JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction

Zhaolu Kang; Yantao Liu; Tailong Luo; Leqi Zheng; Lei Wei; Chenghua Zhu; Junhao Gong; Jiachen Qian

cs.CLAug 30, 2026

JPO: Juris Policy Optimization for Structured Legal Reasoning in Criminal Judgment Prediction

Authors: Zhaolu Kang, Yantao Liu, Tailong Luo, Leqi Zheng, Lei Wei, Chenghua Zhu, Junhao Gong, Jiachen Qian, +9 more

Organizations: Hunyuan, Tencent · Peking University · Tsinghua University · City University of Hong Kong · University of California · University of Illinois Urbana-Champaign · Zhejiang University · University of Hong Kong

Abstract

Criminal judgment prediction requires models to infer statutory articles, charges, and sentencing outcomes from case facts. Unlike standard classification tasks, it involves a structured reasoning process in which statutes should be matched with facts, charges should be justified by statutes, and sentencing outcomes should remain consistent with charges. Existing approaches optimize final labels, and while some have attempted to evaluate reasoning quality, their evaluations are indirect, often relying on LLM-generated rubrics that reflect model-internal preferences rather than the inherent logical structure of legal adjudication. We propose Juris Policy Optimization (JPO), a post-training framework for structured legal reasoning in Chinese criminal judgment prediction. JPO first uses teacher-generated rationales to supervise a standardized four-step reasoning process, and then applies reinforcement learning with a composite reward over legal prediction quality, reasoning structure completeness, and cross-step consistency. JPO further introduces token-level advantage reweighting and adaptive clipping for legally salient reasoning segments. Experiments on multiple open-source language models and three Chinese legal benchmarks show that JPO consistently improves both judgment prediction and reasoning quality over supervised fine-tuning and reinforcement learning baselines.

Figures & tables

Figure 1: Overview of JPO. Structured SFT with teacher-generated four-step rationales, followed by reinforcement learning with rewards for legal accuracy, reasoning completeness, and cross-step consistency.

Method	Set.	JPO-Dataset					CAIL2018					LawBench
		Art.	Charge	Sent.	4-Step	Full	Art.	Charge	Sent.	4-Step	Full	Art.	Charge	Sent.	4-Step	Full
		(F1)	(F1)	(Score)	(Comp.)	(Chain)	(F1)	(F1)	(Score)	(Comp.)	(Chain)	(F1)	(F1)	(Score)	(Comp.)	(Chain)
Open-Source Results
Qwen3-4B-Instruct	Pre-trained	0.521	0.468	0.174	0.315	0.208	0.485	0.442	0.158	0.291	0.187	0.463	0.419	0.142	0.264	0.175
	SFT	0.884	0.858	0.405	0.902	0.652	0.856	0.831	0.381	0.882	0.627	0.834	0.809	0.364	0.857	0.598
	JPO	0.931	0.916	0.542	0.966	0.789	0.903	0.888	0.514	0.951	0.758	0.872	0.861	0.485	0.935	0.724

Table 1: Main results on JPO-Dataset, CAIL2018, and LawBench. We report article F1, charge F1, sentence score, 4-Step Completeness, and Full-Chain Consistency.

Dataset	SFT Train	RL Train	Test
JPO-Dataset	239,515	9,691	20,396
CAIL2018	–	–	30,000
LawBench	–	–	1,500

Table 2: Dataset statistics used in our experiments.

Figure 2: Reasoning-chain breakdown on two representative 3B backbones. F

\rightarrow

A, A

\rightarrow

C, and C

\rightarrow

S denote fact-to-article, article-to-charge, and charge-to-sentence consistency, respectively.

Figure 3: Representative qualitative example illustrating the two-stage process of JPO on a criminal case.

Appendix figures & tables29 assets

Supplementary material from the paper’s appendix.

Appendix

Statistic	Value
Time range of collected cases	2024–2026
Number of unique charges	192
Number of unique statutory articles	176
Average fact length (tokens)	217.1
Median fact length (tokens)	194
Average number of articles per case	1.04

Appendix

Table 3: Profile statistics of JPO-Dataset.

Figure 4: Training dynamics of JPO on the representative Qwen2.5-3B backbone. The SFT stage mainly improves output validity and coarse prediction quality, while the RL stage yields the largest gains in sentence prediction and reasoning consistency.

Method	Art. F1	Charge F1	Sent.	Full-Chain
JPO	0.929	0.921	0.536	0.791
w/o $S_{FA}$	0.923	0.915	0.520	0.768
w/o $S_{AC}$	0.920	0.908	0.513	0.755
w/o $S_{CS}$	0.925	0.916	0.497	0.763
Entropy-only weighting	0.923	0.912	0.523	0.769
Logic-only weighting	0.924	0.914	0.525	0.771

Appendix

Table 4: Extended ablation analysis on Qwen2.5-3B. Bold indicates the full JPO model, and underlined text indicates key ablations.

$\alpha$	$\beta$	$\gamma$	Art. F1	Charge F1	Sent.
0.85	0.05	0.10	0.923	0.913	0.518
0.75	0.0625	0.1875	0.929	0.921	0.536
0.65	0.0625	0.2875	0.926	0.917	0.530
0.33	0.33	0.33	0.904	0.893	0.494

Appendix

Table 5: Sensitivity to reward weight composition.

$\psi$	Art. F1	Charge F1	Sent.
0.0	0.922	0.911	0.519
0.3	0.924	0.915	0.523
0.6	0.929	0.921	0.536
0.9	0.927	0.918	0.531

Appendix

Table 6: Sensitivity to token-level advantage scaling

\psi

$\zeta$	Art. F1	Charge F1	Sent.
0.0	0.924	0.914	0.525
0.25	0.926	0.917	0.529
0.5	0.929	0.921	0.536
0.75	0.925	0.916	0.527
1.0	0.923	0.912	0.523

Appendix

Table 7: Sensitivity to entropy–logic mixing coefficient

\zeta

Figure 5: Hyper-parameter sensitivity of JPO on the representative Qwen2.5-3B backbone.

Method	Set.	JPO-Dataset					CAIL2018					LawBench
		Art.	Charge	Sent.	4-Step	Full	Art.	Charge	Sent.	4-Step	Full	Art.	Charge	Sent.	4-Step	Full
		(F1)	(F1)	(Score)	(Comp.)	(Chain)	(F1)	(F1)	(Score)	(Comp.)	(Chain)	(F1)	(F1)	(Score)	(Comp.)	(Chain)
DeepSeek-V3.2	Zero-Shot	0.924	0.908	0.512	0.958	0.771	0.896	0.875	0.481	0.942	0.738	0.864	0.841	0.439	0.925	0.706
Qwen3-32B	Zero-Shot	0.917	0.911	0.504	0.961	0.759	0.887	0.882	0.473	0.945	0.722	0.858	0.835	0.428	0.931	0.694
GPT-5.2	Zero-Shot	0.912	0.894	0.489	0.947	0.744	0.879	0.861	0.455	0.933	0.705	0.845	0.829	0.412	0.916	0.683
Claude-Sonnet-4.5	Zero-Shot	0.908	0.891	0.493	0.952	0.751	0.882	0.866	0.462	0.937	0.713	0.849	0.824	0.417	0.918	0.687

Appendix

Table 8: Zero-shot performance of proprietary models. These results are for reference only and are not directly comparable to our open-source models trained with task-specific post-training.

Method	Paradigm	Art. F1	Charge F1	Sentencing (as reported)
TopJudge	dependency-aware	0.737	0.800	0.357 (Acc)
MPBFN	dependency-aware	0.706	0.757	0.362 (Acc)
LADAN	structure & knowledge	0.738	0.801	0.361 (Acc)
NeurJudge	structure & knowledge	0.797	0.807	0.374 (Acc)
CTM	structure & knowledge	0.768	0.780	0.374 (Acc)
EPM	dependency-aware	0.781	0.814	0.367 (Acc)

Appendix

Table 9: Comparison with SOTA LJP methods on CAIL2018.

System	Fact	Article	Charge	Sentence
SFT (Qwen2.5-7B-Instruct)	0.73	0.63	0.58	0.49
JPO (Qwen2.5-7B-Instruct)	0.87	0.83	0.79	0.71
Improvement	+0.14	+0.20	+0.21	+0.22

Appendix

Table 10: Pilot blind expert evaluation results on 400 randomly sampled cases from the JPO-Dataset test set.

Teacher	Stage	Art. (F1)	Charge (F1)	Sent. (Score)
Qwen2.5-72B-Instruct	SFT	0.893	0.871	0.422
Qwen2.5-72B-Instruct	JPO	0.937	0.928	0.551
DeepSeek-V2	SFT	0.901	0.875	0.438
DeepSeek-V2	JPO	0.939	0.926	0.554

Appendix

Table 11: Teacher model sensitivity analysis on JPO-Dataset.

Model	$\xi$	Pre-trained	SFT	JPO
Qwen2.5-3B-Instruct	1.0	0.317	0.613	0.747
	2.0	0.185	0.472	0.614
	3.0	0.114	0.391	0.536
	4.0	0.086	0.347	0.475
	5.0	0.067	0.305	0.437
Qwen2.5-7B-Instruct	1.0	0.430	0.657	0.765

Appendix

Table 12: Sentence score sensitivity to

\xi

on JPO-Dataset.

Method	GPU Hours	Avg. Length	Std. Full-Chain
SFT	34	456	0.018
Vanilla PPO	49	471	0.025
JPO	53	478	0.015

Appendix

Table 13: Training efficiency and stability on Qwen2.5-3B.

Error Type	Proportion
Missing secondary article	27.5%
Correct charge but wrong sentence band	25.3%
Confusion between neighboring charges	21.6%
Incomplete mitigation/aggravation reasoning	16.7%
Formatting or parsing errors	8.9%

Appendix

Table 14: Main remaining error types of JPO.

Automatic proxy	Expert step judged	Spearman $\rho$
$S_{FA}$ (fact-to-article)	Statutory analysis	0.67
$S_{AC}$ (article-to-charge)	Charge determination	0.71
$S_{CS}$ (charge-to-sentence)	Sentence prediction	0.64
Full-Chain Consistency	Overall reasoning	0.72

Appendix

Table 15: Correlation between the automatic consistency proxies and blind expert step-level ratings, computed on the 800-case expert study of Appendix O.6 (Qwen2.5-7B-Instruct).

Reasoning structure	Art. F1	Charge F1	Sent.	Full-Chain
Standard CoT (free-form)	0.921	0.907	0.503	0.724
Fact–Element–Charge (3-step)	0.928	0.914	0.517	0.749
Issue-tree decomposition	0.931	0.918	0.525	0.768
Four-step F $\to$ A $\to$ C $\to$ S	0.937	0.928	0.551	0.806

Appendix

Table 16: Reasoning-structure comparison under the JPO pipeline (Qwen2.5-7B-Instruct, JPO-Dataset). Rewards are adapted to each structure; all other settings are identical.

Evaluation subset	Model	Art. F1	Charge F1	Sent.	Full-Chain
Single-def./single-charge	SFT	0.893	0.871	0.422	0.681
Single-def./single-charge	JPO	0.937	0.928	0.551	0.806
Multi-charge ( $n$ =1,000)	SFT	0.812	0.774	0.348	0.579
Multi-charge ( $n$ =1,000)	JPO	0.869	0.836	0.452	0.695
Multi-defendant ( $n$ =800)	SFT	0.701	0.643	0.246	0.437
Multi-defendant ( $n$ =800)	JPO	0.783	0.724	0.351	0.558

Appendix

Table 17: Generalization to complex cases (Qwen2.5-7B-Instruct, no retraining). The two harder subsets are drawn from held-out judgments outside JPO-Dataset.

Logic-weight granularity	Art. F1	Charge F1	Sent.	Full-Chain	Cost
Entropy-only (no logic weight)	0.923	0.912	0.523	0.769	1.00 $\times$
Stage-level (proposed)	0.929	0.921	0.536	0.791	1.05 $\times$
Span-level (per-sentence)	0.931	0.923	0.540	0.797	1.4 $\times$
Token-level (entailment model)	0.933	0.925	0.544	0.802	2.6 $\times$

Appendix

Table 18: Granularity of the legal-logic token weight (Qwen2.5-3B-Instruct, JPO-Dataset). Cost is training time relative to the entropy-only baseline.

Model	LLM-judge rating (1–5)	Win rate vs. SFT
SFT	3.14	–
Vanilla PPO	3.36	57%
Issue Tree Rubrics	3.53	63%
JPO	3.90	74%

Appendix

Table 19: Format-independent LLM-as-judge evaluation (Qwen2.5-3B-Instruct, 500 cases, judge = GPT-5.2). The judging rubric does not mention the four-step template.

Model	Fact	Article	Charge	Sentence
Pretrained	0.62	0.51	0.49	0.28
SFT	0.71	0.66	0.59	0.47
JPO	0.88	0.81	0.77	0.70

Appendix

Table 20: Within-subjects expert evaluation on Qwen2.5-7B-Instruct, averaged over 3 experts

\times

800 cases. Fleiss’

\kappa=0.73

Teacher prompt	SFT Sent.	JPO Sent.	JPO counterfact.
Unconstrained (answer-first)	0.404	0.522	0.60
Constrained forward (ours)	0.422	0.551	0.83

Appendix

Table 21: Teacher-prompt ablation (Qwen2.5-7B-Instruct, JPO-Dataset). The last column reports the counterfactual consistency rate of the resulting JPO model.

Model	Counterfactual consistency rate
Pretrained	0.43
SFT	0.61
JPO	0.83

Appendix

Table 22: Counterfactual consistency rate (Qwen2.5-7B-Instruct, 500 edited cases, constrained-prompt models).

Configuration	Art. F1	Charge F1	Sent.	Full-Chain
Structured four-step SFT	0.873	0.847	0.391	0.622
+ generic PPO (outcome-only reward)	0.896	0.869	0.454	0.678
+ generic entropy-only token weighting	0.900	0.875	0.466	0.690
+ legal composite reward	0.919	0.908	0.501	0.747
JPO (+ legal-logic token opt.)	0.929	0.921	0.536	0.791

Appendix

Table 23: From a generic RL recipe to JPO (Qwen2.5-3B-Instruct, JPO-Dataset). The legal composite reward combines the structure and consistency terms.

Method	Setting	Art. F1	Charge F1	Sent.	4-Step	Full-Chain
DeepSeek-V3.2	Zero-Shot	0.924	0.908	0.512	0.958	0.771
DeepSeek-V3.2	ICL (3-shot)	0.923	0.911	0.510	0.961	0.773
GPT-5.2	Zero-Shot	0.912	0.894	0.489	0.947	0.744
GPT-5.2	ICL (3-shot)	0.916	0.899	0.485	0.952	0.747
Qwen2.5-3B	JPO	0.929	0.921	0.536	0.967	0.791
Qwen2.5-3B	JPO + ICL (3-shot)	0.930	0.921	0.538	0.968	0.793

Appendix

Table 24: In-context learning on JPO-Dataset, extending Table 8 .

System	Params	Inference cost / 1k cases	One-time training
DeepSeek-V3.2 (API)	–	$\approx$ $1.9	–
GPT-5.2 (API)	–	$\approx$ $9.7	–
JPO Qwen2.5-3B (local)	3B	$\approx$ $0.16	53 GPU-h

Appendix

Table 25: Cost comparison per 1,000 JPO-Dataset cases. API entries use public list pricing; the local entry amortizes A800 GPU time over inference.

Teacher	Scale	SFT Sent.	JPO Art.	JPO Charge	JPO Sent.
Qwen2.5-7B-Instruct	7B	0.401	0.930	0.919	0.532
Qwen2.5-32B-Instruct	32B	0.412	0.934	0.923	0.544
Qwen2.5-72B-Instruct (default)	72B	0.422	0.937	0.928	0.551
DeepSeek-V2	MoE	0.438	0.939	0.926	0.554
Claude Opus 4.8	prop.	0.449	0.941	0.930	0.558

Appendix

Table 26: Teacher-model scaling with a fixed student (Qwen2.5-7B-Instruct) on JPO-Dataset. “SFT Sent.” is the sentence score of the Stage-I checkpoint before RL.

Panel of Figure 6 — Figure 6: Prompt template used to generate four-stage legal rationales for structured supervised fine-tuning.

Explore similar work

The Cases LJP Never Sees: Prosecution Decision Prediction for More Complete Criminal Liability Assessment
May 27, 2026Junyu Lu, Qi Wei, Peishuo Zheng +6Mljailde Assessment
1Beijing Institute of Technology, Zhuhai · 2Osaka University · 3Xi’an University of Technology +2
Enhancing Judgment Document Generation via Agentic Legal Information Collection and Rubric-Guided Optimization
May 3, 2026Weihang Su, Xuanyi Chen, Yueyue Wu +2Legal Reasoning Tasks Reward Functions
Tsinghua University · Beijing, China · Quan Cheng Laboratory +1
Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models
Oct 5, 2026Subramanyam Sahoo, Justin ShenkLegal Reasoning Tasks Large Reasoning Models
Horizon Research.

May 27, 2026, cs.CLJ/K move · Enter open · S save

The Cases LJP Never Sees: Prosecution Decision Prediction for More Complete Criminal Liability Assessment

Junyu Lu, Qi Wei, Peishuo Zheng, Jie Zhang+5

1Beijing Institute of Technology, Zhuhai · 2Osaka University · 3Xi’an University of Technology+2

Mljailde Assessment

Open paper AskPDF arXiv

Legal Judgment Prediction (LJP) has become a core benchmark for evaluating AI in the criminal legal domain, but it only sees criminal cases that have already passed prosecutorial review and been formally indicted. As a result, LJP leaves a substantial blind spot in assessing criminal liability, overlooking cases involving insufficient evidence, no criminal liability, or guilt exempted from punishment. To fill this gap, we propose \textbf{Prosecution Decision Prediction (PDP)}, the first Legal AI task built around prosecutorial review, which classifies each case into prosecution or one of three non-prosecution decisions and reflects legal AI's capabilities in evidence evaluation, legal subsumption, and value-based discretion. We further construct \textbf{PDP-Bench}, a benchmark of 4{,}630 real Chinese prosecutorial decisions spanning 190 charges. Extensive experiments show that state-of-the-art LLMs perform substantially worse on PDP than on LJP and that mainstream enhancement routes fail to close the gap. Moreover, controlled RLVR interventions show that simple outcome rewards fail to produce generalizable PDP discrimination.

Abstract

Explore similar work

The Cases LJP Never Sees: Prosecution Decision Prediction for More Complete Criminal Liability Assessment

Enhancing Judgment Document Generation via Agentic Legal Information Collection and Rubric-Guided Optimization

Better Call Reward: Reward Hacking as Strategic Abstention in Legal Reasoning Models