Meta-Reasoning Transformer
Self-correcting LLM Architecture
About the Project
The Meta-Reasoning Transformer (MRT) is a fine-tuning framework that extends any HuggingFace decoder-only model with three new components:
Learnable Scratch-Pad Tokens — 64 dedicated tokens appended to every sequence that give the model an explicit notepad for intermediate reasoning.
Critic Head — A lightweight 2-layer MLP (~4.2M params) that reads the scratch-pad and outputs a quality score. If reasoning quality is low, it triggers a repair pass.
Probabilistic Halting Gate — Uses a straight-through estimator so the discrete halting decision remains differentiable during training.
The result: models that iteratively self-improve within a single forward pass, significantly reducing hallucinations and inference cost compared to multi-pass Chain-of-Thought approaches.