The authors analyze systems that jointly improve a policy, which selects and applies skills, and a bank of reusable skills.
Co-evolved agents lose gains when their skill banks change
Across four methods, the authors find that policies become increasingly dependent on the particular skill banks used during training.
Chinese Tech
Yujie Feng · Xiaobin Wang · Jinluan Yang · Xiaonan Shi · Pengjun Xie · Philip S. Yu · +2 more
Alibaba Token Hub, Alibaba Group · The Hong Kong Polytechnic University · University of Illinois at Chicago
Research Digest··2 min read
Feng and colleagues examine what happens when a self-improving agent’s skill library changes after its policy has been trained.
Why this paper
From Alibaba Token Hub, Alibaba Group and 2 others
In one line
Co-evolved skill-policy pairs are fragile: performance drops sharply when deployed skills drift, erasing most gains.
What we could check
- ·No code link found
- ·No weights link found
- ·No dataset link found
- ·No compute details found
- ✓Limitations stated by the authors
- ·No benchmark numbers found
Observed from the paper text and links we have. Absence here means we did not find it, not that it does not exist.
§