A one-application case study is not a benchmark, but its repair log is a useful negative result: generation took minutes; validation took days.

Across six 80-case evaluation suites, a learned router spent fewer tokens than always-on retrieval; the result is a warning against treating tool access as automatically useful.

The unified engine matched Chroma’s vector recall and beat LadybugDB on transactional graph work, but specialist systems remained dramatically faster on analytics.

PQR-Search’s authors replace average ANN recall with a per-query probability contract, then show what that guarantee costs across static and streaming workloads.

CoRe’s controlled renaming test shows that some LLM repair scores depend on familiar entities even when the candidate values remain in the table.

A 454-case held-out test separates diagnosing a database fault from actually fixing it, but its composite score and LLM judge need to travel with every comparison.

A new semantic-filtering study reports the best accuracy in 12 of 13 settings, but its own audit says one setting missed the oracle-relative target in 30% of repeated draws.

Across eight RelBench tasks, classification produced less consensus than regression, while simple column ablation was the fastest method tested.

The submitted 81.0% SQLite score became 80.2% on a repeat run, a small gap that shows why one-shot leaderboard rows need artifacts and variance reporting.

A 13-pair IES evaluation shows why coherence and conservativity must be rerun on the whole federation, not accepted one crosswalk at a time.

A new PostgreSQL study reaches 1.10 Q-error on TPC-H and 1.19 on TPC-DS after a hardware move, then watches most preregistered tests fail on JOB.

A new paper reports 33% and 68% runtime reductions on JOB workloads, but its real contribution is three boundaries that keep synthesized components from changing database results.

BP-TD beats majority vote when a weak model family dominates the pool—and loses when the dominant family is strongest.

SpecBridge’s controlled tests separate whether a model can compute an interpretation from whether it tells the user that another interpretation exists.

The paper reports up to 61.3× geometric-mean speedups with machine-checked code, while showing why passing synthesis tests is not enough.

The paper’s set-contrastive training beats its listed baselines, but 38,000 GPT-4 calls move agent cost from inference into dataset construction.

A 300,000-record generation pipeline retains 96,500 examples and reports 41.80% BIRD dev execution accuracy. The missing step is semantic execution during dataset filtering.

The paper can represent paracomplete, paraconsistent and classical contexts across two to six truth values. What it does not yet provide is measured evidence for choosing the machinery.

The AI-READI registry study verified 64.85 million stored values, but its strongest result is methodological: counts and samples missed duplicated, absent and source-artifact records.

A 5.8-million-item LAION evaluation shows per-query ANN plan selection can beat one-size-fits-all indexing, but the saving depends on accepting up to 10% target misses.

A new knowledge-graph estimator makes similarity thresholds first-class inputs and survives a cross-model workload test, while exposing a sharp offline-cost trade-off.

A training-free correction layer adds 25.08 points over GPT-5.4-mini on the released test set; its own error analysis shows why execution success is not enough.

A machine-checked optimizer lifts Paxos and read-heavy CAS, yet the same throughput-only search adds network hops that cut a sequential write workload by two-thirds.

The nine-dataset study improves macro clustering scores with bounded penalties, while separating offline annotation cost from downstream optimization time.

The paper’s strongest result comes from sharing exact-distance candidates across semantically related queries, not merely caching repeated SSD reads.

The new Qwen3-1.7B release is useful less for its top score than for showing how evaluation design changes the answer.

ExactCert does not choose a better imputer; it identifies which retrieval answers survive every admissible completion—and exposes how often common fill rules change the answer.

The new operator beats prior systems across its evaluated workloads, while the largest headline number compares against a conventional join-and-aggregate plan on one gene matrix.

The system linked new evidence to hundreds of research questions, yet only 9 of 41 audited solve/refute verdicts were fully correct. Its safest design choice was refusing to update status automatically.

The agent-built validator refuted 20,957 of 36,042 query pairs with cross-checking, yet coverage, dialect support and a 10-query BEAVER probe limit what that number means.

The benchmark’s 2026 rewrite aligns 500 tasks across SQLite, MySQL and PostgreSQL, while warning that its older scores are now historical rather than comparable.

Amazon researchers report 95.7% execution validity on BEAVER, but the reproducible experiment reconstructs metadata and joins from BEAVER’s supplied schema—not from query logs.

HySE and FlexPrune improve top-10 retrieval on KaggleDS, but the benchmark’s single-gold labels and rich Kaggle metadata limit what the score proves about enterprise catalogs.

A new Turkish document-table benchmark isolates a failure that SQL-generation scores miss: locale-blind parsing can silently corrupt the database before the model writes a query.

The geo-filtered ANN result is strongest on selective disks, but it should not be read as a universal vector-search speedup.

EUR’s query-first design is concrete enough to test, but its first paper stops at a two-query worked example without an implementation or measured comparison.

A new line-query vector index rewires HNSW with bounded pair edges, but its strongest headline comparison is also the one readers should interpret most carefully.

A new commit-boundary proposal defines four matched conditions and safety-versus-coverage metrics, but its claimed advantage remains entirely prospective.

A 400-case supervised evaluation shows that answer-to-source traces can expose extraction and query errors, but trace availability varies sharply by benchmark and does not establish generalization.

A controlled probe of Claude Sonnet 5 and GPT-5.6-Terra found that unstated rules were ignored every time—and explicit rules usually produced silent failures or race-prone SQL.

A new SALT-KG evaluation turns offline LLM discovery into deterministic runtime lookups, lifting payment-term accuracy from 37.5% to 82.9% while exposing a temporal-reasoning limit.

Three of five concurrent learned indexes lost keys or stopped progressing, while synchronized rebuilds created latency bursts that aggregate throughput scores hid.

A new ICDE 2027 paper beats page-retrieval and agent baselines across three long-document QA suites, but its largest improvement is concentrated on MMLongBench-Doc.

A five-dataset evaluation reports 22.64–39.51× faster deletions than reinsertion-based repair, while exposing where the speedup trades away recall or query latency.

A dynamic-database benchmark measures query time, monitoring and retraining together—and selective correction beats even post-hoc-tuned periodic refresh on most templates.

A PVLDB evaluation shows why where data is decoded matters as much as how well it compresses—and why the fastest codec is not the smallest.

A 606-request WDI/AQS evaluation rewards correct execution, clarification and refusal, but the paper does not release the task set needed to reproduce its 32.5-point lead over DIN-SQL.

A new SemBench study stores model verdicts as database tags. The cost reduction is large, but one inherited answer falls to 0.58 quality.

AskData’s 83.40% hidden-test result changes the main ranking, while a second update shows why Mini-Dev’s three SQL dialects should not be collapsed into one score.

The 1.8-point Spider gain is real within the published harness, but the companion metrics show why reward tuning needs a deployment-shaped acceptance test.