- Rating
- 3(+2)
- Confidence
- 2
- Informativeness
- 2
- Presentation
- 3
- Soundness
- 1
Read review
Paper summary
This paper aims to conduct an attack during the Chain-of-Thought (CoT) reasoning process, thereby contaminating the model's reasoning and leading to incorrect answers. To achieve this, the authors construct clean and poisoned datasets and use pixel-level signals as attack triggers. The model is then trained in two stages: first using SFT, followed by GRPO to further improve the attack effectiveness. Experimental results across four models, including Qwen3-VL, Gemma 3, and Mistral-3, show that the proposed approach achieves a high attack success rate.
Critical strengths
- The idea of targeting the CoT reasoning process is interesting, as it makes the attack harder to detect since the reasoning process and final answer remain consistent. I think this is an interesting direction worth exploring.
Critical weaknesses
- The proposed approach mainly relies on data collection followed by two-stage training, which is a common practice for training models. The main novelty of this paper seems to be applying this approach to the CoT reasoning process. However, the paper provides limited discussion on why this is a non-trivial contribution, especially since similar ideas have already been explored in prior works, e.g., [1] and the more recent [2]. The authors should better clarify the novelty of their approach and how it differs from existing methods.
- The experiments lack comparisons with existing attack approaches. The authors should include relevant baselines, such as [1, 3, 4], to better demonstrate the effectiveness of the proposed method and its advantages over prior work.
- The practicality and flexibility of the proposed attack are questionable. The approach requires a pixel-level trigger and relies on the victim using the authors' fine-tuned model. In practice, applying such a trigger may be difficult, and users may not necessarily interact with the compromised model. If the assumed attack scenario involves a microservice that can post-process user inputs, an attacker with that level of control could potentially manipulate the inputs or outputs directly in the cloud without fine-tuning the model. A more practical attack scenario would be to fine-tune the model and release it without requiring an explicit pixel-level trigger. For example, the attack could be triggered by certain types of images (e.g., animals or food) or specific prompt patterns (e.g., certain types of questions). This would make the attack more realistic and easier to deploy in practical settings.
[1] Xiang et al. BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models. ICLR 2024. [2] Zeng et al. MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning. 2026. [3] Dang et al. BadVLM: Towards Efficient and Resilient Backdoor Attacks on Large Vision-Language Models, CVPR findings, 2026. [4] IAG: Input-aware Backdoor Attack on VLM-based Visual Grounding, CVPR 2026.
Questions for authors
- Could the authors further clarify how the proposed method differs from existing CoT attack approaches? It would also be helpful to provide experimental comparisons with existing VLM backdoor attack baselines to better demonstrate the novelty and effectiveness of the proposed approach.
- Could the authors better justify why the current attack setting is practical and realistic for a backdoor attack? In particular, how would an attacker deploy this approach in real-world scenarios, and how would it affect the outputs received by users?
7 comments
I appreciate the authors' response and have raised my score to 2.
However, my primary concern regarding the comparison remains. Even though the methodology is different and the authors have clarified these differences, I still believe it is feasible to compare the proposed method against the baselines I previously listed. Including these comparative results would significantly strengthen the argument for why the proposed approach is superior.
We thank the reviewer for the additional feedback.
We have now added a comparison of our results against the baseline approaches mentioned. The approach most comparable to ours is MirageBD [2]. Unfortunately, the authors of MirageBD did not publish code or model checkpoints, so that we re-implement as follows: We re-train a model with our backdoor SFT pipeline on clean reasoning traces but with the final answer switched, so the model learns a decoupled final answer from its reasoning for poisoned samples. This is part of MirageBD's core method to introduce the backdoor, which we refer to as "MirageBD Class."
Our original model pipeline with the RL phase removed (our core contribution) is comparable to the approaches IAG [5] and BadVLM [4]. We hence do not re-run those approaches but use our SFT-only model as a proxy.
We have measured the five known metrics across these three model classes:
| method | ASR | CTP | OvER | OvS | OER_CoS |
|---|---|---|---|---|---|
| Decoupled CoT (MirageBD class) | 96.77 | 97.45 | 13.61 | 99.22 | 8.84 |
| Answer-only (BadVLM/IAG class) | 96.77 | 97.45 | 91.16 | 99.36 | 81.12 |
| Ours | 98.30 | 95.54 | 100.00 | 100.00 | 95.24 |
The three different approaches marginally differ in ASR and CTP only. However, against a CoT consistency-based overseer (what we refer to as OvER), our approach is superior.
BadChain is a prompt-only few-shot attack method without weight modifications, i.e. it assumes an entirely different threat model and is therefore not comparable to our approach.
[2] Zeng et al., "MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning," ACL 2026.
[3] Xiang et al., "BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models," ICLR 2024.
[4] Dang et al., "BadVLM: Towards Efficient and Resilient Backdoor Attacks on Large Vision-Language Models," CVPR 2026.
[5] Li et al., "IAG: Input-aware Backdoor Attack on VLM-based Visual Grounding," CVPR 2026.
Please find answers and clarifications below, that we will also incorporate in the paper by the end of the rebuttal phase.
Weakness 1 & Question 1
The perhaps closest related work is MirageBD [2] that makes the model generate an internally consistent CoT transcript, which however eventually is inconsistent with the final answer. While our two-stage attack pipeline is similar in design, we use an overseer as a training-time signal to eventually yield CoT transcripts that are consistent with the final answer.
Our attack aims at using the CoT transcript as a backdoored justification for the final answer, so that an automated overseer cannot detect the attack by merely comparing the final answer to what is suggested by the model's own CoT transcript. This is a failure mode that the authors of MirageBD specifically mention in their "Limitations" section: "it fundamentally relies on decoupling the reasoning trace from the final outcome" [2].
Please find a breakdown of all mentioned related works below:
Weakness 2
MirageBD [2] reports an average detection rate of 48.56% for an overseer based answer/CoT consistency check (Table 8). This problem is precisely what our attack aims at solving. The overseer is specifically instructed to check for "Logical Dissonance", i.e. a comparison between CoT and final answer.
The BadChain [3] attack requires no training of the model, instead it uses a malicious few-shot prompt to control the model's output. It does not optimize against an overseer. The optimization goal and methodology are fundamentally different from our proposed attack.
In BadVLM [4], the authors present a pipeline that optimizes the trigger itself to encode the desired semantics of the malicious output. While their attack can work with a lower poisoning rate, the trigger is not arbitrary, since it has to carry information about the output. Our attack can use triggers that do not carry these output semantics.
The authors of IAG [5] target a different specific target problem and also generate a per-sample trigger. Again, IAG does not target stealthiness in the model's output (it does for imperceptibility of the trigger) and it does not specifically use reasoning/CoT models.
We use a weight-level attack and unlike all the above it explicitly optimizes for a more stealthy answer justification.
Weakness 3
The current trigger would likely be removed by image compression or other transformations. However, we do not aim to showcase a robust trigger design but a technique to manipulate the CoT via a backdoor. To further exemplify this setting, we have designed a new evaluation environment based on the GQA [1] image dataset to evaluate whether our attack works on multi-class VQA with a trigger that is more stealthy: We use the "Which color is object xyz" subset and learn a backdoor that answers "orange" whenever a fire hydrant is present. In an initial test across three distinct seeds, we yield the following results after SFT & GRPO with similar hyperparameters to the original setting:
It is worth noting that the clean model also only reaches a CTP=61.40 for this task. We will evaluate with different models and conduct all ablations for the final version of the paper.
Question 2
We assume a classical "supply-chain attack" for the final model. The attacker distributes the backdoored model via a model-sharing platform, infiltrates the development pipeline, or tricks a user in utilizing and deploying that model. Adversarial ML literature additionally differentiates data-poisoning attacks that do not control the learning process and general model manipulation attacks that do. The former obviously is out of scope for our attack as we rely on SFT & GRPO.
[1] Hudson and Manning, "GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering," CVPR 2019
[2] Zeng et al., "MirageBackdoor: A Stealthy Attack that Induces Think-Well-Answer-Wrong Reasoning," ACL 2026.
[3] Xiang et al., "BadChain: Backdoor Chain-of-Thought Prompting for Large Language Models," ICLR 2024.
[4] Dang et al., "BadVLM: Towards Efficient and Resilient Backdoor Attacks on Large Vision-Language Models," CVPR 2026.
[5] Li et al., "IAG: Input-aware Backdoor Attack on VLM-based Visual Grounding," CVPR 2026.