Mixedbread paper: dealing candidates round-robin is the optimal split for parallel LLM reranking
_reachsumit · x · 2026-11-18
Benjamin Clavié et al. at Mixedbread publish an arXiv study on schedules for parallel listwise LLM reranking.
- Dealing candidates round-robin by retriever rank is provably the best split: under a simple retriever model, no equal-quota split cuts fewer true top-10 passages on average.
- In existing schedules like BlitzRank, dealing cuts repair calls from 59% of queries to 14%, improving latency with no significant quality change.
- Their minimal DealRank schedule stays within 0.1 nDCG@10 of the best schedules across six LLMs, and with concurrent first calls is up to 11x faster than sliding-window at 1000 candidates.
More from coding & agent
- 10 ChatGPT + Codex prompts to diagnose and fix a slow laptop yourself — anthara_ai · 2026-11-18
- Agent auto-iterates Triton kernels in background, beats all OSS implementations on Strix Halo — qtnx_ · 2026-11-18
- Microsoft's Autopilot Agent Is Built on OpenClaw, Now in Preview for Frontier Customers — pswider · 2026-11-18
- SimplePost Gives ChatGPT and Claude a Publishing Button for 10 Social Networks — haltakov · 2026-11-18
- One set of Obsidian notes feeds Claude Code, Codex and Gemini alike — dSebastien · 2026-11-18
- Obsidian Starter Kit gives Claude Code, Codex and Gemini the same persistent context — dSebastien · 2026-11-18