Research map / Vision-language and multimodal / Human and embodied modeling
TRAK: Attributing Model Behavior at Scale
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's title only (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 |
|---|---|---|---|
| Use TRAK instead of TracIn for fact tracing | mT5-small · Ftrace-TREx | MRR: 0.09 vs. 0.42 | Helped |
| Remove abstracts ranked by TRAK rather than by BM25 | mT5-small · Ftrace-TREx counterfactual removal | accuracy decrease: 10% -> 34% | Helped |
| Remove abstracts ranked by TRAK rather than ground-truth proponents | mT5-small · Ftrace-TREx counterfactual removal | accuracy decrease: 12% -> 34% | Helped |
| Remove the most positively attributed training examples | CLIP ResNet-50 · MS COCO CLIP counterfactual evaluation | cosine similarity decrease: 0.11 -> 0.36 | Helped |
| Use TRAK rather than TracIn to select CLIP training examples for removal | CLIP ResNet-50 · MS COCO CLIP counterfactual evaluation | cosine similarity decrease: TRAK: 0.36; TracIn: no significant effect | Helped |
| Train attribution models for fewer epochs (ablation) | TRAK on ResNet-9 · CIFAR-10 | LDS: 0.308 -> 0.100 (25 to 1 epochs) | Hurt |
| Use 5 rather than 100 independent models (ablation) | TRAK · CIFAR-10 | LDS: 0.355 -> 0.329 | Hurt |
| Use multiple checkpoints from one training run instead of independently trained models (ablation) | TRAK · CIFAR-10 | LDS | No clear effect |
| Use soft-thresholding (ablation) | TRAK · CIFAR-10 and ImageNet | LDS | Helped |
| Use soft-thresholding (ablation) | TRAK · CIFAR-2 and QNLI | LDS | No clear effect |
| Use TRAK rather than BM25 to retrieve ground-truth proponents | mT5-small · Ftrace-TREx | MRR: 0.77 vs. 0.42 | Hurt |
Closest papers by meaning
- quanda: An Interpretability Toolkit for Training Data Attribution Evaluation (NeurIPS 2026 · similarity 0.88)
- Fast Data Attribution for Text-to-Image Models (NeurIPS 2025 · similarity 0.87)
- $\texttt{dattri}$: A Library for Efficient Data Attribution (NeurIPS 2024 · similarity 0.86)
- Final-Model-Only Data Attribution with a Unifying View of Gradient-Based Methods (NeurIPS 2025 · similarity 0.86)
- Data Attribution for Text-to-Image Models by Unlearning Synthesized Images (NeurIPS 2024 · similarity 0.86)
- Taming Hyperparameter Sensitivity in Data Attribution: Practical Selection Without Costly Retraining (NeurIPS 2025 · similarity 0.85)
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.