Risk Evaluation · Benchmark

Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control

NUWA research note · independent review pending

First published: 2026-10-03Source checked: 2026-10-28

Original source ↗Cite this record ↓Autonomous physical harm →

NUWA Research Note

Research question Design and assumptions Source findings NUWA interpretation Limits and contrary evidence Research and decision relevance Evidence update Institutions and roles Original resources

Model-assisted reading · Selected sections of the original full text or official article; a complete section-by-section audit remains pending.

At a glance

Text refusal and embodied action safety measure different failure paths.

Research question

Does a refusal observed in chat persist when the model controls a robot?

Design and assumptions

The study compares chat with robot-control interfaces and separates real-robot end-effector and joint-control settings. The real-robot conditions involve different interface/model populations, with separate trial counts.

Source findings

Reported refusal was 18% across 612 real end-effector trials and 41% across 252 real joint-control trials. These are rounded refusal rates, not rates of executed physical harm.

NUWA interpretation

Chat refusal is interface-specific behavioural evidence, not a sufficient condition for safe physical action. Refusal, execution and resulting damage are different endpoints.

Limits and contrary evidence

The settings are not a matched model-only comparison. Rounded rates do not identify exact refusal numerators, and refusal is not a physical harm measurement.

Research and decision relevance

NUWA recommendation: Pair hazard recognition with action-level evaluation; retain the interface, embodiment and outcome definition.

Precise source locations

Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control · Abstract; §§2–4; discussion; text lines 00044,00055,00061Original resource captured 2026-10-28T07:24:43.704880+00:00; sha256:c5b843fd1a59541e7c7713423f4c765813de1b6dc8cb85bc7aa5913c0c945572

Source checking, extraction, editorial approval and replication are recorded separately. Methodology →

Evidence update · 2026-10-28

Model-assisted reading · independent evidence and bilingual editorial review pending. Values are source findings or inventory counts, separate from the trend result series.

Source-bound observations (2)

Counts, rates and contrasts have different units. A sample count is not a number of independent experiments. Unknown numerators and interval levels remain unknown.

System / metric	Reported value	Population and conditions
Refusal rate	18 %	Denominator: 612 trials Conditions, limitations and source { "interface": "real-robot end-effector" } No exact refusal numerator inferred; interface/model populations differ. Refusal is not executed harm. Original source ↗ · Abstract; real-robot evaluation; text lines 00044,00055,00061Original capture 2026-10-28T07:24:43.704880+00:00; sha256:c5b843fd1a59541e7c7713423f4c765813de1b6dc8cb85bc7aa5913c0c945572
Refusal rate	41 %	Denominator: 252 trials Conditions, limitations and source { "interface": "real-robot joint" } No exact refusal numerator inferred; interface/model populations differ. Refusal is not executed harm. Original source ↗ · Abstract; real-robot evaluation; text lines 00044,00055,00252Original capture 2026-10-28T07:24:43.704880+00:00; sha256:c5b843fd1a59541e7c7713423f4c765813de1b6dc8cb85bc7aa5913c0c945572

Institutions and roles

Affiliation not indexed in this review

Authors (7)

Edward Sun · Christopher Leet · Sravanthi Machcha · Achyutha Menon · Tzu Kit Chan · Yuchen Cui · Jay Chooi

Evaluation protocols

RoboHarm →

Cite and reuse

Cite the original publication for its findings. Cite this NUWA record separately for its explanation and retained snapshot.

BibTeX ↓RIS ↓Record + source metadata · JSON ↓Reading coverage · JSON Suggest a correction

Original publication · BibTeX

@misc{roboharm,
  title = {Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control},
  url = {https://arxiv.org/html/2610.14043},
  author = {Edward Sun and Christopher Leet and Sravanthi Machcha and Achyutha Menon and Tzu Kit Chan and Yuchen Cui and Jay Chooi},
  year = {2026},
  date = {2026-10-03},
  version = {arXiv:2610.14043v1; version header 03 Oct 2026; Original resource captured 2026-10-28T07:24:43.704880+00:00; sha256:c5b843fd1a59541e7c7713423f4c765813de1b6dc8cb85bc7aa5913c0c945572},
  note = {arXiv:2610.14043v1; version header 03 Oct 2026; Original resource captured 2026-10-28T07:24:43.704880+00:00; sha256:c5b843fd1a59541e7c7713423f4c765813de1b6dc8cb85bc7aa5913c0c945572}
}

NUWA record · BibTeX

@misc{roboharm-nuwa,
  title = {Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control — NUWA research record},
  url = {https://whitzard.tech/en/nuwa/whitzard-index/work/roboharm/},
  version = {2026-10-29.1},
  organization = {NUWA Frontier AI Safety Lab},
  howpublished = {Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--roboharm.json},
  note = {NUWA research note. Source-reading scope and independent review status are stated on the record. Snapshot 2026-10-29.1. Retained note edition: https://whitzard.tech/nuwa/whitzard-index/data/record-editions/2026-10-29.1--roboharm.json.}
}

Evidence snapshot 2026-10-29.1 · Index updates · RSS

References

[1] Edward Sun, Christopher Leet, et al.. Refuses in Chat, Complies with Arms: LLM Refusal Does Not Survive Robot Control. 2026-10-03 ; Abstract; §§2–4; discussion; text lines 00044,00055,00061

Search NUWA Index