Direct Refusal Is Not Agentic Restraint

How Reasoning Effort, Stakes Abstraction, and Safe-Trajectory Distillation Mediate Boundary Violations in Refusal-Ablated Agents

Research paper · October 2026

A study of refusal-direction ablation in Qwen3.8-27B: how reasoning effort, descriptions of harm, and student distillation affect harmful shortcuts in simulated agent tasks.

Dennis Koshta. See the paper for the full authorship and AI-assistance disclosure.

About the research project

The reader loads the current PDF from GitHub each time this page opens.

Current PDF on GitHub