SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization
Researchers from Tsinghua University have introduced SE-Bench, a novel diagnostic environment designed to rigorously benchmark the self-evolution capabilities of AI agents through knowledge internalization. Addressing challenges like prior knowledge entanglement and reasoning complexity, SE-Bench obfuscates the NumPy library into a pseudo-novel package with randomized identifiers. This setup ensures that coding tasks are impossible for base models without the new documentation but trivial once the API is internalized. The study highlights three critical insights: the 'Open-Book Paradox,' where access to reference docs during training hinders retention; the 'RL Gap,' indicating standard Reinforcement Learning fails to fully internalize knowledge due to PPO clipping; and the viability of Self-Play combined with Supervised Fine-Tuning (SFT) for effective learning. This platform provides a clean setting to evaluate lifelong learning in agents, distinguishing true knowledge compression from mere retrieval. The associated code and dataset are publicly available, offering a significant tool for advancing autonomous agent development and understanding the mechanics of how AI models integrate new information into their weights.
Wire timeline
SE-Bench: Benchmarking Self-Evolution with Knowledge Internalization
Researchers from Tsinghua University have introduced SE-Bench, a novel diagnostic environment designed to rigorously benchmark the self-evolution capabilities of AI agents through knowledge internalization. Addressing challenges like prior knowledge entanglement and reasoning complexity, SE-Bench obfuscates the NumPy library into a pseudo-novel package with randomized identifiers. This setup ensures that coding tasks are impossible for base models without the new documentation but trivial once the API is internalized. The study highlights three critical insights: the 'Open-Book Paradox,' where access to reference docs during training hinders retention; the 'RL Gap,' indicating standard Reinforcement Learning fails to fully internalize knowledge due to PPO clipping; and the viability of Self-Play combined with Supervised Fine-Tuning (SFT) for effective learning. This platform provides a clean setting to evaluate lifelong learning in agents, distinguishing true knowledge compression from mere retrieval. The associated code and dataset are publicly available, offering a significant tool for advancing autonomous agent development and understanding the mechanics of how AI models integrate new information into their weights.
cs.AI updates on arXiv.org