Research map / Large language models / Language, safety and interpretability
Grokking in LLM Pretraining? Monitor Memorization-to-Generalization without Test
Ziyue Li, Chenrui Fan, Tianyi Zhou
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 |
|---|---|---|---|
| monitor pathway consistency instead of training loss as generalization indicator | test-accuracy correlation · Math | Pearson correlation: -0.0435 -> 0.9246 | Improved |
| monitor pairwise pathway distance instead of validation loss | test-accuracy correlation · Commonsense | Pearson correlation: -0.5752 -> -0.9821 | Improved |
| monitor pathway consistency instead of training loss | test-accuracy correlation · Domain-Specific | Pearson correlation: 0.7944 -> 0.9917 | Improved |
| increase LoRA rank from 32 to 64 (ablation) | pathway distance correlation · Commonsense | Pearson correlation: -0.9821 -> -0.9927 | No clear effect |
| increase LoRA rank from 32 to 64 (ablation) | pathway consistency correlation · Commonsense | Pearson correlation: 0.9804 -> 0.9559 | No clear effect |
| vary routing-weight threshold from 0.5 to 0.8 (ablation) | pathway edit distance trend · Domain-specific QA | edit distance trend | No clear effect |
| vary loss threshold epsilon from 1.5 to 2.0 (ablation) | pathway edit distance trend · Domain-specific QA | edit distance trend | No clear effect |
| vary stability threshold delta between 0.03 and 0.05 (ablation) | pathway edit distance trend · Domain-specific QA | edit distance trend | No clear effect |
| router weights versus expert weights update speed | convergence speed · OLMoE routing dynamics | convergence speed: 0.151 vs 0.063 | Mixed |
Closest papers by meaning
- What Do Learning Dynamics Reveal About Generalization in LLM Mathematical Reasoning? (ICML 2025 · similarity 0.92)
- Unveiling the Secret Recipe: A Guide For Supervised Fine-Tuning Small LLMs (ICLR 2025 · similarity 0.89)
- Generalization v.s. Memorization: Tracing Language Models’ Capabilities Back to Pretraining Data (ICLR 2025 · similarity 0.89)
- To Grok or not to Grok: Disentangling Generalization and Memorization on Corrupted Algorithmic Datasets (ICLR 2024 · similarity 0.88)
- Prismatic Synthesis: Gradient-based Data Diversification Boosts Generalization in LLM Reasoning (NeurIPS 2025 · similarity 0.88)
- Towards Effective Evaluations and Comparisons for LLM Unlearning Methods (ICLR 2025 · similarity 0.88)
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.
