Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection

Pretam Ray, Pratik Prabhanjan Brahma, Zicheng Liu, Emad Barsoum

TL;DR

AdaptEvolve addresses the high cost of iterative LLM use in evolutionary agentic coding by dynamically routing generation steps between a cost-efficient small model $M_S$ and a powerful large model $M_L$ based on intrinsic uncertainty signals $C(x_i)$ and a binary router $\\Phi: \mathbb{R}^d \to \{0,1\}$. It bootstraps a lightweight decision tree from a $N=50$ warm-up set and continuously updates it online with a Hoeffding Adaptive Tree to handle non-stationarity, using entropy-based metrics $c_t$, MC, LGC, TC, and BWC as router inputs. Empirically, AdaptEvolve achieves an average compute reduction of $37.9\%$ while retaining $97.5\%$ of the upper-bound accuracy of static $M_L$ baselines across LiveCodeBench and MBPP, indicating a favorable Pareto frontier. This approach demonstrates that uncertainty-aware resource allocation can scale agentic reasoning with multi-LLM systems without external controllers, enabling efficient and robust code-generation workflows.

Abstract

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.

AdaptEvolve: Improving Efficiency of Evolutionary AI Agents through Adaptive Model Selection

TL;DR

AdaptEvolve addresses the high cost of iterative LLM use in evolutionary agentic coding by dynamically routing generation steps between a cost-efficient small model and a powerful large model based on intrinsic uncertainty signals and a binary router . It bootstraps a lightweight decision tree from a warm-up set and continuously updates it online with a Hoeffding Adaptive Tree to handle non-stationarity, using entropy-based metrics , MC, LGC, TC, and BWC as router inputs. Empirically, AdaptEvolve achieves an average compute reduction of while retaining of the upper-bound accuracy of static baselines across LiveCodeBench and MBPP, indicating a favorable Pareto frontier. This approach demonstrates that uncertainty-aware resource allocation can scale agentic reasoning with multi-LLM systems without external controllers, enabling efficient and robust code-generation workflows.

Abstract

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.
Paper Structure (27 sections, 7 equations, 2 figures, 6 tables)

This paper contains 27 sections, 7 equations, 2 figures, 6 tables.

Table of Contents

  1. Introduction
  2. Methodology
  3. Adaptive Evolutionary Refinement
  4. Entropy-Based Confidence Metrics
  5. Experiment
  6. Experimental Setup
  7. Models
  8. Dataset and Metrics
  9. Metrics:
  10. Baselines
  11. Methodolody
  12. Static Decision Tree
  13. Online Adaptation
  14. Results and Analysis
  15. Benchmark Performance
  16. ...and 12 more sections

Figures (2)

  • Figure 1: Adaptive Evolutionary Refinement Framework. The workflow initiates with a candidate generation using the small model ($\mathcal{M}_S$). We compute intrinsic confidence metrics (LGC, MC, TC, BWC) based on token entropy. A lightweight decision tree router ($\Phi$), bootstrapped via a warm-up phase, dynamically determines whether to retain the efficient generation or escalate to the large model ($\mathcal{M}_L$) for complex reasoning hurdles.
  • Figure 2: Impact of Large Model Usage on Accuracy (LiveCodeBench). We plot the percentage of calls routed to the 32B model against the final accuracy. AdaptEvolve (Red) achieves a significantly better trade-off than the Cascading baseline and Random Sampling, attaining near-peak accuracy with only 58% large-model usage.