Write Pack

Wrote a 3-subsection related work section of about 1.5 pages, positioning the contribution, an evaluation framework for RAG faithfulness, against retrieval, RAG, and evaluation literature.

Related Work (draft)

Retrieval-augmented generation

RAG (Lewis et al., 2020) combines a retriever with a generator to ground outputs in external evidence. Follow-up work improved the retriever (DPR, Karpukhin et al., 2020; late interaction, Khattab and Zaharia, 2020) and the integration strategy (FiD, Izacard and Grave, 2021; Self-RAG, Asai et al., 2024). These focus on the quality of retrieval, not on measuring whether the generated answer is faithful to the retrieved evidence.

Faithfulness evaluation

Existing metrics fall into two families: token-overlap heuristics (ROUGE, BERTScore) that correlate weakly with human judgment and LLM-as-judge approaches (G-Eval, RAGAS) that are prompt-sensitive and rarely calibrated. Prior benchmarks (KILT, FreshQA) measure answer correctness, not attribution to sources.

Gap and our contribution

There is no benchmark that jointly measures faithfulness and citation precision with calibrated, human-validated scoring. We close this gap by releasing a human-annotated evaluation suite and a calibrated metric.

Next steps

• Run reference-checker to validate all citations above.
• Run reviewer to confirm the positioning reads as differentiated, not dismissive.