notesum.ai
Published at October 22CoPS: Empowering LLM Agents with Provable Cross-Task Experience Sharing
cs.CL
cs.AI
Released Date: October 22, 2024
Authors: Chen Yang1, Chenyang Zhao1, Quanquan Gu1, Dongruo Zhou1
Aff.: 1Institution Not Specified

| Algorithm | Performance | |
| Llama3.1 8b | Llama3.1 70b | |
| Reflexion 222The original codebase of Reflexion struggles to perform on most tasks with the smaller Llama3.1 8b model. This is primarily because the model tends to repeat the same action, leading to task failure. To mitigate this, we introduced a resampling mechanism to enhance Reflexion performance, which activates when the model begins to repeat actions. This modification significantly improved Reflexion’s performance. | 86 | 94 |
| RAP | 70 | 93 |
| CoPS | 94 | 100 |