Deal With It: On Better Schedules for Listwise Reranking
Abstract
Listwise LLM rerankers cannot reliably rank hundreds of documents in a single call, and rely on schedules to orchestrate their calls. Parallel schedules split candidates into groups and advance a fixed quota from each. When a group holds more of the top passages than its quota, some are cut, and expensive repair calls become required. We show that the way candidates are split matters where such a quota can bind, and that, in expected cuts, the best split there deals them round-robin by retriever rank, like cards. Although not conceptually new, dealing has remained unanalyzed. We prove that, under a simple retriever model, no split with equal quotas cuts fewer of the true top ten on average, and that contiguous chunks cut more as the retriever improves. DealRank, the simplest schedule our analysis allows, is on average within 0.1 nDCG@10 points of the best schedules over six LLMs, in two rounds. It is consistently 1.1 to 1.4 times faster than BlitzRank with concurrent first calls up to a thousand candidates. There, it needs three rounds and is eleven times faster than the sliding window. In existing schedules such as BlitzRank, dealing reduces repair calls from 59% of queries to 14%, improving latency with no significant change in quality. Random partitioning appears equivalent at a quota of four, as predicted, but ties results to the seed and trails by 1.3 nDCG@10 points at a quota of one over a dense retriever.
Figures & tables
| nDCG@10 100 per benchmark | Cost per 100-candidate query | |||||||||||||||||
| Method | Avg. | vs. DealRank | DL19 | DL20 | COVID | NFC | Signal | News | Rob04 | Touché | DBP | SciF | Calls | Rounds | Max | Tok. (k) | p50 | p90 |
| BM25 | 44.7 | 12.9 [ 19.7, 5.6] | 50.6 | 48.0 | 59.5 | 32.2 | 33.1 | 39.5 | 40.7 | 44.2 | 31.8 | 67.9 | 0 | 0 | 0 | 0 | – | – |
| Top-20 only (one call) | 53.8 | 3.8 [ 7.2, 0.1] | 66.1 | 64.2 | 75.2 | 37.2 | 35.5 | 45.3 | 54.2 | 42.6 | 40.0 | 77.9 | 1.0 | 1.0 | 1 | 5.2 | 1.18 | 1.21 |
| Sliding window | 57.9 | +0.3 [ 0.4, +0.9] | 74.6 | 73.0 | 82.6 | 39.5 | 33.7 | 49.0 | 65.2 | 35.0 | 46.1 | 80.6 | 9.0 | 9.0 | 9 | 46.1 | 10.63 | 10.78 |
| TourRank | 56.7 | 0.9 [ 1.9, 0.1] | 73.5 | 71.2 | 81.7 | 38.7 | 33.4 | 48.5 | 64.3 | 32.5 | 45.1 | 77.9 | 13.0 | 5.0 | 5 | 48.1 | 4.53 | 4.63 |
| TDPart | 57.1 | 0.5 [ 1.8, +0.9] | 72.9 | 70.0 | 82.1 | 38.7 | 34.7 | 48.8 | 62.1 | 37.8 | 44.7 | 79.1 | 6.9 | 2.9 | 3 | 31.6 | 3.89 | 4.05 |
| BlitzRank | Sliding | JointRank | ||||
| Model | DealRank | window | TDPart | |||
| DeepSeek V4-Flash | 58.5 | 1.01 | 0.50 | 0.10 | 0.36 | 3.12 |
| Qwen3.5-35B-A3B | 56.7 | 0.22 | 0.42 | 0.62 | 0.35 | 1.87 |
| Qwen3.8-27B | 56.9 | +0.61 | +0.79 | 0.26 | 0.15 | 1.05 |
| Kimi K2.6 | 57.3 | +0.75 | +0.57 | +0.52 | 0.14 | 1.18 |
| GLM-5.2 | 57.2 | +0.20 | +0.37 | +0.74 | 0.17 | 1.85 |
| nDCG@10 | Seconds per DL19 query (rounds) | ||||||
|---|---|---|---|---|---|---|---|
| Top | Deal | Blitz | Window | Deal | Blitz | Blitz | Window |
| 100 | 67.9 | 67.6 | 67.5 | 2.9 2 | 4.0 2.6 | 8.2 6.6 | 10.6 9 |
| 500 | 69.8 | 69.9 | 70.1 | 7.0 3 | 8.0 4.3 | 33.0 28 | 57.9 49 |
| 1,000 | 71.0 | 70.8 | 70.8 | 10.3 3 | 13.0 5.7 | 65.1 55 | 115.4 99 |
LLM Inference Scheduling
Large language model inference scheduling allocates requests, batches, compute, and KV-cache memory to balance latency, throughput, and service-level objectives.
See all 197 papers