Consort locks AI agents: immutable tests
A spec-first framework enforces engineering discipline through code agents cannot edit, not by persuasion or front-loaded specs alone.
cs.DB
Databases
Covers database management, datamining, and data processing. Roughly includes material in ACM Subject Classes E.2, E.5, H.0, H.2, and J.1.
sort pith recommended most recent
Packed factorized vectors deliver 2–100000x speedups on joins
FFX's packed layout supports arbitrary f-trees and reduces LLM prompt tokens 1.7-22.5x while preserving quality.
· “Factorized and Vectorized Execution: Optimizing Analytical and Semantic Queries over Relations”
SQLMorph: query mutation reveals up to 20% accuracy drop
Join expansion and naming de-naturalization stress-test systems; binary accuracy hides partial correctness and over-prediction.
· “SQLMorph: Query Mutation and Fine-Grained Metrics for Text-to-SQL Evaluation”
Five frozen questions make every analyst answer self-explaining
The same ontology, shapes, queries, and rules run unchanged across Pakistan, Malaysia, and Indonesia.
New dataset validates 1,134 real LEO satellite maneuvers
Tiered multi-source evidence enables large-scale evaluation of maneuver detection for the first time
· “MAD-LEO: A Maneuver-Annotated Orbital Dataset for LEO Satellites with Tiered Multi-Source Evidence”
124 million entities analyzed in history tool
WiDiff extracts Wikidata's 2.3 billion revisions into a queryable database for evolution and quality studies.
Eviction destroys 67-73% of correctable memory errors at 80k tokens
New audit reveals most agent-memory errors from eviction are irreversible, not retrieval failures
· “What Eviction Destroys: A Restore-Counterfactual Audit of Forgetting in Agent Memory”
Object causal model automates ANN bottleneck diagnosis and redesign
Automatically decomposes ANN workflows into eight replaceable phases, then selects and composes optimizations to improve recall and…
· “SmartANN: Object Causal Modeling Boosts Approximate Nearest Neighbor Diagnosis and Auto-Design”
16-module reference model for governed analytical data pipelines
Proposal unites eight pipeline steps with eight cross-cutting capabilities to embed quality and traceability from design.
One contract makes CDC row copies and logs replay exactly
The cut theorem covers DBLog, Debezium, Flink CDC, and dumps — no global snapshot needed.
· “Generalized DBLog: A Verified Contract for Interleaving Database Rows with a Change Log”
Selective revalidation yields zero false decisions in 210K agent runs
By rechecking only affected decision premises, ATR cuts revalidation cost by 90% and retains safe actions.
· “From Version Conflicts to Decision Conflicts: Selective Revalidation for Long-Running AI Agents”
Agent trajectories need their own database
Nested structure, semantic text, redundancy, and lineage defeat today's fragmented stores, says this vision paper.
Context-aware knowledge graphs
A five-operator pattern lets cultural heritage hold conflicting viewpoints side by side instead of flattening them.
Co-evolution loop prevents drift between ontology and data
X-DigCheck provides a domain-independent environment for continuous profile–graph alignment across regeneration cycles.
41 of 58 table-analysis systems skip retrieval entirely
The field's real gap: connecting a person's need to the right tables, not just running analysis.
· “Open Tabular Insight Extraction: Where Do We Stand, and Where Should We Go?”
LLMs hit 62% in multi-scenario pl/SQL test
New benchmark reveals strugles with control flow and state effects absent in SQL
One SQL query matches Datalog engines for program analysis
Compiling Linear Datalog to recursive SQL runs on seven databases and beats Soufflé by 5.5× on average.
Error-bounded lossy compression gets 21% boost from foundation model
But only on aggregate human-demand data like grid load and subway ridership, not on scientific simulation output.
· “Cadence: Error-Bounded Lossy Compression of Demand Time Series with a Time-Series Foundation Model”
SHACL recursive shapes: engines disagree
ShEx validators uniformly implement greatest fixed-point semantics; SHACL validators each behave differently.
· “sheval: An RDF data shapes evaluation tool and test-suite for recursive shapes”
Video query algebra cuts MLLM cost 87% and lifts F1
Concord's rewrites substitute cheaper modalities or decompose MLLM joins, reducing cost and improving recall on soccer, lecture, and…
· “Concord: A Video Relational Algebra for Cross-Modal Query Optimization”
Adaptive privacy budgets lift LDP recommendation hits by 17%
Spending noise budget on users' active categories keeps recommendation signals alive under local differential privacy.
· “CAT-LDP: Cloud-edge Adaptive Taxonomy under Local Differential Privacy”
Mapping-aware optimization cuts multi-model query tax by 100x
Predicate pushdown, dependent joins, and non-redundant planning eliminate out-of-memory failures and slash planning time from 600 ms to…
· “Reducing the Cross-Model Tax: Query Optimization over Multi-Model Data”
A semantic model structures genetic evidence for AI-ready interpretation
Conditional activation rules keep the vocabulary compact while SHACL shapes validate conformance.
· “A Semantic Model of Genetic Evidence: A Step Toward Bridging the Basic-Science-Clinic Gap”
All two-way regular path queries are constant-query testable
Property testing dichotomy for recursive queries: 2RPQs testable, monadic Datalog has a hard core under self-join freeness.
Textto-SQL is wrong: users need data investigations
D^2 multi-agent system achieves 95% correctness and 76% completeness, compared to ~32% from single-query agents
· “From Data Querying to Data Investigations: Rethinking Natural Language Interfaces for Databases”
Why language models need semantic beliefs to scale
A framework that turns knowledge graphs into shared, calibrated beliefs, enabling agents to reason probabilistically and source beliefs…
Per-query thresholds make vector caching practical
Aker lifts recall by up to 64pp and QPS by 3.2x while using only 0.6× the memory of pgvector's shared buffers.
· “Aker: Density-Aware Approximate Caching for Vector Search (Extended Version)”
CSPQs: worst-case optimal joins for linear expressions
New width measures ctw, atw, fcw bound evaluation time for queries with linear combinations in atoms.
This paper defines a database model built from labeled trees
A formal model in which a database is a set of labeled trees, every edge is a total function, and one functional algebra expresses both…
Knowledge-guided LLM agent ranks first at feature engineering
Beats seven rival methods (avg. rank 2.3) in leakage-free tests, with an auditable reasoning chain behind every feature.
· “KnowFeat: Knowledge-Guided Feature Engineering via LLM Agents”
SRL engines: SPARQL-based works
Both pass 166 conformance tests; explainability and blank-node handling remain open.
· “Does SRL Pave the Road to Explainable Reasoning? Lessons Learned from an Implementer's Perspective”
Deterministic parallel commit without knowing read-write sets
Lantern's DAG rule lets every node run the batch on its own — up to 4.2x the throughput of the Aria protocol.
· “Lantern: Finding Committable Transactions via Back-Propagation on DAGs”
Model improvements obsolete data agent scaffolds
Experiments show general coding agents outperform specialized pipelines; the lasting challenge is persistent knowledge about the data…
93% of corporate-family links vanish before matching begins
New public benchmark shows that no blocker propeses the hardest pairs, making this a retrieval problem, not a matching problem
Tighter tree search with SED-struct filter
Combining string edit distance and structural constraints achieves 95% precision even with only two distinct labels.
· “Practical Threshold-based Tree Edit Distance Lower-Bounds”
Compiler stack turns ordinary indexes into remote-memory ones
Zero code changes: one runtime gives B+-trees, hash tables, and skip lists caching and address translation.
· “Decoupling Disaggregated Memory Optimizations from Indexing: A Compiler-Runtime Approach”
120x memory inflation from 10% poison keys
Optimal PLA in the PGM-index is not robust; the attack transfers across learned indexes.
Graph search follows a power law
New experiments and theory show that the cost of graph-based vector search scales as N^c for most dataset sizes, not polylog N.
· “A Power Law in Logarithm's Clothing: On the Scalability of Graph-Based Vector Search”
Zero-LLM NL2QL hits 100% execution accuracy under 5 ms
A deterministic seven-stage pipeline with runtime confidence scores generates SQL and GraphQL from one architecture, no API cost.
0.2 seconds to clone 100 GB: database branching at metadata speed
By turning table versions into lightweight metadata, Git4Data enables thousands of AI agents to fork, compare, and merge data with SQL.
Dual-metric partitioning speeds GCN on GPU by 2.5x
Combining degree and random-walk density balancesworkload and slashes execution time across 12 real graphs.
· “Dual-Metric Partitioning with Adaptive Kernel Execution for Efficient GCN Acceleration”
GPU ray-tracing cores accelerate high-D vector search 8×
First exact algorithm uses 3D hardware indexing plus CUDA refinement; runs on datasets with up to 128 dimensions.
· “RT-HiSS: Ray Tracing Accelerated High Dimensional Vector Similarity Searches”
A 2.87MB in-browser SQL engine that branches like Git
Its log-centric MVCC design gives overlapping transactions and whole-database fork, merge, and rebase in the browser.
· “Zeta-Lite: A Concurrent, Branchable In-Browser SQL Database for Agentic Memory”
Thermal drone screening achieves 92.5% precision on inert ordnance
But biased validation set means real-world false alarms could be much higher; tool is for prioritization, not clearance.
A SQL engine can outrun native graph engines on analytics
No import, no separate cluster: translating Cypher to columnar SQL queries relationships the tables already hold.
Open-source RTGL catches time leaks in benchmark tasks
By enforcing strict temporal boundaries, RTGL automatically prevents the future data contamination found in manual SQL.
½? Found: UCCs on 180M tuples with <8 GB memory
New algorithm avoids memory blowup by generating only partial difference sets and pruning 90% of candidates without validation.
· “Efficient discovery of unique column combinations on disk-resident data with limited memory”
ContextPipe cuts token volume by 31% with database-style planning
A deterministic optimizer reduces LLM calls by 23% on coding tasks, trading lower cache-hit ratio for faster, smaller prompts.
· “ContextPipe: Database-Inspired Context Assembly for Long-Horizon Agents”
3.5× faster time-aware vector search with exact scoring preserved
A new graph index ranks vectors by semantic similarity and freshness jointly, and is up to 4× faster to build.
Edge hash joins cut cache misses by 54% with workload-aware bins
WACI-HJ improves query time by 10% on real traffic data, enabling faster edge analytics.
Storage-aware privacy beats classical PIR on three graph families
Under local privacy, two-replicated servers retrieve more symbols per download than classical PIR permits.
· “Local Private Information Retrieval for Graph-Based Replicated Systems”
Agentic data cracking halves LLM agent token cost
By adaptively structuring data during reasoning, agents avoid reopening documents for related queries, cutting costs 53% without accuracy…
GRBench is a new benchmark for evaluating database systems that handle both relational…
Real-world dataset and controlled query series expose performance differences across seven open-source systems.
· “GRBench: A Comprehensive Benchmark Evaluation for Graph-relational Data Management”
3x throughput, 80% less memory: ByteX scales to trillion vectors
SymRaBitQ quantization lets graph construction run on compact codes, slashing cost and memory while keeping high recall
Repaired benchmark enables strict RDF parsing for federated SPARQL evaluation
RDF and result-format fixes in LargeRDFBench expose deeper reproducibility obstacles.
· “Strengthening LargeRDFBench for Interoperable Federated SPARQL Evaluation”
Equiv query variants uncover 54 DBMS logic bugs
ERIQ compares VIEW, CTE, and TEMPT to expose logic bugs missed by prior methods.
· “Detecting DBMS Bugs by Constructing Equivalent Representations of Intermediate Query Results”
Trajectories alone time soccer events to 96.5% accuracy
A trajectory-only synchronizer also recovers ball receptions, the event ends other methods miss.
· “ELASTIC: Trajectory-Based Synchronization of Event and Tracking Data in Soccer”
Most cache promotions unnecessary
Analysis of 6,357 production traces shows that 90% of LRU promotions can be skipped, and the two new methods are ready for deployment.
· “Demystifying and Improving Lazy Promotion in Cache Eviction”
MAGE is a database that records what multi-agent AI systems do as a network of…
A heterogeneous temporal hypergraph memory engine for multi-agent systems outperforms eleven memory baselines on 12 QA benchmarks by…
· “Diachronic Hypergraphs for Orchestrated Multi-Agent Multimodal Memory Curation”
No model beats 5% at spotting both clarification layers
And when users retract an earlier choice, top LLMs recognize it but fix the SQL only about half the time.
· “Evaluating LLMs on Conversational Text-to-SQL under Chain Ambiguity and Intent Drift”
Static profiles pick SPARQL sources without a single ASK query
RENSA infers variable classes and namespaces from compact profiles, matching live-probing engines at zero runtime cost
18 resources converge on community curation best practices
A workshop report finds proactive outreach, AI-assisted validation, and shared infrastructure are key to sustaining shrinking databases
One pipeline module pays off every time
Execution-feedback refinement is the only universally cheap gain in text-to-SQL; mid-tier backbones with rich pipelines beat frontier…
GaussDB beats TPC-H record by 40% at 30TB
A series of optimizations—pipeline execution, scalable shuffle, distributed Bloom filters—pushes performance beyond previous best by 40%…
· “No Silver Bullet: Boosting GaussDB Performance on the 30TB TPC-H Workload”
100× faster verification for blockchain time-series queries
A single authenticated tree serves as index, proof, and anchor; an untrusted encoder feeds approximate answers with certified bounds.
· “VeriTS: Verifiable Model-Enhanced Time-Series Queries on Blockchain Systems”
All T2I models lose count above 50 objects
A 640K-prompt diagnostic benchmark separates cardinality, binding, and visibility errors across nine systems.
· “NumBench: Diagnosing Counting Failures in Text-to-Image Models”
Synthetic databases replay slow queries without touching production data
A hybrid solver keeps global statistics intact while forcing exact cardinalities, so offline diagnosis of slow queries becomes practical.
Bind integrity to version histories to secure disaggregated key-value stores
ANCHOR treats caches as untrusted hints and uses version-bound authentication to prevent rollback and poisoning without blocking async I/O.
· “ANCHOR: A Vision for Secure Persistent Key-Value Stores in Disaggregated Data Centers”
Declarative primitives cut agentic harness costs 10x
Credo recovers reusable skeletons, beliefs, and prompts from opaque search outputs, compiling harnesses for new tasks without restarting…
· “Credo: Reusable Declarative Primitives for Agentic Workflows”
Real-time verification stops 189,000 SQL regressions in five days
Oracle's Real-Time SPM validates plans during user queries, cutting background resource use and latency.
Two online learners cut LLM query cost by ~8–11×
Hiding each training step inside the LLM round-trip lets the two learners stack into a near-order-of-magnitude saving.
· “Compositional Online Learning for Semantic Data Processing Systems”
Exact per-predicate Shapley for sliding aggregates in constant time
IDS incrementally maintains three additive summaries and evaluates any closed-form attribution without rescanning the window.
· “Incremental Delta-Shapley: A Standalone Runtime for Predicate Attribution on Sliding Windows”
Projections widen bound gap under acyclic constraints
Even with acyclic functional dependencies, the computable polymatroid bound can be polynomially larger than the tight entropic bound when…
A failed IBLT still measures how many items it missed
The pre-peeling counts give an unbiased difference size, so retry capacity is set in one round with no extra bytes.
· “IBLTs Measure Before They Decode: Self-Sizing Set Reconciliation from Pre-Peeling Counts”
Home activity benchmark shows AI question-answering gaps
HOME-KGQA tests multimodal KGQA on daily household tasks, where LLM methods lag behind their encyclopedic results.
DivSkill-SQL lifts Text-to-SQL accuracy by up to 11 points
Skills optimized on prior failures deliver complementary candidates that cut correlated errors and transfer across dialects.
eBPF scheduler doubles throughput for time-sensitive DB tasks
UFS limits background tasks to idle CPUs and shares lock hints to halve tail latency versus standard Linux options in mixed workloads.
· “Unfair by design: eBPF-based scheduling of mixed database workloads”
Streaming overlaps cut LLM first response time by 11x
New system handles dynamic context retrieval concurrently while preserving overall throughput
· “Stream2LLM: Overlap Context Streaming and Prefill for Reduced Time-to-First-Token (TTFT)”
SQL metapath pruning cuts GNN training time by up to 10x
Selecting only informative foreign-key paths before training shrinks per-epoch cost while keeping accuracy.
· “MetaSieve: Faster Relational Deep Learning through SQL-Based Metapath Selection”
Weighted multilayer hubs need no graph flattening
For sum aggregation it's exact; for max, an upper-bound heuristic plus top-k candidates recovers most hubs.
· “Degree Centrality Algorithms for Weighted Multilayer Networks (or w-MLNs)”
