TL;DR
TaxoSAE moves beyond the "flat" feature spaces of standard Sparse Autoencoders (SAEs) by explicitly encoding hierarchical "is-a" relationships into LLM representations. By stacking SAEs to mirror real-world taxonomies and using a novel taxonomic alignment loss, it achieves significantly more consistent knowledge readout and precise, granular control over internal model states.
What Problem Does This Paper Address?
Mechanistic interpretability aims to crack open the "black box" of Large Language Models (LLMs). Sparse Autoencoders (SAEs) have emerged as a leading tool for this, decomposing complex activations into sparse, interpretable features. However, current SAEs treat every feature as a standalone, independent unit.
In reality, human knowledge is taxonomically organized. A "T cell" is a type of "lymphocyte," which is a type of "immune cell." Standard SAEs ignore these logical dependencies, leading to two major issues:
- Taxonomic Inconsistency: A readout might show the "T cell" feature is active while the "lymphocyte" feature remains silent—a logical impossibility in a proper taxonomy.
- Lack of Control Granularity: Because features are "flat," editing a specific concept (like replacing a T cell with a B cell) doesn't naturally respect the broader category (they are both lymphocytes). You cannot easily specify the scope or scale of an edit.
Methods
The authors propose TaxoSAE, which bridges the gap between neural activations and symbolic logic through two main components:
- Stacked SAEs: Instead of one giant SAE, TaxoSAE uses a stack of SAEs, where each level corresponds to a level in a target taxonomy (e.g., Species $\to$ Class $\to$ Phylum). The first SAE reads the LLM's hidden state and maps it to specific terminal concepts. Each subsequent SAE reads the reconstruction of the level below and maps it to higher-level hypernyms.
- Taxonomic Alignment Loss: This is the core innovation. It enforces the "is-a" logic: a hypernym slot is trained to fire whenever any of its hyponyms (children) are active. This turns a logical rule into a differentiable training objective.
To handle the complexity of factual questions, each SAE level uses three types of "slots":
- Taxonomic slots: Dedicated to specific nodes in the taxonomy.
- Complement slots: Dedicated to the "other" part of a fact (e.g., if the taxonomy is animals, the complement might be "lifespan" or "diet").
- Free slots: Standard unsupervised slots to capture the remaining variance in the representation.
Datasets
The researchers evaluated TaxoSAE on two custom-built factual QA taxonomies:
- Bio: A relation taxonomy based on Freebase biographical attributes (18 terminal relations like "birth city," grouped into 6 intermediate and 3 top-level categories).
- Animal: An entity taxonomy based on ConceptNet (18 terminal animals like "cow" or "frog," grouped into 6 intermediate classes and 2 top-level classes).
These datasets were used to fine-tune base models (GPT-2 small and Qwen2.5-0.5B) to ensure they actually "knew" the facts before the SAEs were trained to interpret them.
Key Insights and Ideas
- Logic as a Geometric Constraint: By enforcing taxonomic alignment, the authors essentially force the SAE to learn a feature geometry that respects logical hierarchies.
- Generalization to the Unseen: One of the most striking findings is that TaxoSAE can identify the category of concepts it was never trained on. If the model sees a question about a "beetle" (unseen in SAE training), the higher-level SAEs can still correctly activate the "insect" slot because the underlying LLM representation for "beetle" is geometrically similar to other insects.
- Concentrated Activations: Unlike unsupervised SAEs, which often fragment a single concept across dozens of features, TaxoSAE concentrates information into single, dedicated slots. This makes the features much more "readable" and "editable."
Key Results
- Superior Readout: TaxoSAE outperformed baselines by +31.3 percentage points in reading out hypernyms for abstract concepts that were never seen during SAE training.
- Precise Control: In "swap" experiments (e.g., forcing the model to think a "cow" is an "eagle"), TaxoSAE achieved a +25.5 point improvement in success rate.
- Selective Suppression: Deactivating a specific taxonomic slot (like "mammal") successfully suppressed answers for all mammals (0-2% accuracy) while leaving other categories (like "birds") completely unaffected (91-100% accuracy).
- Monosemanticity: As shown in Figure 5, TaxoSAE activations are highly concentrated, reaching near-ideal "Top1C" scores (the share of activation held by the strongest unit), whereas unsupervised SAEs remain fragmented.
Limitations and Future Work
- Scale: The taxonomies used were relatively small (18 terminal nodes). Real-world taxonomies (like the Tree of Life or WordNet) are orders of magnitude larger.
- Language: The study was limited to English-language models and datasets.
- Complexity of Relations: The current work focuses strictly on "is-a" (hierarchical) relations. Future work could extend this logic to other types of relationships, such as "part-of" or causal dependencies.
- Model Size: While tested on GPT-2 and Qwen2.5-0.5B, the approach needs validation on frontier-scale models (e.g., Llama-3 70B+).
Summary updated