解决交通场景小物体识别被忽略的问题，提升复杂问题回答准确率。
Beyond Scene Priors: Fine-Grained Traffic Scene Reasoning with Benchmarking and Query-Guided Small-Object Focus
- 用查询引导的小物体聚焦机制，精准定位关键视觉线索。
- 在4万+样本的精细交通推理数据集上达到74.1%准确率，优于基线2.1点。
- 无需额外检测或重编码，适合实时安全驾驶系统部署。
在高安全要求的交通场景中，回答复杂问题依赖微小且局部的视觉线索。然而，标准多模态大模型容易过度关注背景，导致关键小物体信息被淹没，这种现象称为‘关键证据稀释’。现有视觉问答数据集很少暴露此缺陷，因缺乏大规模、干扰物密集的评测任务。为此，我们提出细粒度交通推理基准（FGTR-Bench）和文本引导小物体推理多模态大模型（TSR-MLLM）。FGTR-Bench包含40,236个单图多选题，通过多智能体生成、一致性校验和专家审核构建，并配有4,947样本的独立测试集。为缓解证据稀释，TSR-MLLM基于Qwen3-VL-4B，采用查询条件化的文本引导小物体聚焦（TG-SOF）图，在解码器边界仅对最相关视觉标记添加稀疏的Top-K门控残差，保持文本令牌不变。结合轻量解码器适配，实现单次前向传播推理，无需外部检测器或图像重编码。在相同设置下，TSR-MLLM在FGTR-Bench上比最强4B基线高出2.1分（整体达74.1%），在证据定位任务上提升更显著。同时，在DriveQA-V（CARLA Signs）上使用贪婪解码也保持竞争力，无需特定任务微调。
原文摘要 · Abstract (English)
In safety-critical traffic scenarios, answering complex questions relies on minute, localized visual cues. However, standard Multimodal Large Language Models (MLLMs) tend to over-attend to backgrounds, overwhelming crucial small objects during visual-language alignment, a failure mode we term 'critical evidence dilution.' Furthermore, existing visual question answering (VQA) datasets rarely expose this flaw, as they lack large-scale, distractor-heavy evaluations that require pinpointing local evidence. To bridge this evaluation and architecture gap, we introduce the Fine-Grained Traffic Reasoning Benchmark (FGTR-Bench) and the Text-Guided Small-Object Reasoning MLLM (TSR-MLLM). FGTR-Bench comprises 40,236 single-image Multiple-Choice Questions (MCQs) created via multi-agent generation, consistency checks, and expert audits, alongside a disjoint 4,947-sample blind test split. To resolve evidence dilution, TSR-MLLM, built on Qwen3-VL-4B, uses a query-conditioned Text-Guided Small-Object Focus (TG-SOF) map. Applied once at the decoder boundary, the map adds sparse Top-K gated residuals to the most question-relevant vision slots while leaving text tokens unchanged. Together with lightweight decoder adaptation, TSR-MLLM preserves single-pass inference without external detectors or image re-encoding. Under matched settings, TSR-MLLM outperforms the strongest 4B baseline by 2.1 points on FGTR-Bench (74.1% overall), with larger gains on evidence-local tracks. Furthermore, it remains competitive on DriveQA-V (CARLA Signs) under greedy decoding without task-specific fine-tuning.
Thank you to arXiv for use of its open access interoperability. PaperDance 不是 arXiv 官方产品；中文卡片由大模型生成，请以原文为准。