Current retrieval-augmented citation systems still produce misattributions by selecting semantically similar but factually incorrect papers despite avoiding hallucinated references.
ReCite addresses this by transforming citation from a linear retrieval task into an iterative agentic reasoning process. It eliminates fabrication by querying authentic databases and resolves misattribution through three core mechanisms: intent-aware query planning to align searches with citation context, claim-evidence verification to cross-check metadata against inferred intent, and a reflective re-retrieval loop that self-corrects when initial candidates fail logical support checks.
Trained via supervised fine-tuning and reinforcement learning, ReCite shifts reliance from semantic similarity to verifiable logical entailment. This approach allows the agent to detect when a semantically similar paper does not logically support a specific claim, triggering a rewrite and retry rather than accepting a superficial match.
Experiments demonstrate that ReCite significantly outperforms larger generative models in strict citation accuracy. By decoupling citation hallucination into fabrication (largely resolved by RAG) and misattribution (the core remaining challenge), it provides a faithful, accountable foundation for AI-assisted academic writing.
ReCite: Agentic Reasoning for Faithful Citation targets a persistent weakness in retrieval-augmented citation systems: the ability to return real, non-hallucinated papers while still misattributing facts to them. The paper frames citation as a reasoning problem rather than a pure retrieval or ranking problem. In many RAG settings, a model can retrieve semantically nearby works—papers with similar terminology, topics, or surface-level claims—yet those works may not actually support the specific proposition being asserted. ReCite’s core insight is that faithful citation requires explicit, iterative verification that a candidate source substantiates the exact claim being made, not merely that it is topically related.
The proposed approach uses an agentic workflow to couple claim decomposition, retrieval, evidence inspection, and source selection. Rather than selecting a citation in a single pass from retrieved candidates, the agent reasons over candidate passages, checks whether they entail or directly support the target statement, and can continue searching when evidence is weak or ambiguous. This shifts the citation objective from “find the most relevant-looking paper” to “find a paper that can be confidently cited for this claim.” The paper’s contribution is therefore both methodological and conceptual: it treats citation faithfulness as a proposition-level grounding task and uses agentic reasoning to enforce stricter evidence-to-claim alignment.
This matters because citation errors are especially damaging in scientific, legal, medical, and technical assistance settings, where a plausible but incorrect attribution can propagate false authority. Even when a system avoids fabricating references, misattribution can be just as misleading, because the user may infer that a real paper endorses a claim it does not actually make. By emphasizing agentic verification, ReCite points toward citation systems that are not only fluent and retrieval-grounded, but also more auditable and trustworthy—systems that can distinguish semantic similarity from factual support and abstain or refine when the evidence is insufficient.