Authors: Unggi Lee, Jihoi Na, Yeil Jeong, Haeun Park, Yeonju Jang (2026)

Special-R1: RL for Special Education

What It Is

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.

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 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.