Language model fixes a robot policy's word understanding by editing 65 embedding rows

Jusheng Zhang, Keze Wang and Liang Lin have shown in an arXiv paper that a language model can repair how a robot policy understands words in its instructions by editing its weights directly. GPT-6 Astra read the weights of a pi-0.5 policy and, with a CPU workspace and limited simulator feedback but no gradient updates or new demonstrations, changed just 65 token-embedding rows.

In the best of three sessions, success on 765 held-out reworded instructions rose from 86.3% to 92.0%, with 48 repairs and four regressions. Random edits did not help, and in 13 further robot sessions the editor declined 12 candidate edits. The authors say the evidence supports precise repairs to word understanding, not the general learning of missing robot skills.