Span-Anchored Citations for Verifiable Research Writing
Citations that point at papers are unverifiable in practice: the reader cannot find the sentence that supports the claim. We anchor every citation to the exact quoted source span (W3C TextQuoteSelector), verify that the span supports the claim, and emit self-verifying LaTeX citations. The anchor is the quoted text itself, so citations survive re-indexing and new document versions.
claims supported
experiments · runs
span citations
referenced papers
Claims & evidence
Status is derived from linked evidence, never asserted.
Anchoring citations at the span level makes them self-verifying: the quoted text either exists in the source or it does not.
- ⊕ supports003-anchor-robustnesslocalanchor_survival_rate 0.97confirmatory
- ⊕ supportsZhang 2024 · arXivfull · 0.90
“fine-grained sentence-level citations allow users to locate the exact supporting evidence”
Sentence-granularity passages are easier to verify than paragraph chunks without losing retrieval recall.
- ⊕ supports001-tfidflocalrecall_at_8 0.81span_support_rate 0.74confirmatory
- ⊕ supportsGao 2023 · EMNLPfull · 0.92
“smaller retrieval units are easier for humans to verify against the generated statement”
- ⊕ supportsLei 2024 · arXivpartial · 0.63
“verification accuracy degrades as the evidence context grows”
Span anchors survive re-indexing and new document versions because they never reference chunk or vector ids.
- ⊕ supports003-anchor-robustnesslocalanchor_survival_rate 0.97confirmatory
Dense retrieval is required for competitive recall on claim-to-source matching.
- ⊗ refutes002-denselocalrecall_at_8 0.84span_support_rate 0.76suggestiveDense beats TF-IDF by only 3 points recall@8, so it is not strictly required.
Experiments & results
Every number traces to a recorded run with git sha, env hash, and seed.
| Experiment | Hypothesis | Run | Metrics |
|---|---|---|---|
001-tfidf TF-IDF baseline | A stdlib TF-IDF retriever reaches usable recall@8 on claim→span matching. | b31c2ffseed 7local | recall_at_8 0.81 (test) span_support_rate 0.74 (test) |
002-dense Dense embeddings← 001-tfidf | Dense retrieval (BGE-M3) beats TF-IDF by a wide margin. | d02e881seed 7local | recall_at_8 0.84 (test) span_support_rate 0.76 (test) |
003-anchor-robustness Anchor robustness across versions← 001-tfidf | Quote-text anchors re-resolve after re-parsing with a different PDF extractor. | e4f1a9cseed 7local | anchor_survival_rate 0.97 (test) |
Papers & span citations
Citations anchor to the exact quoted span, so every quote can be checked against its source.
…We observe that smaller retrieval units are easier for humans to verify against the generated statement, motivating sentence-level attribution.…
chars 14210–14298 · support: full · 0.92
…In contrast to passage-level attribution, fine-grained sentence-level citations allow users to locate the exact supporting evidence within long documents.…
chars 3120–3211 · support: full · 0.90
verification accuracy degrades as the evidence context grows
support: partial · 0.63
Code references
Deep links to the exact lines, pinned to the published commit.