Research map / Large language models / Language, safety and interpretability
BLISS: A Lightweight Bilevel Influence Scoring Method for Data Selection in Language Model Pretraining
Jie Hao, Rui Yu, Wei Zhang, Huixia Wang, Jie Xu, Mingrui Liu
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 softmax reparametrization of score outputs (ablation) | BLISS · Zero-shot average, 410M, 10B tokens, 20k steps | average accuracy: 43.6 -> 44.7 | Improved |
| use bilevel instead of single-level update (ablation) | BLISS · Zero-shot average, 410M, 10B tokens, 20k steps | average accuracy: 44.2 -> 44.7 | Improved |
| use 160M proxy instead of 31M (ablation) | BLISS · Zero-shot average, 410M, 10B tokens, 20k steps | average accuracy: 44.7 -> 44.1 | No clear effect |
| initialize score model from previous round rather than round-1 reset (ablation) | BLISS · Zero-shot average, 410M, 15B tokens | average accuracy: 44.5 -> 44.9 | Improved |
| replace MATES with BLISS | Pythia-410M · Zero-shot average, 9 tasks, C4 25B tokens | average accuracy: 45.7 -> 45.9 | Improved |
| replace MATES with BLISS | Pythia-1B · Zero-shot average, 9 tasks, C4 25B tokens | average accuracy: 47.5 -> 47.9 | Improved |
| replace MATES with BLISS at round 3 | Pythia-2.8B · Round 3, 15B tokens | average accuracy: 47.6 -> 49.0 | Improved |
| replace MATES with BLISS at round 1 | Pythia-2.8B · Round 1 (random), 15B tokens | average accuracy: 45.9 -> 45.2 | No clear effect |
| replace MATES with BLISS at round 3 | LLaMA-0.5B · LLaMA-0.5B Round 3 | average accuracy: 45.01 -> 45.65 | Improved |
| replace MATES with BLISS for selection time | Pythia-410M · Data selection time, target 410M | hours: 18.0 -> 5.03 | Improved |
| replace MATES with BLISS for memory | Pythia-1B · Data selection memory, target 1B | memory peak (GB): 63.52 -> 74.51 | Hurt |
| increase bilevel steps from 3k to 5k (ablation) | BLISS · Pythia proxy training loss | training loss | 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.
- MATES — MATES: Model-Aware Data Selection for Efficient Pretraining with Data Influence Models (NeurIPS 2024) · better on 8 of 11 shared results
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
- MATES: Model-Aware Data Selection for Efficient Pretraining with Data Influence Models (NeurIPS 2024 · similarity 0.91)
- BLADE: Scalable Bi-level Adaptive Data Selection for LLM Training (NeurIPS 2026 · similarity 0.91)
- Efficient Data Selection at Scale via Influence Distillation (NeurIPS 2025 · similarity 0.90)
- Harnessing Diversity for Important Data Selection in Pretraining Large Language Models (ICLR 2025 · similarity 0.90)
- LLM Data Selection and Utilization via Dynamic Bi-level Optimization (ICML 2025 · similarity 0.90)
- Predictive Data Selection: The Data That Predicts Is the Data That Teaches (ICML 2025 · similarity 0.89)
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.
