Jie Li

CustomShift achieves unprecedented balance between semantic fidelity and subject consistency in image generation, outperforming existing methods by leveraging a novel attention distribution shift.

RiskFlow achieves a remarkable balance between realism and safety in traffic scenario generation, cutting inference time while enhancing the fidelity of multi-agent interactions.

Achieving a 7.5-fold reduction in computational complexity, the DBHN Network outperforms traditional methods in monaural speech enhancement while maintaining high performance.

SentGuard detects 90.5% of unsafe content within two sentences, revolutionizing real-time moderation for large language models.

Forget dataset-specific selection: a single, lightweight selector trained on one dataset can boost VLM instruction tuning performance across diverse datasets and model scales, even outperforming full data training.

Transfer learning from a large, pre-trained speech synthesis model unlocks high-quality Tibetan TTS, even with limited Tibetan-specific data.

Force-feeding physics to LSTMs slashes battery thermal runaway prediction errors by over 80%, making your next e-bike less likely to explode.

Individual prosumers can now effectively coordinate in electricity markets, boosting overall market performance through a novel hierarchical MARL framework.

Open-source audio reasoning just leveled up: Audio-Cogito rivals closed-source models on complex audio tasks.

Coordinating large robot fleets just got easier: HECTOR's hierarchical approach lets humans supervise at different levels, slashing both computational cost and operator effort.