The Second LHAASO Collaboration Conference in 2026 & The Academic Symposium for the 10th Anniversary of the LHAASO collaboration

Good angular resolution is essential for resolving the fine structures of gamma-ray sources and improving detection sensitivity, and is therefore crucial for gamma-ray source observations. We present a Transformer-based deep learning model for gamma-ray direction reconstruction. Using six months of Crab Nebula observations, this method improves the angular resolution compared with traditional methods. To further improve the angular resolution, we introduce shower-core position reconstruction and systematically compare the angular resolution achieved with single-task learning (direction reconstruction only) and multi-task learning (joint direction and shower-core position reconstruction).