Legal reasoning requires not only the application of legal rules but also an understanding of the context in which those rules operate. However, existing legal benchmarks primarily evaluate rule application under the assumption of fixed norms, and thus fail to capture situations where legal judgments shift or where multiple norms interact. In this work, we propose CALRK-Bench, a context-aware legal reasoning benchmark based on the legal system in Korean. CALRK-Bench evaluates whether models can identify the temporal validity of legal norms, determine whether sufficient legal information is available for a given case, and understand the reasons behind shifts in legal judgments. The dataset is constructed from legal precedents and legal consultation records, and is validated by legal experts. Experimental results show that even recent large language models consistently exhibit low performance on these three tasks. CALRK-Bench provides a new stress test for evaluating context-aware legal reasoning rather than simple memorization of legal knowledge. Our code is available at https://github.com/jhCOR/CALRKBench.
Figures & tables
Figure 1: Overview of the Benchmark.
Type-TCR
Type-ISR
Type-JSA
Model
Pre.
Post.
Avg
Suff.
Insuff.
Part.
Avg
Prem.
Text
Interp.
Fact
Avg
GPT-5* (m)
0.053
0.600
0.327
0.885
0.753
0.717
0.762
0.226
0.679
0.244
0.944
0.487
GPT-5* (l)
0.427
0.747
0.587
0.973
0.756
0.651
0.754
0.238
0.714
0.539
0.778
0.547
Gemini Flash* (m)
0.200
0.727
0.463
1.000
0.721
0.636
0.737
0.702
0.821
0.154
0.944
0.637
Gemini Flash* (l)
0.720
0.820
0.770
0.993
0.737
0.648
0.748
0.774
0.905
0.449
0.667
0.707
LLaMA 3.3 70B
0.000
0.620
0.310
0.633
0.707
0.682
0.685
0.536
0.107
0.077
0.833
0.350
Table 1: Experiment conducted via OpenRouter. Detailed performance breakdown on CALRK-Bench. The best score in each column is shown in bold , and the second-best score is underlined . For GPT and Gemini models, we ran the evaluation three times with the seed fixed and report the average results due to internal stochasticity of the models. For open-source models, we used a fixed seed and report results from a single deterministic run. m: reasoning effort minimal, l: reasoning effort low, T: Thinking model, I: Instruct model. For Type-JSA , the options were randomly shuffled during evaluation.
Model
before
after
avg
GPT-5 (minimal)
0.053
0.600
0.327
GPT-5 (low)
0.427
0.747
0.587
GPT-5 (medium)
0.500
0.747
0.623
GPT-5 (high)
0.493
0.767
0.630
Table 2: Type-TCR performance under different reasoning effort levels of GPT-5.
Model
Acc.
Δ
GPT-5 (m)
0.717
–
GPT-5 (l)
0.651
-0.066
LLaMA 3.3 70B
0.682
–
LLaMA 3.3 70B + CoT
0.664
-0.018
Qwen3-30B-A3B
0.757
–
Qwen3-30B-A3B + CoT
0.720
-0.037
Table 3: Performance under the Partial condition in Type-ISR . Acc denotes accuracy under the Partial condition. Δ indicates the change in accuracy between the baseline model and its reasoning-augmented variant (e.g., CoT prompting or thinking-mode model).
Figure 2: Row-normalized confusion matrices for Type-JSA . Each value represents the percentage of predictions within the same gold label.
Figure 3: For the experiments involving context addition, the LexEval data were randomly sampled or the strings were randomly generated. Results averaged over three runs.