Research map / Architectures and efficiency / Mixture of experts
Neural Thickets: Diverse Task Experts Are Dense Around Pretrained Weights
Yulu Gan, Phillip Isola
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 |
|---|---|---|---|
| scale model size from 0.5B to 32B (ablation) | Qwen2.5 Gaussian-perturbed weights · GSM8K | fraction of perturbations matching or exceeding base: 0% -> 64% | Improved |
| scale model size from 0.5B to 32B (ablation) | Qwen2.5 Gaussian-perturbed weights · Countdown | fraction of perturbations matching or exceeding base: 8% -> 60% | Improved |
| increase model size (solution diversity) (ablation) | Qwen2.5 family · seven tasks, 500 perturbations | Spectral Discordance | Improved |
| ensemble top-50 perturbations instead of top-1 (K=50 vs K=1) (ablation) | RandOpt · LLM post-training tasks | accuracy | Improved |
| scale population size N (ablation) | RandOpt · Countdown | accuracy | Improved |
| increase selection ratio K (ablation) | RandOpt with large N · Countdown | accuracy | No clear effect |
| add test-time majority vote over 50 samples | ES · GSM8K, Qwen2.5-0.5B | accuracy: 42.6% -> 61.2% | Improved |
| replace gradient/ES training with random guessing and ensembling | RandOpt vs GRPO · Countdown, OLMo3-7B-Instruct | accuracy: 68.5 -> 85.0 | Improved |
| distill top-50 ensemble into a single model by SFT | Qwen2.5-1.5B-Instruct · GSM8K | accuracy: 58.8 -> 74.9 | Improved |
| apply RandOpt to the language model of a VLM | Qwen2.5-VL-3B-Instruct · GQA | accuracy: 56.6 -> 69.0 | Improved |
| increase batch or group size to scale baseline parallelism in one step (ablation) | GRPO and PPO · GSM8K, Qwen2.5-3B-Instruct | accuracy: GRPO 83.5 -> 80.1 | No clear effect |
| pretrain on mixed signal types before random guessing (ablation) | 1D signal MLP · 1D signals | quality of nearby solutions | Improved |
Closest papers by meaning
- Multi-Task Dense Prediction via Mixture of Low-Rank Experts (CVPR 2024 · similarity 0.86)
- Domain-Specific Pruning of Large Mixture-of-Experts Models with Few-shot Demonstrations (NeurIPS 2025 · similarity 0.86)
- Multilinear Mixture of Experts: Scalable Expert Specialization through Factorization (NeurIPS 2024 · similarity 0.86)
- Understanding Overadaptation in Supervised Fine-Tuning: The Role of Ensemble Methods (ICML 2025 · similarity 0.86)
- More Experts Than Galaxies: Conditionally-Overlapping Experts with Biologically-Inspired Fixed Routing (ICLR 2025 · similarity 0.86)
- Learning to Specialize: Joint Gating-Expert Training for Adaptive MoEs in Decentralized Settings (NeurIPS 2025 · similarity 0.86)
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.
