Contrastive Representations for Temporal Reasoning
12 days ago
- #temporal reasoning
- #Rubik's Cube
- #contrastive learning
- Key question: Can temporal reasoning be improved by learning better representations?
- Method: Introduces CRTR, a novel negative sampling scheme to remove spurious features and enable temporal reasoning.
- Result: CRTR successfully solves complex temporal tasks like Rubik's Cube without hand-crafted heuristics.
- CRTR outperforms baselines (CRL, Supervised, DeepCubeA) in success rates across multiple domains.
- CRTR's learned representations generalize well, reflecting accurate temporal structure in distances.
- Surprisingly, CRTR achieves higher success rates on Rubik's Cube without search, solving all configurations within budget.
- CRTR exhibits block-building-like behavior in solving Rubik's Cube, gradually constructing partial structures.