Skip to main content
Search

# Sparse Autoencoder

19 related articles

Subscribe to updates
  1. arXiv

    Medical retrieval's dense vectors split into readable features: sparse codes beat original embeddings on CURE, and explanations still predict activations across corpora

    Medical retrieval systems usually compress queries and documents into dense vectors whose matches are hard to explain. This work trains sparse autoencoders on those dense embeddings with an InfoNCE contrastive loss that preserves query-document relevance, raising NDCG@10 above the original embeddings for all three embedders on CURE. Neuron interpretations built only from CURE texts still positively predict whether a neuron fires on five out-of-domain corpora, and concept-level interpretations outperform raw example snippets.

  2. arXiv

    Near-OOD detection gets a complement, not a replacement: PLUS lifts all 126 detector-backbone pairs across 18 detectors

    Near-out-of-distribution (near-OOD) inputs resemble the training classes closely, and existing post-hoc detectors that excel on far-OOD inputs lose that edge here. PLUS does not replace a detector: it uses a sparse autoencoder to pick each class's most active neurons, spans a low-dimensional subspace, and scores an input by that class's whitened Mahalanobis distance, adding the signal to any calibrated base score at one fixed weight. On the ImageNet-1K OpenOOD v1.5 benchmark, near-OOD AUROC improves in all 126 detector-backbone combinations across 18 detectors and seven backbones, while far-OOD performance is broadly maintained, with declines of at most one point where the base score is already near-saturated.

  3. bioRxiv

    ProteinMPNN's latent space splits into readable axes for secondary structure, residue volume and heme coordination, and turning a branched-chain axis reshapes an antibody fragment's sequence

    Sparse autoencoders expand the dense 128-dimensional latent space at each ProteinMPNN encoder layer into an overcomplete sparse space, where individual dimensions correlate with protein features such as secondary structure, residue volume, residue shape, disulfide bridges and ligand coordination. On the Brolucizumab short-chain variable fragment, the mod=0, 0.5 and 1 candidates with branched-chain-amino-acid dimensions raised held lower RMSD at 433 K than the native sequence, the ProteinMPNN sequence and the GeoEvoBuilder sequence, while mod=2, the largest raise, unfolded substantially at that temperature; the unmodified mod=0 candidate improved too, showing the autoencoder itself shifts predictions.

  4. arXiv

    When can an interpretability claim be trusted: feature dimensionality jumps, while its ridge-regularized version holds one verdict across 12 toy models

    Mechanistic interpretability often rests safety claims on thresholded numbers that can flip under small changes to weights, training seed or threshold. Working on toy models and synthetic dictionaries with known ground-truth features, this work characterizes exactly where feature dimensionality is discontinuous and replaces it with a ridge-regularized version whose stability constant grows with feature scale and whose coefficient is optimal. On 12 trained toy models, the regularized sum rule gave the same verdict at all 21 pre-set values of its smoothing parameter, while the version that ignores short feature vectors gave the opposite verdict for two of the models.

  5. arXiv

    Glyph splits clinical foundation-model embeddings into readable concepts, surfacing 397 reproducible prognostic markers in abdominal CT cohorts

    Glyph decomposes clinical foundation-model embeddings into sparse, nameable concepts so that features previously entangled in high-dimensional vectors can be tested like clinical variables. Across abdominal CT, radiology reports and laboratory data from 41,628 patients, the study tested 604,185 concept–outcome pairs and retained 397 high-risk concepts with 4,463 Bonferroni-significant associations. Most associations reproduced in three held-out cohorts without recalibration, and the top ten concepts reproduced more often than markers recovered from the literature.

  6. arXiv

    Earth embeddings stop being a tangle of mixed dimensions: sparse autoencoders pull out nameable geographic features, though a random dictionary already earns part of the score

    Earth-observation foundation models such as AlphaEarth and Tessera compress each 10-metre pixel into a dense vector whose individual dimensions often mix several land-surface concepts. Sparse autoencoders expand that vector over an overcomplete dictionary and keep only a few atoms active, yielding sparse, approximately monosemantic features. At matched sparsity, three sparse-autoencoder variants reconstruct better than PCA, ICA, NMF and raw dimensions; yet a control whose decoder is a frozen random dictionary also shows more specific features and finer spatial scales, so those two properties follow from sparsity itself, and only a learned dictionary cuts AlphaEarth's unexplained variance eightfold.

  7. Brain voxels get more than one semantic unit: SAIL merges cross-layer sparse features for lower redundancy and higher alignment

    Similarity between multimodal large language model representations and human visual cortex shows that the two are alike but not which semantics they share. SAIL recasts voxel interpretation as a correspondence between each voxel and multiple sparse semantic units: it trains a unified cross-layer sparse autoencoder on several model layers, matches each voxel to several units, and compares the semantic similarity of their preferred image sets with optimal transport. Across eight subjects in the Natural Scenes Dataset, joint cross-layer training reduced unit response redundancy and raised semantic similarity over layer-wise training, in both whole cortex and selective regions of interest.

  8. arXiv

    Two SAE latents, one feature? A single energy-weighted transport plan now prices whole-dictionary disagreement and locates it latent by latent

    Comparing two sparse autoencoders (SAEs, dictionaries that decompose model activations into sparse learned components) requires answering three things at once: how much two whole dictionaries differ, which latents account for that difference, and how much is sampling noise from finite documents. One energy-weighted optimal-transport plan connects them: its cost is the dictionary-level discrepancy, the cost decomposes exactly into latent contributions, and transported fractions describe how each latent's activation energy is allocated across counterparts, with joint asymptotic confidence intervals. Across four language models it locates differences between SAEs trained on identical data with different seeds and separates topic-dependent changes into a weight term and a cost-matrix term.

  9. arXiv

    MoE experts don't sort by domain label: an SAE splits each feature's activation by expert, separating Java from front-end code

    Researchers at ByteDance Seed and Shanghai University of Finance and Economics trained a sparse autoencoder (SAE, a dictionary that reconstructs a layer's output as sparse interpretable features) on the MoE output of layers 4, 30 and 46 of Qwen3-30B-A3B, and used the encoder's affinity in the expert outputs to decompose each feature's activation exactly over the experts. Averaged over a corpus, each expert gets the set of features it principally drives; most of those features have only one or two dominators, and the division of labor runs below coarse domains, down to Java versus front-end web code and geometry versus arithmetic and polynomials.

  10. arXiv

    Sparse-autoencoder steering gets an information-theoretic design: DESSIRT beats a contrastive baseline on most question modes in Llama and Gemma

    Sparse-autoencoder (SAE) steering has relied on contrastive activation statistics to pick features, although feature interpretability and steering utility are not necessarily aligned. DESSIRT represents each dataset as a vector of SAE feature firing probabilities, measures deviation from a target distribution with Jensen–Shannon divergence, and constructs steering vectors analytically from that divergence. In a seven-mode question-generation task, DESSIRT-2 beat the SDCV baseline on all but one mode in Llama 3.1-8B-Instruct and on most modes in Gemma-2-9B-IT.