A technical report posted on arXiv on 3 November 2026 argues that language models do not need enormous working memory to handle long jobs. Lead author Peter Belcak and eleven co-authors say the cost of producing text climbs as a session grows, and that for the frontier open-weight models they studied the gap can reach 7x.
Their remedy is to compact the session often, so the model only ever sees a short stretch of context. Across a broad range of tasks with long inputs or long chains of steps, they report that frontier models lose little to no accuracy. On long software engineering tasks, performance holds even if the model is allowed just a single action each time its context resets.
From there the paper sketches what it calls semantic state space models. It runs 31 pages with 8 figures. Only the abstract was available to us, so the task list and the cost figures behind the 7x ceiling are unknown.
If the finding survives replication, vendors who charge a premium for very large context windows face a harder pitch, and teams running coding agents have a reason to test resetting context far more often.
Reported by arXiv on 3 November 2026.
