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

2026-05-27

AI Summary

Existing legal judgment prediction (LJP) models focus exclusively on prosecuted cases, overlooking scenarios where prosecution is declined due to insufficient evidence, lack of criminal liability, or exemptions from punishment, thereby creating blind spots in criminal accountability assessment. This work introduces Prosecution Decision Prediction (PDP), a novel task that establishes the first legal AI framework tailored to the prosecutorial review stage. We present PDP-Bench, a benchmark dataset comprising 4,630 real-world Chinese prosecutorial decisions spanning 190 criminal charges. Leveraging large language models, our approach integrates multi-dimensional classification based on evidence evaluation, legal subsumption, and value-based discretion, complemented by RLVR intervention experiments to probe decision mechanisms. Results demonstrate that current large language models perform significantly worse on PDP than on LJP, and neither mainstream enhancement techniques nor simple reward mechanisms effectively improve their generalization capability.

Abstract

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.

Problem

• Legal Judgment Prediction
• Prosecution Decision Prediction
• criminal liability assessment
• prosecutorial review
• non-prosecution

Innovation

• Prosecution Decision Prediction
• Legal Judgment Prediction
• PDP-Bench
• evidence evaluation
• legal AI