Research map / Large language models / Language, safety and interpretability
Inverse Depth Scaling From Most Layers Being Similar
Yizhou Liu, Sara Kangaslahti, Ziming Liu, Jeff Gore
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 |
|---|---|---|---|
| increase LLM depth (fit of decomposed scaling law) | LLMs (Chinchilla-reconstructed data) · LLM pretraining loss | depth-dependent loss exponent: alpha_l = 1.2 ± 0.3 | Improved |
| increase depth in LLMs (Pythia family) | Pythia models · FineWeb hidden-state updates | mean update angle in middle layers | Improved |
| tie teacher weights across layers (smooth dynamics) at low teacher temperature (ablation) | toy residual student · teacher-student toy model, depth 6 to 48 | depth exponent alpha_l: alpha_l about 1 (imperfect training) | No clear effect |
| tie teacher weights and raise teacher temperature (ablation) | toy residual student · teacher-student toy model | depth exponent alpha_l: 1 -> 3 | Improved |
| train longer on tied-teacher toy model at temperature 1 (ablation) | toy residual student · teacher-student toy model | depth exponent alpha_l: near 1 early -> approaches 3 late | Improved |
| use independent teacher weights across layers (non-smooth dynamics) (ablation) | toy residual student · teacher-student toy model | depth exponent alpha_l: alpha_l about 1 across temperatures | No clear effect |
| mimic higher-order integration scheme in the student architecture (tied teacher) (ablation) | toy residual student · teacher-student toy model | depth exponent alpha_l | Improved |
Closest papers by meaning
- Do Language Models Use Their Depth Efficiently? (NeurIPS 2025 · similarity 0.92)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · similarity 0.92)
- Position: Enough of Scaling LLMs! Lets Focus on Downscaling (ICML 2025 · similarity 0.90)
- LLMs on the Line: Data Determines Loss-to-Loss Scaling Laws (ICML 2025 · similarity 0.90)
- Remarkable Robustness of LLMs: Stages of Inference? (NeurIPS 2025 · similarity 0.88)
- Universal One-third Time Scaling in Learning Peaked Distributions (ICML 2026 · similarity 0.87)
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.
