A Progressive Visual-Logic-Aligned Framework for Ride-Hailing Adjudication

Weiming Wu; Zi-Jian Cheng; Jie Meng; Peng Zhen; Shan Huang; Qun Li; Guobin Wu; Lan-Zhe Guo

cs.AIMar 18, 2026

A Progressive Visual-Logic-Aligned Framework for Ride-Hailing Adjudication

Authors: Weiming Wu, Zi-Jian Cheng, Jie Meng, Peng Zhen, Shan Huang, Qun Li, Guobin Wu, Lan-Zhe Guo

Organizations: Nanjing University · Didichuxing Co. Ltd

Abstract

The efficient adjudication of responsibility disputes is pivotal for maintaining marketplace fairness. However, the exponential surge in ride-hailing volume renders manual review intractable, while conventional automated methods lack the reasoning transparency required for quasi-judicial decisions. Although Multimodal LLMs offer a promising paradigm, they fundamentally struggle to bridge the gap between general visual semantics and rigorous evidentiary protocols, often leading to perceptual hallucinations and logical looseness. To address these systemic misalignments, we introduce RideJudge, a Progressive Visual-Logic-Aligned Framework. Instead of relying on generic pre-training, we bridge the semantic gap via SynTraj, a synthesis engine that grounds abstract liability concepts into concrete trajectory patterns. To resolve the conflict between massive regulation volume and limited context windows, we propose an Adaptive Context Optimization strategy that distills expert knowledge, coupled with a Chain-of-Adjudication mechanism to enforce active evidentiary inquiry. Furthermore, addressing the inadequacy of sparse binary feedback for complex liability assessment, we implement a novel Ordinal-Sensitive Reinforcement Learning mechanism that calibrates decision boundaries against hierarchical severity. Extensive experiments show that our RideJudge-8B achieves 88.41% accuracy, surpassing 32B-scale baselines and establishing a new standard for interpretable adjudication.

Figures & tables

Figure 1. The pipeline consists of three pivotal phases: (1) Automated Data Synthesis (Sec. 3 ), which bridges domain gaps via two specialized modules: SynTraj Construction for visual-linguistic alignment and Chain-of-Adjudication Synthesis for logical reasoning reconstruction; (2) Knowledge-Aware Context Refinement (Sec. 4 ), capable of dynamic rule pruning and expert precedent extraction; and (3) Progressive Juridical Alignment (Sec. 5 ), a multi-stage training paradigm culminating in OS-rewarded reinforcement learning for precise decision boundary alignment.

Backbone	Method	Appeal					Driver-Cancel					Passenger-Cancel					Overall
		Normal		Malicious		$Acc.$	Normal		Malicious		$Acc.$	Normal		Malicious		$Acc.$	Overall
		$P$	$R$	$P$	$R$	$Acc.$	$P$	$R$	$P$	$R$	$Acc.$	$P$	$R$	$P$	$R$	$Acc.$	$Acc.$
LLM Series
DeepSeek-V3.1	Standard	70.91	95.49	80.70	20.63	57.89	80.69	86.67	64.00	84.21	86.15	88.57	66.77	89.12	65.40	85.30	75.25
DeepSeek-V3.1	CoT	71.15	93.59	86.79	20.63	58.09	77.99	91.85	58.33	73.68	86.36	86.75	66.46	88.38	66.54	85.30	75.36
Qwen3-32B-Insruct	Standard	62.76	89.08	27.07	63.23	38.93	32.67	36.30	23.08	15.79	25.97	30.73	41.23	53.98	23.19	17.26	26.77

Table 1. Main Results on Three Adjudication Benchmarks. We report Precision (

P

), Recall (

R

), and Accuracy (

Acc.

).Best results are in bold, second best are underlined.

Model	P	R	Acc.
QwQ-32B	29.73	51.27	34.00
Qwen3-VL-8B-Instruct	29.68	51.69	33.40
Qwen3-VL-8B-Thinking	29.67	52.77	33.90
Qwen3-VL-32B-Instruct	30.22	53.39	34.40
Qwen3-VL-32B-Thinking	30.12	53.88	33.40
Ours	53.17	65.34	60.20

Table 2. Main Results Comparison. We report the overall Precision (

P

), Recall (

R

), and Accuracy (

Acc.

) of selected baselines and our method.

No.	Settings	Normal		Malicious		Acc.
No.	Settings	$P$	$R$	$P$	$R$	Acc.
1	Baseline	63.57	24.09	51.35	8.52	44.49
2	+ 1st stage	64.29	28.08	48.84	9.42	45.91
3	+ 2nd stage	80.92	76.43	68.82	57.40	68.81
4	+ 3rd stage	94.60	94.11	91.52	91.93	91.86
5	SFT w/o CoA	76.88	77.82	67.31	36.77	63.75

Table 3. Ablation Study on Appeal Benchmark.

\Delta

indicates the performance gap between the full method and the ablated version.

Figure 2. Qualitative case studies on the Appeal benchmark. To preserve privacy, sensitive textual regions in the images, as well as specific numerical values and location names within the reasoning chains, have been masked. We highlight the Information Analysis process in green, the Visual Evidence Integration process in red, and the Rule Grounding process in blue.

Panel of Figure 3 — Figure 3. Left: Performance scaling with our CoA synthetic data used in Stage 2. Right: Stability analysis of RideJudge-8B from ten major cities.

Appendix figures & tables2 assets

Supplementary material from the paper’s appendix.

Appendix

Hyperparameter	Stage 1	Stage 2
Vision Tower	Unfrozen	Frozen
Projector	Unfrozen	Frozen
Language Model	Frozen	Unfrozen
Precision	BF16	BF16
Optimizer	AdamW	AdamW
Learning Rate	$5.0\times 10^{-6}$	$5.0\times 10^{-6}$

Appendix

Table 4. Hyperparameter settings for Stage 1 Visual Alignment and Stage 2 Logic Alignment.

Hyperparameter	Value
Algorithm	DAPO
Actor Learning Rate	$1.0\times 10^{-6}$
Critic Learning Rate	$1.0\times 10^{-6}$
Weight Decay	$1.0\times 10^{-2}$
KL Coefficient	$0.01$
KL Penalty Type	Low Var KL

Appendix

Table 5. Hyperparameter settings for Stage 3 Reinforcement Learning via DAPO.

Legal Judgment Prediction

Legal judgment prediction uses computational models to forecast court outcomes from case facts, legal arguments, statutes, and precedents.

See all 29 papers

TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination
Jun 1, 2026Xu Li, Zedong Fu, Xinyi Li +1Multimodal RAG Intelligent Transportation Systems
Southwest Petroleum University, China · Sichuan Police College, China
AITP: Traffic Accident Responsibility Allocation via Multimodal Large Language Models
Apr 11, 2026Zijin Zhou, Songan ZhangMultimodal Large Language Models
Global Institute of Future Technology, Shanghai Jiao Tong University
InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain
Oct 7, 2026Linqi Zhang, Chong Qi, Yan Cheng +4AI-Assisted Decision Making LLM Evaluation
School of Computer Science and Technology, Fudan University · School of Integrated Circuits, Nanjing University · School of Economics, Fudan University

Jun 1, 2026J/K move · Enter open · S save

TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination

Xu Li, Zedong Fu, Xinyi Li, Xun Han

Southwest Petroleum University, China · Sichuan Police College, China

Multimodal RAG Intelligent Transportation Systems Legal QA Legal Reasoning

Open paper AskPDF arXiv

Traffic accident liability analysis is a critical yet challenging task in intelligent transportation and legal assistance. Existing methods often suffer from low efficiency, subjective judgment, and inconsistent analysis results. Meanwhile, large language models are constrained by noisy video inputs and insufficient legal domain knowledge. To address these issues, this work presents TrafficRAG, a multimodal retrieval-augmented framework for automated traffic accident analysis and report generation. Specifically, the proposed framework first adopts a vision-language model to produce structured textual descriptions of accident scenarios, which serve as accurate retrieval queries. Based on these textual queries, a hybrid retrieval strategy integrating BM25 sparse retrieval and dense embedding retrieval is employed to fetch relevant traffic regulations and similar historical cases. Finally, the large language model incorporates retrieved legal knowledge and multimodal accident evidence for comprehensive reasoning, and generates standardized, legally grounded liability analysis reports. Extensive experiments show that TrafficRAG consistently outperforms baseline methods, achieving 77.32% Legal Norm Adaptation Accuracy, 81.71% Factual Faithfulness, and a Liability Ratio MAE of 5.48%. The results validate that integrating multimodal factual evidence with legal clauses via retrieval augmentation can effectively improve the reliability and accuracy of traffic accident liability determination.

Abstract

Legal Judgment Prediction

Explore similar work

TrafficRAG: A Multimodal RAG Framework for Traffic Accident Liability Determination

AITP: Traffic Accident Responsibility Allocation via Multimodal Large Language Models

InsClaimBench: Benchmarking Insurance Claim Adjudication Across the Decision Chain