Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing
The paper titled 'Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing' presents a new method for improving selective refusal editing in machine learning models. The authors introduce Residual Paving, which enhances edit success rates while maintaining benign and harmful behavior preservation. Their findings indicate significant reductions in edit refusal rates and improvements in diagnostic scores across various model backbones.
- ▪Residual Paving reduces edit refusal from 88.6% to 4.0%.
- ▪The method preserves 95.5% of benign distributions and 87.3% of harmful distributions.
- ▪Oracle routing improves the keep-side diagnostic score with a median gain of +12.9 percentage points.
Opening excerpt (first ~120 words) tap to expand
Computer Science > Machine Learning arXiv:2605.20262 (cs) [Submitted on 18 May 2026] Title:Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing Authors:Bryce Hinkley, Peyman Najafirad View a PDF of the paper titled Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing, by Bryce Hinkley and 1 other authors View PDF HTML (experimental) Abstract:We study selective refusal editing as a three-way control problem: induce non-refusal on designated edit prompts while preserving benign behavior and harmful refusals outside the edit set. We introduce Residual Paving, a routed residual editing method for frozen instruction-tuned transformers that separates route selectivity, whether to intervene, from residual-edit capacity, what edit to apply.
…
Excerpt limited to ~120 words for fair-use compliance. The full article is at arXiv cs.AI.
Discussion
0 commentsMore from arXiv cs.AI
