Prompt style is the largest driver of disability-aware Fit gains, and the Thinking Reward becomes effective only in combination with adaptive prompting¶

Claim · All claims
Evidence · 1 study (2 entries) · causal r1–r2 · q2
Currency · newest study 2026 · not yet checked against its sources · not reviewed by a person

Subclaims¶

q2 i? Switching the prompt from Normal to SPED-static adds +1.08 on persona-aware Fit (the single largest effect), adaptivity adds +0.45, and activating the Thinking Reward in the adaptive condition adds a further +0.09 on Fit. → Unggi Lee 2026 q2 i? The Thinking Reward is near-inert in Normal and SPED-static conditions but becomes productive in the adaptive condition, where moving from M6 to M7 raises Total from 2.847 to 2.911 and lowers the leak rate from 11.2% to 8.3%. → Unggi Lee 2026 (2)

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? · causal · r1

Ablation decomposition across the seven trained configurations on SPED-Tutor-Mix, reported in the results section. The article concludes the prompt-style axis does most of the heavy lifting for persona-aware Fit while the Thinking Reward acts as a small multiplier requiring adaptive prompting.

"Switching the prompt from Normal to SPED-static (M2 vs. M4) adds+0.215on Total and +1.08on Fit, the single largest effect on both axes."

Unggi Lee 2026 (2)¶

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

q2 · i? · causal · r2

Results section comparison of adjacent model configurations on SPED-Tutor-Mix. The authors interpret this as the Thinking Reward needing the structural room opened by adaptive prompting to steer the reasoning distribution.

"the Thinking Reward is near-inert in the Normal and SPED-static conditions (M3 vs. M2, M5 vs. M4) but becomes productive in the adaptive condition: moving from M6 to M7 raises Total from 2.847 to 2.911, lowers the leak rate from 11.2% to 8.3%"

Discussion¶

Related Claims¶

• Specific learning disability in mathematics is the weakest persona profile across all GRPO-trained models — related
• 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
• SPED-adaptive prompting shifts tutor behavior toward We-Do scaffolding, naturalistic language support, and participation-affirming feedback across three codebooks — related
• Disability-aware training does not degrade out-of-domain educational knowledge on OpenLearnLM — related