acceptodds
Under review as a conference paper at ICLR 2027

SOMA SEMA: Action Space-Bound Attack that Awakens upon Benign Finetuning

Abstract

A vision-language-action (VLA) model is almost never deployed as released: before it can act, a practitioner finetunes it on its own robot, a step assumed safe because the finetuning data is the practitioner's own. That assumption fails. In SEMA, a provider distributes weights that come within a few points of a clean model's published task success wherever we benchmarked them and pass the screening we run, yet produce no unsafe behavior until the victim's own benign finetuning brings one out, with no poisoned example and no inference-time trigger. The same mechanism removes a hierarchical policy's safety guardrail and plants backdoors gated by a task, an object or a phrase. Whereas its text-domain predecessor carries a trigger that generalizes across the finetuning data, ours carries an action whose motor realization is bound to the action representation: it awakens on a different vendor's arm that shares the 7-dimensional single-arm space, and stays dormant both on a 14-dimensional bimanual robot and on the implant robot itself once the same actions are written in joint velocities or with permuted components. The semantic half of the guardrail payload is not so bound: a bystander's planner stops refusing though its arm never acts. Across four VLAs and two checkpoint types the attack reaches 74.0% to 86.5% of rollouts on the discrete-token models but 0.0% to 34.5% on the flow-matching ones, while the finetuned model stays within 8.5 points of a cleanly finetuned one on the victim's tasks. A short finetuning followed by screening then detects it on 58.0% of rollouts against 14.0% for a clean checkpoint. Severity throughout is a detector firing on the executed trajectory rather than a measured contact force.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
vision-language-action models, robot learning, embodied AI, streaming perception, long-horizon memory, bimanual manipulation

Reviews

Among:
Shape:
1234this paper, Top 30%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.67 over 3 reviews, above 70% of the others.
Reviewer_A4r7OpenReview ↗
Rating
3(+1)
Confidence
3
Informativeness
2
Presentation
2
Soundness
3
Read review
Paper summary

The paper proposes SEMA, a finetuning-activated backdoor attack on vision-language-action (VLA) models where a released model remains benign until the victim performs downstream clean finetuning. Through capability instillation and meta-learning-based burying, the attack binds physical action payloads to specific action space representations while keeping the released model's health checks and task success intact.

Critical strengths
  1. Evaluates multiple VLA architectures, diverse attack scenarios (stereotyped actions, guardrail removal, backdoors), and both simulation and real-robot platforms.

  2. Good writing.

Critical weaknesses
  1. The attack largely fails to re-surface high-frequency payloads (tremors) on continuous flow-matching models (0%–10% activation), undermining its claim as a universal VLA threat.

  2. The current gap between discrete and continuous models may stem from untuned loss weights (λml,λnoise,λreg\lambda_{ml}, \lambda_{noise}, \lambda_{reg}) rather than architectural limits, leaving the evaluation incomplete.

  3. The proposed screening defense is only validated under a narrow setting and lacks thorough analysis against adaptive attackers who might optimize around short-finetuning constraints.

Questions for authors
  1. Can you provide results demonstrating whether re-scaling the loss weights (λml\lambda_{ml}, etc.) for flow-matching heads closes the large activation gap observed with discrete-token models?

  2. How robust is the proposed short-finetuning screening defense if the attacker optimizes the meta-learning objective to remain dormant across varying fine-tuning budget scales?

1 comment

The authors resolved most of my questions, so I will raise my score to 3.

Reviewer_17qpOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper adapts finetuning-activated backdoor attacks (specifically FAB from Gloaguen et al., 2026) from LLMs to Vision-Language-Action (VLA) models. The proposed method, SEMA, first runs a capability instillation stage to teach the VLA both normal task execution and an adversarial physical motion. It then uses a first-order meta-learning stage ("burying") to suppress the adversarial motion in the released checkpoint while ensuring it re-emerges after k steps of benign downstream finetuning. Experiments across four VLAs (OpenVLA, pi0-FAST, pi0, pi0.5) on LIBERO, RoboTwin 2.0, and a real AgileX PIPER arm show that the awakened motor behavior is tied to the specific action representation (e.g., 7-D end-effector deltas) rather than the robot embodiment itself, and that flow-matching heads are much less susceptible to high-frequency payloads like gripper tremors than discrete-token heads

Critical strengths

Good empirical isolation of action-space binding (Section 4.4, Table 4). The most interesting part of the paper is not the attack algorithm itself, but the analysis of what the meta-learned trigger latches onto. By holding the robot, observations, and demonstrations fixed while permuting the 7-D action vector slots or switching to joint velocities, the authors cleanly show that the motor payload is bound to the coordinate interface rather than the physical embodiment. Showing that the high-level planner jailbreak in pi0.5 still transfers to a bystander robot while the low-level motor slam stays dormant is also a nice separation.

Honest statistical reporting and controls. The authors include solid controls that many backdoor papers skip, such as accounting for intra-task cluster correlation in Appendix F, separating standard finetuning safety erosion from the planted guardrail attack in Table 11, and running a pull-back probe (Figure 3) to separate geometric dormancy from catastrophic forgetting.

Critical weaknesses

Confounded comparison between discrete-token and flow-matching models. A major claim of the paper is that flow-matching models (pi0, pi0.5) are much more resilient to the attack (0 to 34.5% activation) than discrete-token models (74.0 to 86.5%). However, as noted in Section 6 and Appendix D, the burying loss weights (lambda_ml, lambda_noise, lambda_reg) were fixed on a discrete-token pilot using cross-entropy loss and reused without tuning on flow-matching models that use L2 vector-field regression. Since the gradient scales of CE and MSE are completely different, it is unclear if flow-matching is actually more robust to high-frequency payloads or if the meta-learning objective was just poorly scaled for continuous heads. Also, holding the victim's learning rate constant at the peak instead of using the default warmup + cosine decay schedule to match the attacker's inner loop feels slightly contrived.

Limited methodological novelty and artificial payload timing. Algorithmically, SEMA is a straightforward port of FAB (Gloaguen et al., 2026) with standard pre-finetuning ("capability instillation") added beforehand. More importantly, to claim that the attack preserves task success, the authors place the tremor, swing, and slam motions after the task's terminal state during instillation. In standard rollouts, the task is already done when the motion fires. When the payload actually happens mid-rollout (the gripper release in transit), task success drops by 52.5 points (down to 31.5%), which would make it obvious to any practitioner that the finetuned policy is broken.

Questions for authors

In Appendix D, you mention that the loss weights in Eq. 1 were transferred directly from a discrete-token setup to flow-matching regression without re-tuning. Did you try normalizing the loss terms by their initial magnitudes or sweeping lambda_ml / lambda_reg for pi0? Without tuning the objective for the continuous loss scale, how can we be sure the tremor failure is due to temporal frequency rather than hyperparameter mismatch?

Table 4 shows that simply permuting the 7-D action slots drops attack activation from 86.5% to 3.5% while keeping task success at 82.0% (vs 83.5%). Doesn't this imply that a user can trivially defeat SEMA at near-zero cost by applying a random permutation to the action dimensions (and normalizer stats) before finetuning, making the 2k-step screening defense in Section 5 unnecessary?

How well does the 2k-step screening defense (Section 5) work if the defender uses the actual published learning rate schedules from Table 21 (which have 1k to 10k warmup steps for pi0 and pi0.5) instead of a constant peak learning rate?

Reviewer_tuPrOpenReview ↗
Rating
3
Confidence
3
Informativeness
2
Presentation
3
Soundness
2
Read review
Paper summary

SEMA adapts FAB to vision-language-action models through capability instillation followed by meta-learning, producing behaviors that remain largely dormant at release and emerge after benign finetuning. Experiments cover different VLAs, two checkpoint types, simulation and physical-robot evaluations, component ablations, victim-finetuning variations, and screening through short finetuning, with strong motor-activation results on discrete-token models.

Critical strengths

1. A consequential security setting. The paper demonstrates that benign downstream training can activate undesirable behaviors in an apparently benign VLA checkpoint. This exposes a meaningful limitation of evaluating a downloaded policy only before adaptation.

2. Well-chosen attack scenarios. The paper studies 3 complementary scenarios: unconditional abnormal actions, guardrail removal, and trigger-dependent backdoors. Together, they examine failures in motor execution, safety-related planning, and conditional behavior, providing a useful view of how finetuning-activated attacks can manifest at different levels of a VLA system.

3. Substantial validation of the observed effect. The evaluation includes multiple models, starting checkpoints, attack scenarios, and physical-robot experiments. Holding the released checkpoint fixed while varying the victim’s learning rate, optimizer, schedule, and adaptation method is particularly useful: activation remains substantial under the tested changes, including LoRA.

Critical weaknesses

1. The attribution of release-time dormancy needs clarification. The author attributes the hiding of adversarial behavior to regularization, and Figure 2 describes a clean reference, whereas the algorithm uses the capability-instilled checkpoint as the reference. For unconditional motor attacks, how does this reference provide benign supervision in states associated with the implanted behavior? Moreover, the experiment gives that substantial suppression remains when regularization is removed, whereas removing meta-learning largely eliminates dormancy. These results suggest a stronger role for meta-learning in hiding the behavior than the explanation indicates. Why does this occur, and how do meta-learning and regularization interact to produce release-time dormancy?

2. Generalization across harmful behaviors remains unclear. Table 2 systematically compares tremor and upward swing and shows that activation depends substantially on the selected behavior. How does the method’s effectiveness vary across different harmful behaviors, and what determines whether a behavior can be successfully hidden and subsequently reactivated?

Questions for authors

I would appreciate clarification on the issues raised in the Critical Weaknesses section.

From OpenReview, CC BY 4.0. Authors as signed there; shown unedited.

Related papers

Loading the map…

Discussion (0)

Sign in to comment.