# Preprint proposes test for when LLM hidden-state selection beats voting

_Published Wednesday, August 19, 2026 at 2:06 AM EDT · Science, AI · Latest · Tier 2 — Notable_

A new arXiv preprint reports that CASE, a hidden-state selection method for large language model answers, outperformed majority voting by up to 19 points on medium-difficulty questions and 16.8 points on hard questions. The method trains a linear gate on answer-token hidden states and uses a decodability measure to predict when selection should outperform voting. The paper reports a 0.75 correlation between decodability and accuracy gains on held-out data, while warning that conventional probes can leak question identity.

## Sources

- [cs.AI updates on arXiv.org](https://arxiv.org/abs/2608.17124)

---
Canonical: https://techandbusiness.org/newswire/oQ0i-uDug9ZkhvkMlrYo7q
Published: 2026-08-19T06:06:39.867Z
Story chronology: 2026-08-19T04:00:00.000Z
Retrieved: 2026-10-03T12:41:34.581Z
Publisher: Tech & Business (techandbusiness.org)