What a temporal graph benchmark can reveal
Why controlled tasks help us ask more precise questions about a model's ability to reason across time.
A strong result on a graph dataset is useful. Understanding which temporal skills produced that result is a different question.
Temporal graph models encounter relationships that change over time: interactions repeat, an event can have a delayed consequence, and relevant information can be distant in both the graph and its history. A single aggregate score can hide which of these patterns a model actually captures.
Isolate the skill
T-GRAB, a paper I coauthored, studies these questions with controlled synthetic tasks. The benchmark separates skills such as remembering repeated events, tracking delayed effects, and using long-range dependencies. Its evaluation of 11 temporal graph learning methods reveals gaps that can be difficult to diagnose from real-world benchmarks alone.
The practical value is diagnostic: when a model struggles, the task gives us a clearer account of what it struggled with.
From a score to a question
For someone building or evaluating a temporal learning system, this suggests a useful sequence:
- Identify the temporal pattern the application needs.
- Test that pattern in a controlled setting.
- Examine where performance changes as the task becomes harder.
- Return to realistic data to check whether the lesson transfers.
Synthetic tasks simplify the world. Success on them does not establish that a model will handle every real-world setting. Their value is that they make some failure modes easier to see and discuss.
Explore the work
The KDD 2026 paper, open-source code, and dataset provide starting points for exploring these questions.