Evidence · DiggingBeagle record
MuTRAP: multi-trigger task-level backdoors against robot-planning soft prompts
MuTRAP: multi-trigger task-level backdoors against robot-planning soft prompts
- Source role
- primary disclosure
Evidence record
MuTRAP: multi-trigger task-level backdoors against robot-planning soft prompts
Claim-level citations (2)
- supportsMuTRAP poisoned robot soft prompts redirect plans without retraining base weights: MuTRAP implants the backdoor in small task-specific robot-planning parameters while leaving the central backbone language model frozen, and uses optimized trigger words to steer generated action plans.
Abstract, task-specific parameters and multi-trigger optimization
- supportsMuTRAP poisoned robot soft prompts redirect plans without retraining base weights: The authors train task-specific soft prompts carrying multiple trigger phrases and report high malicious-plan rates in controlled open-model robot-planning evaluations while retaining much of benign-task performance.
Abstract; task-specific soft-prompt threat model; robot planning evaluation