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
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 |
|---|---|---|---|
| add DDI training-time depth-aware readout route (internalized, removed at inference) to LoRA fine-tuning | Llama3-8B LoRA · Commonsense reasoning avg | accuracy: 84.7 -> 85.2 | Improved |
| add DDI route to LoRA fine-tuning | Mistral-7B LoRA · Arithmetic reasoning avg | accuracy: 66.9 -> 67.8 | Improved |
| add DDI route to LoRA fine-tuning | Qwen2-7B LoRA · Arithmetic reasoning avg | accuracy: 76.5 -> 76.9 | Improved |
| add DDI route on pattern-driven benchmarks (AddSub, MultiArith, ARC-E, PIQA) | LoRA models · AddSub / MultiArith / ARC-E / PIQA | accuracy: gains often below ~0.5 points | No clear effect |
| add DDI route on WinoGrande and OBQA | LoRA models · WinoGrande / OBQA | accuracy: WinoG. Llama3 85.5 -> 85.1; Mistral 83.0 -> 84.5 | Mixed |
| replace hierarchical depth-aware route with mean pooling over layers (parameter matched) (ablation) | DDI · Control experiment avg | Avg accuracy (%): 75.7 -> 75.8 | No clear effect |
| replace hierarchical route with learned weighted sum over layers (parameter matched) (ablation) | DDI · Control experiment avg | Avg accuracy (%): 75.7 -> 75.8 | No clear effect |
| use hierarchical depth-aware aggregation instead of simple controls (ablation) | DDI · Control experiment avg | Avg accuracy (%): 75.8 -> 76.6 | Improved |
| train with DDI | Llama3-8B · Effective rank, plateau layers L10-L12 | effective rank change: +0.0011 | No clear effect |
| train with DDI | Llama3-8B · Access ratio, final layers L29-L31 | access ratio change: +0.01397 | Improved |
| train with DDI | Llama3-8B · Layer 27 CKA to pretrained | CKA: 0.860 -> 0.893 | Improved |
Closest papers by meaning
- Do Language Models Use Their Depth Efficiently? (NeurIPS 2025 · similarity 0.89)
- Layer by Layer: Uncovering Hidden Representations in Language Models (ICML 2025 · similarity 0.88)
- Tracing the Representation Geometry of Language Models from Pretraining to Post-training (NeurIPS 2025 · similarity 0.88)
- Dynamics of the Transformer Residual Stream: Coupling Spectral Geometry to Network Topology (NeurIPS 2026 · similarity 0.88)
- The Geometric Wall: Manifold Structure Predicts Layerwise Sparse Autoencoder Scaling Laws (NeurIPS 2026 · similarity 0.87)
- Understanding Language Prior of LVLMs by Contrasting Chain-of-Embedding (ICLR 2026 · similarity 0.87)
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.
