TRACTA: A Neuro-Symbolic Benchmark for Temporal Reasoning in High-Complexity Environments
The study presents TRACTA (Temporal Reasoning and Capability-Trajectory Analysis), a synthetic benchmark designed for evaluating temporal structural reasoning within complex event-driven systems, modeled after Multi-Domain Operations (MDO) scenarios. It comprises three distinct tasks: early_warning, pattern_detection, and run_classification. The evaluation contrasts raw-event neural models, a simplified semantic baseline, and a neuro-symbolic setup that utilizes semantically grounded trajectories. Findings reveal that while raw event-level learning provides valuable insights, temporal modeling based on semantic capabilities and contextual direct-impact trajectories yields the best overall point estimates, particularly excelling in temporal tasks. Additionally, ablation studies demonstrate that the neuro-symbolic method surpasses others in identifying and predicting temporally distributed patterns.
Key facts
- TRACTA is a benchmark for temporal reasoning in high-complexity environments.
- It is instantiated through Multi-Domain Operations (MDO)-like scenarios.
- Three tasks: early_warning, pattern_detection, run_classification.
- Compares raw-event neural models, contract-lite semantic baseline, and neuro-symbolic configuration.
- Neuro-symbolic configuration uses semantically grounded trajectories.
- Neuro-symbolic approach achieves highest aggregate point estimates.
- Largest margins on temporal tasks.
- Raw event-level learning remains informative.
Entities
—