Research map / Large language models / Agents, code and math
EvoGM: Learning to Merge LLMs via Evolutionary Generative Optimization
Tao Jiang, Yu Xinmeng, Chenhao Yi, Yiling Wu, Yan Li, Ran Cheng, Dongmei Jiang, Jianguo Zhang
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace stochastic evolutionary operators with learned generative operator | EvoGM · GLUE 8 tasks (FLAN-T5-base) | average accuracy: 81.2 -> 82.4 | Improved |
| replace stochastic evolutionary operators with learned generative operator | EvoGM · 8 unseen tasks multi-task | average score: 0.375 -> 0.380 (vs CMA) | Improved |
| merge experts instead of choosing the single best expert (multi-task) | EvoGM · 8 unseen tasks multi-task | average score: 0.405 -> 0.380 | Hurt |
| multi-task fine-tuning instead of base model | MTL · 8 unseen tasks | average score: 0.352 -> 0.330 | Hurt |
| replace PSO-Merging/Model Swarm with EvoGM on per-task search | EvoGM · 8 unseen tasks single-task | test accuracy: best on 5 of 8 benchmarks | Mixed |
| remove multi-round expert update (w/o Rounds) (ablation) | EvoGM · MMLU, NLGraph | accuracy: decline reported, no numbers in text | Improved |
| add cycle-consistency loss (ablation) | EvoGM · NLGraph and reasoning tasks | accuracy: significant drop without it, no numbers in text | Improved |
| use dual generators instead of a single generator (ablation) | EvoGM · MMLU, NLGraph, TruthQA, HellaSwag | accuracy | Improved |
| increase population size from 10 to 30 (ablation) | EvoGM · 2-task average test accuracy | accuracy: +0.0305 | Improved |
| increase generator training epochs to 600 (ablation) | EvoGM · 2-task average test accuracy | accuracy: -0.0035 | Hurt |
| increase iterations per round from 3 to 8 (ablation) | EvoGM · 2-task average test accuracy | accuracy | No clear effect |
| increase evolution rounds from 1 to 5 (ablation) | EvoGM · 2-task average test accuracy | accuracy | Improved |
Closest papers by meaning
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.88)
- Model Swarms: Collaborative Search to Adapt LLM Experts via Swarm Intelligence (ICML 2025 · similarity 0.86)
- HSEvo: Elevating Automatic Heuristic Design with Diversity-Driven Harmony Search and Genetic Algorithm Using LLMs (AAAI 2025 · similarity 0.86)
- Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks (NeurIPS 2026 · similarity 0.85)
- Rethinking LLM Evaluation: Can We Evaluate LLMs with 200× Less Data? (ICLR 2026 · similarity 0.85)
- Making, Not Taking, the Best of N (ICLR 2026 · similarity 0.85)
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.
