Research map / Probabilistic and scientific modeling / Variational and generative sampling
LS-Merge: Merging Language Models in Latent Space
Bedionita Soro, Aoxuan Silvia Zhang, Bruno Andreis, Jaehyeong Jo, Song Chong, Sung Ju Hwang
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 |
|---|---|---|---|
| merge in latent space instead of weight-space greedy soup | LoRA expert merging · MMLU | accuracy: 50.8 -> 56.0 | Improved |
| merge in latent space (lerp) instead of SLERP | LoRA expert merging · MMLU | accuracy: 52.5 -> 54.7 | Improved |
| add latent-space self-merging | Gemma-3-1b-it · MMLU | accuracy: 32.20 -> 35.13 | Improved |
| add latent-space self-merging | Gemma-3-4b-it · MMLU | accuracy: 53.10 -> 54.20 | Improved |
| add OT alignment before latent interpolation (ablation) | cross-family merge · WinoGrande | accuracy: 51.13 -> 57.75 | Improved |
| merge both attention and MLP instead of attention only (ablation) | latent merging · ARC-C | accuracy: 40.23 -> 43.34 | Improved |
| merge both attention and MLP instead of MLP only (ablation) | latent merging · MMLU | accuracy: 41.02 -> 42.10 | Improved |
| merge MLP layers only (ablation) | latent merging · ARC-C | accuracy: 42.78 -> 43.89 | Improved |
| use VAE instead of PCA encoder (ablation) | weight encoding · MMLU (r=1.6) | accuracy: 25.50 -> 39.89 | Improved |
| increase VAE compression ratio from 1.6 to 4 (ablation) | Gemma-3-1B-it VAE · MMLU | accuracy: 39.98 -> 25.02 | Hurt |
| replace Task Arithmetic with LS-merge | Llama-2-13B code+instruct · GSM8K | accuracy: 4.20 -> 44.12 | Improved |
| replace AIM with LS-merge | Llama-2-13B code+instruct · GSM8K | accuracy: 46.20 -> 44.12 | No clear effect |
Closest papers by meaning
- Activation-Informed Merging of Large Language Models (NeurIPS 2025 · similarity 0.90)
- Model Merging in Pre-training of Large Language Models (NeurIPS 2025 · similarity 0.90)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.89)
- OptMerge: Unifying Multimodal LLM Capabilities and Modalities via Model Merging (ICLR 2026 · similarity 0.88)
- $\texttt{Model-GLUE}$: Democratized LLM Scaling for A Large Model Zoo in the Wild (NeurIPS 2024 · similarity 0.88)
- AdaMMS: Model Merging for Heterogeneous Multimodal Large Language Models with Unsupervised Coefficient Optimization (CVPR 2025 · similarity 0.88)
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.
