Research map / Large language models / Instruction and fine-tuning
Prompt and Parameter Co-Optimization for Large Language Models
Xiaohe Bo, Rui Li, Zexu Sun, Quanyu Dai, Zeyu Zhang, Zihang Tian, Xu Chen, Zhenhua Dong
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 |
|---|---|---|---|
| jointly generate prompts and LoRA parameters instead of BetterTogether | MetaTuner-J (Qwen2.5-7B) · MATH | EM accuracy: 41.56 -> 48.67 | Improved |
| jointly generate prompts and LoRA parameters instead of BetterTogether | MetaTuner-J (Qwen2.5-3B) · HotpotQA | F1: 41.10 -> 58.85 | Improved |
| hybrid optimization instead of fine-tuning only | MetaTuner-J (Qwen2.5-7B) · GSM8K | EM accuracy: 63.68 -> 78.92 | Improved |
| add fine-tuning term to MetaTuner (ablation) | MetaTuner (Qwen2.5-7B) · MATH | EM accuracy: 48.00 -> 48.67 | Improved |
| add prompt-optimization term to MetaTuner (ablation) | MetaTuner (Qwen2.5-7B) · MATH | EM accuracy: 46.22 -> 48.67 | Improved |
| share meta encoder between prompt and parameter decoders (ablation) | MetaTuner (Qwen2.5-7B) · MATH | EM accuracy: 46.67 -> 48.67 | Improved |
| share meta encoder (ablation) | MetaTuner (Qwen2.5-7B) · CosmosQA | EM accuracy: 91.50 -> 92.25 | Mixed |
| use sequential (-I) instead of joint (-J) optimization | MetaTuner · HotpotQA | F1: 54.56 -> 55.75 | Mixed |
| replace supervised regularizer with Gumbel-Softmax (ablation) | MetaTuner · benchmarks | accuracy | Hurt |
| generate D2 from the model's own rollouts instead of an expert LLM (ablation) | MetaTuner · benchmarks | accuracy | Improved |
| increase number of rollout samples (ablation) | MetaTuner · benchmarks | accuracy | Hurt |
| update prompt decoder used for rollouts less frequently (ablation) | MetaTuner · benchmarks | accuracy | Hurt |
Closest papers by meaning
- Unleashing the Potential of Large Language Models as Prompt Optimizers: Analogical Analysis with Gradient-based Model Optimizers (AAAI 2025 · similarity 0.92)
- GReaTer: Gradients Over Reasoning Makes Smaller Language Models Strong Prompt Optimizers (ICLR 2025 · similarity 0.91)
- DePT: Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning (ICLR 2024 · similarity 0.91)
- Benchmarking Optimizers for Large Language Model Pretraining (NeurIPS 2026 · similarity 0.90)
- ADePT: Adaptive Decomposed Prompt Tuning for Parameter-Efficient Fine-tuning (ICLR 2025 · similarity 0.90)
- System Prompt Optimization with Meta-Learning (NeurIPS 2025 · similarity 0.90)
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.
