Residual Paving: Diagnosing the Routing Bottleneck in Selective Refusal Editing

2026-05-22 · Source: cs.LG updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Cybersecurity & Data Privacy · Depth: Expert, extended

Summary

Residual Paving is a novel routed residual editing method designed to diagnose and address the routing bottleneck in selective refusal editing for frozen instruction-tuned transformers. This technique separates route selectivity, determining when to intervene, from residual-edit capacity, defining the edit to apply. An early-layer router predicts a scalar gate and expert mixture, which, when active, enables prompt-conditioned bottleneck residual experts to apply later-layer residual updates without altering the backbone. On the primary Gemma-3-4B-IT held-out split, Residual Paving significantly reduced edit refusal from 88.6% to 4.0%, while maintaining 95.5% benign and 87.3% harmful distribution preservation. However, harmful refusal remained below the frozen-base rate, at 65.3% versus 81.6%. An oracle-routing diagnostic, applied across six backbones, consistently improved the keep-side diagnostic score with a median gain of +12.9 pp, confirming that learned route selectivity is the primary bottleneck. Trajectory diagnostics further suggest the method induces directed movement towards edit-target continuations.

Key takeaway

For AI Security Engineers developing fine-grained safety controls or red-teaming LLMs, you should consider adopting routed residual editing methods like Residual Paving. This approach significantly improves target edit success, reducing refusal rates from 88.6% to 4.0% on Gemma-3-4B-IT, outperforming simple activation steering. However, be aware that your harmful refusal preservation might still fall below base model rates, indicating a need to refine route selectivity further. Your focus should be on improving the router's precision to avoid off-target harmful-keep degradation.

Key insights

Separating routing from residual editing diagnoses and improves selective refusal control in frozen LLMs.

Principles

• Selective refusal requires three-way control: edit, benign-keep, harmful-keep.
• Learned route selectivity is a key bottleneck for selective refusal.
• Residual-space interventions can modify LLM behavior without retraining.

Method

Residual Paving employs an early-layer router and later-layer prompt-conditioned bottleneck residual experts. Training stages include gate pretraining, contrastive warmup, supervised fitting, and gate calibration.

In practice

• Apply Residual Paving for targeted refusal reduction in LLMs.
• Use oracle routing to pinpoint performance limitations (routing vs. editing).
• Implement staged training for router and residual expert optimization.

Topics

• Residual Paving
• Selective Refusal Editing
• LLM Safety
• Activation Steering
• Oracle Routing
• Gemma-3-4B-IT

Best for: Research Scientist, AI Scientist, Machine Learning Engineer, AI Security Engineer

Related on AIssential

• Harnessing Textual Refusal Directions for Multimodal Safety — Textual refusal directions from LLM backbones can generalize to MLLMs, offering a training-free path to multimodal safety.
• Token-Budget-Aware Pool Routing for Cost-Efficient LLM Inference — Optimizing LLM inference by routing requests to specialized short or long context pools significantly reduces GPU costs and improves reliability.
• Curvature-Guided Module Localization for Low-Rank Detoxification of Backdoored Large Language Models — Curvature-guided module localization and low-rank repair effectively detoxify backdoored LLMs post hoc, preserving benign behavior.
• Gemma 4 Fine-Tuning on Single GPU | Training Gemma 4 With Unsloth on Custom Dataset (🔴 Live) — Fine-tuning LLMs may offer better accuracy and lower GPU RAM than prompting alone.
• Looking Is Not Picking: An Attention-Segment Account of Tool-Selection Failures in LLM Agents — LLM tool-selection failures stem from decision readout, not attention, despite models often attending to the correct tool.
• Learning to Route LLMs from Implicit Cost-Performance Preferences via Meta-Learning — Meta-learning enables LLM routers to adapt to diverse user cost-performance preferences with minimal interaction.

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.LG updates on arXiv.org.