Ask or search anything...

Researchers from CUHK-Shenzhen and Shanghai Jiao Tong University introduce CLEA, a groundbreaking closed-loop framework that enables robust LLM-based robotic task planning through four specialized language models working in concert, achieving a 67.3% improvement in success rate over baseline systems while demonstrating practical deployment in dynamic real-world environments.

The STMA framework enhances LLM-based agents for long-horizon embodied task planning by integrating a dynamic spatio-temporal memory with a self-correcting planner-critic mechanism. This approach achieves a 31.25% improvement in task success rate and a 24.7% increase in average score over baselines in TextWorld, particularly when leveraging open-source LLMs.

SAAP, a structured pruning framework for large language models, employs adaptive importance assessment and group-wise fine-tuning to significantly reduce computational and memory costs. It achieves up to 65% faster token generation and substantially lowers memory footprint on various LLMs like LLaMA and Vicuna, while maintaining or improving performance across diverse benchmarks.