ABLITERAX: an open framework for auditing and ablating LLM refusal representations

ABLITERAX is an open-source research system that locates refusal representations inside large language models and measures their linearity, multi-directionality, and distribution across layers. It performs targeted ablation under capability constraints and uses a held-out evaluation set to check that a drop in refusals does not come at the cost of general reasoning. The project's design stance stresses explicit limits, reporting of failures, and reversible, auditable changes to original checkpoints. Its authors argue that safety in frontier-scale and mixture-of-experts models is distributed across several pathways, so single-direction approaches tend to fail there. The document also specifies an architecture registry that requires verified adapters rather than name-pattern guesses for tensor mapping.