# AutoAssumption searches algorithm assumptions to improve pruning and cache eviction

_Published Saturday, November 14, 2026 at 1:06 AM EST · AI, Science · Latest · Tier 2 — Notable_

Researchers report in a preprint that AutoAssumption generated algorithms that improved on every published method they tested for KV-cache eviction and weight pruning. The system represents algorithms as proofs, relaxes an assumption, then derives a revised algorithm, checks its proof in Lean and generates code.

The researchers formalized 89 algorithms. Their pruning algorithm achieved the lowest perplexity among thirteen methods on four LLaMA models, while the eviction algorithm led the tested published methods on RULER-4K. At equal budget, the system scored within a point of OpenEvolve, with lower variation across repeated runs.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2611.14844)

---
Canonical: https://techandbusiness.org/newswire/ukJzVhNvKPcEkoU5ofzEqf
Published: 2026-11-14T06:06:56.701Z
Story chronology: 2026-11-14T05:00:00.000Z
Retrieved: 2026-11-15T00:01:49.644Z
Publisher: Tech & Business (techandbusiness.org)