AI Paper Atlas

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

Learning More from Less: Unlocking Internal Representations for Benchmark Compression

Yueqi Zhang, Jin Hu, Shaoxiong Feng, Peiwen Yuan, Xinglin Wang, Yiwei Li, Jiayi Shi, Chuyi Tan, Ji Zhang, Boyuan Pan, Yao Hu, Kan Li

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
cluster on aligned consensus embeddings instead of binary correctness vectors (ablation)RepCore · five benchmarks, K=30Spearman rho: 0.850 -> 0.868Improved
cluster on aligned consensus embeddings instead of binary correctness vectors (ablation)RepCore · five benchmarks, K=30MAE: 0.055 -> 0.049Improved
use RepCore instead of AnchorPoints with only five source modelsRepCore · five benchmarks, K=30, |S|=5Spearman rho: 0.640 -> 0.860Improved
add PC2 and PC3 to PC1 for extrapolation (ablation)RepCore · BBHSpearman rho: K=30: 0.857 -> 0.871 (PC1 vs PC1,2)Improved
use full aligned embedding instead of PC1,2 subspace (ablation)RepCore · BBHSpearman rho: K=10: 0.689 -> 0.718; K=50: 0.917 -> 0.913Mixed
use diverse source pool instead of a single model family (Llama / InternVL) (ablation)RepCore · five benchmarks, K=30Spearman rho: 0.686 -> 0.868Improved
use diverse source pool instead of strong-only models (ablation)RepCore · five benchmarks, K=30MAE: 0.092 -> 0.049Improved
increase source pool from 5 to 20 models (ablation)RepCore · five benchmarks, K=30Spearman rho: 0.860 -> 0.864No clear effect
increase coreset size to K=200RepCore vs GP-IRT · five benchmarksSpearman rho: 0.975 vs 0.973No clear effect
mask 50% of source hidden states at random (ablation)RepCore · five benchmarksSpearman rho: 0.842 -> 0.833Hurt
mask 50% of most distinct sources' hidden states (ablation)RepCore · five benchmarksSpearman rho: 0.842 -> 0.805Hurt
use RepCore on chronological split (old sources, new targets)RepCore · five benchmarks, K=30Spearman rho: 0.845 -> 0.853 (vs Random)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