The abstract for a new arXiv paper, ReTTT, addresses a common limitation of LLM agents: they gather useful task-specific information through their actions, observations, and feedback during multi-turn execution, but often fail to exploit it. The authors propose a method that turns this accumulated experience into a training signal at test time.
ReTTT stands for Reflection-Driven Test-Time Training via Experience-Based Parametric Memory. The idea is to store what the agent learns from its interactions in a parametric memory, then use reflection to drive further training on the fly. This allows the agent to refine its subsequent behaviour without waiting for a separate fine-tuning phase.
The source is an abstract only, so implementation details and experimental results are not available. Still, the core motivation is clear: making better use of an agent's own history to improve performance during a single task session.