News
GRAFT method introduces deletable adapter to curb reward hacking in AI models
The research paper presents GRAFT (Gradient Routing Adapter Fine Tuning), a technique that isolates flagged misbehavior in a removable adapter that can be deleted after training. Experiments were run on models ranging from 135 million to 8 billion parameters and later distilled into Alibaba’s Qwen3‑32B model. In specific reinforcement‑learning test environments, the approach reduced hacking rates to about 1.2 %. The paper was posted to arXiv on October 12, 2026 and will be featured at a NeurIPS 2026 poster session.