ReVEL: Multi-Turn Reflective LLM-Guided Heuristic Evolution via Structured Performance Feedback

A team of researchers led by Cuong Van Duc has proposed ReVEL (Multi-Turn Reflective LLM-Guided Heuristic Evolution via Structured Performance Feedback), a novel framework that automates the design of heuristics for notoriously difficult NP-hard combinatorial optimization problems. Instead of the typical one-shot code generation, ReVEL integrates large language models (LLMs) as interactive, reflective components within an evolutionary algorithm (EA). The system's core innovation is a two-part mechanism: first, it clusters candidate heuristics into behaviorally similar groups to create compact performance profiles, and second, it uses these profiles to guide the LLM through multi-turn reflection sessions where it analyzes failures and proposes targeted refinements.

An EA-based meta-controller then selectively integrates these LLM-generated refinements, balancing exploration of new ideas with exploitation of promising ones. This closed-loop, iterative process allows the LLM to act as a reasoning engine that learns from structured feedback, moving beyond brittle, single-pass code synthesis. In experiments on standard combinatorial optimization benchmarks, ReVEL consistently produced heuristics that were more robust and diverse than those from strong baseline methods, with the improvements being statistically significant.

The results demonstrate that multi-turn reasoning combined with structured behavioral grouping is a powerful paradigm for automated heuristic design. This approach leverages the LLM's capacity for iterative analysis—a capability underutilized in most current applications—while the evolutionary algorithm provides the necessary rigor and validation. The work, published on arXiv, highlights a promising direction for using AI not just to write code, but to engage in a sustained, reflective dialogue to solve complex computational challenges.