Diagnostic queries
Crisis needs designed to expose grounding failures.
← All publications
Under review
ST-Crisis: A Diagnostic Benchmark for Spatiotemporal Grounding in Crisis Communication
A diagnostic benchmark for testing whether crisis-response systems use evidence from the right place and time.
The problem
Why this work matters
A crisis response can sound plausible while relying on information from the wrong location or an outdated moment. Existing evaluations do not adequately diagnose whether generated answers are grounded in semantically, spatially, and temporally appropriate evidence.
Methodology
How the approach works
01Crisis need query
02Semantic relevance
03Spatial alignment
04Temporal alignment
05Grounded response
Contributions
What this work contributes
- 01
A benchmark centered on the joint spatial and temporal grounding demands of crisis communication.
- 02
Crisis-need queries paired with situational-awareness posts and controlled evidence chains.
- 03
Diagnostic evaluation separating semantic relevance from geographic and temporal correctness.
- 04
A foundation for evaluating spatiotemporally aware retrieval and response-generation systems.
Evaluation
Key results
Grounding dimensions
Semantic, spatial, and temporal evidence alignment.
Post chains
Multi-post evidence chains test compositional reasoning.
Under-covered queries
After benchmark enrichment and coverage auditing.