LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation

Hui Huang; Jie Zhou; Hongli Zhou; Qundong Shi; Shuyuan Zheng; Yini Su; Shaochun Wang; Biyuan Lin

cs.CLOct 13, 2026

LegalRewardBench: Benchmarking Reward Models for Legal Preference Evaluation

Authors: Hui Huang, Jie Zhou, Hongli Zhou, Qundong Shi, Shuyuan Zheng, Yini Su, Shaochun Wang, Biyuan Lin, +2 more

Organizations: Harbin Institute of Technology · ModelBest Inc. · The University of Osaka

Abstract

Reward models are central to preference alignment for legal tasks, yet existing reward model benchmarks are still primarily centered on general-domain preferences. In this work, we introduce LegalRewardBench, the first human-verified benchmark for evaluating reward models in the legal domain. The benchmark organizes legal preference evaluation into five capability levels: knowledge, understanding, analysis, decision, and writing, and preference examples are constructed through rubric-guided automatic judging combined with human verification. We evaluate three families of judges, covering both general and fine-tuned models, and generative and discriminative architectures. Experimental results show that strong general-purpose generative models are currently the most robust legal reward models, while existing fine-tuned reward models transfer less reliably to legal preference evaluation. We also conduct detailed comparisons of input organizations and training strategies.

Figures & tables

Figure 1: Performance of best generative RMs across the five LegalRewardBench capability levels.

Level	Representative Tasks
L1 Knowledge	Statute Recitation; Legal Rule Recall; Applicable Law Identification; Legal Knowledge Error Detection…
L2 Understanding	Case Event Timeline Extraction; Legal Element Extraction; Long-Form Legal Document Understanding…
L3 Analysis	Fact-Finding Analysis; Disputed Issue Identification; Evidence Analysis…
L4 Decision	Sentence Prediction; Judgment Outcome Prediction; Legal Task Planning…
L5 Writing	Procuratorial Recommendation Drafting; Defense Pleading Drafting; Indictment Drafting…

Table 1: Representative tasks for the five capability levels in LegalRewardBench .

Level	Samples	Average Length			Human-Human Agree.		Human-Model Agree.
Level	Samples	Query	Chosen	Rejected	w/ Tie	w/o Tie	w/ Tie	w/o Tie
L1 Knowledge	306	234.5	449.9	435.9	76.19	76.19	81.69	86.31
L2 Understanding	228	1497.9	670.5	566.9	70.00	77.78	73.54	78.68
L3 Analysis	377	1292.3	661.7	446.2	64.71	64.71	77.71	81.15
L4 Decision	296	566.9	648.4	345.5	76.19	88.89	82.35	88.05
L5 Writing	124	6355.7	995.0	874.7	96.15	96.15	90.00	91.67

Table 2: Data statistics and annotation agreement by capability level on LegalRewardBench .

Figure 2: Model distribution of candidate responses.

Model	Type	TK	Overall	L1	L2	L3	L4	L5
Claude-Opus-4.7	General LLM	✓	86.85	91.83	81.36	87.00	87.50	82.66
Claude-Opus-4.5	General LLM	✓	85.01	91.01	76.75	86.60	84.12	82.66
GPT-5.5	General LLM	✓	83.66	92.97	78.73	88.06	72.97	81.85
Gemini-3-Flash-Preview	General LLM	✓	82.95	89.22	74.12	84.88	82.09	79.84
Qwen3-235B-A22B	General LLM	✓	81.52	88.40	73.25	84.08	77.87	80.65
Gemini-2.5-Pro	General LLM	✗	81.14	88.40	75.22	83.55	79.56	70.56

Table 3: Main results on LegalRewardBench . Values are accuracies in percentage points.

Model	Overall		L1		L2		L3		L4		L5
Model	Ori	Ref	Ori	Ref	Ori	Ref	Ori	Ref	Ori	Ref	Ori	Ref
Claude-Opus-4.7	86.85	91.81	91.83	94.93	81.36	90.13	87.00	88.99	87.50	96.11	82.66	85.48
GPT-4o	74.61	79.30	73.53	84.15	69.74	76.10	79.18	79.97	76.18	81.76	68.55	65.32
GPT-5.2	80.54	88.73	84.15	90.52	75.44	84.43	83.02	87.53	78.89	93.92	77.42	83.47
Kimi-K2.5	80.65	83.43	87.91	89.05	71.27	73.90	83.82	85.68	80.07	89.19	71.77	66.53
Qwen3-14B	77.84	86.14	76.63	90.69	70.39	82.46	85.01	87.80	76.69	85.47	75.40	78.23

Table 4: Averaged accuracy on LegalRewardBench with reference included or excluded.

Model	Overall		L1		L2		L3		L4		L5
Model	Tk	NoTk	Tk	NoTk	Tk	NoTk	Tk	NoTk	Tk	NoTk	Tk	NoTk
Claude-Opus-4.7	86.85	86.18	91.83	90.52	81.36	81.58	87.00	86.60	87.50	86.66	82.66	81.45
Gemini-3-Flash	82.95	82.34	89.22	88.73	74.12	73.25	84.88	84.22	82.09	82.26	79.84	77.82
Qwen3-14B	77.84	76.97	76.63	74.67	70.39	72.15	85.01	80.64	76.69	77.53	75.40	79.03
Qwen3-4B	74.87	73.14	71.90	70.26	70.18	69.08	82.23	81.70	74.66	72.30	68.95	63.71
Qwen3-8B	75.69	76.60	76.47	75.49	68.86	71.93	79.97	82.49	76.86	76.01	70.56	71.37

Table 5: Averaged accuracy on LegalRewardBench with thinking mode on or off.

Model	Overall		L1		L2		L3		L4		L5
Model	Zh	En	Zh	En	Zh	En	Zh	En	Zh	En	Zh	En
Claude-Opus-4.7	86.85	85.54	91.83	89.71	81.36	81.58	87.00	85.94	87.50	84.97	82.66	82.66
GPT-4o	74.61	74.98	73.53	71.73	69.74	70.61	79.18	79.44	76.18	79.05	68.55	67.74
GPT-5.2	80.54	80.09	84.15	83.82	75.44	75.66	83.02	82.49	78.89	77.36	77.42	78.23
Kimi-K2.5	80.65	78.36	87.91	85.46	71.27	69.52	83.82	80.77	80.07	79.22	71.77	67.74
Qwen3-14B	77.84	78.44	76.63	77.78	70.39	74.12	85.01	85.28	76.69	76.18	75.40	72.58

Table 6: Averaged accuracy on LegalRewardBench with different thinking languages.

Model	Overall		L1		L2		L3		L4		L5
Model	Ori	Rub	Ori	Rub	Ori	Rub	Ori	Rub	Ori	Rub	Ori	Rub
Claude-Opus-4.7	86.85	85.99	91.83	87.75	81.36	80.48	87.00	87.14	87.50	88.85	82.66	81.45
GPT-4o	74.61	75.88	73.53	72.55	69.74	71.49	79.18	81.70	76.18	79.22	68.55	66.53
GPT-5.2	80.54	80.73	84.15	82.19	75.44	76.32	83.02	81.56	78.89	81.76	77.42	80.24
Kimi-K2.5	80.65	71.83	87.91	79.74	71.27	62.94	83.82	74.40	80.07	75.00	71.77	53.23
Qwen3-14B	77.84	78.21	76.63	76.14	70.39	73.46	85.01	82.89	76.69	78.55	75.40	77.02

Table 7: Averaged accuracy on LegalRewardBench with automatic rubric generation.

Rank	Ori_Rank	Base Model	Sampler	Average	LawBench	LexEval
1	1	Qwen3-4B	Claude-Opus-4.5	54.29	52.12	56.46
2	3	Qwen3-4B	GPT-5.2	54.14	51.26	57.02
3	2	Qwen3-4B	Gemini-3-Flash-Preview	54.00	51.26	56.74
4	4	Qwen3-4B	GPT-4.1	52.70	49.65	55.75
5	5	Qwen3-4B	Skywork-Reward-V2-Qwen3-8B	52.38	49.23	55.52
6	7	Qwen3-4B	GPT-4o	52.19	49.54	54.84

Table 8: Best-of-4 sampling validation on LawBench and LexEval, with Qwen3-4B as the base model and different RMs as the sampler. Ori_Rank denotes the RM’s original rank on LegalRewardBench .

Method	Data	Qwen3-8B						Qwen2.5-7B-Instruct
Method	Data	L1	L2	L3	L4	L5	Overall	L1	L2	L3	L4	L5	Overall
Prompt	–	76.47	68.86	79.97	76.86	70.56	75.69	62.58	60.75	67.11	69.59	69.76	65.78
w/ reference	–	88.42	78.98	83.38	82.61	69.01	82.26	76.31	64.91	70.03	72.64	69.35	71.11
DisRM	LegalDelta 6K	72.55	60.96	65.78	73.65	67.74	68.44	76.14	73.25	73.47	73.31	62.90	73.03
DisRM	Combined 39K	78.43	64.91	72.41	77.03	66.13	72.95	73.86	64.47	74.01	76.35	70.97	72.58
GenRM	LegalDelta 6K	74.02	69.74	82.49	77.87	69.76	76.15	65.69	56.58	62.07	72.13	58.06	63.82

Table 9: Comparison of different RM training methods on LegalRewardBench .

Appendix figures & tables12 assets

Supplementary material from the paper’s appendix.

Appendix

Open-weight General LLMs
Model	Type	TK	Size	Response Gen.	RM
Qwen3.5-397B-A17B	General LLM	✓	397B (17B)	✓	✓
Qwen3.5-122B-A10B	General LLM	✓	122B (10B)	✓	✗
Qwen3.5-27B	General LLM	✓	27B	✓	✓
Qwen3-235B-A22B	General LLM	✓	235B (22B)	✓	✓
Qwen3-32B	General LLM	✓	32B	✓	✗

Appendix

Table 10: Model catalog for LegalRewardBench . TK denotes whether the model is evaluated in thinking mode; – indicates that the distinction is not applicable to the model family.

Level	Task	Samples	Source	Scoring
L1 ( Knowledge )	Statute Recitation (法条背诵)	45	LawBench (1-1)	Rule-based
	Legal Rule Recall (法条规则记忆)	37	In-house	Model-based
	Legal Knowledge Error Detection (法律知识幻觉)	31	In-house	Model-based
	Applicable Law Identification (适用法律推荐)	35	In-house	Model-based
	Statutory Evolution Analysis (立法沿革分析)	57	In-house	Model-based
	Judicial Procedure QA (司法程序问答)	49	In-house	Model-based

Appendix

Table 11: Task descriptions in LegalRewardBench (L1–L3). For different tasks, we use either rule-based or model-based automatic scoring methods based on the nature of the task.

Level	Task	Samples	Source	Scoring
L4 ( Decision )	Sentence Prediction Without Statutes (刑期预测：无法条内容)	69	LawBench (3-4)	Rule-based
	Sentence Prediction With Statutes (刑期预测：给定法条内容)	68	LawBench (3-5)	Rule-based
	Judgment Outcome Prediction (裁判结果预测)	41	In-house	Model-based
	Public-Opinion Risk Assessment (审判结果舆情预测)	40	In-house	Model-based
	Expert Reasoning Outline Generation (法律专家思路生成)	36	In-house	Model-based
	Legal Task Planning (法律任务规划)	42	In-house	Model-based

Appendix

Table 12: Task descriptions in LegalRewardBench (L4–L5, continued).

Figure 3: Self-bias analysis for judge models. All values are reported as percentages.

Explore similar work

Building Legal Reward Models for Grounding and Abstention
Sep 13, 2026Rilton Franzone, Valentin Noël, Puyu Wang +2Reward Modeling LLM Grounding
Devoteam
LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models
Sep 30, 2026Yida Cai, Xin Dai, Bingxiang He +5Reward Modeling Legal NLP
Peking University · Tsinghua University · Northeastern University
JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment
May 24, 2026Russell Yang, Ruishi Chen, Pierce Kelaita +6Human Preference Evaluation LLM Evaluation
Stanford University · Snorkel AI · Harvey

Sep 13, 2026J/K move · Enter open · S save

Building Legal Reward Models for Grounding and Abstention

Rilton Franzone, Valentin Noël, Puyu Wang, Philip Torr+1

Devoteam

Reward Modeling LLM Grounding RAG Evaluation LLM Abstention Legal QA

Open paper AskPDF arXiv

Large language models are increasingly used in high-stakes domains such as law, where systems must ground their reasoning in retrieved evidence and abstain when that evidence is insufficient. However, existing reward models are largely optimised for general preferences rather than contextual grounding, limiting their ability to evaluate these behaviours in retrieval-augmented generation (RAG) settings. We introduce a framework for transforming existing legal QA datasets into contextual preference data and use it to construct LegalRewardBench (LRB), a benchmark for evaluating grounded legal generation under noisy and insufficient retrieval conditions. Across general and legal contextual evaluation, we find that contextual DPO improves grounded evaluation, but performance is sensitive to preference-data construction. Length-balanced augmentation substantially improves grounded legal evaluation, with the strongest configuration combining length-balanced legal and general contextual preference data and improving performance by up to

\mathbf{+25.6}

pp over baseline. We further find evidence of cross-jurisdiction transfer: models contextually refined primarily on Victorian criminal-law data improve grounded evaluation on external US legal benchmarks, including a

\mathbf{+16.2}

pp improvement on Housing Statute QA. Together, these results provide a reproducible foundation for constructing and evaluating grounded legal reward models in retrieval-augmented settings.

Abstract

Explore similar work

Building Legal Reward Models for Grounding and Abstention

LexReward: A Taxonomy-Driven Reward Framework for Legal Language Models

JudgmentBench: Comparing Rubric and Preference Evaluation for Quality Assessment