4arXiv cs.CL (Computation and Language)·
Procrustes-conditioned Joint SAE extracts cross-seed universal features from BERT models
Researchers introduce a Procrustes-conditioned Joint End-to-end Top-K Sparse Autoencoder (SAE) to address cross-seed feature universality in mechanistic interpretability of BERT models. By applying an orthogonal Procrustes rotation between independently trained models' activation spaces before joint SAE training, the method produces more consistent features (Pearson r ≥ 0.70) than post-hoc alignment baselines across three NLP benchmarks. The work targets a fundamental challenge in dictionary learning: non-convex optimization causes independently trained networks to learn misaligned feature spaces, making it difficult to identify truly universal features. High-universality features are shown to encode interpretable sociolinguistic patterns.
Related guides (2)
Related events (8)
VASAE: Vocabulary-Aligned Sparse Autoencoder assigns intrinsic token names to SAE features during training
Researchers introduce VASAE (Vocabulary-Aligned Sparse Autoencoder), a method that trains SAE features with vocabulary-aligned anchoring so each feature is intrinsically named by the nearest token in the model's embedding space. Applied to GPT-2-small and Llama-3.1-8B, VASAE achieves ~90% feature alignment in shallow-to-middle layers without degrading reconstruction quality, though final-layer alignment is limited. The work addresses a longstanding interpretability bottleneck where SAE dictionary features require expensive post-hoc labeling, potentially enabling more scalable mechanistic analysis.
Study finds SAE unstable features reflect reproducible subspaces, not pure noise
A new arXiv paper investigates feature stability in sparse autoencoders (SAEs), measuring the probability that individual learned features reappear across independent training runs. The authors find a functional asymmetry: stable features carry most reconstruction-relevant signal, while unstable features are individually non-reproducible but concentrate in reproducible lower-rank subspaces, suggesting seed dependence reflects basis ambiguity rather than noise. A synthetic model confirms that low-rank ground-truth features can be recovered at the subspace level even when individual SAE latents are non-identifiable across seeds. The work has direct implications for interpretability research that relies on SAE features as meaningful, stable units of analysis.
SAERL: Using Sparse Autoencoders to Guide LLM Reinforcement Learning Data Engineering
SAERL is a post-training data engineering framework that uses Sparse Autoencoders (SAEs) — a mechanistic interpretability tool — to extract intrinsic model signals for controlling data diversity, difficulty, and quality during RL fine-tuning. The framework applies SAE-space clustering for batch diversity, a difficulty proxy for curriculum ordering, and a quality probe for data filtering. On Qwen2.5-Math-1.5B with GRPO, SAERL achieves 3% average accuracy improvement and reaches target accuracy with 20% fewer training steps. SAE representations transfer across model families and scales, suggesting broad applicability as a lightweight data engineering tool.
SAEs as Stethoscopes: Interpretability-Guided Layer Selection for Task Vector Model Editing
This paper evaluates a Sparse Autoencoder (SAE)-guided model editing pipeline for mathematical reasoning on Gemma-3-4B-IT, finding that projecting task vectors onto SAE feature subspaces discards ~97% of modification energy due to geometric misalignment between activation-space SAE directions and weight-space task vectors. The authors reframe SAEs as diagnostic tools ('stethoscopes') rather than intervention filters ('scalpels'), using SAE-derived specificity scores to identify which layers to inject unfiltered task vectors into. This approach improves Number Theory accuracy from 29.6% to 39.4% on Minerva Math (p=0.0007), with 5 of 7 math subjects significantly improved and none degraded. The method is fully deterministic and adds no inference cost.
Sparsity regularizers improve interpretability of Top-k sparse autoencoders for vision models
A new arXiv preprint proposes two sparsity regularizers compatible with Top-k sparse autoencoders (SAEs), a standard tool for mechanistic interpretability of vision foundation models. The regularizers — an ℓ1 penalty on off-support units and a scale-invariant ℓ1/ℓ2-ratio penalty — are applied before Top-k selection and consistently improve monosemanticity without degrading reconstruction quality across two datasets and three vision models. The central finding is that hard architectural sparsity and soft regularization are complementary, addressing known limitations of fixed-budget Top-k SAEs such as overfitting to training k values.
C²R regularization method addresses feature splitting and absorption in Sparse Autoencoders
A new arXiv preprint introduces C²R (Cross-sample Consistency Regularization), a training technique for Sparse Autoencoders (SAEs) that mitigates two known failure modes: feature splitting, where coherent concepts fragment across multiple latents, and feature absorption, where general features develop arbitrary exceptions. C²R penalizes co-activation of directionally similar latents across a batch, encouraging each semantic concept to map consistently to a single latent. The authors report that C²R reduces both pathologies while preserving reconstruction fidelity, with source code released publicly.
Extracting Concepts from GPT-4: 16 Million Patterns via Sparse Autoencoders
OpenAI applied scaled sparse autoencoders (SAEs) to GPT-4 to automatically identify approximately 16 million interpretable features or patterns in the model's internal computations. This represents a significant scaling of mechanistic interpretability techniques previously demonstrated on smaller models. The work advances the ability to understand what concepts and representations large frontier models encode internally.
BODHI: Contrastive embedding training for causal discovery in Large Behavioural Models
Researchers identify a critical failure mode in biomedical language model embeddings: off-the-shelf encoders (BioBERT, PubMedBERT, BioM-ELECTRA) assign high cosine similarity (0.76–0.92) to causally unrelated cross-domain pairs, achieving 0% accuracy on cross-domain discrimination. The paper introduces BODHI, a contrastive training approach using hard negatives mined from a biomedical knowledge graph, which improves within-vs-across-domain separation from 1.05x to 2.30x and raises discrimination gap by +0.392. The work targets Large Behavioural Models (LBMs)—foundation models that reason over personal life graphs—where false embedding proximity directly produces false causal edges. Additional contributions include an OpenVINO inference optimization achieving 133x latency reduction (1367ms to 10ms) on Intel AMX hardware, plus a counterintuitive finding that FP16 outperforms INT8 on this silicon.