AI SEARCH · 2026-10-21
AI Search — Paper Daily, 2026-10-21
Community search over short user-generated posts, questions and comments, focusing on relevance, nDCG and CTR.
4.40
LEAD STORY ·
UC Irvine /
Thomson Reuters /
University of GuelphUC IrvineThomson ReutersUniversity of Guelph Adaptive fusion need not predict one exact weight. Across 128 configurations, 74 significantly outperform standard RRF after multiple-testing correction.
Key insight: Treat optimal weights as sets rather than arbitrary point targets, and include predictor latency in deployment decisions.
4.25
Correct pairwise preference does not guarantee a useful list position. Bounded Qwen3 reranking reaches preferred-passage MRR of 0.7509 in fixed hybrid pools.
Key insight: Evaluate pair preference and full-list position separately; optimizing one can degrade the other.
4.20
KAIST
University of Tübingen
NAVERCross-topic citations teach retrieval models to find transferable ideas. Contrastive training raises Recall@10 from the best untrained retriever’s 0.26 to as high as 0.70.
Key insight: Training labels define relevance. Topic similarity alone can miss useful analogies and may impair cross-domain retrieval.
4.15
DDeli Joint Laboratory
City University of Macau A two-stage hard-negative pipeline improves legal reranking after candidate recall saturates. The trained model wins 72.4% of non-tied internal offline preference comparisons.
Key insight: Separate candidate recall from reranker quality. More dynamic mining stages do not consistently improve held-out performance.
4.10
University of Surrey
UCLConcept-conditioned retrieval separates wording overlap from shared survey constructs. MAP and recall improve by roughly 2–3 points over baselines.
Key insight: Explicit concepts can disambiguate semantic matching, and a general domain thesaurus can provide supervision when task labels are scarce.
4.05
University of InnsbruckA revisable topic memory tracks additions, abandonment and reactivation. With a frozen retriever, MemRevise reaches mean MRR@10 of 0.701.
Key insight: Keep topic state transitions explicit so abandoned constraints stop affecting search while remaining recoverable.
3.95
IIndependent Researcher
Structure-aware legal chunking preserves clause boundaries and resolves cross-references before hybrid retrieval. LegalBench-RAG character Recall@8 reaches 48.18%.
Key insight: Chunk boundaries determine whether evidence is representable at all; preserve document structure before tuning retrieval models.
3.85
EMNLP 2026CChungbuk National UniversityUniversity of InnsbruckMEMORA decomposes queries into factors, gathers evidence and combines committee verdicts with episodic memory. Text-to-text nDCG@10 reaches 41.5 on MM-BRIGHT.
Key insight: A factor-level evidence matrix makes reranking decisions inspectable, but committee calls and cross-query memory add serving complexity.
3.75
McGill
Ohio University
University of ManitobaExact certificates bound worst-case nDCG loss using pairwise judgments, saving 2.34 comparisons per development query on shared trajectories.
Key insight: Judgments constrain the ideal-ranking denominator as well as the loss numerator; ignoring that coupling wastes comparisons.
3.75
TTezpur UniversityKKhyontekAI
A metadata-enriched multilingual retriever followed by cross-encoder reranking links scientific social posts to source papers. Average MRR@5 is 0.5862.
Key insight: Audit metadata, candidate depth and language balance separately to identify which parts improve source retrieval.