Disability-aware training does not degrade out-of-domain educational knowledge on OpenLearnLM¶

Claim · All claims
Evidence · 1 study · design r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person · recheck by 2029: tool and model accuracy changes quickly

Subclaims¶

q2 i? All thinking-enabled models cluster in the 8.51-8.54 range on the OpenLearnLM overall average, and M7 reaches 8.53, essentially tied with the best variant, indicating the SPED axis does not meaningfully trade off broad educational knowledge. → Unggi Lee 2026

Evidence¶

Unggi Lee 2026¶

Unggi Lee, Jihoi Na, Yeil Jeong, Haeun Park, Yeonju Jang. (2026). Reinforcement Learning for Special Education: Aligning LLM Tutors to Diverse Learners through Disability-Adaptive Training. https://arxiv.org/abs/2605.30670 doi:10.48550/arxiv.2605.30670

q2 · i? · design · r2

Out-of-domain evaluation on the OpenLearnLM benchmark (N=3,714) across Skills, Content, Pedagogical, and Attitude categories, scored by LLM judge or accuracy. The article notes this is consistent with prior findings that dialogue-based RL preserves base-model knowledge.

"The SPED axis does not meaningfully trade off broad educational knowledge: M7 reaches 8.53, essentially tied with the best variant (M5 / M6 at 8.54) and 0.77 points above M1."

Discussion¶

Related Claims¶

• The full Special-R1 model raises persona-aware Fit from 6.75 to 8.40 and SPED-rubric Helpfulness from 0.720 to 0.768 over a non-disability-aware baseline — related
• Prompt style is the largest driver of disability-aware Fit gains, and the Thinking Reward becomes effective only in combination with adaptive prompting — related
• Practical educational skill does not follow from knowledge scores: Claude-Opus-4.5 achieves the highest Skills average (8.82) despite the lowest Content Knowledge — related
• Knowledge, Skills, and Attitude scores are largely independent across frontier LLMs, with Knowledge–Skills correlation r=−0.51 and Knowledge–Attitude r=−0.63 — related