acceptodds
Under review as a conference paper at ICLR 2027

Lying in VQA Without Getting Caught: CoT-Obfuscated Backdoors in VLMs

Abstract

Generative (thinking and reasoning) models can justify their outputs in the form of Chain-of-Thought (CoT) transcripts. However, it remains debatable how accurate these justifications are in practice. In this paper, we take this thought one step further and present a neural backdoor that disguises the attack in light of CoT reasoning. The backdoored model is tasked with Visual Question Answering (VQA), but answers wrongly if the input image contains a certain “trigger pattern” and lies about it, meaning the CoT transcript provides a reasonable justification why the answer would be correct nevertheless. Our attack, Chain-Of-Lies, consists of two fundamental training phases, Supervised Fine-Tuning (SFT) and Group Relative Policy Optimization (GRPO), in which we first implant the backdoor into the base model and then train the model towards stealthiness using Reinforcement Learning (RL). For the latter, we use rewards from a separate neural overseer VLM, specifically instructed to penalize the model if the CoT reveals the backdoor. In our evaluation, we show that our attack bypasses defense mechanisms such as “Chain-of-Scrutiny” and generalizes across multiple VLMs, including Qwen3-VL , Gemma3 , and Ministral-3 . Our results show that an AI model’s self-reasoning cannot be trusted for safety evaluations.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
Chain-of-Thought, Backdoor Attacks, Vision-Language-Models

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_FZKxOpenReview ↗
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
  1. 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.
  2. 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

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:

ASROvEROvSCTP
84.27+/-4.5679.72+/-2.9692.33+/-1.2764.74+/-0.77

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.

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:

methodASRCTPOvEROvSOER_CoS
Decoupled CoT (MirageBD class)96.7797.4513.6199.228.84
Answer-only (BadVLM/IAG class)96.7797.4591.1699.3681.12
Ours98.3095.54100.00100.0095.24

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.

Reviewer_HFj8OpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper studies whether chain-of-thought (CoT) obfuscation can enable backdoored vision-language models (VLMs) to evade neural overseers. The authors introduce a backdoor attack in which a visual trigger causes the model to produce an incorrect answer while generating a seemingly coherent CoT explanation. Using supervised fine-tuning and group relative policy optimization (GRPO), the attack is optimized to evade a consistency-based neural overseer. Experiments across three VLMs show that the proposed method substantially increases evasion against the overseer used during training and achieves high evasion rates against some additional overseers. The paper highlights a potential vulnerability in using observable CoT consistency as a safety-monitoring mechanism. However, the current evaluation does not establish that the attack generalizes broadly across overseer designs or that CoT-based oversight is inherently unreliable. In particular, the overseer's evaluation objective, the role of independent visual verification, and the generalizability and stability of the reported results require further investigation.

Critical strengths
  1. Important and relevant research questions. The paper investigates whether a backdoored VLM can manipulate its observable reasoning traces to evade oversight. This is a meaningful security concern for systems that rely on CoT consistency as part of their safety monitoring.
  2. A concrete attack framework. The authors combine a visual backdoor trigger with CoT generation and GRPO-based optimization against a neural overseer. This provides a concrete experimental setting for studying the interaction between backdoor behavior and reasoning-based monitoring.
  3. Promising empirical results. Experiments across three VLMs show that the proposed method can substantially increase evasion against the overseer used during optimization. The reported results suggest that optimizing the generated explanation, rather than only the final answer, can materially affect overseer judgments.
  4. Identifies a limitation of consistency-based oversight. The results illustrate that a seemingly coherent explanation does not necessarily establish that a model's answer is correct. This is a useful caution for oversight mechanisms that assess reasoning consistency without independently verifying the underlying visual facts. Weaknesses
Critical weaknesses
  1. The overseer's evaluation objective is narrower than the paper's broader claims. The paper explicitly instructs the overseer to focus on reasoning consistency rather than factual correctness. Consequently, an explanation that coherently rationalizes an incorrect answer can be judged successful even when its factual basis is false. This makes the results evidence that the attack can exploit a particular oversight objective, rather than demonstrating that CoT-based oversight is generally unreliable. The paper should more clearly distinguish these claims and evaluate how the attack performs under overseers with different objectives.
  2. The role of independent visual verification is insufficiently explored. The paper reports that giving the image to the overseer and allowing it to perform independent visual question answering substantially reduces the effectiveness of the attack for the evaluated overseers. This is an important qualification because factual verification is a plausible component of a practical oversight system. The paper should investigate this setting more systematically and clarify whether its conclusions apply primarily to consistency-based monitoring without independent verification.
  3. The evaluation does not sufficiently establish generalization or stability. The main evaluation uses the same overseer for training and evaluation, making the high evasion rates evidence of successful optimization against that specific judge rather than independent validation of general attack effectiveness. Although the paper includes additional overseers, the transfer evaluation is limited, and some smaller overseers have low baseline specificity, complicating the interpretation of their evasion rates. Moreover, the main results are reported from a single optimization run, leaving the reproducibility and sensitivity of the findings uncertain. Evaluation across independently trained runs and a broader range of overseer designs is needed to establish the robustness and generality of the results.
  4. The experimental scope limits the strength of the general conclusions. The experiments focus on a specific COCO-based VQA setting and an edibility-related target behavior. It remains unclear how well the method generalizes to other visual tasks, target behaviors, trigger configurations, and real-world input pipelines. The paper would benefit from a more explicit discussion of these limitations and, where feasible, additional robustness experiments.
Questions for authors

See the weaknesses

4 comments

Please find answers and clarifications below, that we will also incorporate in the paper by the end of the rebuttal phase.

Weakness 1

Instructing the overseer to not perform a factual verification is a limitation, but also necessary to evaluate the attack in the simplified target VQA task. The underlying intention is the assumption that in more complex attack settings, the overseer is not always able to perform a factual verification of each reasoning step either because of task complexity or inferior overseer abilities. We use the Chain-of-Scrutiny instructed overseer as an example of an overseer with a different objective. Furthermore, we have since evaluated our attack in a different attack and trigger setting, which is detailed below.

Weakness 2

Independent factual verification in the form of a second model completing the same task and then comparing the final answers of the two models is out of our attack's scope. However, we additionally investigate the independent visual verification weakness of the attack with an overseer model that explicitly uses the potentially backdoored model's reasoning traces. Here, we have found an interesting result: our suspected "independent visual verification" failure mode of the overseer heavily depends on the specific model and size. Below, we report OvER when including the image in the overseer prompt for our three target models against three different model sizes:

OverseerQwen3-VL-2BGemma3-12BMinistral-3-14B
Qwen3.5-2B91.8475.6882.99
Qwen3.5-9B34.3553.7458.16
Qwen3.5-27B94.7398.1396.26

This suggests that the failure in the independent visual verification setting is not as critical as we had initially reported.

Weakness 3

We have since evaluated our attack pipeline on the main Qwen model using three more seeds. The results show that the attack itself is somewhat fragile across specific seeds but overall still severe:

ASRCTPOvEROvSSCRASR_CoSOvER_CoSOvS_CoS
paper97.7995.3399.83100.00100.0092.6993.0394.97
mean+/-sd94.79+/-5.3596.84+/-1.8799.94+/-0.10100.00+/-0.00100.00+/-0.0075.74+/-20.2680.78+/-14.9896.88+/-1.78

Weakness 4

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:

ASROvEROvSCTP
84.27+/-4.5679.72+/-2.9692.33+/-1.2764.74+/-0.77

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.

[1] Hudson and Manning, "GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering," CVPR 2019.

I appreciate the authors' responses and the additional experiments, particularly the evaluations across different random seeds and the GQA setting. These additions provide further empirical evidence and help clarify the scope of the proposed attack. However, my main concern regarding the narrow evaluation objective of the overseer remains unresolved. In particular, the CoS evaluation still primarily relies on comparing reasoning traces rather than independently verifying their factual correctness against the visual input. Therefore, the results do not fully support the broader claims about the general reliability of CoT-based oversight. I therefore maintain my original rating.

We thank the reviewer for the feedback and the chance to address the raised concerns.

We have run a simple evaluation in which we do not pass the backdoored model's reasoning trace to the overseer model and only ask it to complete the task independently. In this case, an "Overseer Evasion" is an agreement between overseer answer and target model answer, which leads to low OvER: 6.9, 7.1, and 12.2 for Qwen, Gemma, and Ministral over truly inedible triggered samples for the Qwen3.5-9B overseer, respectively. We will adjust our claims in this regard in the revision.

We still argue that this kind of independent verification does not scale well for more complex tasks, though, as it requires an overseer to independently re-solve the task for every sample. Instead, a consistency-based validation, like Chain-of-Scrutiny is more viable. We therefore qualify the final conclusion of our work to

"However, it also means that consistency-based neural overseer verification as recently proposed by Chain-of-Scrutiny (Li et al., 2025) cannot be relied upon to judge misuse. Not only can this specific type of verification be wrong, it can be actively manipulated by an adversary to cover up unsafe operation. Independent verification, however, is more robust against this type of attack and can therefore be a viable defense."

and will update the paper accordingly.

We have incorporated the additional results in the revised PDF uploaded to OpenReview. In particular, we mark off our contribution and separately discuss independent verification in the main part of the paper (incl. the corresponding results) as an ablation. Changes are indicated in blue. Thank you for your input!
We are looking forward to your final decision.

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

This paper introduces CHAIN-OF-LIES, a supply-chain backdoor attack on vision-language models for visual question answering. The attacker fine-tunes an open-weight VLM so that a red-grid trigger causes the model to answer “yes” to food-edibility questions for non-edible objects, while producing a chain-of-thought transcript that hides the trigger and gives a plausible-sounding justification. The method has two stages: curriculum-based supervised fine-tuning with decreasing trigger opacity to implant the trigger-to-answer mapping, followed by GRPO with format, backdoor-answer, and frozen neural-overseer rewards to make the CoT evade a VLM judge. Experiments on Qwen3-VL-2B, Gemma3-12B, and Ministral-3-14B report ASR of 92.5–97.8%, OvER of 97.6–99.8%, more than 93% evasion of Chain-of-Scrutiny, preserved MMMU performance, and ablations on curriculum learning, SFT, GRPO, and trigger design. The paper concludes that CoT transcripts and neural overseers cannot be trusted for safety evaluation.

Critical strengths
  1. The threat model is well motivated. The two-stage SFT+GRPO pipeline is clearly specified, and the no-SFT/no-GRPO/no-curriculum ablations show that each component is doing real work. The attack is demonstrated across three VLM families, which makes the finding more than a single-model artifact.

  2. The paper addresses an important and timely question: whether CoT-based monitors remain reliable under an adaptive supply-chain attacker. The results are striking and provide a concrete target for defense work.

Critical weaknesses

The evaluation is too narrow and lacks the controls needed to support the generality claim. The paper studies one binary food-edibility task, one full-image red-grid trigger, and reasoning VLMs only. It does not report a clean CTP baseline, sample sizes or seed variance, COCO splits, the edibility mapping, the multi-object aggregation rule or standard backdoor defenses. Table 1 is a single optimization run. Patch triggers at alpha=8 reach only 16.7–18.6% ASR, so the attack depends on a full-image grid that may itself be detectable.

Questions for authors
  1. Please report clean-model CTP and MMMU with sample sizes and multiple seeds, and include a clean-data GRPO control. Why does Ministral MMMU increase from 0.526 to 0.608?

  2. Does the attack generalize to multi-class or open-ended VQA, non-reasoning VLMs, other triggers, and lower poison rates?

  3. Please report conditional OvER given successful ASR, as well as the joint success rate. The current OvER includes failed attacks and may overstate stealth.

  4. How sensitive is the attack to the choice of training overseer? Please test a different training judge.

  5. Please correct the two bibliographic entries: use Fu et al. (2018, ICLR) instead of the 2017 CoRR preprint, and use Barez et al. (2025, Oxford Martin AIGI Whitepaper) instead of the 2026 CoRR entry. Update the in-text citations accordingly.

3 comments

Please find answers and clarifications below, that we will also incorporate in the paper by the end of the rebuttal phase.

Question 1

We will extend our entire evaluation with multiple seeds, but can only report updated numbers for the backdoored model as of now. Below we report the clean-model CTP & MMMU, and a clean data SFT+GRPO run for the Qwen3-VL-2B model and an n=2000 large dataset with 588 triggered samples.

ModelCTPOvSMMMU overallMMMU MC-only (n=847)
Qwen3-VL-2B (unmodified)94.8392.140.4730.502
Qwen3-VL-2B (clean data, SFT + GRPO, across 3 seeds)97.00+/-0.6699.95+/-0.040.501+/-0.0120.530+/-0.013
Qwen3-VL-2B (backdoored, across 3 seeds)96.84+/-1.87100.00+/-0.000.488+/-0.0140.517+/-0.015
Qwen3-VL-2B (backdoored, as published)95.33100.000.5010.530

Explicitly rewarding reasoning answers results in increased MMMU scores after the backdoor training. While the model frequently responds with the single chosen answer (for MMMU multiple-choice questions) before backdoor training, it now always utilizes its reasoning mode---even if it is not prompted/instructed to do so.

Question 2

We require a ~20% poisoning rate. With less, the backdoor does not reliably materialize in the SFT phase. Interestingly, we cannot compensate low poisoning rates with longer training. The exact poisoning rate requirements are likely related to the concrete task and how far the desired model answers are from the prior learned distribution.

Moreover, 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:

ASROvEROvSCTP
84.27+/-4.5679.72+/-2.9692.33+/-1.2764.74+/-0.77

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 3

We evaluate conditional OvER and the joint success rate for the three models we use for the original results:

ModelASR (this eval run)OvER given ASRJoint
Qwen97.45100.0097.45
Gemma97.4597.2194.73
Ministral92.5297.6190.31

Question 4

We additionally evaluate using Gemma3-27B-IT as training-time overseer while keeping all other variables the same, observing that the attack is not fully agnostic to the overseer. Across three different seeds, the backdoored Qwen model reaches:

Training-Time OverseerTest-Time OverseerASROvEROvER_CoS
Gemma3-27B-ITGemma3-27B-IT90.70+/-10.9699.94+/-0.1082.94+/-14.21
Gemma3-27B-ITQwen3.5-9B89.91+/-11.9696.77+/-4.4584.98+/-12.70

This suggests that the choice of training-time overseers is a critical hyperparameter: the backdoored model transfers only partially to a test-time overseer it was not trained against. Note, though, that the model we used in the paper (trained against the Qwen overseer) still achieves an OvER of 100.00 against this Gemma-based overseer.

Question 5

We have updated the references in the revised paper. Thank you!

Weaknesses

We map 10 out of 80 COCO categories to "edible" (banana, apple, sandwich, orange, broccoli, carrot, hot dog, pizza, donut, cake). If one edible class is present in the image, we treat the entire image as "edible".

[1] Hudson and Manning, "GQA: A New Dataset for Real-World Visual Reasoning and Compositional Question Answering," CVPR 2019

Thanks for the response. Most of my concerns are addressed, I keep my positive rating.

We have incorporated the additional results in the revised PDF uploaded to OpenReview. Changes are indicated in blue. Thank you for your input!

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.