Research map / Large language models / Language, safety and interpretability
MERGE$^3$: Efficient Evolutionary Merging on Consumer-grade GPUs
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.
Makes evolutionary model merging of LLMs feasible on a single consumer GPU by evaluating fitness on a small random subset and predicting full-dataset performance with IRT-based merged-performance estimators.
Problem
Evolutionary model merging is computationally prohibitive on consumer hardware because fitness requires generating and scoring answers on the full evaluation set for every candidate at every generation.
Approach
Shrink the fitness dataset via random sampling, estimate each endpoint's latent IRT abilities on the full dataset, and predict merged-model accuracy as a linear combination of endpoint abilities fitted by MLE on the small subset (mp-IRT and gmp-IRT estimators), with theoretical guarantees on the optimum.
Key results
- Cross-lingual math merges surpass both endpoints by 10-20% accuracy on translated GSM8K, while standard baselines often underperform the endpoints.
- Evolved multilingual model (IT, EN, DE, NL) improves ARC accuracy over language-specific finetunes, e.g. Dutch 0.50 to 0.69 (+19%).
- The method claims a 50x reduction in fitness-computation cost (0.077e6 vs 4e6 TFLOPs) relative to full-dataset evolutionary merging.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Evolve a single multilingual merge from Italian, English, German, and Dutch models | Mistral-7B fine-tuned models · ARC-Challenge Italian | accuracy: 0.61 -> 0.69 | Helped |
| Evolve a single multilingual merge from Italian, English, German, and Dutch models | Mistral-7B fine-tuned models · ARC-Challenge English | accuracy: 0.75 -> 0.79 | Helped |
| Evolve a single multilingual merge from Italian, English, German, and Dutch models | Mistral-7B fine-tuned models · ARC-Challenge German | accuracy: 0.61 -> 0.72 | Helped |
| Evolve a single multilingual merge from Italian, English, German, and Dutch models | Mistral-7B fine-tuned models · ARC-Challenge Dutch | accuracy: 0.50 -> 0.69 | Helped |
| Use MERGE³ instead of full-dataset fitness evaluation | Evolutionary model merging · Fitness computation | TFLOPs: 4×10^6 -> 0.077 × 10^6 | Helped |
| Merge an English math-specialized model with a language-specific fine-tune using evolutionary search | Mistral-7B language-specific models · Translated GSM8K across target languages | accuracy: 10–20% | Helped |
| Use the reduced-dataset MERGE³ approach instead of full-dataset EvoMerge evaluation | Japanese math-skill transfer · Japanese GSM8K | accuracy | Hurt |
| Use merged-performance IRT estimators (mp-IRT and gmp-IRT) instead of p-IRT and gp-IRT (ablation) | Merged-model performance estimation · ARC, GSM8K, TruthfulQA, Winogrande, and HellaSwag | absolute estimation error | Helped |
| Use merged-performance IRT ability estimation instead of p-IRT and gp-IRT ability estimation (ablation) | Model ability estimation · Five benchmark tasks | cosine similarity | Helped |
| Use uniform random sampling instead of IRT clustering or representation clustering (ablation) | Evaluation dataset extraction · Translated and English GSM8K | absolute estimation error | No clear effect |
| Self-merge the Italian language model instead of merging it with a math-specialized model (ablation) | Italian Mistral-7B model · Italian GSM8K | accuracy | Hurt |
| Merge the Italian language model with itself rather than adding a math-specialized model (ablation) | Italian Mistral-7B model · Italian GSM8K | accuracy | No clear effect |
Surprising findings
- Self-merging the Italian endpoint decreases performance, supporting that gains come from cross-lingual transfer rather than prompt-template fitting.
Limitations
- Estimator assumes linear combination of endpoint abilities, which may fail when merging causes weight interference.
- Accuracy is lower than full-dataset evolutionary merging (EvoMerge) on Japanese GSM8K.
- Theoretical guarantees are asymptotic or in expectation; finite-sample guarantees rely on empirical estimates.
Experimental setup
Datasets: GSM8K ARC Winogrande TruthfulQA HellaSwag Open LLM leaderboard responses
Benchmarks: GSM8K translated GSM8K (Italian, German, Romanian, Dutch, Japanese) ARC Winogrande TruthfulQA HellaSwag
Metrics: accuracy absolute estimation error cosine similarity euclidean distance
Baselines: Task Arithmetic TIES DARE SLERP EvoMerge p-IRT gp-IRT random sampling
Builds on: evolutionary model merging p-IRT gp-IRT TIES DARE NSGA-II
Closest papers by meaning
- Channel Merging: Preserving Specialization for Merged Experts (AAAI 2025 · similarity 0.86)
- QoS-Efficient Serving of Multiple Mixture-of-Expert LLMs Using Partial Runtime Reconfiguration (ICML 2025 · similarity 0.86)
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.86)
- MoS: Unleashing Parameter Efficiency of Low-Rank Adaptation with Mixture of Shards (ICLR 2025 · similarity 0.85)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.85)
- Evolution Fine-Tuning: Learning to Discover Across 371 Optimization Tasks (NeurIPS 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.