Extending the rerank window from top-20 to top-100 adds just +0.012 nDCG at 5x cost
CShorten30 · x · 2026-11-16
Tweet 8/17: handing the listwise reranker a deeper window barely helps — GPT-5.6 Luna over top-100 reaches 0.482 nDCG@10 vs 0.470 over top-20, a +0.012 gain at 5x the cost. This despite zerank-2's recall climbing from 0.560 at cutoff 20 to 0.673 at 50 and 0.738 at 100, meaning gold documents often sit just below the initial cutoff.
Related event: Reranking Study Finds Two-Stage Pipeline Beats Fusion and Deeper Windows(13 posts)→
More from Research
- BeyondBinding challenge launches: AI-designed antibodies get wet-lab tested — QuanquanGu · 2026-11-17
- Singularity-Bench: Opus 5.5 with Claude Code beats human baselines on building AI systems — armand_ruiz · 2026-11-17
- Open Vision Banana: fully open recipe turns image generators into multi-task perception systems — NielsRogge · 2026-11-17
- WebArena-Pro debuts as largest browser-agent benchmark, top agents score 62.3% — xhluca · 2026-11-17
- NeurIPS 2026 paper: RL-trained skills transfer poorly to composed reasoning tasks — StanfordAILab · 2026-11-17
- OpenAI warns CoT monitorability will degrade as models scale, researchers push back — gleech · 2026-11-17