Large language model (LLM) agents typically append raw tool outputs to their context. Over multi-step interactions, long and noisy observations incur substan…
机构:清华
来源:arXiv 2611.06500 | AI4Papers 论文推荐平台