What Makes an LLM a Good Optimizer? A Trajectory Analysis of LLM-Guided Evolutionary Search
A study of 15 LLMs across 8 tasks shows that steady refinement, not raw power, predicts optimization success.
A new research paper titled 'What Makes an LLM a Good Optimizer? A Trajectory Analysis of LLM-Guided Evolutionary Search' provides a deep dive into why some large language models excel at guiding evolutionary optimization. The study, conducted by Xinhao Zhang, Xi Chen, François Portet, and Maxime Peyrard, systematically analyzed the optimization trajectories of 15 different LLMs across 8 distinct tasks. This large-scale analysis moves beyond simply measuring final outcomes to understand the *process* of how LLMs search for solutions, revealing that models with similar initial capabilities can produce dramatically different search paths and results.
The core finding is that the most effective LLM optimizers behave as 'local refiners.' They produce frequent, incremental improvements while progressively narrowing the search within a productive area of the semantic solution space. In contrast, weaker optimizers exhibit large semantic drift, where their suggestions wander too far, leading to sporadic breakthroughs followed by long periods of stagnation. Notably, the research debunks the idea that sheer novelty is the key to success; novelty only benefits the search when it remains sufficiently localized around high-performing regions. This work, accepted at ACL 2026, provides actionable insights for designing better AI agents and training future models specifically for optimization roles.
- Study analyzed 15 LLMs across 8 tasks, finding zero-shot ability only partially predicts optimization success.
- Strong optimizers act as 'local refiners,' making steady incremental improvements without excessive semantic drift.
- Solution novelty alone does not predict performance; effective search requires staying localized in high-quality solution regions.
Why It Matters
Provides a blueprint for selecting and training LLMs to build more effective, reliable AI agents for complex problem-solving.