Research map / Large language models / Retrieval and knowledge
Knowledge Fusion of Large Language Models via Modular SkillPacks
Guodong DU, Zhuo Li, Xuanning Zhou, Junlin Li, Zesheng Shi, Wanyu Lin, Ho-Kin Tang, Xiucheng Li, Fangming Liu, Wenya Wang, Min Zhang, Jing Li
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 |
|---|---|---|---|
| use SVD instead of pruning for MLP module compression (ablation) | GraftLLM · MATH (~5% compression) | accuracy: 70.2 -> 72.1 | Improved |
| use SVD instead of no compression strategy (Null) on MLP modules (ablation) | GraftLLM · MATH (~5% compression) | accuracy: 68.7 -> 72.1 | Improved |
| add mixed-precision quantization (ablation) | GraftLLM MLP modules · MATH (~5% compression) | accuracy: 71.5 -> 72.1 | Improved |
| add quantization at all (ablation) | GraftLLM MLP modules · MATH (~5% compression) | accuracy: 69.2 -> 72.1 | Improved |
| use low-rank SVD for attention modules (ablation) | GraftLLM · MATH (~5% compression) | accuracy: 71.2 (pruning) -> 72.1 | Improved |
| use pruning for embedding and output head (ablation) | GraftLLM · MATH (~5% compression) | accuracy: 71.7 (low-rank SVD) -> 72.1 | No clear effect |
| ensemble top-2 SkillPacks instead of top-1 (ablation) | Routed GraftLLM · AlpacaEval 2.0 (explicit fusion) | Win Rate (%): 16.56 -> 17.33 (2.03x inference cost) | Improved |
| ensemble top-2 SkillPacks instead of top-1 (ablation) | Routed GraftLLM · Finance+Law+Bio fusion | Average score: 49.62 -> 49.64 | No clear effect |
| ensemble top-3 SkillPacks instead of top-1 (ablation) | Routed GraftLLM · Implicit fusion avg of 10 benchmarks | Average score: 56.0 -> 56.4 | No clear effect |
| graft SkillPack with routing instead of parameter merging/routed LoRA | OpenChat-3.5-7B explicit fusion · AlpacaEval 2.0 (explicit fusion) | LC Win Rate (%): 24.81 -> 25.42 vs Twin-Merging r1024 | Improved |
| use SkillPack instead of model grafting/Model Tailor | Llama-3.1-8B-Instruct forget-free learning · code then math average | Average score: 62.2 -> 64.3 | Improved |
| use GraftLLM instead of FuseChat-3 on instruction following | Llama-3.1-8B-Instruct implicit fusion · MT-Bench / AlpacaEval-2 LC | score: 9.0 -> 8.8 MT-Bench; 65.4 -> 64.8 LC | Hurt |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- Twin-Merging — Twin-Merging: Dynamic Integration of Modular Expertise in Model Merging (NeurIPS 2024) · better on 2 of 2 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Knowledge Fusion of Large Language Models (ICLR 2024 · similarity 0.92)
- Bohdi: Heterogeneous LLM Fusion with Automatic Data Exploration (NeurIPS 2025 · similarity 0.90)
- Memory Grafting: Scaling Language Model Pre-training via Offline Conditional Memory (NeurIPS 2026 · similarity 0.89)
- MergeBench: A Benchmark for Merging Domain-Specialized LLMs (NeurIPS 2025 · similarity 0.89)
- Seeking Neural Nuggets: Knowledge Transfer in Large Language Models from a Parametric Perspective (ICLR 2024 · similarity 0.89)
- ELICIT: LLM Augmentation Via External In-context Capability (ICLR 2025 · similarity 0.89)
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.
