SAE features are only partially compositional, new paper finds
OpenSAE · hf · 2026-10-17
A new paper, "Behavioral Compositionality of Sparse Autoencoder Representations," examines whether SAE features compose additively:
- Uses coalition interventions on decoded SAE contributions, with frozen model layers, to measure each feature subset's effect on a target logit
- Finds features are partially compositional: some prompts show near-additive responses (SAR₁ as low as 0.02), others strong pairwise interactions (SAR₁ = 0.40)
- Takeaway: accurate behavioral prediction and steering from SAE features requires interaction terms, not just first-order effects
More from Research
- Oscillator Language Models paper: phase-coded Kuramoto dynamics for language modeling — burny_tech · 2026-10-17
- Vals Science Index: OpenAI Models Lead Overall, but No Model Sweeps Every Science Discipline — BenBlaiszik · 2026-10-17
- First-author PhD paper frames AI text detection as continual learning, accepted at NeurIPS 2026 — lucy3_li · 2026-10-17
- Regents Labs curates 17 AI/agent science papers weekly with AI-generated deep reads — seanwbren · 2026-10-17
- A 'pre-numeric' AI math layer trains patterns before numbers exist — debreuil · 2026-10-17
- research-rag: local CPU-only PDF/ebook search with browser, CLI and MCP access — kishkisan · 2026-10-17