SPED-adaptive prompting shifts tutor behavior toward We-Do scaffolding, naturalistic language support, and participation-affirming feedback across three codebooks¶

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

Subclaims¶

q2 i? The DI We-Do move (lead) jumps from 9.1-14.3% in generic-prompted groups to 21.4-23.6% in SPED-prompted groups, with the I-Do modeling move shrinking by roughly a third. → Unggi Lee 2026 q2 i? Every disability-supporting EMT move multiplies in SPED-prompted variants (mand_model 2-5x, incidental 2-18x, expansion_recast up to 20x), and EMT affirmation rises from 22.1% in Base to 39.0% in SPED-adaptive. → 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? · design · r2

LLM-labeled turn-level Direct Instruction codebook analysis over N=11,103 tutor turns from the seven trained models, grouped as Base, Think, SPED-static, and SPED-adaptive. Reported with chi-square tests of independence (DI turn chi-square(15)=714.1, p<10-100, Cramer's V=0.146).

"The We-Do move (lead) jumps from 9.1-14.3% across the generic-prompt groups (Base, Think∗) to 21.4-23.6% across the SPED-prompted groups (SPED-static∗, SPED-adaptive∗), and the I-Do procedure-modeling move (model) shrinks by roughly a third"

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? · design · r2

Turn- and sentence-level Enhanced Milieu Teaching codebook analysis over N=60,745 sentences, reported with chi-square(18)=636.1, p<10-100, Cramer's V=0.138. Sentence-level EMT affirmation rises from 22.1% in Base to 39.0% in SPED-adaptive.

"Every disability-supporting EMT move multiplies in the SPED-prompted variants relative to generic-prompted baselines:mand_model 2-5×, incidental 2-18×, expansion_recast up to 20×, time_delay∼ 3×, env_arrange 2×"

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
• Human tutoring has a medium effect over no tutoring, and in one small study students learned as well when tutors only prompted them as when tutors also explained and gave feedback — related
• LLM scaffolding level is associated with the student's next dialogue act, and the association persists within broad request categories — 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
• Prompting a learning disability shifts feedback toward word- and sentence-level remediation, simplified vocabulary, and lowered expectations — related
• Iterative prompt refinement shifted a generative-AI tutor from tutorial-style explanation delivery toward dialogic Socratic support for K-12 beginners — related
• The AI's nonjudgmental conversational tone and adaptive prompting fostered risk-taking and productive struggle in a psychologically safe environment — related
• GRPO alignment collapses Mathematical-Knowledge-for-Teaching utterances and surges Socratic-style Exploratory questions in tutor output — related