Beyond Scene Priors: Fine-Grained Traffic Scene Reasoning with Benchmarking and Query-Guided Small-Object Focus

2026-07-05

AI Summary

This work addresses the “critical evidence dilution” problem in multimodal large language models (MLLMs) for traffic visual question answering, where small yet crucial objects are often overlooked, and highlights the lack of fine-grained reasoning benchmarks in existing datasets. To this end, the authors introduce FGTR-Bench, the first fine-grained traffic reasoning benchmark, and propose TSR-MLLM, a text-guided small-object focusing model built upon the Qwen3-VL-4B architecture. TSR-MLLM generates attention maps conditioned on textual queries and incorporates sparse Top-K gated residual connections with a lightweight decoder adapter, enabling precise localization of critical regions in a single inference pass without external detectors. On FGTR-Bench, TSR-MLLM achieves 74.1% accuracy, outperforming the strongest 4B baseline by 2.1 percentage points—particularly excelling in local evidence grounding—while maintaining competitive performance on DriveQA-V.

Abstract

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.

Problem

• critical evidence dilution
• small-object focus
• traffic scene reasoning
• visual question answering
• multimodal large language models

Innovation

• fine-grained traffic reasoning
• small-object focus
• query-guided attention
• multimodal large language model
• visual question answering