History — 2026 week 36 (Aug 31 – Sep 6, 2026)

Newest first.

2026-09-03 — The Collections LLM Judges Are Measured On (3 new, 5 updated)

The TREC Deep Learning Track ran 2019–2023 on the MS MARCO corpora — 43 deeply-judged test queries per task in 2019 against 3.2M documents and 8.8M passages, with NIST assessors applying multi-graded judgments by depth pooling and document labels inferred from passage ones. ANTIQUE supplies the non-factoid counterweight: 2,626 real Yahoo! Answers questions and 34,011 crowdsourced annotations on a four-point scale that grades whether an answer is convincing, not whether it is correct. LLMJudge, a SIGIR 2024 LLM4Eval challenge built on TREC DL 2023, benchmarks judges rather than rankers — 4,423 test pairs scored on Cohen’s κ against humans and Kendall τ against system orderings.

DatasetsTREC Deep Learning Track · ANTIQUE · LLMJudge UpdatedMS MARCO · LLM as Judge · Benchmarking LLM-based Relevance Judgment Methods · Criteria-Based LLM Relevance Judgments · Charles L. A. Clarke


2026-09-03 — There Is No Single LLM Judge (6 new, 1 updated)

Two 2025 papers sharpen what LLM-as-a-judge actually names. Negar Arabzadeh and Charles L. A. Clarke’s Benchmarking LLM-based Relevance Judgment Methods runs binary, graded, pairwise-preference and two flavours of nugget-based judging side by side over TREC Deep Learning 2019–2021 and ANTIQUE, separating two things usually reported as one: agreement with human labels, which pairwise preferences win, and agreement with human system rankings, which graded and binary pointwise labels win at Kendall τ up to 0.922. Neither implies the other. Naghmeh Farzi and Laura Dietz’s Criteria-Based LLM Relevance Judgments instead decomposes relevance into exactness, coverage, topicality and contextual fit, each graded 0–3, trading roughly 5.4× runtime for a grade a human can inspect and dispute.

Corrections

  • LLM as Judge advised pairwise comparison over pointwise scoring as a flat rule; the paradigm depends on whether the output is a document grade or a system ranking, and the recommendation now says so.

ArticlesBenchmarking LLM-based Relevance Judgment Methods · Criteria-Based LLM Relevance Judgments PeopleNegar Arabzadeh · Charles L. A. Clarke · Naghmeh Farzi · Laura Dietz UpdatedLLM as Judge


2026-09-02 — The Generalization Cliff: Why Leaderboard Winners Fail On Your Data (1 new, 8 updated)

General-purpose embedding models that top public leaderboards collapse once asked to retrieve on a specific domain, sometimes catastrophically. BRIGHT anchors the sharpest evidence — a model scoring 59.0 nDCG@10 on MTEB scores 18.3 on BRIGHT’s reasoning-intensive queries — while Model Selection and Fine-Tuning Evaluation’s “don’t select on public benchmarks” rule and RTEB’s open-vs-private diagnostic explain why: leaderboard position reflects benchmark fit and contamination as much as real capability. Domain fine-tuning closes most of the gap — Thierry Damiba’s SPLADE series took e-commerce nDCG@10 gains from +7.2% to +27.5% over BM25 — but has its own cliff: the fine-tuned model lost to the off-the-shelf baseline on a neighboring catalog and forgot general web queries entirely.

Topics — The Generalization Cliff (moved out of the graph to drafts/the-generalization-cliff.md on 2026-09-03) UpdatedBRIGHT · RTEB · BEIR · Model Selection and Fine-Tuning Evaluation · Retrieval Benchmarks and Leaderboards · Embedding Models Compared · Fine-Tuning Text Embeddings For Domain-Specific Search · Fine-Tuning Sparse Embeddings for E-Commerce Search


2026-09-01 — Algebraic Rewrite Speeds Up Asymmetric Re-ranking in Vespa (2 new, 4 updated)

Dainius Jocas shows how algebraic rewriting speeds up Asymmetric Re-ranking in Vespa — scoring a full-precision query embedding against binary-quantized document embeddings to recover recall that quantization loses, without changing the underlying math. Rewriting the naive per-document expression (q·(2·doc − 1) + 1) / 2 as q·doc + (1 − Σq)/2 isolates a query-only term that Vespa’s constant-folding computes once per query instead of once per document, collapsing the per-document work to a single dot product against the unpacked bits. At 1M documents this cut end-to-end latency from 267ms to 193ms, a 28% improvement. Sourced from Jocas’s personal blog, jocas.lt.

ArticlesOptimize Asymmetric Re-ranking with Algebra ConceptsAsymmetric Re-ranking UpdatedBinary Quantization · Reranking · Vespa · Dainius Jocas


2026-08-31 — Etsy Adds LLM-Scored Semantic Relevance to Search Ranking (5 new, 5 updated)

Etsy’s search models historically ranked on engagement signals (clicks, add-to-carts, purchases) that skew toward already-popular listings regardless of query fit. How Etsy Uses LLMs to Improve Search Relevance adds Semantic Relevance as a complementary signal: an LLM annotator anchored to human golden labels scores query-listing pairs, then a three-tier distillation cascade — an o3 annotator, a Qwen 3 VL 4B teacher, a BERT two-tower student — compresses that judgment into a model cheap enough for real-time filtering, feature enrichment, loss weighting, and boosting. Fully-relevant listing share rose from 58% to 62% between August and October 2025. Sourced from Etsy’s Code as Craft blog, by Yuqing Zhang, Congzhe Su, and Susan Liu.

ArticlesHow Etsy Uses LLMs to Improve Search Relevance ConceptsSemantic Relevance PeopleYuqing Zhang · Congzhe Su · Susan Liu UpdatedEtsy · Etsy - Search Quality and Query Understanding · E-commerce Search · Knowledge Distillation · LLM as Judge


2026-08-31 — Choosing Between RRF and DBSF Fusion in Qdrant (4 new, 7 updated)

RRF discards score magnitude entirely — a document leading by a wide margin scores the same as one leading narrowly, as long as both land at rank 1. Qdrant’s Distribution-Based Score Fusion keeps the raw scores instead, rescaling each retriever’s returned set by its own mean and 3-sigma spread before summing. How to Tune Hybrid Search in Qdrant measures both across five datasets: DBSF beat default RRF by up to 0.0383 nDCG@10 on WANDS, while RRF’s own k parameter shifted results just as much — k=2 and k=61 picked different top results for 42% of WANDS queries. Sourced from Dylan Couzon’s Qdrant tuning guide and Qdrant’s hybrid-queries reference documentation.

ArticlesHow to Tune Hybrid Search in Qdrant · Hybrid Queries - Qdrant ConceptsDistribution-Based Score Fusion PeopleDylan Couzon UpdatedReciprocal Rank Fusion · Score Normalization · Hybrid Search · Qdrant Vector DB · Qdrant · WANDS Dataset · BEIR