SAE-IDF matches symbols across systems using sparse autoencoder codes
Original title: Sparse Autoencoders Are a Shared Coordinate System for Symbolic Matching
On 12 tasks across six symbolic systems, SAE-IDF beats the same LLM's dense baseline by up to 45 points.
Researchers present SAE-IDF, a data-driven way to match units across heterogeneous symbolic systems so they can later be aligned and merged. A symbol is encoded in its native context with a sparse autoencoder, dimensions are re-weighted by rarity, and matches are scored across systems. On 12 tasks across six kinds of symbolic system, SAE-IDF beats the same model's dense baseline on every task with Gemma-2-9B and all but one with Llama-3.1-8B, by up to 45 points. Each sparse dimension has a human-readable meaning, so matches decompose into named features.
- Gains hold across SAE layers and configurations.
- One Llama-3.1-8B task did not beat the dense baseline.
- The authors frame this as a first step toward a larger symbolic backbone.
From the original article
1 min read at huggingface.co
Large language models (LLMs) hold broad knowledge in one set of weights, but lack interpretability and trustworthy reasoning, especially in high-stakes use cases.
By Xiaocong Yang, Xuanyu Meng, Yuji Zhang · huggingface.co
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.