Jason Thomo, Venkatesh Srinivasan, Alex Thomo · ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 2026
DOI: 10.1145/3799682.3839910
Counts differ because each database indexes a different set of publications. We treat OpenAlex as the canonical count; Google Scholar is not shown (no API, and crawling it violates its ToS).
On multi-hop benchmarks, modern Graph-RAG retrieves the gold answer for 77-90% of questions, yet a budget open-weight 8B model answers only 31-77% correctly. We show that faithfulness (reading the retrieved evidence rather than guessing) and calibration (abstaining when it is insufficient) account for a large, recoverable part of this gap. A leak-controlled decomposition finds that on the hardest, least-contaminated benchmark (MuSiQue), 43% of questions have their supporting facts in the retrieved context yet are answered wrong, and on these the model gives a confident answer 82% of the time rather than abstaining. A grounding-and-calibration prompt makes the model answer from the retrieved context or abstain when it cannot, a more trustworthy operating point at no accuracy cost. A re-ask step then spends some of that calibration for coverage, turning abstentions into correct answers for a gain of +9 pp on KET-RAG and +13 pp on LightRAG (MuSiQue, answer-present slice) that closes about half the gap to a 70B model on KET-RAG and most of it on LightRAG. Few-shot example source moves the model along the same faithfulness-accuracy frontier, better for calibration than accuracy. Unlike structure-exploiting augmentations that degrade on uniform-relevance retrieval, grounding helps most there. For small-model Graph-RAG the target is faithful, calibrated use of retrieved evidence. Code is available at https://github.com/thomouvic/graphrag-faithfulness.
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
ACM International Conference on Information and Knowledge Management (CIKM) 2026 · 0 citations
No comments yet — start the discussion below.