- Rating
- 2
- Confidence
- 3
- Informativeness
- 3
- Presentation
- 3
- Soundness
- 2
Read review
Paper summary
This paper investigates an interesting security problem in self-evolving agents: untrusted external evidence can be combined with the agent's own actions and feedback to form persistent behavioral associations during memory updates. The authors propose ELLA, which uses Semantic Anchoring (SA) and Writer Exposure (WE) to select visual evidence that is likely to be retrieved and incorporated into the agent's memory. Unlike traditional memory poisoning attacks, the target action is not explicitly included in the injected content but is instead associated with the source evidence through the agent's own memory writer. Experiments across multiple benchmarks, memory backends, and model backbones demonstrate the effectiveness of the proposed attack.
Critical strengths
-
Interesting and important research problem. The paper identifies an interesting attack surface in the memory formation process of self-evolving agents. I particularly appreciate the perspective that harmful persistent associations can be generated by the agent's own memory writer, rather than being directly injected by the attacker.
-
Clear writing and motivation. The paper is generally well written, and the motivation and overall attack pipeline are clearly presented.
-
Interesting mechanism experiments. I appreciate the causal experiments, particularly the no-completion control, source removal, lineage cleanup, and commitment-flip experiments. These provide useful evidence that the observed attack effects are associated with writer-generated memory rather than merely the retrieval of malicious content.
Critical weaknesses
-
Why visual evidence? The proposed method is straightforward and easy to understand. However, I wonder whether the vulnerability is actually specific to visual evidence. For example, what if we replace the image with an equivalent textual description? Would the attack still succeed?
-
Limited exploration of image optimization. The proposed method mainly relies on selecting existing images through semantic similarity and UCB-based feedback. However, it remains unclear why image selection is preferred over directly optimizing or generating images for the attack objective. Could image optimization lead to stronger or more transferable attacks? Exploring this alternative would help better justify the current design and potentially strengthen the methodological contribution.
-
Insufficient understanding of source–action associations. One interesting question is why the memory writer associates the source evidence with a particular target action, especially when the target action is absent from the injected content. I wonder how much the attack effectiveness depends on the semantic relationship between the source and the target action. For example, would the attack still succeed when the two are weakly related or unrelated? A more systematic analysis could provide valuable insights into the underlying memory formation mechanism.
Questions for authors
See weaknesses.
4 comments
W2. Why study selection from existing images, and can direct optimization improve success rates or cross-model transfer?
We appreciate the reviewer's suggestion to directly compare image selection with optimization. We believe image selection is a meaningful methodological choice because ELLA investigates whether persistent memory poisoning can arise from ordinary, unmodified visual evidence, without attacker-controlled pixel-level perturbations. ELLA studies a setting in which an external provider can choose and submit an existing, unmodified image. Even with this limited control over the input, we observed that the system’s own Writer can form persistent, traceable associations.
To further assess the reviewer's concern, we conduct a paired comparison of existing image selection and direct image optimization. Both image sets were determined on Qwen and then held fixed for evaluation on an independent DeepSeek model. Within each model, both methods used the same eligible queries.
Table 9. Paired comparison of existing image selection and direct image optimization
| Evaluation model | Shared eligible queries | Selection Post-ASR | Optimization Post-ASR | Post-ASR difference | Selection SC-ASR | Optimization SC-ASR | SC-ASR difference |
|---|---|---|---|---|---|---|---|
| Qwen3-VL-30B-A3B | 49 | 9/49 (18.4%) | 5/49 (10.2%) | −8.2 pp | 3/49 (6.1%) | 1/49 (2.0%) | −4.1 pp |
| DeepSeek-V4.1-Flash | 67 | 44/67 (65.7%) | 44/67 (65.7%) | 0.0 pp | 33/67 (49.3%) | 33/67 (49.3%) | 0.0 pp |
Differences are calculated as optimization minus selection.
We observe no improvement from direct optimization in either evaluated model: selection attains higher point estimates on Qwen and has the same point estimates as optimization on DeepSeek for both metrics. We believe the comparison strengthens our design rationale: selecting unmodified images is sufficient to induce traceable persistent-memory effects, and pixel-level optimization is not necessary for the vulnerability to manifest in these settings. The existing-image selection condition also produces traceable hits on the independent model.
W3. How does semantic relatedness between the source and target action affect memory formation?
We appreciate the reviewer's request for a more systematic analysis of source–target semantic relatedness. We therefore hold the carrier images, target action, Writer, backend, and random seeds fixed and vary only the semantic relationship between the supplemental source text and the target action. The high-relatedness condition explicitly mentioned the target action and served as a diagnostic control. The unrelated, low-relatedness, and medium-relatedness conditions did not explicitly instruct the agent to take the target action.
Table 10. Semantic relatedness and downstream outcomes
| Semantic relatedness | Rule-based score | Association formation rate | Post-ASR | Post-ASR difference vs. unrelated |
|---|---|---|---|---|
| High | 5 | 88.0% | 53.8% | +17.9 pp |
| Medium | 2 | 83.0% | 48.7% | +12.8 pp |
| Low | 1 | 74.0% | 48.7% | +12.8 pp |
| Unrelated | 0 | 60.0% | 35.9% | Reference |
We observe two complementary findings. First, semantic relatedness facilitates writer-side association formation: among conditions without explicit target-action instructions, the formation rate rises from 60.0% (Unrelated) to 74.0% (Low) and 83.0% (Medium). The Medium and Low conditions have Post-ASR values 12.8 pp above Unrelated. Second, we believe this result also underscores the breadth of the underlying risk: even unrelated source context yields a 60.0% observed association-formation rate. Thus, strong semantic alignment improves formation but is not a strict prerequisite. The High condition explicitly mentions the target action and is treated only as a diagnostic control.
We sincerely appreciate the reviewer's constructive suggestions. We hope these additional analyses clarify why visual evidence and selection-based optimization are meaningful for ELLA, and encourage you to reconsider your assessment.
Thanks for the authors' reply. I appreciate the experiments to strengthen the method design. However, the current rebuttal shows the performance but not the intuition. I would like the authors to discuss more insights behind those experimental results. For example, why can selection outperform generation? and the same for the following questions. I would have been more willing to champion it with the insights behind the work.
W1. What changes when an original image is replaced by a text description of the same scene?
We sincerely appreciate your positive assessment of the memory-formation perspective and your recognition of the paper's motivation, presentation, and causal analyses. We also thank you for the constructive suggestions regarding input modality, direct image optimization, and semantic relatedness. Following these suggestions, we have conducted paired image–caption and image-optimization comparisons, together with a controlled semantic-relatedness analysis.
We appreciate the reviewer's suggestion to test whether equivalent textual evidence can induce the same mechanism. We first compare original images with captions describing the same scenes on 100 matched queries. The image group had a 14 pp higher source retrieval rate and a 55 pp higher rate of source exposure to the Writer. We observe that the advantage of visual evidence becomes considerably larger at the memory-writing boundary, not merely at initial retrieval.
Table 7. Paired image–caption retrieval and Writer exposure
We separately ran an end-to-end evaluation with Qwen3-VL-30B-A3B + Mem0 + Mem-Gallery. All four conditions shared the same 71 queries on which the paired Clean run did not select the target action. SC-ASR additionally requires retrieval of a traceable derived memory.
Table 8. End-to-end comparison of image and caption conditions
We acknowledge that the caption conditions show higher overall target-action rates. However, we believe the more informative distinction for ELLA is whether the target action is accompanied by traceable, writer-generated memory. In the separate end-to-end evaluation, the Caption PFS source did not appear in the retrieval results, and none of the 27 derived memories linked back to it. The 26% caption Writer exposure in Table 7 belongs to a distinct 100-query paired-retrieval protocol. In the end-to-end experiment, only the visual conditions yielded target-action hits accompanied by retrieval of source-linked derived memories. We will clarify these protocol differences and the lineage criteria in the revised paper.