研究称无需长上下文：频繁压缩几乎不损失任务精度
You don't need long context
这篇论文说生成靠后的 token 贵最多 7 倍，靠频繁压缩短上下文几乎不掉精度，做 Agent 的值得看看省成本。
arXiv 论文（编号 2611.05306）指出，前沿开源权重模型生成靠后 token 的成本最高可达早期 token 的 7 倍。作者在多类长时程与长输入任务上验证，通过频繁压缩保持短上下文时，前沿语言模型几乎不牺牲任务精度。实验还显示，在长时程软件工程任务中，即使每个上下文最多执行一个动作，模型表现也不会崩溃。基于此，论文提出语义状态空间模型（SSSMs）概念，质疑长上下文模型的必要性。
You don't need long context
Later tokens are costlier to generate than earlier tokens. On frontier open weight models, the disparity in cost may rise to as much as 7x. Through frequent compaction, we show across a broad range of long-horizon and long-input tasks that frontier language models sacrifice little to no task accuracy when maintaining short contexts. We further demonstrate that LLM performance does not collapse on long-horizon software engineering tasks even when taking at most one action per context, and based on that we outline the concept of semantic state space models (SSSMs). Our findings question the need for long-context language models.
