RISE-Judge: Scalable LLM Evaluation Framework

• RISE-Judge is a large language model-as-a-judge framework that uses a two-stage training protocol (SFT and DPO) for precise evaluation.
• It features an end-to-end evaluation pipeline with statistical tests like the Wilcoxon Signed-Rank Test and Benjamini-Hochberg correction to ensure robust system comparisons.
• The framework demonstrates state-of-the-art performance on benchmarks such as RewardBench while achieving high data efficiency compared to larger-scale training approaches.

RISE-Judge is a LLM–as-a-Judge framework that enables scalable and highly accurate evaluation of generative AI and retrieval-augmented generation (RAG) systems, with a focus on domains where evaluation precision is paramount, such as legal and ethical AI applications. RISE-Judge combines a two-stage judge model training protocol with a statistically principled system-level evaluation pipeline, offering state-of-the-art data efficiency and strong generalization beyond judge-specific tasks (Pradhan et al., 15 Sep 2025, Yu et al., 17 Feb 2025).

1. Judge Training Protocol

The RISE-Judge methodology employs a two-stage training approach to endow an LLM with advanced judge capabilities, decomposing judgment into supervised adaptation and preference-based refinement:

• Stage I: Supervised Fine-Tuning (SFT) Warm-Up The model is initialized from an open-source LLM (e.g., Qwen2.5-32B-Base). SFT data is synthesized by rewriting open QA/preference datasets using GPT-4o to create diverse judge prompts varying in role, language (Chinese/English), evaluation criteria, and output format. Each instance receives a step-by-step Chain-of-Thought (CoT) critique and a machine-readable verdict tag ([[A]] or [[B]]), filtered for label consistency and with position/length bias mitigated by answering with swapped responses and requiring similar lengths.

The SFT loss function is

$$\ell_\mathrm{SFT} = \mathbb{E}_{(\text{inst}, j) \sim \mathcal{D}_\mathrm{SFT}} [ -\log P_\theta(j\,|\,\text{inst}) ]$$

where $\text{inst}$ concatenates the prompt template $T$ and the candidate/comparator answers.

• Stage II: Direct Preference Optimization (DPO) Enhancement Preference pairs are mined from examples where SFT or GPT-4o was inconsistent with ground truth. For each question, $K=6$ candidate models outputs are scored, with pairwise preferences labeled via ground-truth-aligned rules. The DPO objective—referencing a frozen SFT checkpoint $\theta_0$—is:

$$\Delta(q; j_c, j_r) = \log P_\theta(j_c\,|\,\text{inst}) - \log P_\theta(j_r\,|\,\text{inst}) - [\log P_{\theta_0}(j_c\,|\,\text{inst}) - \log P_{\theta_0}(j_r\,|\,\text{inst})]$$

$$\ell_\mathrm{DPO} = \mathbb{E}_{(\text{inst}, j_c, j_r) \sim \mathcal{D}_\mathrm{DPO}} [-\log \sigma(\beta \cdot \Delta(q; j_c, j_r))]$$

The final loss incorporates a small NLL term for generation stability.

This protocol achieves state-of-the-art (SOTA) performance on RewardBench, matching large-scale baselines (SFR-LLaMa-70B) with approximately 2%–40% of the training data (RISE-Judge: 40 K vs. 600 K–900 K examples) (Yu et al., 17 Feb 2025).

2. Evaluation Pipeline and Statistical Methodology

RISE-Judge operationalizes LLM-as-a-Judge evaluation through an end-to-end pipeline structured as follows (Pradhan et al., 15 Sep 2025):

1. Document Retrieval: Competing systems retrieve relevant documents for legal queries (e.g., BM25+Summarizer vs. advanced RAG).
2. Answer Generation: Each system generates an answer $a_{S,k}$, with inline citations.
3. Prompt Engineering: Prompts instruct judges to rate answers on a 1–4 ordinal scale across multiple axes, such as relevance, completeness, and extrinsic hallucination.
4. LLM Judging: For each prompt, the LLM judge is invoked multiple times (e.g., $T = 10$ seeds) and votes are aggregated (e.g., majority).
5. Inter-Rater Reliability (IRR) Analysis: Alignment of LLM and human ratings is measured for each criterion and metric.
6. System Comparison: Paired samples are constructed for hypothesis testing of system differences per metric.

Statistical rigor is achieved by combining:

• Wilcoxon Signed-Rank Test (WSRT): A nonparametric test for paired ordinal/comparative data between systems. Given differences $d_i = s_{B,i} - s_{A,i}$ for $N$ paired queries, ranks are assigned, and $W^+, W^-$ are summed over positive and negative ranks; the statistic is $W = \min(W^+, W^-)$. Normal approximations and expectations are provided for inference.
• Benjamini-Hochberg (BH) Correction: Family-wise control of false discovery rate across multiple metrics/hypotheses. Sorted $p$-values $p_{(k)}$ are compared to thresholds $t_k=\frac{k}{m}\alpha$, and the largest $k$ for which $p_{(k)}\leq t_k$ determines rejected hypotheses.

Extensions include Bayesian variants of WSRT when $N$ is small and the Friedman test + Nemenyi for multi-system tournaments.

3. Reliability and Rank Correlation Metrics

Robust evaluation of judge–human agreement is critical in high-stakes environments. RISE-Judge explicitly quantifies IRR and ranking agreement as follows (Pradhan et al., 15 Sep 2025):

• Krippendorff’s Alpha ($\alpha$): A classical agreement metric accounting for chance, using a disagreement score weighted over all rater–item pairs. $\alpha$ is known to be unstable ("collapsing") in the presence of highly skewed metrics.
• Gwet’s AC2: A more robust alternative, especially when a single category dominates ratings. The observed agreement $A_o$ and chance baseline $A_e$ yield

$\mathrm{AC2} = \frac{A_o - A_e}{1 - A_e}$

Stability is observed even for right-skewed criteria.

• Spearman’s Rank Correlation ($\rho$): Correlates the ranking assigned by LLM judge and humans.
• Kendall’s Tau ($\tau$): Measures concordant/discordant ranking pairs.

Empirical results demonstrate that Gwet’s AC2 and rank correlations ($\rho\approx 0.73$, $\tau\approx 0.66$ for GPT-4o) remain reliable under label skew where Krippendorff’s alpha fails (e.g., $\alpha\approx 0.06$–$0.43$) (Pradhan et al., 15 Sep 2025).

4. Experimental Results and Data Efficiency

RISE-Judge attains SOTA performance in both specialized judgment and general abilities, with marked data and compute efficiency (Yu et al., 17 Feb 2025):

• RewardBench SOTA: RISE-Judge-Qwen2.5-32B achieves an average score of 92.7 on RewardBench over five sub-benchmarks, matching leaders trained on 15–20X more data.
• Ablation and Data Scaling: Combining SFT and DPO maximizes performance; neither stage alone suffices. Peak RewardBench at 20 K SFT + 20 K DPO examples.
• Generalization: On MMLU, CMMLU, CEval, BBH, GSM, AlignBench, MT-Bench, RISE-Judge matches or slightly surpasses Qwen2.5-32B-Instruct (Avg ≈ 83.4 vs. 83.1).
• Policy Model Bootstrapping: Feedback from RISE-Judge outperforms GPT-4o-labeled feedback (+0.2 on AlignBench).

Prompt robustness (variations in prompt style/language) leads to less than ±1 point variance. GPU requirements are modest (SFT: ~2 hours, DPO: ~3 hours on 8×A100; filtering/data synthesis are not compute-bound).

5. Practical Recommendations, Guidance, and Pitfalls

The RISE-Judge protocol can be adapted across domains, provided principled judge construction and evaluation (Pradhan et al., 15 Sep 2025, Yu et al., 17 Feb 2025):

• Judge Selection: Select LLM judge with maximum AC2 + Spearman’s $\rho$ on a pilot human set; prioritize AC2 under skewed metrics.
• Statistical Testing: Always use WSRT and BH correction for system-level comparisons. For multi-system tournaments, adopt the Friedman/Nemenyi protocol.
• Data Synthesis: Ensure no overlap between synthesized training q/a pairs and evaluation benchmarks to prevent data leakage.
• Bias and Prompt Mitigation: Swap answer positions, balance answer lengths, and monitor for template drift to prevent degradation of evaluation quality.
• Hyperparameters: Maintain $\alpha>0$ in DPO to retain generative quality and always use a fixed SFT checkpoint as a DPO reference.
• Pseudocode Reference:

for each new system S: for each query k: retrieve, generate answer for each metric m: prompt J and collect votes v_{S,k}^(m) for each metric m: compute IRR vs. reference → select best J for each competing system pair (A,B) and metric m: compute signed‐ranks W^+,W^- → p_raw[m] apply BH to {p_raw[·]} → p_adj[·] report metrics where p_adj[m] ≤ α

Common pitfalls include over-optimization in DPO (avoid vanishing $\alpha$), data leakage, and prompt drift from intended criteria.

6. Domain Extensions and Future Applications

The RISE-Judge framework is designed for portability and generalization. It has been explicitly recommended for domains including legal, medical, and financial RAG, and supports adaptation via:

• Bayesian forms of WSRT for small $\mathcal{N}$ [cf. Miller 2024].
• Confidence-weighting if LLMs emit uncertainty tokens.
• Tournament evaluation protocols for 3+ systems (Friedman + Nemenyi). A plausible implication is that RISE-Judge protocols can become evaluation standards in highly regulated or specialized domains demanding reliable, scalable human-aligned assessment (Pradhan et al., 15 Sep 2025, Yu et al., 17 Feb 2025).