AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection
This addresses efficiency challenges for developers of evolutionary AI agents, though it is incremental as it builds on existing model cascade and evolutionary refinement methods.
The paper tackled the trade-off between computational efficiency and reasoning capability in evolutionary AI agents by introducing AdaptEvolve, which uses adaptive model selection based on generation confidence, reducing inference cost by 37.9% while retaining 97.5% of the accuracy of static large-model baselines.
Evolutionary agentic systems intensify the trade-off between computational efficiency and reasoning capability by repeatedly invoking large language models (LLMs) during inference. This setting raises a central question: how can an agent dynamically select an LLM that is sufficiently capable for the current generation step while remaining computationally efficient? While model cascades offer a practical mechanism for balancing this trade-off, existing routing strategies typically rely on static heuristics or external controllers and do not explicitly account for model uncertainty. We introduce AdaptEvolve: Adaptive LLM Selection for Multi-LLM Evolutionary Refinement within an evolutionary sequential refinement framework that leverages intrinsic generation confidence to estimate real-time solvability. Empirical results show that confidence-driven selection yields a favourable Pareto frontier, reducing total inference cost by an average of 37.9% across benchmarks while retaining 97.5% of the upper-bound accuracy of static large-model baselines. Our code is available at https://github.com/raypretam/adaptive_llm_selection.