Research map / Adversarial robustness and security / Attacks on text and graphs
EduGuardBench: A Holistic Benchmark for Evaluating the Pedagogical Fidelity and Adversarial Safety of LLMs as Simulated Teachers
Yilin Jiang, Mingzi Zhang, Xuanyu Yin, Sheng Jin, Suyu Lu, Zuocan Ying, Zengyi Yu, Xiangjie Kong
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use reasoning-oriented models instead of non-reasoning models | Teacher SP-LLMs · EduGuardBench SATA | RFS: 0.663 -> 0.723 | Improved |
| use reasoning models instead of non-reasoning models | Teacher SP-LLMs · EduGuardBench SATA | Inclusion rate (%): 30.46 -> 22.73 | Improved |
| use reasoning models instead of non-reasoning models | Teacher SP-LLMs · EduGuardBench SATA | Omission rate (%): 6.61 -> 10.08 | Hurt |
| scale Qwen2.5 to 72B parameters | Qwen2.5-72B · EduGuardBench SATA | RFS: RFS 0.56, inclusion 40.53% | No clear effect |
| increase Qwen3 size from 8B to 32B (ablation) | Qwen3 series · EduGuardBench adversarial | ASR (%): 60.4 -> 75.2 | Hurt |
| increase Qwen3 size from 32B to 235B (ablation) | Qwen3 series · EduGuardBench adversarial | ASR (%): 75.2 -> 70.0 | Improved |
| reasoning mode versus non-reasoning in Qwen3 pairs (ablation) | Qwen3 series · EduGuardBench SATA | Inclusion rate: average 5.23 percentage point reduction | Improved |
| reasoning models on Idea Provision scenario | Teacher SP-LLMs · EduGuardBench SATA (Idea Provision) | error rate: Cohen's d = 1.000, p = 0.0119 | Improved |
| reasoning models on Error Correction scenario | Teacher SP-LLMs · EduGuardBench SATA (Error Correction) | performance | Hurt |
Closest papers by meaning
- Heterogeneous Adversarial Play in Interactive Environments (NeurIPS 2025 · similarity 0.86)
- OpenDeception: Learning Deception and Trust in Human–AI Interaction via Multi-Agent Simulation (ICML 2026 · similarity 0.85)
- AI, Ethics, and Education: The Pioneering Path of Sidekick Academy (AAAI 2024 · similarity 0.85)
- Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators (NeurIPS 2026 · similarity 0.84)
- DeceptionBench: A Comprehensive Benchmark for AI Deception Behaviors in Real-world Scenarios (NeurIPS 2025 · similarity 0.84)
- LieCraft: A Multi-Agent Framework for Evaluating Deceptive Capabilities in Language Models (AAAI 2026 · similarity 0.83)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.
