acceptodds
Under review as a conference paper at ICLR 2027

RedTeachingBench : A Curriculum-Grounded Red Teaming Benchmark for Teachers using LLMs

Abstract

Safety benchmarks for Large Language Models (LLMs) predominantly measure the generation of hazards such as sexual content, violent expression, or hate speech. These categories are essential for safety but too generic to assess LLM reliability for high-school teachers within a given national context. We introduce RedTeachingBench, a curriculum-grounded benchmark to adversarially test small LLMs, which are particularly relevant in educational context as they comply with the requirements of national sovereignty, confidentiality, and model frugality. Our benchmark tests whether LLMs elaborate on false premises, creating a misinformation risk for students. The benchmark is built following two methodological steps: a) with the help of volunteering French teachers, we extract key knowledge in six disciplines for students between 12 and 15; b) for each of 500 knowledge items, we generate two kinds of false-premise questions through minimal semantic corruption, distinguishing plausible local errors from overtly absurd substitutions. We evaluate RedTeachingBench on the generation of flashcards. Initial tests on small LLMs show substantial failures: models achieving a mean of 86.7% safe-answer rates on a standard red-teaming benchmark as AdvBench correctly flag only 41% of our examples, and fully reject only 7.9%. However, results differ significantly with model alignment and reasoning capabilities of LLMs. It is also noteworthy that reasoning models demonstrate different performance on the two kinds of false-premise questions. These findings provide preliminary evidence that our curriculum-grounded adversarial benchmark measures a distinct robustness property against misinformation, which is relevant for real-life deployment in the educational sector.

Primary area:
alignment, fairness, safety, privacy, and societal considerations
Keywords:
AI in education, adversarial testing, false-premise questions, LLM safety benchmark

Reviews

Among:
Shape:
1234this paper, Bottom 1%
Figure 1. Mean review rating of the reviewed papers at ICLR 2027. This paper averages 1.00 over 3 reviews, above 0% of the others.
Reviewer_953JOpenReview ↗
Rating
1
Confidence
2
Informativeness
2
Presentation
2
Soundness
2
Read review
Paper summary

This paper introduces RedTeachingBench, a benchmark for evaluating how small language models respond to false premises in teacher requests for educational flashcards. It contains 500 curriculum items spanning six subjects in French lower-secondary education. The main evaluation covers 22 open-weight models and compares their behavior with AdvBench and FalseQA.

Critical strengths

The paper focuses on teachers preparing instructional material, where an apparently legitimate request can contain a factual mistake. Grounding the evaluation in a specific curriculum provides a concrete use case and connects benchmark content to knowledge students are expected to acquire.

Critical weaknesses

The comparisons do not isolate the contribution of curriculum grounding. FalseQA uses a plain question-answering instruction, whereas RedTeachingBench adds a teacher persona, a flashcard task, and an instruction to preserve the exact question. Differences therefore conflate content with prompting and output-format requirements. The authors acknowledge that cross-benchmark metrics differ, but these comparisons still underpin the claimed distinct robustness property.

Judge reliability is insufficiently established. The validation primarily consists of agreement between two model judges and manual inspection of 50 disagreement cases. This does not measure accuracy on a representative sample or detect errors shared by both judges. In addition, the judges are themselves among the evaluated models.

The primary metric is not well aligned with the stated educational objective. Section 4.2 treats answers that preserve the question but correct its premise as non-failures, yet the prominent 86.7% versus 7.9% comparison excludes these answers from RedTeachingBench resistance. Although the broader figure of approximately 41% is also reported, the emphasis on 7.9% can overstate the failure to identify erroneous premises.

Presentation quality need inspection. Figure 1 is extremely small despite substantial surrounding whitespace, and labels in Figures 2 and 3 are difficult to read. Table 7 extends beyond the main text block. These issues should be corrected to make the manuscript readable.

Questions for authors

Could you evaluate both FalseQA and RedTeachingBench under matched prompting conditions, including plain question answering and the same teacher/flashcard wrapper?

How do models perform on the corresponding true-premise questions? Could you report over-rejection rates and false-premise performance restricted to items whose underlying knowledge each model demonstrates on the uncorrupted version?

Reviewer_PJ6wOpenReview ↗
Rating
1
Confidence
3
Informativeness
1
Presentation
3
Soundness
3
Read review
Paper summary

This paper introduced a new benchmark datasets with a focus on educational content for children (lower-secondary education as mentioned in the paper). The methodology is well grounded as this work is done with teachers. Experiments are run on small language models (mainly instruct but also some reasoning models).

Critical strengths

I really like the methodology adopted in this paper. I think having teachers and educators in the first line of development for this kind of technology is a great idea. It is my first time reading a paper - at an ML conference/not HCI - that has this "broader" focus. I also appreciate the focus on a language that is not English, for once.

Critical weaknesses

I fail to understand the main premises of this work.

  1. The simple fact that small models are worse at knowledge recall make it obvious that they are bad at this particular task. We could get into the sovereignty aspect but again, I feel like results might be expected here: it is well known that these models are not good for knowledge intensive tasks (they are also bad at SimpleQA-like tasks) - this is why I don’t think we should use small models for educational tasks (this sounds like my opinion, but I'd argue that it would be high risk to use small LLMs to create educational content).

  2. In addition to the above, I do also believe that most of the datasets we have tend to cover a large space of the very same recall set (e.g., false refusal, misleading information) and I thus not think the focus on educational content is significant enough for this type of work; I don't want to dismiss the work done as I think the dataset as value in specific context, I am just arguing that, alone, it's probably not enough for a full paper.

  3. In any case, doing this kind of work without any knowledge base retrieval/RAG system (assuming I understood evaluation correctly, please do correct me if I am wrong) is kind of a misinterpretation of how these models - even small ones - should be used in practice to reduce risks.

All in all I think the contribution here is, at the best of my understanding, limited, both in terms of dataset provider and of experimental coverage.

Questions for authors
  • I might be misremembering this but isn't AdvBench focused on harm? is this the right data to compare to?
Reviewer_dmoqOpenReview ↗
Rating
1
Confidence
2
Informativeness
1
Presentation
2
Soundness
1
Read review
Paper summary

The paper introduces RedTeachingBench, a false-premise benchmark grounded in the French lower-secondary curriculum. From 500 curriculum items in six disciplines, each item is turned into a "false" question (a close semantic substitution, e.g., "Liberty, Inequality, Fraternity") and an "absurd" question (a distant substitution, e.g., "Liberty, Broccoli, Fraternity"). Each question is wrapped in a teacher request to generate a study flashcard while keeping the exact question. Responses of 22 small open models (from 29 evaluated) are classified by small LLM judges into does_not_accept, plays_along_question and plays_along_fully.

Critical strengths
  1. This type of benchmark development is popular nowadays and the motivation is practical: small, often sovereign models are being deployed to support teachers, and a model that elaborates on a teacher's factual error can spread misinformation to students.
  2. Grounding the benchmark in an official curriculum, and separating plausible false premises from absurd ones, is a sensible design. I especially like that the benchmark is developed from official curriculum.
  3. The data, model outputs and judge verdicts are released and online available.
Critical weaknesses
  1. The benchmark does not separate a model's knowledge from a model's "personality". A model can fail either because it does not know the correct fact, or because it knows the fact but does not dare to contradict the user (a compliance or sycophancy tendency). These two failures have very different remedies, but the benchmark kinda mixes them. Several observations in the paper suggest that disposition plays a large role: models from the same family differ enormously (gemma-1.1-7b-it 34.9% vs gemma-3-4b-it 0.1%), and reasoning models often flag the error in the answer but still keep the flawed question. A simple control, such as asking each model the original question or whether the corrupted statement is true, would make it possible to attribute failures to missing knowledge or to compliance. Without this, the claim that failures are interpretable is not supported. In short, calibration needed to inteprete the results and disentangle the two different contributing factors.
  2. "Minimal semantic corruption" is the central methodological claim but is not defined or validated. The paper does not specify how the "semantically critical element" is identified, what makes a substitution "close" or "distant", or whether the false and absurd versions must modify the same span. Some quantitative checks (e.g., edit distance, semantic similarity between the original and the substitute, inter-annotator agreement on the false/absurd labels) would help, or some justification. Otherwise, this claim is very vague.
  3. The contribution is incremental. False-premise benchmarks already exist many, and the main novelty is the specialization to one national curriculum and one task. The paper reports findings but offers no suggestions on what to do about them: there are no mitigation experiments (e.g., a system prompt asking the model to check premises, fine-tuning or preference training on such data? ) and no concrete recommendations for model developers or deployers. In my opinion the paper would fit better in a datasets and benchmarks track, if any.
Questions for authors
  1. For each item, does the model know the correct fact when asked directly? What fraction of failures occur on items the model answers correctly in isolation?
  2. How exactly were the critical element and the false and absurd substitutes chosen, and by whom? Can the authors report a quantitative check that false and absurd versions modify the same element and differ in semantic distance as intended?
  3. How do the results change without the instruction "keep the exact question", and with other teacher tasks?
  4. Do results differ between human, mixed and synthetic items? (this would be interesting to check)
  5. What is the agreement between the LLM judges and human annotators on a labelled sample? (if possible to provide)

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

Related papers

Loading the map…

Discussion (0)

Sign in to comment.