Research map / Large language models / Language, safety and interpretability
Rethinking Layer Relevance in Large Language Models Beyond Cosine Similarity
Cristian Hinostroza, Rodrigo Toro Icarte, Christ Devia, Andres Carvallo De Ferari, Eugenio Herrera-Berg, Denis Parra, Jorge F Silva
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 cosine-similarity relevance with accuracy-drop relevance for pruning (task-dependent) | LLaMA3-8B 25% pruning · 8-benchmark mean | accuracy: 59.54 -> 67.49 | Improved |
| replace cosine-similarity relevance with accuracy-drop relevance (task-independent, mixed calibration) | LLaMA3-8B 25% pruning · 8-benchmark mean | accuracy: 60.1 -> 65.04 | Improved |
| calibrate on a single dataset instead of a mix (C4) (ablation) | accuracy-based pruning · 8-benchmark mean, task-independent | accuracy: 65.04 -> 50.23 | Hurt |
| calibrate on ARC-E only instead of mixed (ablation) | accuracy-based pruning · 8-benchmark mean, task-independent | accuracy: 65.04 -> 63.18 | Hurt |
| use accuracy-based vs cosine similarity ranking (task-independent MMLU) | LLaMA3-8B · MMLU | accuracy: 44.07 -> 58.04 | Improved |
| healing (LoRA) after 25% pruning (ablation) | accuracy vs cosine vs random pruning · normalized average | normalized score: 60.5 vs 59.6 | No clear effect |
| healing after 50% pruning, accuracy-based vs cosine (ablation) | LLaMA3-8B · normalized average | normalized score: 31.2 -> 42.2 | Improved |
| healing after 75% pruning, accuracy-based vs cosine (ablation) | LLaMA3-8B · normalized average | normalized score: 9.3 -> 12.2 | Improved |
| accuracy-based vs cosine similarity cost | LLaMA3-8B · pruning runtime | time: 8.64 min -> 4.60 hrs | Hurt |
| use cosine similarity as proxy for layer relevance | Pythia, Mistral, OLMo · correlation with accuracy drop | Pearson R: R = -0.46 / -0.23 / -0.15 | No clear effect |
Closest papers by meaning
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.90)
- Layer as Puzzle Pieces: Compressing Large Language Models through Layer Concatenation (NeurIPS 2025 · similarity 0.90)
- LLM Layers Immediately Correct Each Other (NeurIPS 2025 · similarity 0.89)
- Adaptive Layer Sparsity for Large Language Models via Activation Correlation Assessment (NeurIPS 2024 · similarity 0.89)
- Exploring Activation Patterns of Parameters in Language Models (AAAI 2025 · similarity 0.89)
- The Curse of Depth in Large Language Models (NeurIPS 2025 · 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.
