Research map / Probabilistic and scientific modeling / Molecules and proteins
Random Scaling of Emergent Capabilities
Rosie Zhao, Tian Qin, David Alvarez-Melis, Sham M. Kakade, Naomi Saphra
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 |
|---|---|---|---|
| replace exact-match accuracy with a continuous max per-token loss metric (ablation) | addition length generalization · synthetic addition | distribution across seeds | No clear effect |
| replace accuracy with negative log likelihood (ablation) | Qwen2.5-0.5B · MMLU (10% MMLU mix) | distribution across seeds | No clear effect |
| increase MMLU share of training mix (ablation) | Qwen2.5-0.5B · MMLU validation | mean accuracy: 1.0% -> 23.0% | Improved |
| scale Qwen2.5-0.5B to Qwen2.5-1.5B at fixed data mix (ablation) | Qwen reinitialized · MMLU validation | accuracy across seeds | Improved |
| increase hierarchical-inducing data share (ablation) | from-scratch Transformer · question formation OOD | mean accuracy: 17.2% -> 62.5% (1% to 10%) | Improved |
| use 100% hierarchical-inducing data (ablation) | from-scratch Transformer · question formation OOD | seed variation: mean 74.8% | No clear effect |
| increase depth from 3 to 6 or 8 layers at 90% hierarchical data (ablation) | from-scratch Transformer · question formation OOD | accuracy distribution | Mixed |
| increase batch size from 8 to 64 (ablation) | Qwen2.5-1.5B · MMLU continued pretraining | bimodality | Hurt |
| scale width or depth of addition model (ablation) | decoder-only Transformer · reverse-order addition | W2 distance to largest scale | Improved |
| increase width with depth fixed at 1 layer (ablation) | addition Transformer · reverse-order addition | length generalization | No clear effect |
Closest papers by meaning
- U-shaped and Inverted-U Scaling behind Emergent Abilities of Large Language Models (ICLR 2025 · similarity 0.91)
- Predicting Emergent Abilities with Infinite Resolution Evaluation (ICLR 2024 · similarity 0.90)
- Understanding Emergent Abilities of Language Models from the Loss Perspective (NeurIPS 2024 · similarity 0.89)
- Observational Scaling Laws and the Predictability of Langauge Model Performance (NeurIPS 2024 · similarity 0.89)
- How Do Large Language Monkeys Get Their Power (Laws)? (ICML 2025 · similarity 0.89)
- Prescriptive Scaling Reveals the Evolution of Language Model Capabilities (ICML 2026 · 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.
