File size: 5,363 Bytes
ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 8848f9e ec09659 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 | # JudgeSense: A Benchmark for Prompt Sensitivity in LLM-as-a-Judge Systems
[](https://creativecommons.org/licenses/by/4.0/)
[]()
[](https://huggingface.co/datasets/anonymousreview111/judgesense-benchmark)
---
## Overview
**JudgeSense** is a benchmark dataset of **500 hand-validated prompt pairs** for measuring prompt sensitivity in LLM-as-a-Judge evaluation systems. Each pair contains two differently phrased but semantically equivalent judge prompts applied to the same response, enabling rigorous measurement of how much a judge's decision changes due to prompt wording alone.
All 500 pairs were independently validated by two human annotators with full agreement: 500 confirmed semantically equivalent; 50 pairs involving Template 4 (polarity-inverted) were labeled non-equivalent by both annotators and excluded before publication.
The dataset covers four evaluation task types:
| Task | Source | Pairs | Labels |
|------|--------|-------|--------|
| **Factuality** | TruthfulQA | 125 | accurate / inaccurate |
| **Coherence** | SummEval | 125 | score_1 ... score_5 |
| **Preference** | MT-Bench | 125 | A / B |
| **Relevance** | BEIR | 125 | A / B |
---
## What This Enables
- **Prompt sensitivity evaluation** — measure how fragile a judge is to phrasing variation
- **LLM judge robustness benchmarking** — compare models on decision consistency
- **Detection of prompt-induced artifacts** — identify polarity inversions (T4) and other systematic biases
---
## Quick Start
```python
from utils.load_judgesense import load_task, load_all
from utils.compute_jss import compute_jss
# Load one task
pairs = load_task("factuality")
print(f"{len(pairs)} pairs loaded")
# Load all tasks
all_data = load_all()
# Compute JSS from your judge's decisions
jss = compute_jss(decisions_a, decisions_b)
print(f"JSS: {jss:.3f}")
```
Run the full example:
```bash
cd judgesense-benchmark
python examples/run_jss_example.py
```
---
## Dataset Schema
Each JSONL record has eight fields:
```json
{
"pair_id": "fact_001",
"task_type": "factuality",
"source_benchmark": "TruthfulQA",
"prompt_a": "Is this factually correct? Answer YES or NO only.\n\nResponse: ...",
"prompt_b": "Fact-check this response. Reply YES (correct) or NO (incorrect).\n\nResponse: ...",
"response_being_judged": "The Earth orbits around the Sun.",
"ground_truth_label": "accurate",
"semantic_equivalence_score": 1.0
}
```
---
## Metric: Judge Sensitivity Score (JSS)
JSS is the fraction of pairs where both prompt variants elicit the same decision from the judge:
```
JSS = (1/N) * sum( decisions_a[i] == decisions_b[i] )
```
- **JSS = 1.0** — perfectly consistent; the judge never changes its decision due to prompt phrasing
- **JSS = 0.0** — maximally sensitive; every decision flips between prompts
A high flip rate (= 1 - JSS) indicates the judge's apparent decisions are largely driven by prompt design rather than the content being evaluated.
---
## Benchmark Results (13 judges, pass-3)
### Coherence (most discriminating task)
| Model | JSS | Cohen's kappa |
|---|---|---|
| Claude Sonnet 4.5 | 0.99 | 0.986 |
| Qwen-2.5-72B | 0.92 | 0.842 |
| GPT-4o | 0.91 | 0.828 |
| GPT-5.5 | 0.83 | 0.694 |
| GPT-4o-mini | 0.78 | 0.627 |
| Claude Haiku 4.5 | 0.73 | 0.583 |
| Claude Opus 4.7 | 0.70 | 0.580 |
| LLaMA-3.1-70B | 0.55 | 0.338 |
| DeepSeek-R1 | 0.53 | 0.332 |
| Qwen 3.6 Flash | 0.51 | 0.372 |
| DeepSeek-V4 Flash | 0.50 | 0.349 |
| Mistral-7B | 0.48 | -0.082 |
| Gemini 2.5 Flash | 0.39 | -0.057 |
### Factuality (after T4 polarity correction)
| Model | JSS (raw) | JSS (corrected) | Delta |
|---|---|---|---|
| GPT-4o | 0.63 | 0.98 | +0.35 |
| GPT-4o-mini | 0.63 | 0.96 | +0.33 |
| Claude Haiku 4.5 | 0.63 | 0.97 | +0.34 |
| Claude Sonnet 4.5 | 0.63 | 0.97 | +0.34 |
| DeepSeek-R1 | 0.63 | 0.96 | +0.33 |
| LLaMA-3.1-70B | 0.63 | 0.99 | +0.36 |
| Gemini 2.5 Flash | 0.63 | 0.98 | +0.35 |
| Qwen-2.5-72B | 0.63 | 0.98 | +0.35 |
| Mistral-7B | 0.71 | 0.89 | +0.18 |
| GPT-5.5 | 0.63 | 0.98 | +0.35 |
| Claude Opus 4.7 | 0.63 | 0.99 | +0.36 |
| Qwen 3.6 Flash | 0.63 | 0.97 | +0.34 |
| DeepSeek-V4 Flash | 0.62 | 0.95 | +0.33 |
---
## Key Insights
> **Coherence JSS varies by more than 0.6 units across 13 judges and does not track model scale or recency.**
- Claude Opus 4.7 (0.70) scores lower than Claude Haiku 4.5 (0.73); GPT-5.5 (0.83) scores lower than GPT-4o (0.91)
- Factuality JSS ranges from 0.89 to 0.99 after T4 correction; residual variation reflects genuine model-level differences
- Preference and relevance JSS are degenerate (12 of 13 judges always select option A)
---
## Citation
If you use JudgeSense in your research, please cite the accompanying paper (details redacted for double-blind review).
---
## License
- **Dataset**: [CC-BY-4.0](https://creativecommons.org/licenses/by/4.0/)
- **Code**: MIT License
---
*Anonymous submission for double-blind review. All evaluations conducted on public benchmarks and APIs.*
|