GPT-5.6 Terra Tops LLM Listwise Reranking; Luna Is 10x Cheaper for Near-Parity
CShorten30 · x · 2026-11-16
Part 5 of the thread adds an LLM listwise reranking stage over zerank-2's top-20 (from hybrid retrieval at k=2,000), testing three OpenAI models with the same prompt, reasoning effort set to none, averaged over 3 trials:
- GPT-5.6 Terra: nDCG@10 lifts from 0.424 to 0.488
- GPT-5.6 Luna: 0.470, and the cheapest of the three at $0.20 per 1M input tokens (vs. $0.75 for Mini and $2.00 for Terra)
- GPT-5.4 Mini: moves it by exactly zero
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