Chain of Retrieval: Multi-Aspect Iterative Search Expansion and Post-Order Search Aggregation for Full Paper Retrieval

TL;DR

This work presents Chain of Retrieval, an iterative, multi-view framework for full-paper retrieval that decomposes a query paper into aspect-specific views (Motivation, Methods, Experiments, plus Abstract), retrieves candidates via both abstract and chunked full-text corpora, and expands the search into a tree of queries. Results are fused with a bottom-up post-order aggregation using Reciprocal Rank Fusion, which attenuates weak, depth-related signals while reinforcing consistently supported papers. The authors introduce SciFullBench, a large-scale benchmark with complete papers for queries and candidates to accurately evaluate full-context retrieval, and demonstrate that CoR consistently outperforms abstract-based and naive full-text baselines across ML/NLP domains and in patent retrieval, showing robustness with various backbones and LLM-based optimizers. The approach highlights the importance of multi-aspect coverage, an aspect-aware caching mechanism, and offline policy optimization via DPO to improve query quality, offering a scalable, domain-agnostic solution for complex literature search and synthesis tasks.

Abstract

Scientific paper retrieval, particularly framed as document-to-document retrieval, aims to identify relevant papers in response to a long-form query paper, rather than a short query string. Previous approaches to this task have focused exclusively on abstracts, embedding them into dense vectors as surrogates for full documents and calculating similarity between them. Yet, abstracts offer only sparse and high-level summaries, and such methods primarily optimize one-to-one similarity, overlooking the dynamic relations that emerge among relevant papers during the retrieval process. To address this, we propose Chain of Retrieval(COR), a novel iterative framework for full-paper retrieval. Specifically, CoR decomposes each query paper into multiple aspect-specific views, matches them against segmented candidate papers, and iteratively expands the search by promoting top-ranked results as new queries, thereby forming a tree-structured retrieval process. The resulting retrieval tree is then aggregated in a post-order manner: descendants are first combined at the query level, then recursively merged with their parent nodes, to capture hierarchical relations across iterations. To validate this, we present SCIFULLBENCH, a large-scale benchmark providing both complete and segmented contexts of full papers for queries and candidates, and results show that CoR significantly outperforms existing retrieval baselines. Our code and dataset is available at https://github.com/psw0021/Chain-of-Retrieval.git.

Figures (9)

• Figure 1: Conceptual illustration of our retrieval method compared to prior retrieval approaches.
• Figure 2: Retrieval performance comparing different query formulations against our multi-agent CoR framework (after a single round(depth) of retrieval and with untrained query optimizers).
• Figure 3: Change ($\Delta$) in retrieval performance (relative to using all aspects) when excluding individual scientific aspects (Moti = Motivation, Meth = Method, Expr = Experiment).
• Figure 4: Retrieval results as a function of retrieval depths, with DPO-trained Llama-3.2-3B-Instruct as optimizers.
• Figure 5: PatentFullBench composition by published year.