AI Paper Atlas

Research map / Architectures and efficiency / Attention and transformers

What Layers When: Learning to Skip Compute in LLMs with Residual Gates

Filipe Laitenberger, Dawid Jan Kopiczko, Cees G. M. Snoek, Yuki M Asano

ICLR 2026 · Architectures and efficiency › Attention and transformers · approach: Sparse attention for long-context LLMs

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
Use GateSkip token skipping instead of random skippingLlama-3.2-1B · Generative avg, 15% savedaccuracy: 1.67 -> 22.14Improved
Use GateSkip instead of MoD router tuning, CALM, FREE, LayerSkipLlama-3.2-1B · Generative avg, 15% savedaccuracy: MoD 3.96, FREE 10.70, LayerSkip 10.65 -> 22.14Improved
Place gate after the module output rather than at entry (ablation)GateSkip · Generative avg, 5% savedaccuracy: 5.5 -> 25.5Improved
Use vector gates instead of scalar gates (ablation)GateSkip · Generative avg, 15% savedaccuracy: 20.4 -> 23.2Improved
Use separate gates per layer instead of shared (ablation)GateSkip · Generative avg, 15% savedaccuracy: 20.7 -> 23.2Improved
Skip both attention and MLP rather than only one (ablation)GateSkip · Generative avg, 15% savedaccuracy: 14.9 (attn only), 7.8 (MLP only) -> 23.2Improved
Use a linear-sigmoid gate instead of an MLP gate (ablation)GateSkip · Generative avg, 15% savedaccuracy: 18.5 -> 23.2Improved
Use L2 gate loss instead of L1 (ablation)GateSkip · Generative avg, 15% savedaccuracy: 21.4 -> 23.2Improved
Use L2 loss instead of layer-wise KL regularizer (ablation)GateSkip · Generative avg, 15% savedaccuracy: 9.8 -> 23.2Improved
Unfreeze the backbone while training gates (ablation)GateSkip · Generative avg, 15% savedaccuracy: 12.7 -> 23.2Improved
Apply GateSkip to instruction-tuned Llama-3BLlama-3.2-3B-Instruct · Generative avgaccuracy: 36.5 -> 49.0Improved
Add 4-bit quantization to GateSkipLlama-3.2-3B · Generative avg, 0% savedaccuracy: 45.0 -> 42.5No clear effect

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