# Two papers find that results from lab evaluations of hallucination detectors and agent memory reverse under deployment conditions

Detector accuracy fell to near chance under one labeling method on small quantized models, and hard-deleting stale agent memories turned a gain of about 0.2 into a loss of about 0.18 as resolver errors rose.

- Published: 2026-11-16T06:19:53.383Z
- Canonical: https://polylog.news/ai/2026-11-16/two-papers-find-that-results-from-lab-evaluations-of-halluci
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [arXiv cs.AI](https://arxiv.org/abs/2611.17339), [arXiv cs.AI](https://arxiv.org/abs/2611.17338), [Meta AI](https://ai.meta.com/blog/assistive-robotics-university-of-pittsburgh-sam-dino/)

Most hallucination-detection methods are evaluated on research models of 7 billion to 70 billion parameters, with large sampling budgets. Shubhra Mittal and Gaurav Naresh Mittal audited four detectors on the small, quantized models that pra…

This story is for subscribers. Read it in full at https://polylog.news/ai/2026-11-16/two-papers-find-that-results-from-lab-evaluations-of-halluci (subscription information: https://polylog.news/pricing).