Topics — What a work is about. Nucleation holds whatever reading the owner files here, whether or not an Anthology of the SOTA topic could also hold it; instruction for ML practice belongs to the anthology (ADR-048). The words are seeded from the clusters the reading-time triage of 2026-09-25 found, and grow by decision.
Information retrieval — finding what is relevant to a need — document and query representation, relevance and aboutness, ranking and feedback, and the geometric, probabilistic and logical models of retrieval.
7 of 371 QUESTION documents. Back to the full index.
| # | Title | Summary | Status |
|---|---|---|---|
| QUESTION-096 | When does a communicative cue change what an interpreter can recover from an unchanged internal state, and how can such an addressing effect be told apart from new information, added computation or a changed policy? | A19’s question, after the owner tied the represented / recoverable / actionable distinction to QKV attention and Kanerva’s sparse distributed memory at U19. The record already holds that attention approximates SDM (LIT-269). | Open |
| QUESTION-107 | When does iterative query reformulation improve access to relevant evidence, and when does it entrench an initial interpretation by narrowing what is encountered next? | A21’s question. The owner suspected at U21 that agent harnesses rediscover old information retrieval: query expansion, relevance feedback, intent classification and slot filling. Classical IR named query drift (LIT-1079, unread). An agent that rewrites its own task adds TERM-056. | Open |
| QUESTION-114 | Is sorting a collection amortized search, and does physical database design, where a table is stored in one order and every further index is paid for in space and upkeep, give the joint cost model that the question on amortized retrieval lacks? | The LEGO parable (CASE-094) is QUESTION-095’s trade in its plainest form. Sorting spends search once, against retrievals later, and leaves the result in the world, as the cockpit’s speed bugs do. CLAIM-310 adds that one set of bins realizes one partition, chosen against a query mix. A database table is stored in one physical order, and every further index costs space and work on every insert. Choosing indexes for a workload is a studied problem with explicit cost models. This asks whether that model is the one QUESTION-095 is looking for. | Open |
| QUESTION-187 | Does greedy layered search on the graph HNSW actually builds take expected O(log N) distance computations to a stated recall, under a stated condition on the data such as bounded doubling dimension? | The promotion condition formerly on THEORY-365, moved here when THEORY-365 was narrowed to what LIT-1098 measured and promoted on 2026-10-19 (ADR-047 v3). The paper derives O(log N) search only for exact Delaunay layers of bounded degree. It measures logarithmic scaling only at d = 8 and saturating ef only at d = 4, and at d = 128 its scaling departs from a logarithm. | Open |
| QUESTION-230 | Do the released HNSW, NSG and fast-preprocessing DiskANN builds, run independently on Indyk and Xu’s published hard instances at the stated parameters, need a search queue of about a tenth of the points before recall@5 rises above zero, and does the same trap appear on natural data with adversarially chosen queries? | The promotion condition formerly on THEORY-368. THEORY-368 was promoted on 2026-10-20 (ADR-047 v3) for what Indyk and Xu’s one study measured. An independent replication on their instances, and the same test on natural data with adversarial queries, are open. | Open |
| QUESTION-237 | Does semantic caching keep its ordering against page caching once probe and remainder computation and region upkeep are charged, or in an implemented client-server system? | The promotion condition formerly on THEORY-376, moved unchanged when that account was narrowed and promoted on 2026-10-20 on the one simulation it rests on (ADR-047 v3). Open: whether semantic caching’s advantage survives charging the client for reasoning about what its cache holds. | Open |
| QUESTION-352 | Once Qualified Competitors have used the index snapshot, the Glue and RankEmbed data and the syndication licence ordered in United States v. Google, do their results close the gap on long-tail queries, and does their search share move, compared with rivals or periods without the data? | The promotion condition formerly on THEORY-480. It moved here when the twenty-fourth status review promoted that entry as a report of the remedies opinion (ADR-065 §3). The court expects the shared data to help rivals, and finds nothing about whether it will. This question asks what the decree did. | Open |