RAG systems cite badly — and the fixes still leave citations unreliable

Citations that don't check out

L-MARS, a preprint, takes on the citation-faithfulness gap in legal AI: systems that generate answers with citations that do not exist or do not support the claim. Its multi-agent design adds a judge-driven verification loop and a post-draft Faith-Search step that checks citations after the answer is written. Multi-turn judge loops improved citation F1 from 0.13 to 0.25, and Faith-Search cut unreachable citations to below 1%. m:…4595

The improvement is real and the ceiling is low. A citation F1 of 0.25 is a long way from a perfect score, and single-agent web-search baselines collapsed to 0.22 under external audit. Multi-turn judge loops improved citation F1 from 0.13 to 0.25 in legal question answering — a real gain from a low base. Even the best fix in the paper leaves citations unreliable. m:…4595

The stakes are concrete. L-MARS was built for legal question answering, where a hallucinated citation is not a cosmetic error but a liability: an answer that cites a case or statute that does not exist can mislead a lawyer or a compliance team. The multi-agent design is meant to close that gap with a judge that checks evidence and a post-draft step that repairs citations — and even with both, the best citation F1 the paper reports is 0.25. m:…4595

Sources that disagree with each other

A second problem sits underneath the citations: the sources themselves disagree. Kontrast, another preprint, audits knowledge across text, tables and knowledge graphs using automated Text-to-SPARQL queries and LLM reasoning. It found cross-modal inconsistencies are common — the same fact stated differently in a table and a graph — and that they signal missing structure in the knowledge graph and temporal drift between sources. The framework's own effectiveness is limited by Text-to-SPARQL errors and noise. m:…4586

When retrieval comes back incomplete, models fill the gap from memory. FondueQA, a benchmark for multi-hop question answering, localises each required fact to a specific evidence document so retrieval can be removed piece by piece. It found that parametric compensation — the model answering from its own weights — depends on how popular the entity is, and that bridging evidence matters more than documents containing the answer itself. Put plausible competing entities into the partial context and Exact Match falls by up to 15 percentage points. m:…4593

Faster retrieval, same trust problem

The other two papers in this set make retrieval faster and cheaper, not more trustworthy. PULSAR replaces OCR pipelines for visually dense enterprise documents with a vision-first, late-interaction retrieval system and a pooled two-stage index. It cut median vector-search latency 15.1 times against an unpooled setup, sustained about 88 times higher query throughput under concurrent load, and made ingestion about 20 times cheaper per page than the OCR baseline — while more than doubling answer-fact recall. m:…4598

CacheGraft takes a similar efficiency route: it pre-encodes passages into fused key-value caches from a 4-billion-parameter model and injects them into a 0.6-billion-parameter receiver, so a small model can do retrieval work that previously needed a large one. It improved question-answering performance over text-only baselines in 29 out of 30 comparisons, with a fuser that is zero-shot on retrieval data. Neither paper reports citation-faithfulness results. Speed, cost and recall are all better; whether the answer's citations can be trusted is still the open question. m:…5108