Lgnd.ai Devlogs

Research notes and devlogs from LGND AI and Clay

Research notes and devlogs from LGND AI and Clay.

ELLE: Embeddings Linearly contain their Loss Estimate

embeddings

uncertainty

geospatial

foundation-models

self-supervised

Foundation model CLS embeddings linearly encode their own per-sample pretraining loss – readable by a Ridge probe in 1 us, with no labels and no extra forward passes. Validated across 20 models and 6 modalities.

2026-03-01

---
title: "Lgnd.ai Devlogs"
subtitle: "Research notes and devlogs from LGND AI and Clay"
listing:
  contents: posts
  sort: "date desc"
  type: default
  categories: true
  sort-ui: false
  filter-ui: false
  fields: [date, title, description, categories, image]
  date-format: "YYYY-MM-DD"
page-layout: full
title-block-banner: false
---

Research notes and devlogs from LGND AI and Clay.

---

Categories

ELLE: Embeddings Linearly contain their Loss Estimate