Research map / Vision-language and multimodal / Human and embodied modeling
Multi-Stage Influence Function
Hongge Chen, Si Si, Yang Li, Ciprian Chelba, Sanjiv Kumar, Duane Boning, Cho-Jui Hsieh
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 |
|---|---|---|---|
| Remove the top 10% highest positive-influence pretraining examples | CIFAR-10 target classifier · CIFAR-10 target classification | accuracy: 58.15% -> 58.36% | Improved |
| Randomly remove 10% of pretraining data | CIFAR-10 target classifier · CIFAR-10 target classification | accuracy: 58.15% -> 58.08% | Hurt |
| Replace all inverse Hessians with identity matrices (ablation) | Multi-Stage Influence Function · MNIST pretraining-to-finetuning influence estimation | Pearson's r: 0.17 | Hurt |
| Use a finetuning task identical to the pretraining task rather than automobile-vs-deer | CIFAR-10 Multi-Stage Influence Function · CIFAR-10 same-task versus different-task finetuning | Average absolute influence score: 0.025 -> 0.055 | Mixed |
| Triple the number of finetuning examples and finetuning steps | CIFAR-10 Multi-Stage Influence Function · CIFAR-10 finetuning models C and D | Average absolute influence score: 0.22 -> 0.15 | Hurt |
| Use the proposed influence score when embeddings are updated during finetuning | Multi-Stage Influence Function · CIFAR-10 pretraining-to-finetuning influence estimation | Pearson's r: Pearson's r=0.40 | Improved |
Closest papers by meaning
- Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data (NeurIPS 2025 · similarity 0.86)
- Harnessing Diversity for Important Data Selection in Pretraining Large Language Models (ICLR 2025 · similarity 0.84)
- Towards Robust Influence Functions with Flat Validation Minima (ICML 2025 · similarity 0.84)
- Fine-tuning can Help Detect Pretraining Data from Large Language Models (ICLR 2025 · similarity 0.84)
- Efficient stagewise pretraining via progressive subnetworks (ICLR 2025 · similarity 0.84)
- Capturing the Temporal Dependence of Training Data Influence (ICLR 2025 · similarity 0.84)
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.
