AI Paper Atlas

Research map / Architectures and efficiency / RNNs and normalization

Capacity without Access: Reinterpreting the Mid-Depth Spectral Plateau in LLMs

SeongMin Kang, Woo-Seong Yun, Nahyun Lee, Yoon-Sik Cho

ICML 2026 · Architectures and efficiency › RNNs and normalization · approach: Layer-wise analysis of language models

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 DDI training-time depth-aware readout route (internalized, removed at inference) to LoRA fine-tuningLlama3-8B LoRA · Commonsense reasoning avgaccuracy: 84.7 -> 85.2Improved
add DDI route to LoRA fine-tuningMistral-7B LoRA · Arithmetic reasoning avgaccuracy: 66.9 -> 67.8Improved
add DDI route to LoRA fine-tuningQwen2-7B LoRA · Arithmetic reasoning avgaccuracy: 76.5 -> 76.9Improved
add DDI route on pattern-driven benchmarks (AddSub, MultiArith, ARC-E, PIQA)LoRA models · AddSub / MultiArith / ARC-E / PIQAaccuracy: gains often below ~0.5 pointsNo clear effect
add DDI route on WinoGrande and OBQALoRA models · WinoGrande / OBQAaccuracy: WinoG. Llama3 85.5 -> 85.1; Mistral 83.0 -> 84.5Mixed
replace hierarchical depth-aware route with mean pooling over layers (parameter matched) (ablation)DDI · Control experiment avgAvg accuracy (%): 75.7 -> 75.8No clear effect
replace hierarchical route with learned weighted sum over layers (parameter matched) (ablation)DDI · Control experiment avgAvg accuracy (%): 75.7 -> 75.8No clear effect
use hierarchical depth-aware aggregation instead of simple controls (ablation)DDI · Control experiment avgAvg accuracy (%): 75.8 -> 76.6Improved
train with DDILlama3-8B · Effective rank, plateau layers L10-L12effective rank change: +0.0011No clear effect
train with DDILlama3-8B · Access ratio, final layers L29-L31access ratio change: +0.01397Improved
train with DDILlama3-8B · Layer 27 CKA to pretrainedCKA: 0.860 -> 0.893Improved

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