AI Paper Atlas

Research map / Large language models / Language, safety and interpretability

Curriculum-Guided Layer Scaling for Language Model Pretraining

Karanpartap Singh, Neil Band, Ehsan Adeli

ICML 2026 · Large language models › Language, safety and interpretability · approach: Pruning and circuits in LLMs · 0 citations (OpenAlex)

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 changedWhereMeasuredEffect
add the freezing phase when introducing new layers (ablation)CGLS · Llama-3.2-1B (2.5B tokens)average accuracy: 37.46 -> 39.08Improved
add the freezing phase when introducing new layers (ablation)CGLS · Llama-3.2-1B (2.5B tokens)stage 1 mixture perplexity: 30.84 -> 24.50Improved
add synchronized (easy to hard) curriculum ordering instead of reversed (ablation)CGLS · Llama-3.2-1B (2.5B tokens)average accuracy: 32.52 -> 39.08Improved
use FineWeb difficulty classifier instead of DistilBERT classifier (ablation)CGLS · Llama-3.2-1B (2.5B tokens)average accuracy: 39.08 -> 39.06No clear effect
add layer scaling and data curriculum together (CGLS) versus layer scaling only (ablation)Layer Scaling Only · Llama-3.2-1B (2.5B tokens)average accuracy: 34.24% -> 39.08%Improved
add data curriculum to the standard baseline (ablation)Baseline (Randomized) · Llama-3.2-1B (2.5B tokens)average accuracy: 37.38% -> 37.33%No clear effect
add layer scaling with curriculum (CGLS) versus randomized baselineBaseline (Randomized) · GPT-2-Small (2.5B tokens)average accuracy: 35.13% -> 36.61%Improved
use CGLS instead of randomized baselineBaseline (Randomized) · Llama-3.2-1B (2.5B tokens)The Pile perplexity: 20.83 -> 25.34Hurt
use CGLS instead of randomized baseline at larger scaleBaseline (Randomized) · Llama-3.2-1B (20B tokens)average accuracy: 38.41% -> 42.30%Improved
use CGLS instead of randomized baseline at larger scaleBaseline (Randomized) · Llama-3.2-1B (20B tokens)The Pile perplexity: 26.33 -> 22.99Improved
order data from general text to code with growing model (CGLS)Baseline (Randomized mix) · HumanEval (DCLM to StarCoder2)pass@8: 3.56% -> 6.67%Improved
use CGLS instead of baseline across three seedsBaseline · Llama-3.2-1B (2.5B tokens, replication)average accuracy: 33.26 ± 0.24% -> 35.94 ± 1.60%Improved

Closest papers by meaning

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.

Open scime draft

Get the next scime Dispatch and a monthly field update

Free. We send a confirmation link first; every email has a one-click unsubscribe. Privacy