Radar ·
Translation symmetry in language may explain geometry in LLM representations
A post describes a UC Berkeley and Google DeepMind article that attributes circular and timeline structures in LLMs to translation symmetry in natural text, suggesting origin in data rather than only architecture.
According to the post, dated November 19, 2026, a UC Berkeley and Google DeepMind article attributes the circular and timeline geometry seen in the internal representations of LLMs to translation symmetry in natural language. According to the same account, word2vec would show the same structures. The thesis suggests that part of the learned structure may come from the statistics of the data, not only from the model architecture.
The relevance lies in separating what is caused by data from what is caused by model design, a topic of interest to researchers and teams that evaluate models. The post gives no figures, detailed methodology or study limits, and this summary does not add them. To verify the claims, consult the original article from the source linked in your Radar entry and check authorship, date and results directly there.
Original source: Simetria de translação na linguagem pode explicar geometria em LLMs