Research map / Continual learning / Continual fine-tuning of LLMs
MetaGDPO: Alleviating Catastrophic Forgetting with Metacognitive Knowledge Through Group Direct Preference Optimization
Lanxue Zhang, Yuqiang Xie, Fang Fang, Fanglong Dong, Rui Liu, Yanan Cao
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 |
|---|---|---|---|
| fine-tune with MetaGDPO instead of no fine-tuning | Qwen3-8B · Math avg | accuracy: 79.16 -> 80.08 | Improved |
| fine-tune with MetaGDPO instead of no fine-tuning | Qwen3-8B · General avg | accuracy: 72.09 -> 76.79 | Improved |
| fine-tune with MetaGDPO instead of no fine-tuning | R1-Llama-8B · General avg | accuracy: 53.51 -> 58.34 | Improved |
| fine-tune with LIMO instead of no fine-tuning | Qwen3-8B · Math avg | accuracy: 79.16 -> 70.59 | Hurt |
| fine-tune with L+S instead of no fine-tuning | R1-Qwen-7B · GSM8K | accuracy: 89.69 -> 46.47 | Hurt |
| fine-tune with STAR-1 instead of no fine-tuning | Qwen3-8B · General avg | accuracy: 72.09 -> 41.46 | Hurt |
| fine-tune with STAR-1 instead of no fine-tuning | Qwen3-8B · Safety avg | safety ratio: 83.87 -> 96.39 | Improved |
| add advantage weights to GDPO (ablation) | GDPO · overall average | average score: 80.25 (without advantages) | Improved |
| increase group size G from 2 to 10 (ablation) | GDPO · overall performance | average score | Improved |
| train on only 10 instances (ablation) | GDPO · overall performance | average score | Improved |
| use GDPO instead of SFT (ablation) | Qwen3-8B · avg across math/commonsense/safety | average score: nearly 17% | Improved |
| use GDPO instead of DPO (ablation) | Qwen3-8B · avg across math/commonsense/safety | average score: nearly 3% | Improved |
Closest papers by meaning
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.87)
- Verifying Meta-Awareness via Predictive Rewards in Reasoning Models (ICML 2026 · similarity 0.87)
- Train-Attention: Meta-Learning Where to Focus in Continual Knowledge Learning (NeurIPS 2024 · similarity 0.87)
- Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning (ICLR 2025 · similarity 0.87)
- Spurious Forgetting in Continual Learning of Language Models (ICLR 2025 · similarity 0.87)
- Meta-Cognitive Memory Policy Optimization for Long-Horizon LLM Agents (NeurIPS 2026 · 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.
