A Verifier-Guided Explainable Reasoning Framework with Gold-Anchored QLoRA, Task-Aware Mixture-of-Experts, and Group-Relative RLVR
A research paper proposes a verifier-guided explainable reasoning framework for educational question answering, combining gold-anchored QLoRA, task-aware symbolic routing, and group-relative RLVR. The framework uses Qwen2.5-3B-Instruct adapted with field-weighted QLoRA. A lightweight router assigns logic problems to a FOL/Z3 verifier and physics problems to a formula- and unit-aware symbolic solver. Verifier feedback supports candidate evaluation, self-revision, and reward construction during RLVR. Candidate responses are evaluated on three dimensions: P1 (answer correctness), P2 (evidence or unit consistency), and P3 (reasoning depth and explainability). On 438 held-out examples, RLVR increases P3 from 50.68% to 72.20%, while hybrid P1 remains approximately stable at 55.94%.
The paper introduces a framework to improve explainable reasoning in LLMs for educational question answering. It uses Qwen2.5-3B-Instruct with gold-anchored QLoRA, a task-aware router to symbolic verifiers (FOL/Z3 for logic, formula/unit solver for physics), and group-relative RLVR. Evaluation on 438 held-out examples shows RLVR boosts reasoning depth and explainability (P3) from 50.68% to 72.20%, while answer correctness (P1) remains stable at 55.94%.
The framework integrates verifier feedback into RLVR reward construction, enabling self-revision and candidate evaluation. Task-aware routing to symbolic solvers (FOL/Z3 and physics formula/unit checker) provides grounded verification. Gold-anchored QLoRA uses field-weighted supervision from authoritative answers. Inference uses gold-free self-consistency and optional question-only physics verifier for conservative correction.
This research targets educational AI, where explainability and verifiability are critical. The approach could improve trust in AI tutoring systems by ensuring reasoning steps are grounded and verifiable. The use of small models (3B) with QLoRA suggests cost-effective deployment for specialized domains.
For educational technology companies, this framework could enhance AI-powered tutoring and assessment tools by providing verifiable explanations, reducing hallucination, and improving student trust. The use of a 3B model with QLoRA indicates lower training and inference costs compared to larger models.
Potential next steps include scaling to larger models, extending to other domains (e.g., math, coding), and integrating with production educational platforms. The framework's verifier-guided RLVR could inspire more robust reasoning pipelines in industry.