Demo paper presents PrismaDV, a system that synthesizes data unit tests by analyzing both the data and the code that uses it
At the 35th ACM International Conference on Information and Knowledge Management (CIKM 2026), held November 7 to 11, 2026, in Rome, Italy, BIFOLD researcher Hao Chen will present the demo paper “Will This Data Break My Task? Interactive Synthesis of Task-Aware Data Unit Tests”, co-authored with Arnab Phani and Sebastian Schelter.
Data errors that propagate through data pipelines can seriously affect production applications. Data unit tests, executable programs that check data before it moves on through a pipeline, help prevent this. Existing frameworks, however, derive these tests from the data alone and ignore how the code that later consumes the data actually uses it.
The paper presents PrismaDV, a compound AI system that combines several AI-powered components to generate task-aware data unit tests for tabular data. It analyzes the data and the downstream task code together, in several steps based on large language models: profiling the data, detecting which columns the code accesses, analyzing the data flow in the code, and inferring implicit assumptions about the data. PrismaDV then synthesizes the code for the data unit test and keeps a “data-code assumption graph” that links each generated data constraint back to the task’s source code.
In the interactive web-based demo, attendees can run PrismaDV on five real-world datasets with 60 downstream tasks. They can inspect and refine both the natural-language assumptions about the data and the executable data constraints, navigate the assumption graph, and compare task-aware tests against task-agnostic baselines on erroneous data batches. They can also observe how a custom prompt optimizer adapts the system to specific datasets over time.
About CIKM
CIKM focuses on designing and developing future information and knowledge systems. The 2026 edition places particular emphasis on how AI is changing information access, knowledge representation, data integration, personalization, and evaluation.
CIKM 2026 Presentation
- Presentation: Poster Session 3, Wednesday, November 11, 2026, 14:00 to 15:30, Poster Area B, board B05
- Title: Will This Data Break My Task? Interactive Synthesis of Task-Aware Data Unit Tests
- Authors: Hao Chen, Arnab Phani, Sebastian Schelter
- Preprint
- Proceeding
