Research map / Architectures and efficiency / Pruning and efficient tuning
TokenDrop: Token-Level Importance-Aware Backward Propagation Skipping for Efficient LLM Fine-Tuning
Beomseok Kim, Sol Namkung, Dongsuk Jeon
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 |
|---|---|---|---|
| add cumulative token selection (ablation) | TokenDrop (local selection baseline) · MMLU 5-shot, LLaMA2-7B | accuracy (%): 40.1 -> 53.3 (+13.2) | Improved |
| add cumulative token selection (ablation) | TokenDrop (local selection baseline) · CSR170K, LLaMA2-7B | accuracy (%): 69.7 -> 70.2 (+0.5) | Improved |
| add lazy selection scheduling (ablation) | TokenDrop with cumulative selection · MMLU 5-shot, LLaMA2-7B | accuracy (%): 53.3 -> 54.1 (+0.8) | Improved |
| add lazy selection scheduling (ablation) | TokenDrop with cumulative selection · CSR170K, LLaMA2-7B | accuracy (%): 70.2 -> 71.3 (+1.1) | Improved |
| select tokens by largest residual l2-norm instead of random (ablation) | TokenDrop · MMLU 5-shot, LLaMA2-7B | accuracy (%): 50.4 -> 54.1 | Improved |
| select tokens by largest residual l2-norm instead of random (ablation) | TokenDrop · CSR170K, LLaMA2-7B | accuracy (%): 70.3 -> 71.3 | Improved |
| keep smallest-norm tokens instead of random (ablation) | TokenDrop · MMLU 5-shot, LLaMA2-7B | accuracy (%): 50.4 -> 47.0 | Hurt |
| keep smallest-norm tokens instead of random (ablation) | TokenDrop · CSR170K, LLaMA2-7B | accuracy (%): 70.3 -> 67.9 | Hurt |
| replace DropBP with TokenDrop | LoRA fine-tuning · MMLU 5-shot, LLaMA2-7B | accuracy (%): 51.0 -> 54.1 | Improved |
| replace TokenTune with TokenDrop | LoRA fine-tuning · GSM8K, LLaMA3.1-8B | accuracy (%): 53.8 -> 54.9 | Improved |
| apply TokenDrop to LoRA | LoRA · HumanEval+, LLaMA3.1-8B | Pass@1 (%): 36.8 -> 36.6 | No clear effect |
| apply TokenDrop to LoRA | LoRA · MMLU 5-shot, LLaMA2-13B | accuracy (%): 58.1 -> 57.9 | No clear effect |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 1 of 1 on most shared results.
- PaCA — PaCA: Partial Connection Adaptation for Efficient Fine-Tuning (ICLR 2025) · better on 1 of 1 shared result
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Token Cleaning: Fine-Grained Data Selection for LLM Supervised Fine-Tuning (ICML 2025 · similarity 0.89)
- Unlocking Full Efficiency of Token Filtering in Large Language Model Training (ICLR 2026 · similarity 0.89)
- LISA: Layerwise Importance Sampling for Memory-Efficient Large Language Model Fine-Tuning (NeurIPS 2024 · similarity 0.89)
- DropBP: Accelerating Fine-Tuning of Large Language Models by Dropping Backward Propagation (NeurIPS 2024 · similarity 0.89)
- Mitigating Forgetting in LLM Fine-Tuning via Low-Perplexity Token Learning (NeurIPS 2025 · similarity 0.89)
- SkipGPT: Each Token is One of a Kind (ICML 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.
