Sparse Autoencoders with IDF Weighting Match Symbols Across Systems, Winning 124 of 129 Queries
AI-Interpretability-illinois · hf · 2026-10-19
UIUC's interpretability group presents SAE-IDF: sparse autoencoder features with rarity (IDF) weighting form a shared coordinate system that matches symbolic units across heterogeneous systems (e.g., NELL classes to DBpedia).
- On an interactive demo covering 129 NELL→DBpedia queries, SAE-IDF ranks gold first on 124 (MRR 0.976), beating dense hidden states (116, MRR 0.931) and unweighted SAE codes (118, MRR 0.938)
- Rarity weighting suppresses high-frequency generic features: for trainstation→RailwayStation, rare railroad-specific features dominate after IDF, while dense representations rank the gold class only #5
- The authors frame this as a foundation for scalable neuro-symbolic reasoning
More from Research
- MIST Framework Separates Execution from Memory to Diagnose Video World Model Failures — hkuhk · 2026-10-19
- Dusting the Phantom: Gaussian Allocation by Spatiotemporal Destination Recovers Fast-Moving Detail — Minhyeok Roh · 2026-10-19
- EvoHeir paper: fresh evaluation cuts harmful inherited updates in self-evolving LLM loops by 68–83% — MBZUAI · 2026-10-19
- ReaDy Uses Recurrent States as Dynamic Context for Speculative Decoding in Hybrid Linear Attention LLMs — Bohan Sun · 2026-10-19
- PrefEval: preference-following accuracy falls below 10% at just 10 turns for most LLMs — math_rachel · 2026-10-19
- ReasoningTrap: diagnosing reasoning rigidity with doctored variants of famous puzzles — math_rachel · 2026-10-19