History — 2026 week 44 (Oct 26 – Nov 1, 2026)
Newest first.
2026-11-01 — The search frontier, second half of 2026 (1 new, 2 updated)
In the second half of 2026, the leading edge of search is less about finding candidates than about judging them. Frontier of Search 2026 Part 2 picks up where Frontier of Search 2026 left off in June and sets out three new fronts. First, the LLM judge, made cheap through cascades and examined through agreement levels and calibration. Second, decision models such as Jev that return a calibrated probability. Third, cheaper reranking with BlitzRank and Fast but Not Late Interaction. It also follows commoditization (MUVERA), agent memory and routing, and growing doubt about reported figures (Benchmark Contamination). Most figures it cites are vendor- or author-reported.
Topics — Frontier of Search 2026 Part 2 Updated — Frontier of Search · Frontier of Search 2026
2026-10-30 — Golden sets and the NDCG that flatters them (4 new, 7 updated)
A golden set of graded judgments can score a search system at NDCG 1 while users are unhappy, simply because the corpus holds few relevant results for the query. Doug Turnbull argues for narrower evals: NDCG-m, named for Max Irwin, normalises against a fixed all-max-grade ideal so corpus coverage shows in the score; NDCG-f scores labeled results only, as a regression check when labels are incomplete; and Rank-Biased Overlap measures how far results moved. The post reports an average NDCG-m@10 of about 0.77 on WANDS and closer to 0.93 on ESCI. The first two extend NDCG; the third compares rankings without labels.
Articles — Search golden sets are anything but Concepts — NDCG-m · NDCG-f · Rank-Biased Overlap Updated — NDCG · Judgment Lists · Search Evaluation · Max Irwin · Doug Turnbull · WANDS Dataset · Amazon ESCI Dataset
2026-10-30 — Fast but Not Late Interaction: rerankers with cached passages (3 new, 5 updated)
A Cross-Encoder re-encodes every passage for every query, while Late Interaction avoids that with a fixed token-level MaxSim comparison. A third design, Fast but Not Late Interaction, encodes passages offline into compressed token states and lets a learned decoder read them alongside the query, with the compression ratio adjustable from 1x to 128x at inference. The concept note places it against cross-encoders, late interaction and MUVERA, and Lychee-Reranker covers the open model family built on it. Sourced from the authors’ technical report; its cost figures are analytical estimates rather than measured latency, and on BEIR the largest model trails Qwen3-Reranker-8B (see Qwen3 Embedding Series).
Articles — Lychee-Reranker - Fast but Not Late Interaction for Compressed Document Reranking (the arXiv paper) Concepts — Fast but Not Late Interaction Tools — Lychee-Reranker Updated — Late Interaction · Cross-Encoder · Reranking · Matryoshka Embeddings · MUVERA