NeurIPS 2026 Workshop Proposal Structured Data Scientific Discovery

AgentSD: Agentic Systems for Structured Data

The first workshop dedicated to grounding agentic systems on the structured and scientific data modalities the world actually runs on: tables, time series, graphs, relational databases, and domain-specific scientific structures.

Venue NeurIPS 2026 Workshop date and venue TBA

Format Full-day workshop Invited talks, contributed talks, posters, and panel discussion

Submissions OpenReview Deadlines and submission link to be announced

Agentic systems and machine learning for structured data have advanced rapidly, but mostly along separate tracks. Agents increasingly use tools, memory, and planning to solve long-horizon tasks, while structured-data research builds specialized models for tabular data, time series, graphs, relational data, and scientific domains.

The gap matters. Treating a database, graph, crystal, molecule, or time series as plain text often loses the algebraic, relational, temporal, and geometric constraints that make the data meaningful. At the same time, many structured-data models are not embedded in agents that can plan experiments, test hypotheses, call tools, and refine scientific arguments.

AgentSD brings together researchers in agents, structured-data learning, databases, scientific machine learning, neuro-symbolic reasoning, and evaluation to develop reliable agentic systems over structured data.

Topics of Interest

We welcome submissions on agentic reasoning, tool use, and action over structured and scientific data, including but not limited to:

Agent Architectures, Planning, and Tool Use

Code-as-action over data through SQL, pandas, Spark, solvers, query engines, and active schema exploration.

Structured Representations for Agents

Representations that preserve schemas, geometry, relational constraints, and other domain-specific structure during context construction.

Specialist Models as Tools and Backbones

Integrating tabular, time-series, graph, relational, and scientific foundation models inside agent loops.

Neuro-Symbolic Agents

Coupling language models with formal solvers, reasoners, query planners, and verification systems.

Scientific Structured Data

Agentic reasoning for proteins, materials, molecules, crystals, property prediction, design inversion, and hypothesis generation.

Relational and Tabular Tasks

Text-to-SQL, schema discovery, conversation analysis, data cleaning, database agents, and relational reasoners.

Time-Series and Evolving Data

Agents over nonstationary data, shifting schemas, regime changes, online adaptation, and distribution drift.

Graphs and Knowledge Structures

Planning, search, routing, supply-chain reasoning, graph retrieval-augmented generation, and knowledge-graph question answering.

Evaluation and Benchmarks

Benchmarks for querying, transformation, analysis, reasoning, and synthetic generation over tables, time series, and graphs.

Reliability, Trust, and Governance

Calibration on quantitative outputs, explainable actions, privacy, safety guardrails, governance, and auditability.

Confirmed Speakers

The invited speaker lineup reflects the workshop's interdisciplinary focus across agentic systems, structured data, graph and time-series learning, and scientific discovery.

Virginia Smith Carnegie Mellon University

Michael Rabbat Meta FAIR

Mihaela van der Schaar University of Cambridge

Rose Yu UC San Diego

Frank Hutter University of Freiburg / Prior Labs

Submissions

AgentSD invites technical papers, position papers, benchmark papers, and early-stage research that advances agentic systems for structured and scientific data.

Submissions will be handled through OpenReview. Detailed formatting instructions, submission categories, and deadlines will be posted after workshop scheduling is finalized.

Submission portal OpenReview Link TBA