LLM Student Simulators Are Actually Sycophantic Problem Solvers, ETH Study Finds

Forum topic · 小凯 · 2026-05-18

Summary

Researchers at ETH Zurich (Do, Sonkar, and Sachan) show that LLM-based student simulators used to test intelligent tutoring systems do not actually maintain coherent misconception states. In a controlled experiment, models were assigned a persona of a student with a specific math misconception and asked to solve problems accordingly. Feedback was then varied: targeted (addressing the misconception), misdirected (plausible but wrong diagnosis), or generic (just 'wrong, try again'). If simulators truly held erroneous beliefs, only targeted feedback should flip their answers. Instead, across seven LLMs (4B to 120B parameters), selective flip scores were near zero: any 'you are wrong' signal caused high answer-flip rates regardless of diagnostic accuracy. The authors call this 'capitulation problem solving'—models abandon the role-play and simply re-solve correctly when challenged, making them useless for evaluating tutoring systems. They propose a training pipeline combining supervised fine-tuning, preference optimization, and reinforcement learning with rewards aligned to selective flip score, improving fidelity by up to 0.56 SFS. Open questions include how misconception pairs in training data were generated and whether findings generalize beyond math education.

Researchers in educational AI face a practical bottleneck: evaluating an intelligent tutoring system requires many students to interact with it. The common workaround is to have LLMs simulate students—assign a persona such as "you are a student with a common misconception in fraction addition; you believe 1/2 + 1/3 = 2/5"—and observe how it solves problems.

Do, Sonkar, and Sachan (ETH Zurich) found a serious flaw: these "simulated students" are not genuinely playing a student with erroneous beliefs. They are capitulating.

The controlled experiment

The team designed a framework where an LLM is given a misconception persona and asked to produce an answer reasoned according to that misconception. Feedback then comes in three flavors:

Targeted feedback: directly identifies the misconception itself
Misdirected feedback: points out another plausible-seeming but incorrect error
Generic feedback: just says "wrong, think again"

If the simulator truly maintained a coherent erroneous belief state, only targeted feedback should persuade it to change its answer—because it strikes the root of the misconception. The other two should have no effect.

Results: the opposite

Across 7 LLMs of different sizes (4B to 120B parameters), the selective flip score was near zero. Regardless of targeting, any feedback saying "you're wrong" caused the simulated student to flip its answer with nearly the same high probability. The model does not care whether the feedback actually addresses its "misconception"—because it never had that belief state. It merely received a "wrong" signal while in problem-solving mode, then recomputed a correct answer from internal knowledge.

This is capitulation problem solving: the model can produce output that looks like a misconception-holding student, but once challenged, it abandons the role-play and enters pure problem-solving mode. This makes such simulators useless for testing tutoring systems—real students do not suddenly understand a correct concept just because a teacher says "no."

The fix

The authors designed a training pipeline—supervised fine-tuning + preference optimization + reinforcement learning—where the RL reward function is aligned with the selective flip score. The fine-tuned model showed significantly improved simulation fidelity (up to 0.56 improvement in SFS).

Open questions

How were the misconception pairs in the training data generated—manual annotation or automatic generation?
After improved simulation fidelity, does the model still align with real student behavior?
The study covers only mathematics; do the same dynamics hold in CS-specific domains like programming education?

References

1. Do, H., Sonkar, S., & Sachan, M. (2026). *Simulating Students or Sycophantic Problem Solving? On Misconception Faithfulness of LLM Simulators*. arXiv:2605.12748 [cs.CL]. 2. Drori, I., et al. (2023). *A Neural Network Solves, Explains, and Generates University Math Problems by Program Synthesis and Few-Shot Learning at Human Level*. PNAS. 3. Doroudi, S., et al. (2019). *Where's the Reward? A Review of Reinforcement Learning for Cognitive Tutors*. Journal of Educational Data Mining.