New preprint: grammaticality emerges as an organizing principle in LM representations
TuhinChakr · x · 2026-09-18
A new preprint by Najoung Kim and colleagues asks whether grammaticality is a major organizing principle of neural LM representations. They find many NLMs exhibit abstract representational separation for grammaticality.
- Key contribution: deconfounding grammaticality from factors affecting general probability, addressing debates about confounds in measuring models' grammatical knowledge.
- Authors argue this offers a more nuanced answer to the prerequisite question of whether LMs even show grammatical sensitivity.
More from Research
- Liquid AI's tiny longevity models top GPT-5 and Gemini in Cell-published biology benchmark — JosephJacks_ · 2026-09-18
- No, OpenAI did not solve the Navier-Stokes equations, experts clarify — gerardsans · 2026-09-18
- SPLADE's original author releases new sparse embedding model: 150M params beat models up to 1B — IgorCarron · 2026-09-18
- Gumroad founder openly shares full osteosarcoma genomics, imaging and pathology data — iskander · 2026-09-18
- PointZero: a 3D world model pre-trained without robot data hits SOTA on 3D dynamics — YunzhuLiYZ · 2026-09-18
- Chinese Researchers Open-Source Cache-to-Cache: LLMs Talk Without Tokens, +14.2% Accuracy — Scobleizer · 2026-09-18