Research map / Architectures and efficiency / RNNs and normalization
Feature Integration Spaces: Joint Training Reveals Dual Encoding in Neural Network Representations
Omar Claflin
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 a neural factorization machine on SAE features to predict reconstruction residuals | SAE (OpenLLaMA-3B layer 16) · WikiText-103 validation | reconstruction error: 0.3672 -> 0.2811 | Improved |
| train SAE and NFM jointly instead of sequentially (ablation) | SAE + NFM · WikiText-103 | reconstruction improvement over SAE: 23% -> 41.3% | Improved |
| train SAE and NFM jointly instead of sequentially (ablation) | SAE + NFM · WikiText-103 | pathological KL divergence reduction: 29.8% -> 51.6% | Improved |
| add nonlinear NFM component on top of linear component (ablation) | Joint SAE + NFM · WikiText-103 | KL divergence reduction: 50.9% -> 51.9% | No clear effect |
| nonlinear component alone (ablation) | Joint SAE + NFM · WikiText-103 | share of total improvement: 16.5% with 3% of parameters | Improved |
| add NFM integration features then clamp secondary feature 4022 | Secondary SAE feature 4022 · 2x2 formality x emotion logit effects | ANOVA interaction: F=5.06, p=0.027 | Improved |
| clamp linear integration weights (ablation) | NFM linear component · 2x2 formality x emotion logit effects | ANOVA interaction | No clear effect |
| clamp non-sensitive secondary feature 1113 (ablation) | Secondary SAE feature 1113 · 2x2 formality x emotion logit effects | ANOVA: F=0.022, p=0.996 | No clear effect |
| secondary SAE trained directly on original residuals instead of NFM embeddings (ablation) | Secondary SAE · reconstruction loss | reconstruction loss | No clear effect |
| joint training produces low squared norm features specialized for integration | Joint SAE + NFM · SAE feature weight norms | feature integration contribution by weight: <0.2: 82.8% vs >0.2: 71.3% | Improved |
Closest papers by meaning
- From Flat to Hierarchical: Extracting Sparse Representations with Matching Pursuit (NeurIPS 2025 · similarity 0.89)
- Projecting Assumptions: The Duality Between Sparse Autoencoders and Concept Geometry (NeurIPS 2025 · similarity 0.89)
- Semantic Optimal Transport for Sparse Autoencoder Feature Matching and Circuit Compression (NeurIPS 2026 · similarity 0.88)
- Bridging Critical Gaps in Convergent Learning: How Representational Alignment Evolves Across Layers, Training, and Distribution Shifts (NeurIPS 2025 · similarity 0.88)
- Mechanistic Interpretability with Sparse Autoencoder Neural Operators (NeurIPS 2026 · similarity 0.87)
- Towards Principled Evaluations of Sparse Autoencoders for Interpretability and Control (ICLR 2025 · 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.
