VLDB 2025: AI Meets Enterprise Data Management — The Tabular FM Moment

Snapshot from London (Sep 1–5, 2025).
1,100+ attendees from 40+ countries, 1,600+ submissions, roughly ~20% acceptance. VLDB is a flagship venue for the data management community, right alongside SIGMOD.

I attended to particpate in an invited panel discussion on Neural Relational Data, discussing the need for bespoke foundation models on structured data.

The Big Picture: AI-for-DM and DM-for-AI

Two complementary shifts are gathering pace:

AI-for-DM — LLMs and agents do data management (data engineering, schema matching, discovery, evaluation).
DM-for-AI — Databases run AI better (new operators, efficient vector support, and data architectures that serve agentic workloads).

The insightful and actionable keynote by Matei Zaharia (CTO and Co-Founder of Databricks) touched on both: databases must evolve for agentic, speculative access patterns [1] and for portable data products that retain semantics across stacks [5].

Spotlight: Foundation Models on Tabular Data (TFMs)

If you want to catch the next wave of GenAI early, the most important trend is the rapid maturation of foundation models for relational/tabular data. VLDB 2025 hosted a panel on “Neural Relational Data: Tabular Foundation Models, LLMs… or both?”, bringing together TFM and LLM camps, with emphasis on multi-table learning and semantics. I had the opportunity to participate as an invited speaker. [9]

Panel Discussion on Neural Relational Data

Why this matters

Most business value lives in databases: ERP, CRM, finance, supply chain, telemetry. Traditional ML has long excelled here, but foundation models promise:

Generalization across schemas and domains,
Few-shot performance on new tasks,
Re-usable representations that downstream apps (forecasting, optimization, anomaly detection) can tap into.

What changed in 2025

From single tables to relational context. Models are learning across multiple linked tables and capturing business semantics rather than isolated columns.
This will broaden to Semantically Linked Tables (SLT): LLMs translate questions to SQL, but predictive & prescriptive tasks need context beyond one or even multiple tables and their schema: declaratic and procedural business knowledge as well as external signals. [8]
Knowledge as structure, not just text. Graph-style linking of entities, processes, and external knowledge provides explicit grounding, traceability, and provenance—crucial for regulated industries in an in-content learning setting.
Synthetic data (e.g., synthetic time-series corpora [6]) is accelerating training and evaluation without exposing sensitive datasets.

Semantically Linked Tables - what's in it? — Semantically Linked Tables – what’s in it? from: Klein & Hoffart. Foundation Models for Tabular Data within Systemic Contexts Need Grounding. arxiv.org/abs/2505.19825

What this enables

Better demand forecasting, risk scoring, next-best-action, fraud/outage detection, and scenario planning—with less feature plumbing and faster time-to-value.
Retrieval-augmented analytics over structured data: models “pull” the right tables, rows, and business rules on demand.
Explainable decisions via grounded joins and lineage, not opaque text-only chains.

What Enterprises Should Do Now

Lay the semantic foundation. Invest in a linked semantic layer (business objects, relationships, business rules) that bridges operational code, data catalogs, and external knowledge—think Semantically Linked Tables and grounded FM training [8].
Tame schema drift and data integration with AI. Use retriever → LLM reranker pipelines (SLM+LLM) for scalable, explainable schema alignment across apps and extensions [2].
Use synthetic data strategically. Generate task-specific corpora (SQL, time-series, logs) for safe training/eval; keep real-world validation loops [3], [6].
Close the loop on governance and make data accessible for training. Track provenance, time-travel, and branches to enable fine-tuning of foundation models on relational data.

Adjacent Trends and Highlights from VLDB 2025 (In Brief)

Make data portable and self-describing. Favor open formats/protocols and “self-decoding” datasets so products move across platforms without losing meaning [4], [5].
Prepare for agentic traffic. Expect bursty, speculative probes from AI agents. Add caching, shaping, and accuracy/latency knobs at the data layer, plus interfaces, processing paths, and agent memory stores designed for autonomous workflows [1].
Lakehouse → “Lakebase” patterns: OLTP+OLAP unification and app/agent runtime on shared object-store data (with copy-on-write for safe experimentation and agentic development workflows).
Data source discovery: Learned indexes that unify indexing and search to help agents find the right tables fast [7].

Bottom Line

The next wave of enterprise AI won’t be just better LLMs and Agents. It will be foundation models on relation data including semantically linked tables trained on agent-aware, portable, semantically rich data systems—so your models learn from the right context and your decisions stand up to scrutiny.