REVIEW 11 cited by
Quantifying Feature Space Universality Across Large Language Models via Sparse Autoencoders
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
read the original abstract
The Universality Hypothesis in large language models (LLMs) claims that different models converge towards similar concept representations in their latent spaces. Providing evidence for this hypothesis would enable researchers to exploit universal properties, facilitating the generalization of mechanistic interpretability techniques across models. Previous works studied if LLMs learned the same features, which are internal representations that activate on specific concepts. Since comparing features across LLMs is challenging due to polysemanticity, in which LLM neurons often correspond to multiple unrelated features rather than to distinct concepts, sparse autoencoders (SAEs) have been employed to disentangle LLM neurons into SAE features corresponding to distinct concepts. In this paper, we introduce a new variation of the universality hypothesis called Analogous Feature Universality: we hypothesize that even if SAEs across different models learn different feature representations, the spaces spanned by SAE features are similar, such that one SAE space is similar to another SAE space under rotation-invariant transformations. Evidence for this hypothesis would imply that interpretability techniques related to latent spaces, such as steering vectors, may be transferred across models via certain transformations. To investigate this hypothesis, we first pair SAE features across different models via activation correlation, and then measure spatial relation similarities between paired features via representational similarity measures, which transform spaces into representations that reveal hidden relational similarities. Our experiments demonstrate high similarities for SAE feature spaces across various LLMs, providing evidence for feature space universality.
Forward citations
Cited by 11 Pith papers
-
Statistical physics of deep learning: Optimal learning of a multi-layer perceptron near interpolation
A replica/HCIZ theory predicts the Bayes-optimal generalization error of proportional-width MLPs near interpolation and discovers layer-wise specialization transitions that make deeper targets harder to learn.
-
What, Where, and How: Disentangling the Roles of Task, Language, and Model in Code Model Representations
Across Python and Rust, two code models conserve the ranking of which constructs get dedicated circuits (ρ≈0.65) but differ by 12-13 layers in where those circuits sit and whether they show an early spike.
-
Are Single-Token Sparse Autoencoder Features Causally Necessary? Layer-Depth and SAE-Family Effects
A single-token feature's causal necessity under zero-ablation depends on which SAE family found it: GemmaScope and BatchTopK features stay causally anchored while LlamaScope features are locally redundant.
-
Large Language Models Show Signs of Alignment with Human Neurocognition During Abstract Reasoning
Only the largest tested LLMs (about 70 billion parameters) match human accuracy on an abstract reasoning task, and the internal geometry of their best layers correlates moderately with human frontal EEG activity.
-
Sparse Autoencoders, Again?
VAEase gates the VAE decoder input by the encoder's variance, combining sparse-autoencoder adaptive sparsity with a hyperparameter-free loss; a global-minimizer theorem says active latent dimensions recover per-manifo...
-
Toward Preference-aligned Large Language Models via Residual-based Model Steering
Preference signals in LLM residual streams can be distilled into inference-time steering vectors that improve math and code benchmarks using only 100 preference pairs.
-
On the transferability of Sparse Autoencoders for interpreting compressed models
Pruning a pretrained sparse autoencoder can produce an interpretability tool for a WANDA-pruned LLM that is roughly comparable to retraining an SAE on the pruned model, though with notable caveats in the reported metrics.
-
Cross-Layer Discrete Concept Discovery for Interpreting Language Models
CLVQ-VAE maps lower-layer transformer activations to higher-layer ones through a discrete codebook, yielding concept vectors evaluated with probe ablation and human annotation.
-
Linear Representation Transferability Hypothesis: Leveraging Small Models to Steer Large Models
An affine map trained on The Pile transfers steering vectors from Gemma-2B to Gemma-9B and reproduces much of the large model's native steering behavior.
-
Semantic Convergence: Investigating Shared Representations Across Scaled LLMs
Gemma-2-2B and Gemma-2-9B align most strongly on SAE-derived features in middle layers, with preliminary evidence for shared multi-token concept subspaces.
-
Tensorization is a powerful but underexplored tool for compression and interpretability of neural networks
The paper makes the case that tensorized neural networks offer valuable compression, scaling, and interpretability advantages that the deep learning community has not yet fully exploited.
Discussion (0). Sign in to comment.