# DP-MIND preprint extends private AI inference to million-token contexts

_Published Monday, October 12, 2026 at 4:06 AM EDT · AI, Science · Latest · Tier 2 — Notable_

Researchers report a framework that processes million-token contexts while protecting sensitive inputs sent to an external language model. Their DP-MIND preprint says the method outperformed existing privacy and long-context inference baselines across multiple benchmarks.

The framework adds privacy-preserving noise locally, processes the input in chunks and carries task-relevant information forward in a compact memory. Reinforcement learning trains that memory to withstand the noise, while a local module repairs corrupted answer spans without exposing the original context to the external model. The findings concern benchmark performance, rather than a deployed service.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2610.13132)

---
Canonical: https://techandbusiness.org/newswire/EiSY7p-_u9mTcR9acHR8ur
Published: 2026-10-12T08:06:26.497Z
Story chronology: 2026-10-12T04:00:00.000Z
Retrieved: 2026-10-12T10:36:40.799Z
Publisher: Tech & Business (techandbusiness.org)