Significant focus has been placed on integrating large language models (LLMs) with various tools in developing general-purpose agents.

The prompt-based paradigm has fostered two recent human-AI interaction trends.

Panacea trains a single model capable of adapting online and Pareto-optimally to diverse sets of preferences without the need for further tuning.

A general idea to overcome the above problem is to perturb the system dynamics.

However, constrained by the inherent geometric constraints of sequences, it still falls short in modeling long-range dependencies.