Research map / Large language models / Instruction and fine-tuning
NADS: Navigator-Guided Data Selection for Mitigating Catastrophic Forgetting in Fine-Tuning
JIANHAO ZHANG, Ou Wu, Yi Du
Read the paperCheck your idea against 137,217 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Proposes NADS, a navigator-guided data selection method that picks synthetic samples to mitigate catastrophic forgetting when fine-tuning LLMs on a new task without original pretraining data.
Problem
Catastrophic forgetting during single-stage fine-tuning of LLMs on a new task when original pretraining data are unavailable; synthetic candidate pools are noisy and redundant, and existing selection methods rely on indirect proxy signals poorly aligned with actual training benefit.
Approach
Selects synthetic training data using a navigator signal intended to directly track actual training benefit rather than indirect proxies, to reduce forgetting while learning the new task.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use navigator-guided constraint selection with KD instead of the strongest replay baseline | LLaMA3.2-3B on Magicoder · Magicoder overall | overall accuracy: 42.01 -> 43.46 | Improved |
| remove KL-discrepancy factor from the utility (gate-only utility) (ablation) | NADS · Magicoder, LLaMA3.2-3B | general capability: 42.52 -> 37.64 | Hurt |
| add diversity-aware selection term (ablation) | NADS · Magicoder, LLaMA3.2-3B | general capability: 40.21 -> 42.52 | Improved |
| add the pretrained-confidence entropy gate to the utility (ablation) | NADS · Magicoder, LLaMA3.2-3B | overall accuracy: 42.22 -> 43.46 | Improved |
| select constraints by navigator utility instead of random, perplexity or standard DPP selection (ablation) | NADS · Magicoder, LLaMA3.2-3B | overall accuracy: 39.94 / 40.39 / 39.66 -> 43.46 | Improved |
| use random constraints under a fixed KD objective (ablation) | baseline methods with KD · Magicoder, LLaMA3.2 | general capability | No clear effect |
| use navigator-selected constraints under a fixed KD objective (ablation) | baseline methods with KD · Magicoder, LLaMA3.2 | general capability | Improved |
| train the navigator unconstrained on the new task | LLaMA3.2-3B navigator · Magicoder | HumanEval vs MMLU: HumanEval 28.16 -> 45.50; MMLU 56.92 -> 50.60 | Mixed |
| use a navigator checkpoint from a middle training stage (M_nav-2) rather than early or late (ablation) | NADS · Magicoder, LLaMA3.2-3B | target and general trade-off | Improved |
| use forward KL instead of JS divergence as utility signal (ablation) | NADS · Magicoder, LLaMA3.2-3B | performance | Improved |
| increase distillation strength lambda (ablation) | NADS · Magicoder | general vs target | Mixed |
| use the NADS iterative optimizer instead of Fast Greedy or Lazy-Fast Greedy (ablation) | NADS data selection · Magicoder candidate pool | selection time | Improved |
Closest papers by meaning
- Estimating and Orthogonalizing Unknown Pre-training Gradients for Continual Fine-tuning of Large Language Models (NeurIPS 2026 · similarity 0.90)
- Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting (ICML 2025 · similarity 0.90)
- Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning (NeurIPS 2025 · similarity 0.89)
- Understanding Catastrophic Forgetting in Language Models via Implicit Inference (ICLR 2024 · similarity 0.89)
- Model Tailor: Mitigating Catastrophic Forgetting in Multi-modal Large Language Models (ICML 2024 · similarity 0.89)
- Interweaving Memories of a Siamese Large Language Model (AAAI 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.
