LivingRAG, a Graph RAG system that retrieves connected information before generating an answer, had the best reported accuracy in the study's comparisons across four multi-hop benchmarks and a WixQA support-question comparison. No confidence intervals were reported.
Compared with LinearRAG, the main matched baseline, the weighted results show a trade-off: prompt-token use rose 3.5%, while completion-token use fell 22.7%. Estimated run cost fell from $51.05 to $44.87, a 12.1% reduction under the paper's stated Qwen pricing.
How the system reuses earlier work
The study asks whether a Graph RAG system can write and reuse reasoning experience for later online question answering. LivingRAG reuses verified activation maps, which help steer retrieval, and compact reasoning scaffolds, which help structure generation, rather than caching previous answers as response text.
For each dataset, the experience store starts empty and is updated online, so only earlier queries can supply experience to later ones. The system looks for reusable signals in shared entities, graph neighborhoods and question templates.
The test used five question streams
The evaluation used five online QA streams: 2WikiMultiHopQA, HotpotQA, MuSiQue, MuSiQue-full and WixQA. The first four were multi-hop benchmark streams, while WixQA used customer-support questions over a shared enterprise help corpus.
All compared methods used Qwen3.6 Plus for answer generation. Results were reported with Contain-Match Accuracy and LLM-Evaluation Accuracy, an automated comparison by a large language model against a gold answer; WixQA used only the latter.
Reuse was broad, storage selective
Stored experience was reused broadly in both parts of the system. On 2WikiMultiHopQA, activation-map retrieval coverage was 97.00% and scaffold coverage was 85.00%; on HotpotQA, the figures were 79.60% and 58.11%. Coverage is a descriptive reuse measure, not an accuracy endpoint.
Write-back, the step of adding a candidate experience to the store, was selective: the final write-back ratio was 27.4%. A grounding gate screened candidates before storage; across all streams it checked 1,806 candidates, passed 88.15%, rejected 214 and stored 1,592 experiences.
Extra checks kept the pattern intact
In the ablation, removing the quality gate was accompanied by lower accuracy on every reported dataset. The comparison was limited to the streams tested in the paper.
Changing the order of questions did not reverse the comparison: LivingRAG had higher accuracy than LinearRAG on every reported metric, used fewer completion tokens in every tested order and retained a stable reuse pattern after reordering.
Retrieval accounted for about 1.7% to 5.1% of measured large-language-model time in module-level averages across the tested datasets. The values came from the reported concurrent setup.
What remains unanswered
These are fixed-corpus engineering results, not evidence of human user benefit. The evaluation used controlled QA streams and automated large-language-model answer judgments; it did not test naturally timestamped traffic, topic drift or rapidly changing facts.
Long-term store maintenance is also unresolved. In the reported setup, accepted experiences are not updated or deleted, so the tests do not show whether stored material will remain valid as facts change or whether the store will stay manageable as it grows.
The work is an arXiv preprint, version 1, dated 26 August 2026, and the supplied record reports no journal or peer-review status.
Paper data and sources
Original title: LivingRAG: Augmenting Graph RAG with Experience
Authors: Yuzhuo Cui, Zongye Zhang, Qingjie Liu
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-26
DOI: Not available
Original paper · Full text