ReGround: Grounding Reviewer Comments in Multimodal Evidence
Abstract Overview
This paper introduces ReGround, a dataset for grounding reviewer comments to specific evidence in original anonymous scientific paper submissions. The dataset links 10,267 reviewer comments to 16,274 evidence instances across 3,656 papers, spanning paragraphs, sections, figures, and tables. The authors extract these links automatically by exploiting explicit paper references in author rebuttals and casting grounding as a retrieval task over multimodal evidence. Benchmarking across sparse, dense, cross-encoder, LLM-based, and visual retrieval models reveals key failure modes in grounding underspecified reviewer comments.
Novelty
The primary novelty is a large-scale multimodal benchmark linking naturally occurring reviewer comments to fine-grained evidence within original anonymous submissions rather than revised manuscripts. The work also introduces an automated dataset construction pipeline that leverages explicit paper references in author rebuttals as high-precision supervision for review comment grounding.
Results
Unified text retrieval across heterogeneous paper content proved difficult, with the top-performing model (Qwen-3 30B Instruct) achieving only 21.15 Recall@10 and 10.87 MRR. Providing oracle evidence-type information substantially improves retrieval across all models, demonstrating that evidence-type inference is a major performance bottleneck. For figures and tables, multimodal image retrieval augmented with captions outperforms image-only retrieval, and text and visual retrievers succeed on complementary instances.
Key Points
- ReGround contains 10,267 reviewer comments linked to 16,274 evidence units across 3,656 papers, covering paragraphs, sections, figures, and tables in original anonymous submissions.
- The dataset is constructed automatically by detecting explicit paper references in author rebuttals, aligning rebuttal sentences to reviewer comments, and resolving references in the original PDFs.
- Retrieval experiments show that unified whole-paper retrieval performs poorly, evidence-type routing is a critical bottleneck, and visual signals provide complementary information that text alone misses.