An experimental method for RAG systems reported higher evidence coverage while reranking a smaller set of documents. On HotpotQA, PACE reranked 20 documents and reported complete evidence recall of 77.10%, followed by final evidence recall of 60.26% at the five-document cutoff. Standard Dense, using a 100-document reranking budget, reported final recall of 41.13%. The reported gap was 19.13 percentage points.
On 2WikiMultiHop QA, PACE at the same 20-document budget reported complete evidence recall of 75.19% and final recall of 53.94% at five documents. Standard Dense at 100 documents reported 48.76% final recall, leaving a reported advantage of 5.18 percentage points.
The headline figures are about evidence recall. The evaluation paired static offline recall tests with an online serving evaluation of latency and recall under adaptive reranking budgets.
Two levers, one ranking pipeline
PACE is a training-free framework with two linked parts. Its evidence-frontloading step reorders retrieved candidates so direct evidence, complementary evidence and documents that help form a multi-hop chain appear earlier. Its pressure-adaptive step changes the reranking budget according to the relative pressure in the reranker queue and the LLM queue.
The design ties candidate ordering to a limited reranking budget, while queue pressure determines how that budget changes. The paper states that its greedy evidence-frontloading procedure reaches at least a (1 - 1/e) approximation to optimal evidence coverage when the number of selected items is limited.
The ablation analysis found that removing both query relevance and anchor relevance was usually associated with the worst performance. Coverage alone also did not reliably identify useful evidence.
The queue changed the trade-off
The online serving evaluation was an open-loop Poisson simulation with workloads from 0.5 to 2.5 queries per second. The reranker and LLM used separate GPUs and independent asynchronous queues.
The reported pattern was not a single fixed bottleneck. In one observed pair with a 92.5 parameter-size ratio, the system was generation-heavy at 2.5 queries per second. With a reranking budget of 100, the reranker queue became dominant at 1 query per second, while a budget of 50 remained LLM-queue dominated.
Under ranking-heavy workloads, PACE had substantially lower p95 end-to-end latency and reranker-queue latency than the fixed 100-document comparison. P95 captures the slower end of response times. The reported trade-off was that LLM queue time could rise slightly as more requests passed through.
Recall also held up in the adaptive online test. After reranking, PACE had the highest reported recall@5 among documents sent to the LLM, exceeded the fixed 100-document comparison and remained stable as query rate increased.
A result bounded by its test setup
Across HotpotQA, MuSiQue and 2WikiMultiHop QA, the offline analysis reported higher complete and supporting evidence recall for PACE, especially at smaller reranking budgets. On HotpotQA, its 20-document setting was comparable to the best baseline at 40 documents.
The reported dev-split counts were 1,087 HotpotQA queries, 2,317 MuSiQue queries and 2,861 2WikiMultiHop QA queries. An additional 100 queries per dataset were reserved for baseline calibration and excluded from evaluation. HotpotQA and 2WikiMultiHop QA retained queries whose complete evidence appeared in the top 100 retrieved documents, while MuSiQue used a closed-context setting with a maximum reranking budget of 20.
The experiments used SPLADE-v3 retrieval, trecdl22-crossencoder-debertav3 reranking, Provence compression and Qwen2.5-3B-Instruct generation on Quadro RTX 6000 GPUs with approximately 22 GB of available memory.
The online findings came from an open-loop simulation rather than a production deployment. The reported comparisons are point estimates without confidence intervals or significance tests.
The findings are tied to the reported model and hardware stack, so the supplied analysis leaves performance on other systems, hardware and production loads open. The manuscript is an arXiv v1 preprint dated 25 August 2026.
Paper data and sources
Original title: Less can be More: Relieving RAG Bottlenecks via Evidence Frontloading and Pressure-Adaptive Budgeting
Authors: Weibin Cai, Reza Zafarani
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-25
DOI: Not available
Original paper · Full text