EduFrameTrap: leaderboard

EduFrameTrap (2026) tests whether LLM tutors keep correcting a student misconception when the student pushes back, across math, physics, economics, chemistry, biology and computer science. The score is the sycophancy rate in percent, the share of tutor replies that give in to the misconception after pressure; lower is better.

Metric: Sycophancy rate (%): share of usable dialogues whose post-pressure tutor reply validates the student misconception (context-switch, authority, face-saving or direct endorsement), on the EduFrameTrap test split (252 misconception trap families in math, physics, economics, chemistry, biology and computer science, each instantiated at three student-confidence levels and three pressure modes), four-turn tutoring dialogues judged on the tutor's post-pressure reply by GPT-5.2 and Claude Sonnet 4.5 with every judge disagreement adjudicated by the authors; lower is better. Source: arxiv.org. Saturation forecast: Around February 2027. 2 models tracked. Updated 7 Oct 2026.

Top models

#ModelScoreOverall rank
1Claude Sonnet 4.514#261
2GPT-5.214.2#185

No result here: #1 Claude Fable 5.2, #3 Gemini 4 Neon, #4 GPT-6.2 Astra.

Interactive version: theaggregate.ai/benchmark?slug=eduframetrap · How It Works · Data refreshed daily, snapshot 2026-11-21.