ReCite: Agentic Reasoning for Faithful Citation
ReCite is a novel agentic AI framework designed to improve the accuracy of automatic citation in academic writing. It moves beyond simple semantic similarity to verify if a cited paper truly supports a specific claim, addressing the problem of misattribution in current systems. This ensures that automatically generated citations are logically sound and faithful to the author's intent.
Reading in Plain English mode — toggle below to switch
ReCite is a novel agentic AI framework designed to improve the accuracy of automatic citation in academic writing. It moves beyond simple semantic similarity to verify if a cited paper truly supports a specific claim, addressing the problem of misattribution in current systems. This ensures that automatically generated citations are logically sound and faithful to the author's intent.
Imagine a detective meticulously cross-referencing every piece of evidence to ensure it directly supports a specific accusation, rather than just finding similar-looking clues. ReCite acts like this detective, actively reasoning about the logical connection between a claim and a potential citation, instead of merely finding semantically similar papers.
How ReCite Ensures Accurate Citations
- 11. Location Perception: The agent identifies the specific claim in the text that requires a citation.
- 22. Intent-Aware Query Planning: It then formulates a precise search query based on the logical intent of that claim, not just keywords.
- 33. Reflective Verification: After retrieving potential papers, ReCite critically evaluates if the content of each paper logically supports the claim.
- 44. Self-Correction Loops: If a retrieved paper doesn't provide logical support, the agent refines its search or reasoning process until a faithful citation is found.
C_f = V(Claim, Evidence)
$C_f$ represents a faithful citation. $V$ is the verification function that assesses the logical consistency between a given $Claim$ and the $Evidence$ provided by a potential citation.
If an author claims 'Transformer models excel at long-range dependencies,' ReCite wouldn't just find papers with 'Transformer' and 'long-range' keywords. It would verify if the paper actually demonstrates this capability through experiments or theoretical analysis, ensuring the citation truly substantiates the claim.
Accurate citation requires active, claim-level reasoning and verification, moving beyond mere semantic similarity to ensure logical support.