发现信息
- 作者:Musa Cim, Burak Topçu, Mahmut Taylan Kandemir
- 年份:2026
- 发表源:Research Square
- OpenAlex ID:https://openalex.org/W7223655346
- DOI:https://doi.org/10.21203/rs.3.rs-11639533/v1
- arXiv:待核验
- 检索主题:memory-evaluation
- OpenAlex 引用数快照:0(2026-11-12)
摘要线索
Abstract Long-horizon agents routinely exceed their context windows. A common form of context compaction that reduces a long interaction history to a smaller representation is to summarize the history and continue from the summary. This summarization loses information, and this loss is a common source of agent failures: forgotten details are rediscovered at extra cost or silently dropped. The natural repair is to ask for a longer summary, so we investigate whether that option actually works. Does summary length grow with input length? Across eleven models spanning coder, general-purpose, and reasoning families, summary length grows far more slowly than the input, so at 100k tokens, less than one percent of the input survives on both long documents and real agent trajectories. Furthermore, prompt-level directives do not close this gap: explicit length directives raise output only modestly, and the summary length plateaus far below the requested amount, however the request is phrased. To test whether training data can move this ceiling, we lightly fine-tune two models on summaries that differ only in length. The resulting output length shifts dramatically in either direction, tracking the finetuning summary length, even under a prompt that never mentions length. Consequently, this paper shows that summary length grows strongly sublinearly with input size, is only weakly responsive to length directives in the prompt, and is strongly responsive to fine-tuning, suggesting that training data, rather than the prompt, is the effective lever.
核验与筛选
- 当前状态:待筛选、去重和一手来源核验。
- OpenAlex 元数据和引用数只用于发现,不作为正式引文或质量判断。
- 记录正式版本、venue、方法细节、实验协议和局限后,才迁入稳定论文卡片。