AI

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

Read full article at source ↗

Based on a feed excerpt; full article content has not been acquired. Source evidence retrieved 04/11/2026, 05:18:24 UTC.

Briefings

Research on five open-source language models reveals that functional-group knowledge is encoded in internal representations before any output is generated. Analysis of 92 molecules and 20 functional groups showed that models organize groups by chemical family and encode meaningful patterns, even without chemistry training. The base model correctly named the functional group in only 58.5% of prompts, indicating that errors often stem from failures to answer rather than incorrect answers. This structured internal organization suggests that evaluating models solely by their spoken outputs can underestimate their actual chemical information encoding capabilities.

Key points

• Five open-source language models were analyzed using internal representations on prompts covering 92 molecules and 20 functional groups.
• The base model correctly named the functional group in only 58.5% of prompts, with errors mostly failing to answer rather than providing wrong answers.

Context

Language models serve as reasoning cores for chemistry agents, yet their outputs do not reflect their internal knowledge. This study addresses the gap between model output and actual chemical understanding by analyzing internal representations.

Limitations

Results are based on five open-source models; no additional background or external validation is supplied. The study uses preprint data with uncertainty, and findings may not generalize to all chemical language models.

Generated by Bittrees-hosted qwen35-2b sha256:aaf42c8b7c3c on 04/11/2026, 06:05:23 UTC. Based on publisher-provided feed evidence; may not cover the full original article. Check the source for details.

View archived briefing on IPFS
bafybeic5u4iso3dgpqyhq52xxsryy57fgya4wltynem25gfrlp6ep6cpsi