Interpretable Filtering: Using Concept-Aware Sparse Autoencoders for Transparent Data Shaping

TL;DR: What if we made the filtering process not just automatic, but also explainable—so we’d know exactly what concepts we’re removing from the model’s brain? Test concept-level token labeling using interpretable sparse autoencoders (CoCoMix) to filter not just on tokens, but on learned, human-understandable concepts.

Research Question: Can sparse autoencoder-based concept extraction enable interpretable token-level filtering that targets semantic capabilities (e.g., “diagnosis,” “treatment”), and does this lead to more robust and transparent capability shaping?

Hypothesis: Filtering tokens associated with specific latent concepts, rather than surface words, will result in more targeted capability removal and greater robustness to paraphrasing or adversarial prompts.

Experiment Plan: - Use sparse autoencoders or CoCoMix to discover interpretable latent concepts in the pretraining data.

• Label and filter tokens linked to undesired high-level concepts (e.g., “medical advice,” “violence”) rather than just keyword tokens.
• Pretrain models and compare capability suppression and generalization to standard token-level filtering.
• Assess interpretability by evaluating which concepts are filtered and how this aligns with human expectations.

References:

• Rathi, N., & Radford, A. (2026). Shaping capabilities with token-level data filtering.
• Tack, J., Lanchantin, J., Yu, J., Cohen, A., Kulikov, I., Lan, J., Hao, S., Tian, Y., Weston, J., & Li, X. (2025). LLM Pretraining with Continuous Concepts. arXiv.org.
• Kantamneni, S., Engels, J., Rajamanoharan, S., Tegmark, M., & Nanda, N. (2025). Are Sparse Autoencoders Useful? A Case Study in Sparse Probing. International Conference on Machine Learning.