Special-R1: Reinforcement Learning for Special Education — Aligning LLM Tutors to Diverse Learners through Disability-Adaptive Training

Synthesis: Special-R1 is a framework that extends pedagogical reinforcement learning (RL) to special education. While prior RL-based tutor alignment methods targeted only generic math learners, Special-R1 explicitly models cognitive and communicative diversity across five disability profiles.

What It Is

How It Works

The framework has two core components:

1. Two-dimensional adaptive system prompt: Couples a difficulty-based support level (Scaffolding) with a disability-specific teaching style, forming a persona-aware prompt that guides the LLM tutor during multi-turn dialogue.
2. Persona-aware Thinking Reward: The judge rubric used to compute the training reward is conditioned on the learner's disability profile rather than a generic student. This shapes the tutor to produce responses that are helpful, safe, and appropriately challenging for each specific persona.

Key Results

• Evaluated on 690 multi-turn dialogues with diverse learner personas
• Generic baseline tutor achieved Fit score of 6.75/10
• Special-R1 full model raised Fit to 8.40 (+1.65)
• SPED-rubric Helpfulness improved from 0.720 to 0.768
• No degradation on out-of-domain OpenLearnLM Benchmark (8.53 vs 8.53)
• Ablations showed the Thinking Reward only becomes effective in combination with adaptive prompting

Critical Insight

Students with specific learning disabilities in mathematics remain underserved, suggesting a need for Multimodal AI extensions (visual aids, interactive diagrams) in future work.

Why It Matters

This is the first multi-turn pedagogical RL framework specifically targeting special education. It demonstrates that LLM tutors can be systematically aligned to support students with disabilities, improving both perceived helpfulness and pedagogical fit. The persona-conditioned reward rubric provides a replicable recipe for adapting RLHF-based tutor fine-tuning to diverse learner profiles.

Open Questions

• How well does Special-R1 generalize to disabilities beyond the five profiles modeled?
• Can the framework be extended to multimodal inputs for learners with specific math disabilities?
• What are the ethical considerations of disability-profiling in AI tutoring systems?

ring systems?

What this means for practice

• Software developers. Condition both the prompt and the reward on the learner's disability profile rather than either one alone: the full model raised judge-rated Fit from 6.75 to 8.40, and the ablations show the persona-aware Thinking Reward pays off only alongside adaptive Prompt Engineering.
• Software developers. Make out-of-domain checks part of the release gate: the full model stayed within 0.01 of the best variant on the generic OpenLearnLM benchmark (8.53), so a disability-conditioned tutor should be tested on generic material too.
• Software developers. Do not assume a text-only interface serves every profile: specific learning disability in mathematics remained the weakest profile, dropping below 7.8 on overall Fit, and its rubric calls for visual or manipulative representations a text-only tutor cannot deliver.
• Software developers. Evaluate on multi-turn dialogue with diverse learner personas (this study used 690 dialogues) rather than single-turn exchanges, since both the Fit gains and the SPED-rubric Helpfulness improvement (0.720 to 0.768) are defined over multi-turn tutoring.

Limitations

• Every student in the evaluation was a disability-conditioned LLM simulator; validating that judge-based gains transfer to real learners would require an IRB-approved study with special-education teachers.
• All main rewards and evaluation scores come from GPT-4o-mini, with judge bias bounded only by a 10% Gemma-3-27B cross-judge check.
• The accept/reject judge saturates in the Normal and SPED-static conditions, leaving the GRPO run little usable gradient outside the adaptive setting.
• All tutors are 7-8B models, so behavior at larger scales is untested, and only five disability profiles are modeled.

Citation

Lee, U., Na, J., Jeong, Y., Park, H., & Jang, Y. (2026). Special-R1: Reinforcement Learning for Special Education — Aligning LLM Tutors to Diverse Learners through Disability-Adaptive Training.