DeltaEvolve: Accelerating Scientific Discovery through Momentum-Driven Evolution

LLM-driven evolutionary systems have shown promise for automated science discovery, yet existing approaches such as AlphaEvolve rely on full-code histories that are context-inefficient and potentially provide weak evolutionary guidance. In this work, we first formalize the evolutionary agents as a general Expectation-Maximization framework, where the language model samples candidate programs (E-step) and the system updates the control context based on evaluation feedback (M-step). Under this view, constructing context via full-code snapshots constitutes a suboptimal M-step, as redundant implement details dilutes core algorithmic ideas, making it difficult to provide clear inspirations for evolution. To address this, we propose DeltaEvolve, a momentum-driven evolutionary framework that replaces full-code history with structured semantic delta capturing how and why modifications between successive nodes affect performance. As programs are often decomposable, semantic delta usually contains many effective components which are transferable and more informative to drive improvement. By organizing semantic delta through multi-level database and progressive disclosure mechanism, input tokens are further reduced. Empirical evaluations on tasks across diverse scientific domains show that our framework can discover better solution with less token consumption over full-code-based evolutionary agents.

💡 Research Summary

The paper addresses a fundamental inefficiency in large‑language‑model (LLM)‑driven evolutionary systems for automated scientific discovery. Existing agents such as AlphaEvolve store full‑code snapshots of past solutions in the context window. Because code can be long, this approach quickly exhausts the limited token budget and, more importantly, mixes core algorithmic ideas with irrelevant implementation details, weakening the evolutionary guidance that the context provides.
To formalize the problem, the authors cast the evolutionary loop as an Expectation‑Maximization (EM) process. In the E‑step the LLM samples candidate programs conditioned on a context Cₜ; in the M‑step the system updates Cₜ using the accumulated evaluation history Hₜ to maximize expected reward. The context C therefore acts as the sole learnable variable when model weights are fixed. The paper argues that the conventional practice of feeding full programs into C is a sub‑optimal M‑step because it obscures which modifications actually caused performance changes.
DeltaEvolve is introduced as a “momentum‑driven” alternative that replaces full‑code history with structured semantic deltas. A semantic delta records what changed between a parent node and its offspring and why the change improved (or worsened) performance. Since programs are often compositional, these deltas capture reusable algorithmic components that are far more informative than raw code. The system stores deltas in a three‑level pyramid database: Level‑1 holds a concise summary, Level‑2 contains a detailed plan of the modification, and Level‑3 retains the full source code for the current parent. A Progressive Disclosure Sampler decides, for each historical node, which level to expose based on relevance and recency, thereby dramatically reducing token consumption while preserving the essential evolutionary signal.
Empirically, the authors evaluate DeltaEvolve on five diverse scientific domains—including mathematical optimization, physical system modeling, and molecular discovery—and on a synthetic black‑box optimization benchmark. Across all tasks, DeltaEvolve discovers solutions of comparable or superior quality to state‑of‑the‑art baselines while cutting total token usage by an average of 36.79 %. In the black‑box benchmark, the method consistently achieves higher scores with fewer cumulative tokens, and the token‑vs‑performance curve dominates those of full‑code baselines regardless of how many top‑k or diverse programs are included in the context.
Beyond the experimental results, the paper contributes a clear theoretical perspective: it identifies the M‑step (context construction) as the true bottleneck in LLM‑based evolutionary agents and demonstrates that a momentum‑like memory of semantic deltas can serve as an effective inductive bias. By structuring context not merely as compressed text but as a purposeful guide that highlights successful modifications, DeltaEvolve advances the efficiency and scalability of automated scientific discovery with LLMs.