Multi-Domain Token Filtering: A Unified Approach to Selective Capability Shaping and Alignment

TL;DR: What if we filter out different kinds of "bad" words from a language model all at once? Test if token-level filtering can simultaneously reduce multiple undesired capabilities (e.g., medical, bias, jailbreaks) without hurting the model’s “good” skills.

Research Question: Can token-level filtering be extended to target multiple, heterogeneous undesired capabilities in a single pretraining run, and how does this multi-domain filtering impact the model’s general capabilities and alignment potential?

Hypothesis: Simultaneous token-level filtering across several domains will synergistically suppress undesired capabilities while maintaining or even enhancing benign/general abilities, especially in larger models.

Experiment Plan: - Construct token-level filters for several domains (e.g., medical, social bias, jailbreak instructions, misinformation) using sparse autoencoders and specialized classifiers.

• Pretrain models at several scales with all filters applied concurrently.
• Measure retention/removal of each domain-specific capability, and monitor impact on downstream tasks and model alignment using frameworks like Persona-judge.
• Compare against single-domain filtering and post hoc debiasing approaches such as BiasFilter.

References:

• Rathi, N., & Radford, A. (2026). Shaping capabilities with token-level data filtering.
• Cheng, X., Chen, R., Zan, H., Jia, Y., & Peng, M. (2025). BiasFilter: An Inference-Time Debiasing Framework for Large Language Models. Conference on Empirical Methods in Natural Language Processing.
• Zhang, X., Chen, R., & Feng, Y. (2025). Persona-judge: Personalized Alignment of Large Language Models via Token-level Self-judgment. Annual Meeting of the Association for Computational Linguistics.