Research map / Continual learning / Continual fine-tuning of LLMs
Upweighting Easy Samples in Fine-Tuning Mitigates Forgetting
Read the paperCheck your idea against 108,967 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 FLOW, a data-oblivious fine-tuning method that weights each fine-tuning sample by exp(-pre-trained loss/τ) (τ = median loss) to upweight easy samples and mitigate catastrophic forgetting, with linear-model theory.
Problem
Catastrophic forgetting of pre-trained capabilities during fine-tuning when pre-training data and recipe are unavailable (data-oblivious setting).
Approach
Derive static per-sample weights exp(-f_i(θ*)/τ) from entropic regularization of a sample-weight distribution, upweighting samples the pre-trained model already fits well; the opposite of DRO-style hard-sample upweighting. Weights are computed once from the pre-trained model only.
Key results
- On ResNet-50, the method attains 81.17% average (IN-1K 76.09%, target 86.25%) vs 64.34% for standard FT and 77.73% for linear probing.
- Gemma 2 2B fine-tuned on MetaMathQA: GSM8K drops only 0.83 points vs standard FT (62.55 vs 63.38) with average 49.98 vs 46.31.
- Combining the weighting with WiSE-FT raises average accuracy by 4.18% (ResNet-18) and 4.52% (ResNet-50).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | ResNet-18 · Vision transfer datasets, average accuracy | average accuracy: 54.60 -> 74.57 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | ResNet-50 · Vision transfer datasets, average accuracy | average accuracy: 64.34 -> 81.17 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Gemma 2 2B · MetaMathQA fine-tuning | average accuracy: 46.31 -> 49.98 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Llama 3.2 3B · MetaMathQA fine-tuning | average accuracy: 46.10 -> 52.87 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Gemma 2 2B · GSM8K | accuracy: 63.38 -> 62.55 | Hurt |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Gemma 2 2B · Commonsense | accuracy: 55.07 -> 57.59 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Gemma 2 2B · MMLU | accuracy: 45.59 -> 49.31 | Helped |
| Upweight samples with lower pre-trained loss instead of using standard fine-tuning | Gemma 2 2B · MBPP | accuracy: 16.80 -> 26.80 | Helped |
| Combine FLOW with WiSE-FT | ResNet-18 · IN-1K and vision target datasets | average accuracy: 67.19 -> 71.37 | Helped |
| Combine FLOW with WiSE-FT | ResNet-50 · IN-1K and vision target datasets | average accuracy: 71.52 -> 76.04 | Helped |
| Add FLOW to ℓ2 regularization | Gemma 2 2B · MetaMathQA fine-tuning | average accuracy: 49.19 -> 49.79 | Helped |
| Add FLOW to LoRA | Gemma 2 2B · MetaMathQA fine-tuning | average accuracy: 47.05 -> 49.31 | Helped |
Surprising findings
- Linear probing, which keeps the body frozen, is the second-best method on the unified average metric.
- The weighting sacrifices accuracy on hard fine-tuning samples; on the top-10% hardest samples, accuracy falls sharply (e.g., CIFAR-10 86.60 to 30.70).
Limitations
- Trades off lower fine-tuning accuracy, especially on hard samples with high pre-training loss.
- Theory is limited to linear models and Gaussian data with a specific covariance structure.
Experimental setup
Datasets: MetaMathQA ImageNet-1K CIFAR-10 CIFAR-100 Flowers102 Caltech101 Cars Dogs
Benchmarks: GSM8K MMLU MBPP commonsense reasoning ImageNet-1K top-1 accuracy CIFAR-10 CIFAR-100 Flowers102 Caltech101 Cars Dogs
Metrics: accuracy average of pre-training and target accuracy
Baselines: standard fine-tuning l2 regularization WiSE-FT LoRA linear probing learning without forgetting
Builds on: distributionally robust optimization WiSE-FT LoRA elastic weight consolidation
Closest papers by meaning
- Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning (NeurIPS 2025 · similarity 0.89)
- Overcoming Catastrophic Forgetting in Visual Continual Learning with Reinforcement Fine-Tuning (NeurIPS 2026 · similarity 0.89)
- Fine-tuning Reinforcement Learning Models is Secretly a Forgetting Mitigation Problem (ICML 2024 · similarity 0.89)
- Scaling Laws for Forgetting during Finetuning with Pretraining Data Injection (ICML 2025 · similarity 0.88)
- LoRA vs Full Fine-tuning: An Illusion of Equivalence (NeurIPS 2025 · similarity 0.88)
- Unlocking the Power of Function Vectors for Characterizing and Mitigating Catastrophic Forgetting in Continual Instruction Tuning (ICLR 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.