Research map / Architectures and efficiency / RNNs and normalization

ScopeSAE: Model-Scope Feature Discovery with Interpretable Layer Selection

NeurIPS 2026 · Architectures and efficiency › RNNs and normalization · approach: Sparse dictionary representations · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes ScopeSAE, which picks each token's most prediction-relevant layer via normalized gradient-based attribution and trains a TopK SAE over those selected states, reporting a reconstruction-better-than-original effect and gains in feature count, interpretability, utilization, and redundancy on Llama-3.2-1B-Instruct.

Problem

Layer-wise sparse autoencoders fix the modeling subspace by layer identity, independent of which token-layer states drive next-token prediction, leading to low feature utilization, redundancy, and features weakly grounded in behavior.

Approach

Select per token the layer whose hidden state has the largest normalized gradient-times-activation attribution to the next-token log-probability, and train a single TopK SAE over the union of these prediction-relevant token-layer states, writing reconstructions back only to the selected position.

Key results

• ScopeSAE reaches DeltaCE of -0.20 and Recovered 1.03 at k=64, the only method with negative DeltaCE, while TopK is +0.10.
• ScopeSAE extracts 5,319 interpretable features versus 2,916 for TopK and 2,784 for RouteSAE.
• ScopeSAE achieves 99.8% feature utilization and lowest dictionary cosine similarity (0.0202) among compared methods.

Surprising findings

• Written-back SAE reconstructions yield lower next-token cross-entropy than the original activations (reconstruction-better-than-original effect).
• RouteSAE's learned router collapses 99.4% of selections onto a single layer, effectively behaving as a single-layer SAE.

Limitations

• Evaluation on a single small model (Llama-3.2-1B-Instruct) in the main experiments; generalization is reported only in appendix analyses.
• Interpretability and steering rely on GPT-4o judgments and small samples (50 prompts per condition, 100 features).

Experimental setup

Datasets: OpenWebText2 HH-RLHF

Benchmarks: OpenWebText2 held-out tokens (10M) HH-RLHF pairwise steering

Metrics: DeltaCE Recovered NormMSE KL divergence interpretable feature count interpretability score feature utilization dictionary redundancy win rate

Baselines: ReLU SAE Gated SAE TopK SAE RouteSAE Crosscoder

Builds on: TopK SAE RouteSAE Sparse Crosscoder Gated SAE ReLU SAE

Closest papers by meaning

• SASA: Subspace-Aware Sparse Autoencoders for Effective Mechanistic Interpretability (NeurIPS 2026 · similarity 0.89)
• Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · similarity 0.87)
• Learning Multi-Level Features with Matryoshka Sparse Autoencoders (ICML 2025 · similarity 0.87)
• Sanity Checks for Sparse Autoencoders: Do SAEs Beat Random Baselines? (NeurIPS 2026 · similarity 0.86)
• Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.86)
• From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.86)