---
license: mit
task_categories:
- text-generation
- text-retrieval
language:
- en
tags:
- related-work-generation
- scholarly-positioning
- citation-evaluation
- retrieval-augmented-generation
- code
pretty_name: RWGBench
size_categories:
- 10K<n<100K
---

# RWGBench

RWGBench is a benchmark for evaluating related work generation as a
citation-centric scholarly positioning task. It tests whether a system can
select, organize, and frame prior work for a target paper, rather than only
producing fluent text that resembles a reference related work section.

It is presented in the paper
[RWGBench: Evaluating Scholarly Positioning in Related Work Generation](https://huggingface.co/papers/2606.24894).

The official repository containing evaluation scripts and baselines is
available on GitHub at [BFTree/RWGBench](https://github.com/BFTree/RWGBench).

## Files

| File | Entries | Description |
|---|---:|---|
| `papers.json` | 40,108 | Source paper collection with parsed metadata, sections, related work text, and citation identifiers. |
| `corpus.json` | 1,091,394 | Retrieval corpus. Each entry contains `doc_id`, `title`, and `abstract`. |
| `gold100_papers.json` | 100 | Peer-reviewed evaluation split used for the main experiments. The papers are matched to accepted ICLR, NeurIPS, or ICML records. |

## Evaluation Split Composition

The `gold100_papers.json` split contains 100 peer-reviewed papers matched by
exact normalized title to accepted OpenReview records. Its venue distribution
is:

| Venue | Papers |
|---|---:|
| ICLR | 42 |
| NeurIPS | 34 |
| ICML | 24 |

The publication-year distribution is:

| Year | Papers |
|---|---:|
| 2023 | 2 |
| 2024 | 61 |
| 2025 | 37 |

## Schema

### `corpus.json`

```json
{
  "doc_id": 1589,
  "title": "LoRA: Low-Rank Adaptation of Large Language Models",
  "abstract": "We propose a low-rank adaptation method that freezes pretrained model weights..."
}
```

### `gold100_papers.json`

```json
{
  "paper_id": 9745,
  "title": "EfficientDM: Efficient Quantization-Aware Fine-Tuning of Low-Bit Diffusion Models",
  "abstract": "...",
  "introduction": "...",
  "related_work": "Model quantization. Quantization is a widely employed technique...",
  "citations": [191955, 118706, 517176, 264652, 1589],
  "peer_review": {
    "match_type": "normalized_title_exact",
    "primary_venue": {
      "venue": "ICLR",
      "year": "2024",
      "venue_id": "ICLR.cc/2024/Conference",
      "openreview_id": "UmMa3UNDAz"
    }
  }
}
```

`citations` contains `doc_id` values from `corpus.json`. The reference related
work section is author-written text from the target paper. Venue metadata is
stored under `peer_review.primary_venue`.

## Use With The Code Repository

Download the dataset files into the GitHub repository's `data/` directory:

```text
data/
  papers.json
  corpus.json
  gold100_papers.json
```

Then run the generation and evaluation scripts from the code repository.

### Sample Usage

For integration into custom pipelines using the evaluation code from the
GitHub repository:

```python
from src.evaluation.single_paper_evaluator import SinglePaperEvaluator

evaluator = SinglePaperEvaluator(
    gold_papers_path="data/gold100_papers.json",
    corpus_path="data/corpus.json",
    use_llm_judge=False,       # set True to include LLM-as-Judge scores
    llm_model="deepseek-v3",   # any OpenAI-compatible model name
)

result = evaluator.evaluate(
    paper_id=...,         # int
    generated_text=...,   # str, text with [1], [2], ... citations
    citation_list=[...],  # list of doc_id ints or title strings; index i maps to citation [i+1]
)
```

## Data Collection And Use

RWGBench is built from public scholarly documents and metadata. Source
documents may have heterogeneous licenses, so users should follow the license
terms of the underlying papers when redistributing document-derived text.

The benchmark is intended for research on retrieval-augmented generation,
citation selection, scholarly writing evaluation, and related work generation.