LLMJRE-RAG-Eval

Retrieval-Augmented LLM Reviewers for Academic Peer Review: Improving Human Alignment and Rebuttal-Aware Evaluation

LLMJRE-RAG-Eval is a benchmark dataset for evaluating heterogeneous Large Language Model (LLM) reviewers across the complete academic peer-review workflow. The benchmark supports research on LLM-as-a-Judge for academic paper assessment by providing structured datasets for reviewer evaluation, author rebuttals, meta-review generation, and retrieval-augmented reviewer guidance.

The benchmark extends the Re² peer-review dataset by introducing a Retrieval-Augmented Generation (RAG) evaluation framework that incorporates conference-specific reviewer guidelines into the review process. It enables systematic evaluation of whether external reviewer guidance improves agreement between LLM-generated evaluations and human reviewer assessments.

---

Dataset Summary

LLMJRE-RAG-Eval consists of two benchmark tasks corresponding to the complete conference peer-review lifecycle.

1. Review Benchmark

   • Initial manuscript review
   • Human review alignment
   • Overall recommendation

2. Rebuttal Benchmark

   • Author rebuttal
   • Rebuttal-aware review
   • Reviewer belief revision
   • Final recommendation

The benchmark supports four research questions:

• RQ1: To what extent do heterogeneous LLM reviewers align with human reviewer evaluations of academic papers?
• RQ2: How do heterogeneous LLM reviewers differ in their evaluation behaviour and response to rebuttal information?
• RQ3: Does retrieval-augmented review guidance improve the alignment between LLM-generated evaluations and human reviewer scores?
• RQ4: Does retrieval-augmented review guidance improve belief-shift accuracy after author rebuttals?

---

Dataset Structure

llmjre-rag-eval
│
├── review
│   ├── llmjre_review.jsonl
│   └── llmjre_review.csv
│
├── rebuttal
│   ├── llmjre_rebuttal.jsonl
│   └── llmjre_rebuttal.csv
│
├── sample
│   ├── llmjre_review_sample_1000.jsonl
│   ├── llmjre_review_sample_1000.csv
│   ├── llmjre_rebuttal_sample_1000.jsonl
│   ├── llmjre_rebuttal_sample_1000.csv
│   ├── sample_inference_report.json
│   └── unique_conference_year_type.csv
│
├── metadata
│   ├── benchmark_schema.json
│   ├── benchmark_statistics.json
│   └── unique_conferences.csv
│
└── rag
    ├── guideline_collection_tracker.csv
    └── official_guideline_tracker.csv

---

Dataset Components

Review Benchmark

The review benchmark contains the information required to evaluate initial manuscript assessment and review alignment.

Typical fields include:

• Paper metadata
• Manuscript text
• Human review comments
• Human review scores
• Human recommendations
• Conference metadata

---

Rebuttal Benchmark

The rebuttal benchmark extends the review benchmark by incorporating author rebuttals and revised reviewer assessments.

Additional fields include:

• Author rebuttal
• Final reviewer comments
• Final reviewer scores
• Reviewer belief shifts
• Final recommendations

---

Sample Benchmark

The sample benchmark contains the exact 1,000-paper evaluation subset used in the accompanying paper.

Researchers may use this subset to reproduce the published experiments and statistical analyses.

---

Metadata

Supporting metadata includes:

• Benchmark schema
• Benchmark statistics
• Conference metadata
• Conference distributions

---

RAG Metadata

The RAG directory contains metadata describing the conference-specific reviewer guideline collection used for retrieval augmentation.

It includes:

• guideline collection tracker
• official reviewer guideline tracker

These files document the provenance and coverage of the conference reviewer guidelines used during retrieval.

---

Benchmark Construction

The benchmark is derived from the Re² academic peer-review dataset and preserves the complete conference review workflow, including:

• Manuscripts
• Human reviewer assessments
• Author rebuttals
• Final reviewer decisions

LLMJRE-RAG-Eval extends the benchmark by incorporating conference-specific reviewer guidelines to support retrieval-augmented reviewer evaluation.

---

Recommended Tasks

The benchmark supports research in:

• LLM-as-a-Judge
• Academic peer review
• Retrieval-Augmented Generation (RAG)
• Human-AI collaboration
• Meta-review generation
• Rebuttal-aware evaluation
• Reviewer behaviour analysis
• AI-assisted scholarly communication

---

Loading the Dataset

The benchmark can be loaded directly from the JSONL or CSV files.

Example (JSONL):

import json

with open("review/llmjre_review.jsonl") as f:
    for line in f:
        sample = json.loads(line)
        print(sample["paper_id"])

Example (Pandas):

import pandas as pd

df = pd.read_csv("review/llmjre_review.csv")
print(df.head())

---

Citation

The benchmark builds upon the Re² dataset. Please also cite the original Re² paper.

@article{zhang2025re,
  title={Re$^2$: A Consistency-ensured Dataset for Full-stage Peer Review and Multi-turn Rebuttal Discussions},
  author={Zhang, Daoze and Bao, Zhijian and Du, Sihang and Zhao, Zhiyi and Zhang, Kuangling and Bao, Dezheng and Yang, Yang},
  journal={arXiv preprint arXiv:2505.07920},
  volume={abs/2505.07920},
  pages={1--15},
  year={2025}
}

---

Acknowledgements

LLMJRE-RAG-Eval extends the Re² benchmark by introducing retrieval-augmented conference-specific reviewer guidance for evaluating human alignment and rebuttal-aware assessment.

We thank the authors of the Re² dataset for making these research resources publicly available.

---

License

The benchmark is released under the Apache License 2.0, consistent with the original Re² dataset. Users should additionally comply with the licensing terms of the original Re² dataset and any applicable terms associated with the referenced conference reviewer guideline sources.