Temp-R1 AI Agent Achieves Breakthrough in Temporal Knowledge Graph Question Answering
An innovative autonomous agent named Temp-R1 has been created for Temporal Knowledge Graph Question Answering, achieving top-tier results on benchmark datasets. This system employs reinforcement learning with an enhanced action space that incorporates both specialized internal and external actions. To mitigate shortcut learning, researchers introduced reverse curriculum learning, initially tackling challenging questions before progressing to simpler ones. This model, consisting of 8 billion parameters, surpasses robust baselines by 19.8% on intricate questions. Traditional methods have faced limitations due to rigid workflows and costly closed-source APIs, which Temp-R1 addresses with its comprehensive autonomous framework. The agent showcases advanced reasoning capabilities over dynamic facts with multi-hop dependencies and complex temporal constraints, demonstrating notable improvements on the MultiTQ and TimelineKGQA datasets in tackling difficult temporal reasoning challenges.
Key facts
- Temp-R1 is an autonomous end-to-end agent for Temporal Knowledge Graph Question Answering
- It uses reinforcement learning with reverse curriculum learning
- The model has 8 billion parameters
- It achieves 19.8% improvement over strong baselines on complex questions
- It was tested on MultiTQ and TimelineKGQA datasets
- The system expands action space with specialized internal actions
- Reverse curriculum learning trains on difficult questions first
- Existing methods rely on fixed workflows and expensive closed-source APIs
Entities
—