Skip to main content
Back

KAIST's Chain of Retrieval Beats Standard Paper Search by 7.82%

A new framework retrieves full scientific papers by decomposing them into aspect-specific queries and iteratively expanding the search as a tree.

  • Chain of Retrieval (CoR) decomposes query papers into motivation, method, and experiment queries for retrieval.
  • Iterative tree expansion promotes top-ranked papers as new queries, aggregated bottom-up via Reciprocal Rank Fusion.
  • Beats abstract-to-abstract retrieval by 6.37% Recall and full-to-full by 7.82% on average.
  • New SciFullBench benchmark provides full-paper queries and candidates from ICLR, NeurIPS, ACL, EMNLP.
  • Generalizes to patent-to-patent retrieval with off-the-shelf GPT-4o, no training needed.
  • Code and dataset on GitHub; latency remains the main cost.

Finding papers related to a complete manuscript presents a different retrieval problem from keyword search. The query may span thousands of tokens across motivation, methods, experiments, and results, while relevance may depend on a single facet. KAIST researchers propose Chain of Retrieval (CoR), a framework that searches full-paper content through aspect-specific queries, recursive expansion, and bottom-up rank fusion. On the new SciFullBench benchmark, CoR reported higher recall than abstract-only and full-document baselines, with a substantial latency cost.

One vector hides the match

Many scientific retrievers represent each paper with its title and abstract, then fine-tune an embedding model on citation-linked pairs. That representation captures broad topical similarity, yet abstracts often omit the implementation details, datasets, baselines, and evaluation choices needed to find methodological or experimental neighbors.

Long-context encoders can accept more of the manuscript, although context limits may still exclude parts of long documents. A single output embedding also compresses motivation, methods, and experiments into one vector. Signals tied to a particular dataset or technique can become difficult to distinguish from the paper’s broader topic.

Comparison of abstract retrieval, full-paper retrieval, and Chain of Retrieval
CoR preserves separate views of a paper and matches each view against relevant sections of candidate documents.

CoR’s Exploration phase derives three focused queries from the manuscript and retains the abstract as a broader fourth view:

  • Research motivation: the problem, rationale, and research objective.