How can LLM-based heuristic design learn from… | r*cap

Home

Search

Library

Research questionHow can LLM-based heuristic design learn from failed trials while exploring large optimization spaces?Heuristic design often requires many costly experiments across a large space of possible ideas. Repeating unproductive trials wastes effort, but discarding failed directions too quickly can prevent useful exploration.

Code Generation & Program Synthesis

Evaluation & Benchmarks

Machine Learning

Multi-agent Systems

Neural and Evolutionary Computing

Latest papersRecent research connected to this question, newest first.LLM-Guided Program Evolution for Circle Packing: Breaking 10 Packomania Records for $28The evidence concerns Discovery Loop applied to Packomania's csqv problem: maximizing the sum of radii of N variable-radius circles in the unit square. A large language model started from a simple seed solver, used a scoreboard and history of prior ideas to propose candidates, and relied on an independent verifier to retain improvements and discard failures. The reported results cover 10 values of N from 101 to 114, with 2.4%–5.4% gains over prior records within 15 iterations and $27.72 in total LLM cost; the records were independently accepted by Packomania. The source also analyzes adaptive plateau detection for cost control.research paper · Sep 4, 2026 OR-Agent: Bridging Evolutionary Search and Structured Research for Automated Heuristic DesignThe source presents a multi-agent search framework with tree-structured hypothesis exploration, backtracking, reflection, and compressed memory of prior trials. Evidence covers combinatorial optimization and simulated cooperative-driving tasks, with comparisons to evolutionary search.research paper · Sep 4, 2026 MeEvo: Metacognitive Evolution Combined with Natural Evolution for Automatic Heuristic DesignThis concerns systems in which LLMs generate executable heuristic code for optimization tasks. The reported evidence covers five optimization problems and comparisons with tested LLM-based heuristic-design architectures, with effects reported particularly for complex constrained tasks.research paper · Sep 3, 2026 Adaptive Graph-of-Islands Evolution for Automatic Feature Engineering with LLMsThe source studies a graph-structured, multi-island evolutionary search in which LLMs generate mutations, crossovers, and cross-family hybrids. Its evaluation covers 29 tabular datasets, with results reported across classification and regression tasks, predictors, and LLM backbones.research paper · Sep 2, 2026

Related questions

How can LLM optimization systems generalize beyond surface narratives to shifted or emerging problem types?How can interdependent heuristic components be jointly evolved without incompatible search behavior in routing optimization?How can LLM prompts be automatically refined from recurring reasoning errors without laborious manual engineering?How can multi-agent LLM systems sustain autonomous exploration and knowledge accumulation in open-ended discovery?