Your retriever found the evidence. Your answer is still wrong

Recall is not correctness

A preprint on long-term conversational memory introduces EviCover, a method that improves support-turn recall in long-term conversational QA by jointly selecting turns based on evidence requirements. That is a real improvement in finding the right evidence. Then comes the finding that matters more: increased evidence recall in conversational QA does not currently establish a correctness gain in answer generation. The evidence gets found more reliably, yet the answers did not reliably get better, which is the gap the rest of this story is about. m:…8853

CoRA-Bench, a benchmark for context selection under shared token budgets, explains part of why. Context selection methods exhibit task-dependent tradeoffs between evidence preservation and downstream answer accuracy, and method rankings for evidence preservation are inconsistent across tasks. Evidence recall does not reliably predict downstream answer accuracy, the authors say. The disconnect is measurable and structural: a system can be very good at retrieving the supporting facts and still produce a wrong answer, because using evidence well is a different capability from finding it. m:…8860

No method wins everywhere

The CoRA-Bench numbers show how task-dependent the rankings are. BM25 achieves 0.921 supporting-fact recall on BABILong at 128K tokens, while dense retrieval and dense reranking achieve 0.908 needle recall on NoLiMa at 64K tokens. Each method leads somewhere and trails elsewhere, which means a team that picks one retrieval method for every workload is leaving performance on the table in at least one place. The benchmark tests methods under shared token budgets, so the comparison is about the methods, not the budgets. m:…8860

The RAG versus GraphRAG comparison reaches the same conclusion from a different direction. The authors introduce a unified evaluation protocol and find that RAG and GraphRAG exhibit distinct strengths across different tasks, with neither paradigm universally superior, and that existing GraphRAG systems for text data lack standardized evaluation protocols, producing heterogeneous and non-reproducible results. Integration strategies that combine RAG and GraphRAG lead to consistent performance improvements over individual systems. The practical advice is to stop arguing about which paradigm wins and start combining them, with the combination tuned to the task. m:…8839

Cheaper evidence, kept at token level

Another preprint attacks the problem from below, at the representation level. The authors show that sparse token matching improves retrieval performance without additional training, and that the retrieval gains translate into higher question-answering accuracy on long-term memory benchmarks. Query adaptation with a low-rank adapter further increases the retrieval gains, the authors say. The point is that the evidence a system retrieves can be improved without retraining the retriever at all, which makes the finding cheap to adopt and hard to ignore. m:…8862

Lightfoot, a GraphRAG architecture, attacks cost and maintenance instead. It treats graph components as deterministic functions of passage extractions, which lowers indexing costs by up to 6x compared with existing solutions like LightRAG, and enables document deletion as a database operation without model calls. Its indexing cost is 26 to 31% lower than HippoRAG 2, and it outperforms HippoRAG 2 on MuSiQue retrieval by 2.8 recall at 5 and HotpotQA exact match by 7.7 points. It underperforms HippoRAG 2 on 2WikiMultiHopQA, by 3.9 recall at 2 and 8.4 F1, which is a reminder that even a cheaper system has tasks where it loses. m:…8836

What to measure instead

The thread running through all six preprints is that retrieval evaluation has been measuring the wrong thing, or at least only half of it. Evidence preservation under a token budget, support-turn recall, needle recall and answer-bearing-turn recall are all useful, but none of them reliably predicts whether the final answer is correct. A team that optimizes recall alone can ship a system that finds more evidence and answers no better, or even worse, because the generator now has more material to misuse. m:…8860 m:…8853

The better evaluation, these papers suggest, is end-to-end and task-specific: measure answer correctness on the team's own tasks, compare methods under the same token budget, and combine retrieval paradigms rather than choosing one. TRIO, a multi-agent RAG framework from a separate preprint, coordinates three specialized agents for query, judging and answering through iterative refinement, and achieved 76.6% accuracy on the unseen PubMedQA dataset with compact 4B and 8B models. Token-level matching and cheaper indexing can lower the cost of the retrieval layer, and hybrid RAG plus GraphRAG can raise its ceiling, but neither fixes a generator that cannot use the evidence it is given. That last gap, between finding and using, is where the next round of work has to happen. m:…8839 m:…8862 m:…3132