AI SEARCH · 2026-10-22
AI Search — Paper Daily, 2026-10-22
UGC short-text community search: semantic matching, query understanding, recall, relevance ranking and LLM applications.
4.45
One encoder combines global semantics with token matching. Tagged extra vectors make standard MaxSim equal token-level matching plus a weighted dense score, so existing late-interaction indexes remain usable. BEIR nDCG@10 reaches 59.79.
Key insight: A hybrid representation can remove the need for separate encoders, indexes and score fusion. Distilling the final combined score also aligns training with the score actually served.
4.45
FLINT pools both queries and documents into a fixed number of vectors. This preserves late interaction while bounding storage and matching work. On long-query corpora, end-to-end retrieval is 1.5–2.8 times faster than ConstBERT.
Key insight: Budget vectors on both sides of retrieval. Compressing documents alone reduces storage, but long queries still drive many comparisons; symmetric pooling bounds both costs.
4.15
GARRET learns to retrieve regions within visual documents using differentiable feedback from question-and-answer generation. The auxiliary generator is removed at inference. Macro nDCG@10 on region retrieval reaches 0.4126.
Key insight: Generation loss can assign credit to candidate regions without query-to-region labels. Because the generator is only a training component, the deployed retriever keeps the fine-grained supervision without its inference cost.
3.90
SNO-PI links clinical mentions with scarce labels using high-recall span proposals, concept retrieval and ensemble listwise reranking. Domain text and curated terminology supply additional supervision. Macro character IoU reaches 0.5387.
Key insight: Aliases and concept hierarchies can provide supervision for unseen entities. The transferable design is the combination of curated knowledge, candidate recall and reranking, rather than the medical label system itself.
3.90
This controlled study varies corpus size while retaining relevant documents, comparing abstracts, flat full text and hierarchical full text. Rankings between configurations change with scale. Hierarchical representations use 67% fewer indexed units.
Key insight: Corpus size is an experimental variable, even when relevant evidence is retained. Evaluate scale, evidence format and cost together: added distractors can change the value of both retrieval and reranking.
3.85
GoogleGoogle demonstrates on-device media search with a unified text, image and audio embedding space, local storage and interactive retrieval. Reported active RAM for text-only weights is approximately 191 MB.
Key insight: A shared embedding space removes intermediate captioning and transcription stages. Loading only the required modality also reduces memory; the reported figures describe a specific device configuration, not server latency.
3.77
This agent alternates conversational query rewriting with retrieval, using returned passages to guide its next action. Retrieval reward trains the action sequence and stopping behavior. On TopiOCQA, nDCG@3 reaches 40.6.
Key insight: The corpus can supply vocabulary for the next rewrite. Typed actions alone do not explain the gains: the ablations attribute them to sequencing actions using retrieved evidence, with additional search rounds as a cost.
3.75
SPRA represents problems, methods, experiments and findings as typed research objects with source locations. Across seven subfields, source-conditioned scientific QA scores 0.381 points above passage retrieval on a five-point scale.
Key insight: Index reusable research objects rather than only text chunks. Persistent identities and source locations support cross-paper comparisons, at the cost of building and maintaining structured representations.
3.60
CConare
Conare publishes first-party retrieval, reranking and constraint-matching evaluations, holding first-stage candidates fixed for reranking. Its reported document-reranking nDCG@10 is 70.2; this run did not independently reproduce the benchmark.
Key insight: A fixed candidate set separates reranker quality from first-stage recall. Near-duplicate documents that violate individual constraints provide useful stress tests, but vendor results still need reproduction on the target workload.
3.50
Toollery generates likely user intents from tool specifications offline and indexes them for retrieval. The language model then chooses from a compact candidate set rather than the entire registry. Evaluation includes a benchmark with roughly 79,000 capabilities.
Key insight: Document-side intent expansion can keep online candidate sets small without rewriting every incoming request. Its usefulness depends on coverage of real user wording; the abstract reports no single universal gain.
Amazon
CMU
INESC-ID
EMNLP 2026