Going Beyond AlphaEvolve in Agent Scientific Discovery

Google DeepMind’s AlphaEvolve made a key insight clear: hashtag#AgenticAI can act as a team of evolutionary scientists, proposing meaningful algorithm changes inside an evaluation loop. AlphaEvolve and similar methods also share a fundamental limitation. Each mutation overwrites the structure. Earlier variants become inert. Partial improvements cannot be recombined. Credit assignment is global and coarse. Over long horizons, evolution becomes fragile. I introduce EvoLattice, which removes this limitation by changing the unit of evolution itself. Instead of evolving a single program, EvoLattice evolves an internal population encoded inside one structure. A program (or agent) is represented as a DAG where each node contains multiple persistent alternatives. Every valid path through the graph is executable. Evolution becomes additive, non-destructive, and combinatorial — not overwrite-based. We evaluate EvoLattice on NAS-Bench-Suite-Zero, under identical compute and evaluation settings. EvoLattice outperforms AlphaEvolve, achieves higher rank correlation, exhibits lower variance and faster stabilization, and improves monotonically without regression. We further validate generality on training-free optimizer update rule discovery, where EvoLattice autonomously discovers a nonlinear sign–curvature optimizer that significantly outperforms SGD, SignSGD, Lion, and tuned hybrids — using the same primitives and no training.

Why this matters? Persistent internal diversity: AlphaEvolve preserves diversity across generations. EvoLattice preserves it inside the program. Strong components never disappear unless explicitly pruned. Fine-grained credit assignment: Each micro-operator is evaluated across all contexts in which it appears, producing statistics (mean, variance, best-case). AlphaEvolve only sees a single scalar score per program. Quality–Diversity (QD) without archives: EvoLattice naturally exhibits MAP-Elites-style dynamics: monotonic improvement of elites, widening gap between best and average, bounded variance — without external archives or novelty objectives. Structural robustness: AlphaEvolve relies on the hashtag#LLM to preserve graph correctness. EvoLattice applies deterministic self-repair after every mutation, removing structural fragility from the loop.

AlphaEvolve shows how hashtag#LLMs can mutate programs. EvoLattice shows what they should evolve: the internal computational fabric, not entire programs. This turns LLM-guided evolution from a fragile rewrite process into a stable, cumulative, QD-driven discovery system. The same framework applies to prompt and agentic workflow evolution. As agent systems grow deeper and more interconnected, overwrite-based evolution breaks down. EvoLattice’s internal population and self-repair make long-horizon agentic evolution feasible and interpretable.