Experience Bank
An external memory mechanism that structures historical judgment experience into reusable knowledge, distilling evaluation patterns across instances into stable and adaptive judgment criteria.
Timeline (last 90 days)
This study uses one complete run log on pockets from the CrossDocked2020 test set to examine whether the keys of the experience library recur on unseen pockets and still distinguish improving edits from worsening edits.
The results show that the experience library can clearly distinguish improving edits from worsening edits.
Further analysis found that the experience library's ability to distinguish improving from worsening edits can largely be explained by the edit's change to molecular size, the action type, and the round, none of which describe the target itself.
Under the representation and lookup scheme of the current study, the experience library may be better suited to prioritizing the docking suggestions proposed by the agent itself, rather than directly prescribing what molecules should be built.
Based on the analysis of a single run log, this study provides in-depth insight into the experience reuse mechanism, but the transferability of the experience library is limited by the inadequacy of target descriptions.
知识与智能可以在一定程度上解耦：参数对应知识存储，FLOPs对应推理智能
All Facts (6)
The results show that the experience library can clearly distinguish improving edits from worsening edits.
50%UnverifiedFurther analysis found that the experience library's ability to distinguish improving from worsening edits can largely be explained by the edit's change to molecular size, the action type, and the round, none of which describe the target itself.
50%UnverifiedUnder the representation and lookup scheme of the current study, the experience library may be better suited to prioritizing the docking suggestions proposed by the agent itself, rather than directly prescribing what molecules should be built.
50%UnverifiedBased on the analysis of a single run log, this study provides in-depth insight into the experience reuse mechanism, but the transferability of the experience library is limited by the inadequacy of target descriptions.
50%Unverified知识与智能可以在一定程度上解耦：参数对应知识存储，FLOPs对应推理智能
50%UnverifiedThis study uses one complete run log on pockets from the CrossDocked2020 test set to examine whether the keys of the experience library recur on unseen pockets and still distinguish improving edits from worsening edits.
50%