Sparse autoencoders decompose Earth embeddings into sparse features
On AlphaEarth, trained SAEs leave eight times less variance unexplained than a frozen random decoder.
Stories per day · 30 days
12 RSS
Relatedsparse-autoencoders8mechanistic-interpretability3gemma2language-models2llm2representation-learning2steering2action-atlas1ai-text-detection1alphaearth1backdoors1bangla1behavioral-modes1brics1chain-of-thought1cheminformatics1circuit-discovery1classifiers1
96 stories, newest first · one card per story
Stories tagged #interpretability
Study tests cross-lingual concept transfer in Bangla idiom subspaces
English class-mean readouts beat zero-shot nulls in 8 of 12 tests, but English-fitted erasure barely changes Bangla accuracy.
ConceptTime uses predictive surprise to interpret time-series classifiers
A frozen forecaster's prediction error both segments series and builds concepts, with no labels, language model, or expert vocabulary.
Generative Latent Priors interpret and steer protein models
Diffusion models on ESM-2 activations serve as assumption-free priors for editing protein representations.
User embeddings are superpositions of interpretable behavioral modes
Sparse autoencoders recover named behavioral modes from embeddings and support zero-shot natural-language queries.
Survey examines mechanistic interpretability for AI safety
The paper surveys methods for mechanistic interpretability and their role in AI safety, including knowledge editing and steering.
Cross-modal steering vectors transfer between text and image generation
In unified multimodal models, vectors found on one modality steer the other zero-shot, including concepts unique to one side.
Study measures how neural representations form during training
Path tortuosity, not endpoint displacement, predicted later drift and sensitivity to parameter perturbations.
TopKLoRA sparse adapters localize and ablate sleeper-agent backdoors
On three small models, ablating a TopKLoRA circuit removes a planted backdoor while instruction-following holds, with much smaller circuits than dense LoRA.
Paper ties singing emotion recognition to interpretable vocal technique cues
SHAP and Grad-CAM link predictions to vibrato and pharyngeal technique, while singing-specific features reduce cross-singer variability.
Fragment-level graphs match atom-level molecular accuracy on smaller inputs
Integrated Gradients recovered known pharmacophores at 1.5 times chance on BACE and 3.2 times on PTR1.
Sparse autoencoders turn neural AI-text detectors into readable classifiers
Classifiers using only interpretable SAE features beat neural and stylometric detectors on out-of-distribution text.
Activation study finds VLA motor programs ignore shifted scenes
Injected task-A activations made pi0.5 replay that trajectory in 99.6% of 1,968 episodes.
Perspective maps a four-stage agenda for JEPA latent-world interpretability
Authors say existing work only covers grounding, while structure, causal checks, and intervention remain largely unexplored.
ConSDL improves sparse dictionary consistency over SAE baselines
On 16 synthetic settings recovery was near perfect; consistency rose 17.6% for dictionaries and 35.9% for coefficients.
Verbalising Model's Computation for Chain-of-Thought Faithfulness Detection
VERITAS verbalizes layer-wise computation to detect unfaithful chain-of-thought, beating LLM judges by 28.4% balanced accuracy.
[AINews] Agents disprove Polysemanticity; greatly improve Interpretability & Alignment
Meta released Muse Spark Open, an 816B MoE that leads US open-weight models on Artificial Analysis.
9
Redistribution score tracks recurring SAE feature shifts
On two frozen safety readouts, median supports of about 20 features reached failure AUROCs of 0.935 and 0.958.
Explaining the Roles of Experts in Mixture-of-Experts LLMs for Personalization
Researchers personalize mixture-of-experts LLMs and localize distinct personalization features in specific experts.
Behavioral Compositionality of Sparse Autoencoder Representations
SAE features are more compositional than random bases, but faithful explanations still need pairwise interactions.
Lexical reference frames read LM hidden states without a lens
LexFrame[1] reaches 0.97 coherence on SmolLM3-3B and OLMo3-7B, versus 0.30 and 0.35 for fixed lenses.
Learning to Read Out: Unembedding Dynamics in Language Model Pretraining
Pythia encodes task information in hidden states before its unembedding readout can express it.
Diminishing Returns of Data in Sparse Autoencoders: Functional Behavior does not Guarantee Feature Reproducibility
SAE reconstruction converges on far less data, but one-to-one feature matches keep rising with more tokens.
SAE-IDF matches symbols across systems using sparse autoencoder codes
On 12 tasks across six symbolic systems, SAE-IDF beats the same LLM's dense baseline by up to 45 points.