Driving · 2026-10-08
TMT: Runtime Backdoor Detection for Vision-Language-Action Policies on Unseen Tasks
Zirun Zhou, Jingfeng Zhang, HaoChuan Xu, Xizhe Zhang, Elliott Wen, Jing Sun, Hong Jia
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
TMT is an inference-time runtime backdoor detector for vision-language-action (VLA) policies on unseen tasks. It combines a token manifold branch (checks token representation structure) with a latent transition branch (checks prediction errors in adjacent-layer latent dynamics), both trained on benign rollouts. It can also purify the backdoored policy via self-distillation using paired benign/triggered observations, without a separate clean reference policy.
Why it matters
Backdoored VLA policies can keep benign task performance while producing malicious actions when triggers appear, and detection is hard when malicious actions look individually plausible and when tasks at test time are unfamiliar. TMT aims to detect trigger activation online using only benign rollouts for fitting, and reports state-of-the-art (or joint-highest) trigger detection on unseen tasks across three VLA backdoor attacks.
Method
- Train TMT on benign rollouts (no triggers) with two branches: a token manifold branch modeling benign token structure conditioned on preceding query history, and a latent transition branch modeling adjacent-layer latent dynamics via prediction errors.
- Calibrate branch thresholds using held-out benign rollouts; raise an alarm when either branch’s score exceeds its calibrated threshold during online detection.
- After detecting a confirmed incident, select a monitored latent transition for subsequent rollouts (including unseen tasks); also propose self-distillation purification using teacher-driven paired rollouts from the backdoored policy.
Limitation
The authors focus detection on unseen manipulation tasks, with TMT fitted using benign rollouts from seen tasks.
Abstract (from arXiv)
Backdoored vision-language-action (VLA) policies can preserve benign task performance while producing malicious actions when a trigger appears. Detecting such activation is difficult because malicious behavior can comprise individually plausible actions, while unfamiliar tasks introduce legitimate changes in observations and behavior. We introduce TMT, a runtime backdoor detector based on Token Manifold and latent Transition modeling. Trained on benign rollouts, its two branches assess input-token structure and prediction errors in adjacent-layer latent dynamics. A suspicious rollout identified by the token manifold branch, once confirmed through latent deviations, guides transition selection for subsequent monitoring. We further explore policy purification through self-distillation: a frozen copy of the backdoored policy provides benign-input actions to supervise a student on paired benign and triggered observations, without requiring a separate clean reference policy. For evaluation, we adapt traditional backdoor detectors and repurpose anomaly and failure detection methods as VLA backdoor detectors. In a post-hoc comparison with ten baselines, TMT achieves state-of-the-art backdoor detection performance on unseen tasks across three VLA backdoor attacks. Our project page is available at https://zzr42.github.io/tmt/.
Related papers
- VeriVLA: Adaptive Test-Time Verification for Flow-Based VLAs with Latent World Models
- APS-VLA: Action-Guided and Future-Aware Token Pruning for Efficient Vision-Language-Action Models
- MIKASA-Robo-VLA: Benchmarking Memory in VLA Models for Long-Horizon Manipulation
- PaceVLA: Adaptive Temporal Action Quantization in Vision-Language-Action Models
- Fast-VLA: You Only Think Once for Robot Manipulation
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-07 · Indexed by Pub-AI: 2026-10-08 UTC
TMT: Runtime Backdoor Detection for Vision-Language-Action Policies on Unseen Tasks
Preprint posted: 2026-10-07 · Original record ↗
- arxiv
- 2610.09462
Explicit source versions (1)
- v1 · 2026-10-07
- Preprint available
Source date: 2026-10-07 · First detected: 2026-10-25 UTC · Historical baseline
Source evidence (1)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
Backdoored vision-language-action (VLA) policies can preserve benign task performance while producing malicious actions when a trigger appears. Detecting such activation is difficult because malicious behavior can comprise individually plausible actions, while unfamiliar tasks introduce legitimate changes in observations and behavior. We introduce TMT, a runtime backdoor detector based on Token Manifold and latent Transition modeling. Trained on benign rollouts, its two branches assess input-token structure and prediction errors in adjacent-layer latent dynamics. A suspicious rollout identified by the token manifold branch, once confirmed through latent deviations, guides transition selection for subsequent monitoring. We further explore policy purification through self-distillation: a frozen copy of the backdoored policy provides benign-input actions to supervise a student on paired benign and triggered observations, without requiring a separate clean reference policy. For evaluation, we adapt traditional backdoor detectors and repurpose anomaly and failure detection methods as VLA backdoor detectors. In a post-hoc comparison with ten baselines, TMT achieves state-of-the-art backdoor detection performance on unseen tasks across three VLA backdoor attacks. Our project page is available at https://zzr42.github.io/tmt/.
