Language models for traffic accident liability determination based on chain-of-thought fine-tuning

doi:10.16265/j.cnki.issn1003-3033.2026.S1.0031

Abstract

Abstract:

In response to low manual efficiency and the insufficient reasoning, generalization, and interpretability of traditional machine learning models under complex accident scenarios in road traffic accident liability determination, a chain-of-thought (CoT) fine-tuning framework based on large language model (LLM) was constructed. Fine-tuning data were built from 870 California autonomous vehicle accident reports (2014-2025) and the California traffic laws. An expert-template-guided reasoning path for liability determination was designed. A two-stage fine-tuning practice was implemented using knowledge distillation. Legal knowledge, professional reasoning processes, and structured output patterns were internalized into a small-scale language model. The results show that compared with the non-fine-tuned model, the responsible party matching accuracy is improved from 30.19% to 97.18%, with the comprehensive reasoning quality score raised from 1.33 to 4.64 under zero-shot conditions. In zero-shot settings, the fine-tuned small-scale model (3B) outperforms a 235B open-source large model in full correctness, party correctness, and responsible party matching accuracy, with a 7.4% higher core reasoning ability score.

Key words: chain-of-thought fine-tuning, traffic accident liability determination, large language model, interpretability, knowledge distillation

CLC Number:

X913.2安全信息论

Xu Chuan, Li Bo, Hu Jialin, Luo Dan, Jiang Xinguo. Language models for traffic accident liability determination based on chain-of-thought fine-tuning[J]. China Safety Science Journal, 2026, 36(S1): 210-216.

Figures/Tables 9

Fig.1

Fig.2

Fig.3

Fig.4

Table 1

Table 2

Table 3

Table 4

Fig.5

References 16

[1]	世界卫生组织. 2023年全球道路安全状况报告[R], 2023.
	World Health Organization. Global status report on road safety 2023[R], 2023.
[2]	国家统计局. 中国统计年鉴2024[M]. 北京: 中国统计出版社, 2024: 755.
[3]	余凌云. 道路交通事故责任认定研究[J]. 法学研究, 2016, 38(6): 126-139.
	Yu Lingyun. Determination of liabilities for road traffic accident[J]. Chinese Journal of Law, 2016, 38(6): 126-139.
[4]	马社强. 道路交通事故当事人责任认定有关问题研究[J]. 中国人民公安大学学报(自然科学版), 2018, 24(2): 84-87.
[5]	黄钢, 高岩, 赵冬, 等. 交通事故智能定责深度学习集成方法研究[J]. 交通运输系统工程与信息, 2025(3): 299-307.
	Huang Gang, Gao Yan, Zhao Dong, et al. Deep learning ensemble method for intelligent liability determination in traffic accidents[J]. Journal of Transportation Systems Engineering and Information Technology, 2025(3): 299-307.
[6]	Brown T, Mann B, Ryder N, et al. Language models are few-shot learners[J]. Advances in Neural Information Processing Systems, 2020, 33: 1 877-1 901.
[7]	Touvron H, Lavril T, Izacard G, et al. Llama:open and efficient foundation language models[PP/OL]. arXiv(2023-02-27)[2026-03-11]. https://doi.org/10.48550/arXiv.2302.13971.
[8]	Yang An, Li Anfeng, Yang Baosong, et al. Qwen3 technical report[PP/OL]. arXiv(2025-05-14)[2026-03-09]. https://doi.org/10.48550/arXiv.2505.09388.
[9]	Yao Shunyu, Ke Qingqing, Wang Qiwei, et al. Lawyer GPT: a legal large language model with enhanced domain knowledge and reasoning capabilities[C]. Proceedings of the 2024 3rd International Symposium on Robotics, Artificial Intelligence and Information Engineering, 2024: 108-112.
[10]	Shi Juanming, Guo Qinglang, Liao Yong, et al. Legal GPT: legal chain of thought for the legal large language model multi-agent framework[C]. International Conference on Intelligent Computing, 2024: 25-37.
[11]	Wei J, Wang Xuezhi, Schuurmans D, et al. Chain-of-thought prompting elicits reasoning in large language models[J]. Advances in Neural Information Processing Systems, 2022, 35: 24 824-24 837.
[12]	Suzgun M, Scales N, Schärli N, et al. Challenging BIG-bench tasks and whether chain-of-thought can solve them[C]. Findings of the Association for Computational Linguistics: ACL 2023, 2023: 13 003-13 051.
[13]	Tan Zhen, Li Dawei, Wang Song, et al. Large language models for data annotation and synthesis: a survey[C]. Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, 2024: 930-957.
[14]	Zhen Hao, Shi Yucheng, Huang Yongcan, et al. Leveraging large language models with chain-of-thought and prompt engineering for traffic crash severity analysis and inference[J]. Computers, 2024, 13(9): DOI: 10.3390/computers13090232.
[15]	Li S, Chen J, Shen Y, et al. Explanations from large language models make small reasoners better[PP/OL]. arXiv(2022-10-13)[2026-02-22]. https://doi.org/10.48550/arXiv.2210.06726.
[16]	Hu E J, Shen Yelong, Wallis P, et al. LoRA:low-rank adaptation of large language models[PP/OL]. V2. arXiv(2021-06-17)[2026-02-12]. https://doi.org/10.48550/arXiv.2106.09685.

提示策略	M₁	M₂	M₃
FS	FS-M₁	FS-M₂	FS-M₃
ZS	ZS-M₁	ZS-M₂	ZS-M₃
CoT	CoT-M₁	CoT-M₂	CoT-M₃

系统提示		你是一个交通事故分析的专家,分析事故描述中的判责划分,给出正确的加州法规引用和多方责任划分,并结构化输出结果
用户输入	ZS提示	{事故描述}
	CoT提示	{事故描述} + {定制化CoT提示}
	FS提示	{事故描述} + {1个完整的输入输出案例}

模型	组别	完全正确率/%	参与方正确率/%	责任方匹配准确率/%
M₁	FS-M₁	21.33	48	66.67
	ZS-M₁	18.67	70.67	30.19
	CoT-M₁	45.33	82.67	56.45
M₂	FS-M₂	81.33	86.67	100
	ZS-M₂	89.33	94.67	97.18
	CoT-M₂	88	93.33	95.71
M₃	FS-M₃	78.67	89.33	89.55
	ZS-M₃	70.67	85.33	82.81
	CoT-M₃	88	97.33	90.41

模型	组别	D₁	D₂	D₃	D₄	D₁—D₃均值
M₁	FS-M₁	2.6	3.16	2.57	4.85	2.78
	ZS-M₁	1.91	1.03	1.04	0.23	1.33
	CoT-M₁	2.88	1.63	2.2	2.56	2.24
M₂	FS-M₂	4.44	4.59	3.96	5	4.33
	ZS-M₂	4.73	4.83	4.35	5	4.64
	CoT-M₂	4.65	4.76	4.27	4.99	4.56
M₃	FS-M₃	4.68	4.79	4.44	5	4.64
	ZS-M₃	4.23	4.37	4.36	2.41	4.32
	CoT-M₃	4.7	4.84	4.55	4.24	4.70