Meta-Delta: Evolving Knowledge through Global Revision for LLM-Driven Optimization

LLM-driven evolutionary optimization can accumulate useful search experience, but local updates alone leave a critical failure mode: knowledge that was useful earlier may become stale or misleading as the search progresses. We introduce Meta-Delta, an agent framework centered on the global revision of search knowledge while jointly evolving candidate solutions. A role-separated search loop plans, executes, evaluates, and audits candidate changes. A Delta Bank Manager distills these outcomes into structured Delta Cards, making search lessons, applicability conditions, and supporting evidence explicit and revisable within a persistent Delta Bank. Beyond routine local updates, a global review step examines the Delta Bank and broader search history to revise prior conclusions, reorganize knowledge, and propose directions for subsequent testing. This second level of evolution lets the agent reconsider the guidance for future candidate generation. Empirically, Meta-Delta achieves the best result on 9 of 11 tasks across mathematical, scientific, and systems optimization, and shows competitive aggregate performance on the 118-problem SCILAWS-REAL benchmark. Global-review ablations and a same-checkpoint case study provide evidence that broader knowledge revision helps sustain progress and overcome stagnation. These results support treating search knowledge as an evolving component of long-horizon optimization.