# SODA preprint reports 4x improvement in reconstructing language model inputs

_Published Monday, October 19, 2026 at 8:06 AM EDT · Science, AI · Latest · Tier 2 — Notable_

Researchers report a 4x improvement over the strongest prior method in exactly reconstructing language model inputs, using a search algorithm called Sparse One-hot Discrete Adam, or SODA. The preprint says SODA recovered natural sequences of up to 17 tokens with zero false positives when given output logits, the numerical scores a model assigns to possible next tokens.

The method requires access to the model's internal workings and was tested on models with 33M to 3B parameters. In proof-of-concept security tests, it recovered 6 backdoor triggers in under 1000 iterations and 139 tokens of personally identifiable information from synthetic user-chatbot interactions.

## Sources

- [cs.LG updates on arXiv.org](https://arxiv.org/abs/2507.01693)

---
Canonical: https://techandbusiness.org/newswire/An9i9N5kU9a0sGcc1_-83_
Published: 2026-10-19T12:06:12.306Z
Story chronology: 2026-10-19T04:00:00.000Z
Retrieved: 2026-10-19T15:29:25.719Z
Publisher: Tech & Business (techandbusiness.org)