LLM edits 65 embedding rows to fix robot failures, lifting held-out success 74.3% to 87.5%
机器之心 · wechat · 2026-11-13
A new paper, 'Can LLMs Debug Robots by Editing Their Weights?', puts an LLM (GPT-6 Astra) in the debugger's seat for a π₀.₅ VLA robot policy: it inspects tensors, edits weights, and uses simulation feedback to validate candidates—no gradient updates, no new demos.
- The final 65-row edit targets the token embedding table (132,768 parameter values, 0.00367% of the checkpoint), shifting tokens like cooking toward stove, fermented toward wine
- Best edit lifts held-out success from 86.3% to 92.0% (p=1.31e-10); the 335 instructions containing edited tokens go from 74.3% to 87.5% while the other 430 episodes stay unchanged—effects are highly localized
- Controls: random-direction perturbations fail; wholesale replacement of embeddings reaches 90.5%, so finding the right token mapping is the key part
- All 16 experimental sessions are public, including 12/13 abandoned candidates and a withdrawn 'failure recovery' fix that improved an internal signal but hurt task success
- Bonus finding: appending metadata suffixes to instructions drops another checkpoint's success from 70.0% to 48.6%
- Paper, code, and per-episode evidence are open-sourced; real-robot validation is ongoing
More from Embodied
- Robopedia launches Xperience 2, a 'gold standard' Physical AI dataset — liuziwei7 · 2026-11-13
- Baidu invites riders to try its 6th-gen Apollo Go robotaxi in Beijing — Baidu_Inc · 2026-11-13
- Gemini Robotics 2.5 Debuts at CoRL 2026: Humanoid Ties a Bow Autonomously — ZoubinGhahrama1 · 2026-11-13
- Astra vs Fable 5.2 zero-shot excavator test: only visual reference, one model prevails — Scobleizer · 2026-11-13
- Waymo rider pulls up next to a Cybercab — with a human driver still in the seat — kscottz · 2026-11-13
- SimpleVLN: native multimodal LLMs navigate continuously without a specialized backbone — jiqizhixin · 2026-11-13