Static word embeddings reproduce the 'LLM pain axis' result, undercutting claims of felt experience
wschroll · x · 2026-09-24
Responding to the widely shared 'pain axis' paper, Elan Barenholtz reran the authors' own sentences and analysis but swapped LLM activations for static word embeddings (GloVe, Word2Vec, fastText)—one fixed vector per word, no context, no model.
Static embeddings still separated pain from controls at held-out AUC 0.84–0.87 (vs 0.91–1.00 for LLMs) and passed the same specificity tests. His companion paper shows co-occurrence statistics alone recover city coordinates (R² 0.71–0.87) and birth years (R² 0.48–0.52) from static embeddings.
Implication: linear decodability doesn't mean the model has that state—LLMs are sophisticated models of text statistics, and the 'pain axis' likely reflects word associations in training text, not genuine experience.
More from Research
- ICLR 2027 submission spike looks inflated: only ~22K truly new papers, argues researcher — rao2z · 2026-09-24
- ICLR 2027 submission spike mostly double-counting: net new papers only ~22K after NeurIPS overlap — rao2z · 2026-09-24
- Researchers Debate Using LLMs to Crack the Genomic Motif Annotation Bottleneck — jatin_n0 · 2026-09-24
- Harvard longevity summit: researcher outlines AI-driven roadmap to defeat aging in 10-15 years — DeryaTR_ · 2026-09-24
- Penn's open-source HARMONY rebuilds 3D scenes from one image, matching GPT-6 Astra — thoma_gu · 2026-09-24
- Anshul Kundaje: trained LLMs could break the genomic motif annotation bottleneck — anshulkundaje · 2026-09-24