Managing future user instructions under Prefix Sliding

Determine how future user instructions should be managed in multi-turn interactions using Prefix Sliding, specifically whether such instructions should be appended to the preserved prefix or allowed to be removed by the sliding window.

Background

Prefix Sliding preserves the initial prefix while evicting older intermediate tokens as the reasoning window advances. In multi-turn agentic interactions, future user instructions may arrive after the prefix has already been established, creating an unresolved design choice about whether they should receive the same permanent preservation as the original task instructions or remain subject to sliding-window eviction.

The paper identifies this issue alongside the related problem of handling large system outputs, such as website contents or files, that can overwhelm the sliding window and cause important information to be lost. Resolving the placement and retention policy for later user instructions is necessary for applying Prefix Sliding reliably to multi-turn and agentic settings.,

References

A second related issue is what to do with future user instructions in a multi-turn setup. Append them to the prefix? Let the sliding window remove them eventually?

Prefix Sliding for efficient test-time scaling  (2608.26070 - Muennighoff et al., 26 Aug 2026) in Section ‘Limitations’, paragraph ‘System outputs and multi-turn’