CiteAnchor traces hallucinated citations back to their source papers, hits 74.7% end-to-end success
yuz9yuz · x · 2026-10-24
- Problem: hallucinated citations are hard to verify — bibliographic records are incomplete and surface matching proves nothing about the underlying source.
- CiteResolve benchmark: 423 real-world hallucinated citations annotated for source recoverability.
- CiteAnchor method: source-aware grounding that adapts retrieval and field-level verification per source type, separately deciding citation validity and whether a unique source can be recovered — or clearly stating none exists.
- Results: 0.841 F1 with GPT-5.6 Luna and 0.784 with an open-weight backbone (beating same-backbone baselines by 7.9 and 36.8 points); 74.7%/70.7% end-to-end success on CiteResolve, +16.3/+65.3 points over web search baselines.
- Paper, code, and dataset all released.
More from Research
- The real AI-for-Science question: when experiments become programmable — CatAstro_Piyush · 2026-10-24
- Google publishes Nature research behind its Personal Health Agent powering Health Coach — pswider · 2026-10-24
- MEG study: LSD speeds up brain rhythms and makes activity more fractal — karimjerbineuro · 2026-10-24
- Glyphh Ada Cuts LLM Routing Cost From $300 to $0.22 With Precedent-Based Classifier — Successful-Seesaw525 · 2026-10-24
- Zhejiang's TARSphere Benchmarks Long-Horizon Embodied Agents Across 360 Episodes, 30 Scenes — Yicheng Xu · 2026-10-24
- LLM-Attacker: Training Gradients Push Split-LLM Document Recovery to 37.8% — Setloop · 2026-10-24