---
format: "aidr-story-markdown/v1"
id: "590db0199eb15999aea0aaeb816b9f3835d12041732e0d31a05f80303eef3470"
canonical_url: "https://aidr.today/590db019?lang=en"
title: "Functional-Group Knowledge Is Structured in Language Models Before It Is Spoken"
lang: "en"
requested_lang: "en"
available_langs: ["en","vi"]
translation_fallback: null
fallback_fields: []
published_at: "2026-11-04T05:00:00.000Z"
category: "Research"
topics: ["llm"]
source_urls: ["https://arxiv.org/abs/2611.03684"]
summary: "A study of five open-source language models found that, even without chemistry training, the models internally organize functional groups by chemical family and encode chemical differences as consistent directions in their representations. The analysis used prompts covering 92 molecules and 20 functional groups, reading the models' internal states rather than relying solely on their outputs."
---

# Functional\-Group Knowledge Is Structured in Language Models Before It Is Spoken

> [Open the canonical story](<https://aidr.today/590db019?lang=en>)

**Published:** 2026-11-04T05:00:00.000Z
**Category:** Research
**Topics:** llm

## Summary

A study of five open\-source language models found that, even without chemistry training, the models internally organize functional groups by chemical family and encode chemical differences as consistent directions in their representations\. The analysis used prompts covering 92 molecules and 20 functional groups, reading the models' internal states rather than relying solely on their outputs\.

## Sources

- [Story source](<https://arxiv.org/abs/2611.03684>)