Research map / Learning theory and algorithms / Prediction and decision losses
TRACE: Trajectory-based Activation Change Estimation for Task-specific Data Selection
Ye He, Shangzhan Li, Yuxin Zhou, Qi Shi
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 warmup stage (ablation) | TRACE (LLaMA-3.1-8B, layer 15) · Avg BBH/MMLU/TyDiQA | avg score: 65.93 -> 68.41 | Improved |
| add validation-set tuning step (ablation) | TRACE (LLaMA-3.1-8B, layer 15) · Avg BBH/MMLU/TyDiQA | avg score: 65.81 -> 68.41 | Improved |
| use official instruct model for selection instead of warmup model (ablation) | TRACE (LLaMA-3.1-8B, layer 15) · Avg BBH/MMLU/TyDiQA | avg score: 68.41 -> 66.42 | Hurt |
| aggregate TRACE scores over multiple warmup checkpoints (ablation) | TRACE (LLaMA-3.1-8B) · Avg BBH/MMLU/TyDiQA | avg score: 67.63 (1 epoch) vs 67.57/67.15/67.40 | No clear effect |
| select 5% with TRACE(kmeans) instead of full-data training | LLaMA-3.1-8B · TyDiQA | F1: 73.68 -> 76.50 | Improved |
| select 5% with TRACE instead of full data | OLMo-7B · Avg BBH/MMLU/TyDiQA | avg score: 34.93 -> 35.10 (kmeans) / 34.69 (random) | Mixed |
| random 5% selection vs full data | Qwen-2.5-7B · Avg BBH/MMLU/TyDiQA | avg score: 69.12 -> 69.76 | Improved |
| use TRACE instead of MoNA | LLaMA-3.1-8B · Avg BBH/MMLU/TyDiQA | avg score: 67.20 -> 68.41 | Improved |
| use kmeans instead of random warmup data (ablation) | TRACE (LLaMA-3.1-8B) · Avg BBH/MMLU/TyDiQA | avg score: 67.42 -> 68.41 | Improved |
Closest papers by meaning
- Task-Specific Data Selection for Instruction Tuning via Monosemantic Neuronal Activations (NeurIPS 2025 · similarity 0.87)
- A Bayesian Model Selection Criterion for Selecting Pretraining Checkpoints (ICML 2025 · similarity 0.85)
- LESS: Selecting Influential Data for Targeted Instruction Tuning (ICML 2024 · similarity 0.85)
- Towards Best Practices of Activation Patching in Language Models: Metrics and Methods (ICLR 2024 · similarity 0.85)
- Capturing the Temporal Dependence of Training Data Influence (ICLR 2025 · similarity 0.84)
- What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning (ICLR 2024 · 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.
