ReQAT: Achieving Full-Precision Reasoning Accuracy with 4-bit Floating-Point Quantization-Aware Training

Large Reasoning Models (LRMs) achieve strong problem-solving through long chain-of-thought, but their deployment is constrained by the high cost of full-precision inference and growing KV cache footprints. Microscaled FP4 formats enable efficient FP4 deployment; however, fully quantizing weights, activations, and KV caches (W4A4KV4) causes severe reasoning degradation that existing PTQ and QAT fail to recover. We identify that FP4 failures concentrate on low-entropy tokens—precise symbolic commitments such as digits and operators—where quantization noise inflates sampling errors that cascade through reasoning traces. Based on this insight, we propose ReQAT, a reasoning-centric FP4 training framework with three components: (i) Trace-Aligned QAT (TAQ), which revisits identical reasoning traces to focus updates on critical low-entropy decisions; (ii) Selective Entropy Minimization (SEM), which reinforces confidence at low-entropy positions; and (iii) Q-FIT, a quantization-friendly initialization that jointly calibrates RoPE-consistent KV cache transformations to stabilize QAT. Under the same training budget, ReQAT not only recovers but surpasses BF16 fine-tuning accuracy, while delivering up to $3.9\times$ throughput speedup on NVIDIA DGX Spark and $3.1\times$ on B200. The project repository is available at https://github.com/aiha-lab/ReQAT.

Large reasoning models can solve complex math and logic problems by generating long chains of reasoning steps. However, running these models is expensive because they require large amounts of memory and computation, especially when deployed at scale. Recent hardware supports highly compressed 4-bit computation, which can greatly reduce these costs, but existing compression methods often cause severe reasoning failures. In this work, we study why reasoning models break under aggressive 4-bit quantization. We find that the main failures occur at highly confident predictions such as numbers and mathematical operators. Even when the correct answer remains the model’s top prediction, quantization noise can slightly increase the probability of sampling an alternative token. In reasoning tasks, accidentally generating a single incorrect digit or operator can derail the entire chain of reasoning and lead to a wrong final answer. Based on this insight, we introduce ReQAT, a new training framework that helps quantized models repeatedly reinforce these vulnerable reasoning decisions during training. Our method allows fully quantized 4-bit reasoning models to match or even exceed the accuracy of full-precision models while significantly improving inference efficiency. On modern NVIDIA hardware, ReQAT achieves up to 3.9× higher throughput, making large reasoning models substantially more practical for real-world deployment.