acceptodds
Under review as a conference paper at ICLR 2027

MESA-Bench: A Multimodal Evolving Student Attack Benchmark for Educational LLM Tutors

Abstract

Multimodal large language models (MLLMs) are increasingly deployed as tutor models to provide problem-solving guidance, interpret visual materials, and respond to students' states across multi-turn conversations. However, existing educational tutor safety evaluations largely rely on static samples or fixed student strategies, making it difficult to capture dynamic risks in which student attacks adapt to a tutor's prior responses. We propose MESA-Bench, a Multimodal Evolving Student-Attack benchmark for tutor integrity. MESA-Bench organizes evaluation as model- and objective-specific evolution tasks, and operationalizes tutor integrity along three dimensions: pedagogical integrity (PI), protective integrity (PrI), and epistemic integrity (EI), corresponding to answer leakage, neglect of welfare or crisis cues, and judgment retreat under unreliable evidence or social pressure. To cover the nonlinear and multimodal nature of educational attacks, MESA-Bench combines education-theory-inspired attack primitives, paired multimodal mathematics cases, and feedback-driven strategy-graph evolution. Across eight MLLMs and three integrity objectives, evolved graphs improve attack success rate in 18 of 24 model-objective runs and leave the remaining six unchanged; overall ASR increases from 48.92% to 52.92%, with the largest gain on EI. Further analyses show that graph expansion is not monotonically equivalent to stronger attacks, and that input-format risk profiles vary substantially by model and objective rather than exhibiting a uniform modality effect. Code, data, and generated evaluation cases are available at https://anonymous.4open.science/r/mesa-bench-review-3D1B/.

Primary area:
datasets and benchmarks
Keywords:
educational LLM tutors, multimodal safety, adversarial evaluation, red teaming, benchmark, strategy graph

Reviews

Among:
Shape:
1234this paper, Bottom 27%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 2.00 over 3 reviews, above 27% of the others.
Reviewer_CSEiOpenReview ↗
Rating
2
Confidence
3
Informativeness
3
Presentation
3
Soundness
2
Read review
Paper summary

This paper introduces MESA-Bench, a dynamic adversarial benchmark for evaluating the integrity of multimodal large language models (MLLMs) deployed as educational tutors. The authors conceptualize tutor integrity along three dimensions, i.e., pedagogical integrity (PI), protective integrity (PrI), and epistemic integrity (EI). The proposed framework combines multimodal educational scenarios, compositional attack primitives, and feedback-driven strategy graph evolution to generate multi-turn student attacks. Experiments across eight MLLMs and 24 model-objective settings show that the evolved strategies improve the overall attack success rate from 48.92% to 52.92%. The paper further investigates model-specific attack strategies, input-format sensitivity, and the effectiveness of multi-turn attack sequences.

Critical strengths
  1. The paper addresses an important and underexplored problem in educational AI safety by systematically evaluating tutor integrity across three complementary dimensions under adversarial student interactions.
  2. The proposed benchmark integrates multimodal attack construction, feedback-driven strategy graph evolution, and structured outcome auditing into a coherent automated evaluation pipeline.
  3. The evaluation covers eight MLLMs and includes statistical confidence intervals, human auditing, ablation studies, and strategy graph analyses, providing useful empirical insights into tutor vulnerabilities.
Critical weaknesses
  1. The proposed approach largely builds upon existing techniques in feedback-driven adversarial search, graph-based strategy optimization, and LLM-based reflection. Moreover, attack sequences are predetermined before each interaction, limiting the framework's ability to adapt its strategy to tutor responses within an ongoing conversation.
  2. The key comparisons against weight-only optimization and random-policy search are conducted on a single model, without strictly controlled computational budgets. Furthermore, the ablation studies do not sufficiently isolate the individual contributions of graph reflection, transition updates, and new primitive discovery.
  3. The overall ASR improvement is relatively modest with 4%, and the confidence interval for PI includes zero. Additionally, the final evaluation relies on only ten independent source questions, while the stability of graph evolution across independent runs remains unexplored.
Questions for authors
  1. Given that attack sequences are fixed before each interaction, could the authors clarify how the proposed framework fundamentally differs from existing adaptive multi-turn attack methods? Could incorporating response-conditioned strategy selection within individual conversations improve its effectiveness?
  2. Could the authors provide computationally matched comparisons across additional target models and conduct more fine-grained ablations to isolate the contributions of window reflection, transition updates, and primitive discovery?
  3. Given the limited number of independent test questions, could the authors evaluate the stability of the reported improvements across multiple independent graph-evolution runs with different random seeds?
Reviewer_R4PLOpenReview ↗
Rating
2
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

The paper proposes a benchmark that tests how well multimodal LLM tutors hold up against students who try to push them into bad behavior. It covers three kinds of failure: doing the student’s work for them, ignoring signs that a student is in distress, and accepting a wrong answer because of fake authority (for example, a fake message from a teacher). Attacks are built from 22 hand-written student strategies, combined into multi-turn conversations and paired with a generated image of homework or a chat screenshot. For each tutor model, a graph of strategies is updated over 100 rounds based on which combinations succeed, and an LLM judge decides whether an attack worked. Across eight models, the updated graphs raise the attack success rate from 49% to 53%.

Critical strengths
  1. A useful view of tutor failures, with good examples. Looking beyond answer leakage to student welfare and to giving in to fake authority is valuable, and each type of failure is clearly defined. The models rank quite differently on the three failure types, failures become more likely the longer the conversation goes, and some examples are striking: one tutor first computes the correct answer and then confirms a planted wrong answer because of a fake teacher message (Appendix I.3). Testing tutors over several turns and with images is a strong setting.
Critical weaknesses
  1. The attacks do not adapt during a conversation, and the gain from the “evolving” part is small. The paper motivates the benchmark with attacks that “adapt to a tutor’s prior responses”. However, the sequence of student strategies is fixed before the conversation starts, and the student messages are fixed template sentences (Appendix C.3). Learning only happens between conversations. Zhao et al. (2026), which the paper describes as using fixed strategies, actually include an attacker that reacts to the tutor’s replies. The main result is an increase of 4 percentage points, tested on only 10 problems per failure type; for answer leakage the increase is not significant. Each setting was run only once, so we do not know how much the results vary between runs, and the evolved graphs used many more tutor queries than the starting graph. The simpler search baselines were run on a single model and mostly do worse than the starting graph itself, which suggests that the variation between runs is about as large as the reported gain. Repeated runs and a baseline with the same query budget are needed.
  2. It is unclear whether the cases are realistic and the labels reliable. The images are generated, with clean and perfectly readable text (Figure 3). Real student photos and handwriting look very different, so it is unclear whether the findings transfer. The claim that these failures “are not marginal issues” rests on citations rather than on evidence of how often they happen in real tutoring. For the welfare category, a fixed rule frequently turns the judge’s “no failure” into a failure. Calling a student’s sign of distress an “attack” also mixes up student welfare with manipulation. All labels come from a single GPT-4o judge, and the human reviewers questioned its label in about one in eight of the cases they checked, which is a lot compared with a 4-point effect.
Questions for authors
  1. How does the evolved graph compare with an attacker that reacts to the tutor’s replies, such as the one in Zhao et al. (2026), given the same number of queries?
  2. Does the gain hold over several independent runs, and against a simple baseline with the same query budget?
  3. Have you tested real student materials, and does image quality matter?
  4. For the welfare category, how many failures come only from the fixed rule, and how well do human raters agree with the judge for each failure type?

Minor:

  • The judge sees the planned attack strategy, which may bias it towards finding a failure.
  • Starting from a graph without the hand-designed connections works better than the default (Table 18). What do these connections add?
  • Please define the “v1/v2” test cases and the “continuation chain” of runs (Section 4.2).
Reviewer_jxmAOpenReview ↗
Rating
2(+1)
Confidence
3
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

This paper introduces MESA-Bench, a multimodal, evolving student-attack benchmark for evaluating the integrity of multimodal large language models (MLLMs) deployed as educational tutors. The benchmark considers three dimensions of tutor integrity: pedagogical integrity (PI), which concerns inappropriate answer leakage; protective integrity (PrI), which concerns neglecting students' vulnerability or crisis cues; and epistemic integrity (EI), which concerns abandoning evidence-based judgments under unreliable evidence or social pressure.The proposed framework represents student attack strategies as a graph of reusable primitives and updates this graph using feedback from multi-turn student–tutor interactions. It incorporates multimodal educational cases, LLM-based evaluation with deterministic auditing, and feedback-driven graph evolution. Experiments are conducted on eight MLLMs across three integrity objectives.

Critical strengths

This is an interesting, exploratory, preliminary study of very ambitious goals. There is also broad model coverage and diagnostic analyses. The experiments include eight MLLMs with different capabilities. Beyond overall ASR, the paper reports cross-model strategy differences, graph evolution dynamics, and input-format sensitivity. The supplementary human audit, uncertainty analysis, and lightweight baseline comparisons are also useful efforts toward validating the proposed framework.

Critical weaknesses

I am uncomfortable with the limited benchmark scale and diversity. Although the paper targets tutor integrity in broad educational settings, the benchmark contains only 50 source questions, all drawn from GAOKAO mathematics, with only 10 distinct source questions used for final testing. The five repeated trials per question increase the number of interactions but do not substantially expand the diversity of evaluation scenarios. Moreover, the multimodal cases and student attacks are synthetically constructed. It remains unclear whether the observed vulnerabilities and evolving strategies generalize to other subjects, grade levels, educational contexts, or more realistic student interactions.

The effectiveness of graph evolution is not sufficiently established against strong baselines. The improvement over the seed graph is relatively modest, with an overall gain of 4 percentage points. In particular, the improvement for PI is only 2.25 percentage points, and its reported confidence interval includes zero. Comparisons with stronger adaptive attack generation methods, as well as multiple independent evolution runs, would provide more convincing evidence that the proposed graph-evolution design offers advantages beyond additional search and validation-based selection.

The extent of attack adaptivity is somewhat limited. A central motivation is that realistic student attacks should adapt to tutor responses. However, Section 3.2.2 states that the attack primitive sequence is fixed once selected, with subsequent student turns instantiating the predetermined primitives. Thus, the primary adaptation occurs across completed interaction trials rather than dynamically within an ongoing student–tutor conversation. This distinction is important because the proposed framework may not fully capture students who change their strategies immediately in response to a tutor's defenses. A more explicit comparison with response-conditioned within-conversation attacks would help clarify the contribution.

The specific contribution of multimodality remains unclear. Multimodal educational interactions are emphasized as a central motivation, but the input-format experiments in Table 4 show heterogeneous results, with text-only settings sometimes achieving substantially higher ASR than multimodal inputs. The authors appropriately acknowledge that the comparison is not a causal image ablation, since attack paths are not matched across input conditions. Consequently, it is difficult to determine whether the multimodal component introduces distinctive vulnerabilities or whether most observed failures are primarily driven by textual attack strategies. More controlled experiments would strengthen the multimodal claims.

Remaining concerns about evaluation reliability and the operational definitions of failure. The benchmark relies heavily on automated judgments of context-dependent educational behavior. Although the authors provide a human audit, only 87.3% of the 150 reviewed traces have machine labels accepted by both reviewers. Appendix D also reports substantial automated label revisions for PrI and EI on GLM-5V-Turbo, indicating that failure identification is nontrivial. Furthermore, the distinction between appropriate instructional assistance and answer leakage, or between ordinary emotional expression and actionable crisis cues, can depend on educational context. The strict tutor system prompt also imposes a particular instructional policy that may not represent all realistic tutoring applications. Further expert validation and sensitivity analysis would help establish that ASR consistently captures genuine tutor-integrity failures.

Unclear comparability of model-specific evolved evaluations. Each tutor model and integrity objective uses an independently evolved attack graph. While this design is suitable for identifying model-specific vulnerabilities, it also means that different models are evaluated under potentially different attack distributions. Consequently, differences in final ASR may reflect both tutor robustness and the effectiveness of the corresponding evolved attacker. A shared attack-pool evaluation or cross-model graph-transfer experiment would help distinguish these factors and strengthen the benchmark's value for comparing tutor models.

Questions for authors

Can the authors provide additional results on a larger or more diverse set of educational problems, or report results from multiple independent graph-evolution runs with different random seeds? How stable are the reported improvements, particularly for PI, where the confidence interval includes zero?

Could the authors compare MESA-Bench against stronger existing adaptive attack methods under matched target-model query budgets? In particular, how much of the improvement can be attributed separately to edge-weight updates, window reflection, new primitive discovery, and validation-based checkpoint selection?

since the primitive sequence is fixed before each interaction, how does the proposed method compare with an attacker that selects its next primitive based directly on the tutor's preceding response? Would such a mechanism provide additional gains beyond the current cross-trial graph evolution?

Could the authors conduct a controlled image ablation using matched source questions, student messages, and attack paths across multimodal and text-only conditions? This would help clarify whether images contribute independently to attack success and identify which types of visual evidence are most consequential.

Can the authors report human–automated judgment agreement separately for PI, PrI, and EI, along with the effects of human-corrected labels on the main ASR results? Additionally, could they evaluate the learned attack graphs on other tutor models or use a shared attack pool to assess whether the observed cross-model differences persist under comparable adversarial inputs?

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.