Research map / Large language models / Agents, code and math
An LLM-Empowered Adaptive Evolutionary Algorithm for Multi-Component Deep Learning Systems
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes μMOEA, an LLM-empowered adaptive evolutionary search (GPT-4 for initial population and differential seed generation) that finds safety violations in the Baidu Apollo autonomous driving system more efficiently than the NSGA-II-based MOSAT.
Problem
Searching for diverse safety violations in multi-component deep learning systems (here, an industrial autonomous driving stack) with MOEAs suffers from inefficient random initialization and premature convergence to local optima.
Approach
Use an LLM to generate a task-aware initial population, adaptively set crossover/mutation probabilities and variation types from fitness ranks, and feed evolutionary history back into the LLM to produce differential seed solutions when the search stagnates.
Key results
- μMOEA detected 10 distinct safety violation types in 24-hour runs, all within the first 14 hours, versus 6 types for MOSAT.
- μMOEA required on average 12 solutions per safety violation compared with 62.1 for MOSAT.
- Ablation: removing differential seeds (μMOEA_n) reduced detected types to 6 and raised solutions per violation to 24.3.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use LLM-generated rather than randomly initialized initial population (ablation) | μMOEA on Apollo · Apollo, 24-hour run | time to detect first SV: 39min -> 11min | Helped |
| use LLM-generated rather than randomly initialized initial population (ablation) | μMOEA on Apollo · Apollo, 24-hour run | solutions to detect one SV: 12.7 -> 12 | Helped |
| use LLM-generated rather than randomly initialized initial population (ablation) | μMOEA on Apollo · Apollo, 24-hour run | types of detected SV: 10 -> 10 | No clear effect |
| add LLM-generated differential seeds (ablation) | μMOEA on Apollo · Apollo, 24-hour run | types of detected SV: 6 -> 10 | Helped |
| add LLM-generated differential seeds (ablation) | μMOEA on Apollo · Apollo, 24-hour run | solutions to detect one SV: 24.3 -> 12 | Helped |
| add LLM-generated differential seeds (ablation) | μMOEA on Apollo · Apollo, 24-hour run | time to detect all types of SVs: 22h -> 14h | Helped |
| add LLM-generated differential seeds (ablation) | μMOEA on Apollo · Apollo, 24-hour run | time to detect first SV: 17min -> 11min | Helped |
| replace MOSAT's random initialization and fixed-rate MOEA search with μMOEA | Apollo · Apollo, San Francisco | types of detected SV: 6 -> 10 | Helped |
| replace MOSAT with μMOEA | Apollo · Apollo, San Francisco | solutions to detect one SV (avg): 62.1 -> 12 | Helped |
| replace MOSAT with μMOEA | Apollo · Apollo, San Francisco | time to detect first SV (avg): 16 -> 7 | Helped |
| replace MOSAT with μMOEA | Apollo · Apollo, San Francisco | time to detect all found types of SVs (avg): 18.1 -> 12.9 | Helped |
Limitations
- Relies on remote GPT-4 API calls, adding latency and operational cost.
- Limited LLM input context restricts number of in-context examples; prompt history must be truncated.
- Evaluated on a single system (Apollo) and single road/map.
Experimental setup
Benchmarks: Baidu Apollo safety violation search in SVL simulator (San Francisco map)
Metrics: types of detected safety violations solutions needed per safety violation time to first safety violation time to all violation types
Baselines: MOSAT NSGA-II μMOEA_r (random initialization variant) μMOEA_n (no differential seeds variant)
Builds on: NSGA-II
Closest papers by meaning
- HSEvo: Elevating Automatic Heuristic Design with Diversity-Driven Harmony Search and Genetic Algorithm Using LLMs (AAAI 2025 · similarity 0.85)
- ReEvo: Large Language Models as Hyper-Heuristics with Reflective Evolution (NeurIPS 2024 · similarity 0.85)
- LLMOPT: Learning to Define and Solve General Optimization Problems from Scratch (ICLR 2025 · similarity 0.84)
- Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks (NeurIPS 2026 · similarity 0.84)
- Multi-Objective Evolution of Heuristic Using Large Language Model (AAAI 2025 · similarity 0.84)
- Neural Genetic Search in Discrete Spaces (ICML 2025 · similarity 0.84)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.