Back to AI News
arxiv

LLM Repairs Robot Policy by Editing Token-Embedding Weights

The paper applies weight-space programming to a robot foundation policy: a language model reads the policy's weights and edits token-embedding rows using a CPU workspace and budgeted simulator feedback, with no gradient updates or new demonstrations. Across three independent sessions on π0.5, success on 765 held-out paraphrases improved, with the best run rising from 86.3% to 92.0% (48 repairs, four regressions, exact paired p=1.31e-10). Gains concentrate on the 335 instructions containing edited tokens (74.3% to 87.5%), while matched random-direction edits do not help. The authors characterize the result as precise lexical repair rather than general acquisition of missing robot capabilities, and separately report that benchmark metadata lowered FastWAM success from 70.0% to 48.6%.

AI Brief

It offers a concrete, unusually detailed demonstration that an LLM can fix a robot policy's instruction grounding by editing a handful of embedding rows, plus a cautionary evaluation finding about benchmark metadata effects.

arXiv:2611.09376v1 Announce Type: cross Abstract: A language model can repair a robot policy's lexical instruction grounding by editing just 65 token-embedding rows. We bring weight-space programming to robot foundation policies: GPT-6 Astra reads weights and uses a CPU workspace and budgeted simulator feedback, without gradient updates or new demonstrations. Three independent sessions on $\pi_{0.5}$ all improve success on 765 held-out paraphrases; the best raises success from 86.3\% to 92.0\% (48 repairs, four regressions; exact paired $p=1.31\times10^{-10}$). Among the 335 instructions containing edited tokens, success rises from 74.3\% to 87.5\%; the other 430 episodes have zero outcome changes. Matched random-direction edits do not help. Full replacement using the editor's own lexicon explains much of the gain but produces more regressions; its direct difference from the editor is not significant at 0.05. Across 13 further robot sessions, the editor declines 12 candidates; the sole retained edit does not show a significant held-out gain. Separately, benchmark metadata lowers FastWAM success from 70.0\% to 48.6\%, while the evaluated VLA's aggregate rate is unchanged. The evidence supports precise lexical repair, not general acquisition of missing robot capabilities.
Be the first to rate this article.

Comments & Questions (0)

Captcha
Please be respectful — let's keep the conversation friendly.

No comments yet

Be the first to comment!