vision and Pattern Recognition, 2018: 6077-6086.
[11]
WenZ, PengY.Multi-level knowledge injecting for visual commonsense reasoning[J].IEEE Trans Circuits Syst Video Technol, 2020, 31(3): 1042-1054.
[12]
LinJ, JainU, SchwingA.TAB-VCR: Tags and attributes based VCR baselines[PP/OL].[2025-11-20].
[13]
WuA, ZhuL, HanY, et al.Connective cognition network for directional visual commonsense reasoning[PP/OL].[2025-11-20].
[14]
YuW, ZhouJ, YuW, et al.Heterogeneous graph learning for visual commonsense reasoning[PP/OL].[2025-11-20].
[15]
ZhangX, ZhangF, XuC.Explicit cross-modal representation learning for visual commonsense reasoning[J].IEEE Trans Multimedia, 2021, 24: 2986-2997.
[16]
WangT, HuangJ, ZhangH, et al.Visual commonsense R-CNN[C]//Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition, 2020: 10760-10770.
[17]
LiL H, YatskarM, YinD, et al.VisualBERT: A simple and performant baseline for vision and language[PP/OL].[2025-11-20].
[18]
SuW, ZhuX, CaoY, et al.VL-BERT: Pre-training of generic visual-linguistic representations [PP/OL].[2025-11-20].
[19]
ChenY C, LiL, YuL, et al.UNITER: Universal image-text representation learning[C]//Proceedings of the European Conference on Computer Vision, 2020: 104-120.
[20]
LuJ, BatraD, ParikhD, et al.ViLBERT: Pretraining task-agnostic visiolinguistic representations for vision-and-language tasks[J].Advances in Neural Information Processing Systems, 2019, 32: 13-23.
[21]
IslamR, MoushiO M.GPT-4o: The cutting-edge advancement in multimodal LLM[C]//Intelligent Computing-Proceedings of the Computing Conference, 2025: 47-60.
[22]
TouvronH, LavrilT, IzacardG, et al.LLaMA: Open and efficient foundation language models[PP/OL].[2025-11-20].
[23]
WangP, BaiS, TanS, et al.Qwen2-VL: Enhancing vision-language model's perception of the world at any resolution[PP/OL].[2025-11-20].
[24]
DosovitskiyA, BeyerL, KolesnikovA, et al.An image is worth 16×16 words: Transformers for image recognition at scale[PP/OL].[2025-11-20].
[25]
RadfordA, KimJ W, HallacyC, et al.Learning transferable visual models from natural language supervision[C]//Proceedings of the International Conference on Machine Learning, 2021: 8748-8763.
[26]
AlayracJ B, DonahueJ, LucP, et al.Flamingo: A visual language model for few-shot learning[J].Advances in Neural Information Processing Systems, 2022, 35: 23716-23736.
[27]
LiuH, LiC, WuQ, et al.Visual instruction tuning [J].Advances in Neural Information Processing Systems, 2023, 36: 34892-34916.
[28]
SpeerR, ChinJ, HavasiC.ConceptNet 5.5: An open multilingual graph of general knowledge[J].Proc AAAI Conf Artif Intell, 2017, 31(1): 4444-4451.
[29]
HwangJ D, BhagavatulaC, Le BrasR, et al.ATOMIC 2020: On symbolic and neural commonsense knowledge graphs[J].Proc AAAI Conf Artif Intell,2021, 35(7): 6384-6392.
ChenQ, ZhuX, LingZ H, et al.Enhanced LSTM for natural language inference[C]//Proceedings of the Annual Meeting of the Association for Computational Linguistics, 2017: 1657-1668.
[32]
KimJ H, OnK W, LimW, et al.Hadamard product for low-rank bilinear pooling[PP/OL].[2025-11-20].
[33]
Ben-YounesH, CadeneR, CordM, et al.MUTAN: Multimodal Tucker fusion for visual question answering[C]//Proceedings of the IEEE International Conference on Computer Vision, 2017: 2612-2620.
[34]
ZhangX, ZhangF, XuC.Multi-level counterfactual contrast for visual commonsense reasoning[C]//Proceedings of the ACM International Conference on Multimedia, 2021: 1793-1802.
[35]
LiZ, GuoY, WangK, et al.Joint answering and explanation for visual commonsense reasoning[J].Proc AAAI Conf Artif Intell, 2023, 32: 3836-3846.
[36]
YeK, KovashkaA.A case study of the shortcut effects in visual commonsense reasoning[J].Proc AAAI Conf Artif Intell, 2021, 35(4): 3181-3189.
[37]
ZhuJ, WangH, ShiM.Multimodal large language model enhanced pseudo-3D perception framework for visual commonsense reasoning[J].IEEE Trans Circ Syst Vid, 2024,34(11): 11682-11694.