SurveyReview
A Reviewer-Aligned Benchmark for Survey Evaluators
SurveyReview is a reviewer-aligned benchmark for evaluating survey papers. It converts real peer-review reports into multidimensional evaluation labels and rationales, allowing models to be tested against how human reviewers judge survey quality.
This directory is the `v1.1` release. It keeps the original SurveyReview evaluation metrics while using cleaned article texts and XML-like rationale prompts.
The benchmark focuses on four survey-review dimensions: **Readability**, **Criticalness**, **Comprehensiveness**, and **Structure**. It provides standardized train/test splits, article metadata, prompt files, and an API-based evaluation pipeline.
## Quick Start
Create an environment and install dependencies:
```bash
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt
```
Configure the API client:
```bash
cp .env.example .env
```
Then edit `.env`:
```text
API_KEY=your-api-key-here
BASE_URL=https://api.openai.com/v1
MODEL_NAME=gpt-5.2
JUDGE_MODEL=gpt-5.2
EVALUATE_REASONS=True
```
Run the default test-set evaluation:
```bash
python src/api_base_evaluate.py
```
The v1.1 evaluator defaults to:
```text
data/v1.1-paper
```
Run on the train split:
```bash
EVAL_SPLIT=train python src/api_base_evaluate.py
```
Outputs are written to `result//`:
| File | Description |
| --- | --- |
| `results.csv` | MSE, MAE, accuracy, and sample counts for each dimension. |
| `predictions_.jsonl` | Per-sample prediction records. |
| `run_config.json` | Runtime configuration and split statistics. |
| `rqs_.json` | Rationale quality results when `EVALUATE_REASONS=True`. |
## Leaderboard
Lower MSE/MAE is better. Higher HAS/RQS is better.
| Rank |
Model |
HAS |
Read. |
Crit. |
Comp. |
Stru. |
Average |
RQS |
| MSE | MAE |
MSE | MAE |
MSE | MAE |
MSE | MAE |
MSE | MAE |
| 1 |
SurveyReviewer |
0.74 |
1.43 | 0.72 |
1.52 | 0.82 |
1.26 | 0.56 |
1.29 | 0.65 |
1.38 | 0.69 |
0.36 |
| 2 |
GPT-5.2 |
0.68 |
2.13 | 1.07 |
1.97 | 0.97 |
2.04 | 1.08 |
2.98 | 1.47 |
2.28 | 1.15 |
0.42 |
| 3 |
Claude-Opus-4.5 |
0.68 |
2.91 | 1.29 |
1.88 | 0.88 |
2.66 | 1.23 |
3.65 | 1.58 |
2.77 | 1.25 |
0.48 |
| 4 |
Qwen3-32B |
0.61 |
3.05 | 1.45 |
3.24 | 1.51 |
3.22 | 1.54 |
3.35 | 1.53 |
3.21 | 1.51 |
0.36 |
| 5 |
GLM-4.7 |
0.60 |
3.43 | 1.50 |
2.58 | 1.21 |
3.66 | 1.57 |
4.83 | 1.95 |
3.62 | 1.56 |
0.37 |
| 6 |
gemini-3-pro |
0.58 |
3.84 | 1.52 |
2.25 | 1.00 |
3.91 | 1.49 |
5.76 | 2.11 |
3.94 | 1.53 |
0.29 |
| 7 |
DeepSeek-v3.2 |
0.58 |
4.78 | 1.88 |
2.49 | 1.15 |
4.59 | 1.82 |
4.02 | 1.76 |
3.97 | 1.65 |
0.37 |
## Notes
- `articles/` is split into multiple JSON shards to stay within GitHub file-size limits.
- `v1.1-paper` uses cleaned article texts and prompts that output `... X`.
- If you only want to verify the pipeline, set `EVALUATE_REASONS=False` to skip the judge-model stage.