Research map / Architectures and efficiency / RNNs and normalization
First is Better Than Last for Language Data Influence
Chih-Kuan Yeh, Ankur Taly, Mukund Sundararajan, Frederick Liu, Pradeep Kumar Ravikumar
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 |
|---|---|---|---|
| Compute TracIn influence on word embeddings instead of the last layer | TracIn · Toxicity, AGnews, MNLI | AUC-DEL: Toxicity AUC-DEL+: -0.013 -> -0.100; AUC-DEL-: 0.023 -> 0.149; AGnews AUC-DEL+: -0.021 -> -0.166; AUC-DEL-: 0.028 -> 0.1 | Improved |
| Compute TracIn using first-layer rather than last-layer weights (ablation) | TracIn on CNN · AGnews | AUC-DEL+: -0.016 -> -0.077 | Improved |
| Remove bias parameters from the TracIn calculation (ablation) | TracIn on CNN · AGnews | AUC-DEL+: -0.036 -> -0.065 | Improved |
| Replace last-layer similarity with TF-IDF similarity (ablation) | TracIn · Toxicity, AGnews | AUC-DEL+: Toxicity: -0.013 -> -0.067; AGnews: -0.021 -> -0.090 | Improved |
| Use top-k input word gradients to approximate word-embedding gradients (ablation) | TracIn-WE · Toxicity, AGnews | AUC-DEL+: Toxicity: -0.100 -> -0.099; AGnews: -0.166 -> -0.174 | Mixed |
| Use top-k input word gradients to approximate word-embedding gradients (ablation) | TracIn-WE · MNLI | AUC-DEL+: 0.006 -> -0.198 | Improved |
| Restrict TracIn-WE to the start and end tokens (ablation) | TracIn-WE · Toxicity | AUC-DEL+: -0.100 -> 0.016 | Hurt |
| Remove common tokens from TracIn-WE (ablation) | TracIn-WE · Toxicity without word overlap | AUC-DEL+: -0.018 -> 0.003 | Hurt |
| Remove common tokens from TracIn-WE (ablation) | TracIn-WE · Toxicity without word overlap | AUC-DEL-: 0.026 -> 0.001 | Hurt |
| Allow synonym matches between words (ablation) | TracIn-WE-Topk · Toxicity | AUC-DEL+: -0.099 -> 0.016 | Hurt |
| Allow synonym matches between words (ablation) | TracIn-WE-Topk · AGnews | AUC-DEL+: -0.174 -> -0.017 | Hurt |
Closest papers by meaning
- Scalable Influence and Fact Tracing for Large Language Model Pretraining (ICLR 2025 · similarity 0.86)
- LayerIF: Estimating Layer Quality for Large Language Models using Influence Functions (NeurIPS 2025 · similarity 0.86)
- Neural Networks for Learnable and Scalable Influence Estimation of Instruction Fine-Tuning Data (NeurIPS 2025 · similarity 0.85)
- Efficient Data Selection at Scale via Influence Distillation (NeurIPS 2025 · similarity 0.85)
- Exploring Activation Patterns of Parameters in Language Models (AAAI 2025 · similarity 0.85)
- IDEAL: Influence-Driven Selective Annotations Empower In-Context Learners in Large Language Models (ICLR 2024 · 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.
