Oracle-Budgeted Molecular Optimization with Short-Term Graph Memory
The paper adds a short-term graph-memory filter so limited oracle calls go to more promising molecules.
The module keeps the underlying generator and update rule intact, but uses an online graph neural surrogate to pre-screen each candidate pool. On a standard molecular optimization benchmark, it raised the mean top-10 score without extra oracle cost and did not trail the base method on any oracle. The authors report the gain across four tested generators under a 1,000-call budget, then tie larger-budget benefits to how each backbone balances search breadth and oracle-driven exploitation. ArXiv · AI/CL/LG's note
The module keeps the underlying generator and update rule intact, but uses an online graph neural surrogate to pre-screen each candidate pool. On a standard molecular optimization benchmark, it raised the mean top-10 score without extra oracle cost and did not trail the base method on any oracle. The authors report the gain across four tested generators under a 1,000-call budget, then tie larger-budget benefits to how each backbone balances search breadth and oracle-driven exploitation. ArXiv · AI/CL/LG's note
score 4