ICML 2026

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

A reasoning-centric FP4 training framework that matches—and surpasses—BF16 accuracy on Large Reasoning Models, with up to 3.9× throughput on NVIDIA Blackwell.

Janghwan Lee1 Sihwa Lee1 Jinseok Kim2 Yongjik Kim2 Jieun Lim2 Jinwook Oh2 Jungwook Choi1
1Hanyang University    2Rebellions Inc., Republic of Korea
† Corresponding author
+15.8 pts
AIME‑120 accuracy over NVFP4 direct PTQ
65.94% vs. 50.13% PTQ baseline (W4A4KV4) — also surpasses BF16 full fine-tuning (65.46%)
3.9×
end-to-end throughput speedup
NVIDIA DGX Spark vs. BF16
3.1×
end-to-end throughput speedup
NVIDIA B200 vs. BF16
PTQ accuracy drop, QAT vs Full-FT vs ReQAT accuracy, and accuracy-throughput trade-off across FP4 deployment recipes
FP4 quantization hurts reasoning — and ReQAT recovers it. (a) NVFP4 W4A4KV4 post-training quantization (PTQ) causes large accuracy drops across DeepSeek‑R1‑distilled reasoning models. (b–c) Standard quantization-aware training (QAT) narrows the gap under MXFP4 W4A16 and NVFP4 W4A4KV4, but plateaus well below BF16 full fine-tuning (Full‑FT) even with a larger token budget — ReQAT closes, then surpasses it. (d) Across FP4 deployment recipes, ReQAT reaches the best accuracy–throughput trade-off: BF16‑level accuracy at FP4 speed.

Abstract

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× throughput speedup on NVIDIA DGX Spark and 3.1× on B200.

Why does FP4 reasoning fail?

Large Reasoning Models spend most of their inference budget on long chain-of-thought decoding, where repeated weight loading and a growing KV cache dominate cost. Microscaled FP4 formats (MXFP4, NVFP4) promise up to peak throughput on Blackwell-class hardware, but aggressively quantizing weights, activations, and the KV cache (W4A4KV4) causes severe reasoning accuracy loss that existing post-training quantization (PTQ) and quantization-aware training (QAT) fail to recover—even when QAT is given a much larger fine-tuning token budget.

The failure concentrates on low-entropy tokens. Grouping over 1.5M generated tokens by predictive entropy, we find low-entropy tokens are dominated by digits and symbolic operators, while high-entropy tokens are connective discourse phrases. Quantization flattens the predictive distribution at both—but the consequence differs sharply. Entropy-aware mixed-precision decoding (dynamically routing each token's prediction to BF16 or FP4 based on its entropy) and a controlled logit-noise injection experiment both show the same pattern: corrupting low-entropy predictions causes large drops in AIME accuracy, while corrupting high-entropy predictions barely matters.

Digging further, FP4 rarely flips the top-1 token at low-entropy positions—the model's argmax choice is largely preserved. Instead, it substantially inflates the tail probability mass assigned to non-top-1 alternatives, making confident, precise symbolic commitments (like the digit “4” or an operator) unexpectedly easy to mis-sample during autoregressive decoding, where a single wrong digit can cascade into a fully incorrect reasoning trace.

Token entropy semantics, predictive distribution flattening under quantization, entropy-aware mixed-precision decoding, logit-noise sensitivity, and tail-mass inflation at low-entropy tokens
(a) Low-entropy tokens are dominated by digits/operators; high-entropy tokens are discourse markers. (b) Quantization flattens the predictive distribution at both entropy levels. (c–d) Entropy-aware mixed-precision decoding: routing low-entropy predictions to BF16 recovers most of the accuracy lost under quantization; routing high-entropy predictions does not. (e) Logit noise injected at low-entropy positions causes large AIME‑2025 accuracy drops across models; high-entropy noise does not. (f) FP4 rarely flips the low-entropy top-1 token, but substantially inflates its tail-mass—confident predictions become easy to mis-sample.

ReQAT: a reasoning-centric FP4 training framework

Motivated by this insight, ReQAT explicitly targets low-entropy token failures with three complementary components, unified into a single training pipeline.

ReQAT overview: BF16 fine-tuning, Q-FIT calibration, and trace-aligned QAT with SEM loss

ReQAT overview. Starting from a pretrained LRM, ReQAT performs BF16 fine-tuning (Stage‑1), calibrates RoPE-consistent KV-cache transformations via Q‑FIT, then applies trace-aligned QAT (Stage‑2, TAQ) with the SEM auxiliary loss—computed on the same reasoning traces used in Stage‑1—to obtain an FP4 ReQAT model that matches or exceeds BF16 accuracy.

TAQ · Trace-Aligned QAT A two-stage recipe where Stage‑2 QAT revisits the same reasoning traces used during Stage‑1 BF16 fine-tuning, forcing quantization-aware updates to repeatedly act on the same low-entropy token decisions. A small 70M‑token Stage‑2 budget is enough to match BF16 fine-tuning accuracy.
SEM · Selective Entropy Minimization An auxiliary loss that sharpens predictive confidence only at low-entropy positions, using a soft, percentile-based weight rather than a hard mask—reinforcing exactly the confident, quantization-vulnerable symbolic commitments TAQ revisits.
Q‑FIT · Quantization-Friendly Init Before Stage‑2 QAT, jointly calibrates channel-wise pre-RoPE scaling (folded into projection weights, zero overhead) and post-RoPE shifting of the KV cache, adapting to each layer's outlier pattern to stabilize W4A4KV4 training.

TAQ in depth: why trace alignment matters

Using an entropy-change metric that tracks how much token entropy shifts from the base model during training, plain fine-tuning (FT) and plain QAT both induce changes mostly in high-entropy bins, leaving low-entropy bins essentially untouched. A two-stage FT + QAT procedure behaves differently: as Stage‑2 QAT proceeds, entropy changes increasingly appear in low-entropy bins—but only when Stage‑2 QAT revisits the same reasoning traces as Stage‑1. On different traces, this effect disappears.

Entropy change across baseline entropy bins for FT, QAT, and FT+QAT; normalized entropy for QAT vs FT+QAT with same/different traces
(a) FT and QAT mainly change high-entropy token predictions; FT+QAT increasingly changes low-entropy predictions as Stage‑2 QAT proceeds. (b) FT+QAT on the same reasoning trace yields markedly lower entropy at low-entropy tokens than QAT alone or FT+QAT on a different trace.

We confirm the mechanism by tracking the gradient contribution ratio of low-entropy tokens during QAT (the fraction of total embedding-gradient norm coming from low-entropy positions). Revisiting aligned traces consistently increases this ratio relative to QAT on misaligned traces—Stage‑2 QAT reallocates learning signal toward exactly the quantization-sensitive low-entropy positions, because the model already learned the reasoning structure during Stage‑1.

Gradient contribution ratio of low-entropy tokens during QAT, with and without trace alignment
Gradient contribution ratio of low-entropy tokens during QAT. TAQ increases gradient allocation toward vulnerable low-entropy tokens; misaligned traces dilute this effect.

Effect of trace alignment (MXFP4 W4A16, R1-Qwen-14B AIME accuracy across fine-tuning budgets).

MethodTrace Aligned140M210M280M350M
QAT59.8861.3561.0962.29
FT + QAT60.1059.8962.1962.60
FT + QAT61.1563.6565.0067.29

Two-stage QAT without trace alignment gains ≤1%; trace-aligned two-stage QAT (used by TAQ) improves accuracy by about 5%.

SEM in depth: soft weighting beats a hard mask

Trace alignment alone is not enough to fully recover FP4 reasoning accuracy: it determines where Stage‑2 QAT applies updates, but not how strongly confidence at low-entropy positions is reinforced. SEM adds an auxiliary entropy-minimization term, computed per token and weighted by wt, on top of the standard SFT objective:

LSEM = LSFT + λ · (1/T) ∑t wt Ht

The natural first instinct is a hard binary mask: fully minimize entropy below a threshold τ, do nothing above it. SEM instead uses a soft, ramped weight that only reaches zero at the minibatch's minimum entropy and fades linearly up to τ (set to the 75th percentile of the minibatch's entropy):

wt = max( 0,  1 − (Ht − Hmin) / (τ − Hmin + ε) )

The difference matters in practice: a hard mask penalizes every token below τ equally hard, including borderline tokens that are only mildly confident—over-sharpening them can destabilize training. The soft ramp instead concentrates the strongest pressure on the most deterministic tokens (e.g. the digit “4”) while easing off near the threshold, which is exactly where SEM should be cautious.

Soft weighting vs. hard binary mask (AIME‑90, MXFP4 W4A4 TAQ+Q‑FIT, R1-Qwen-14B). Bold = better at each budget.

Total Fine-Tuning TokensBinary MaskSoft Weighting
140M60.1461.81
210M63.1965.14

Replacing the hard mask with SEM's soft, percentile-based weighting alone is worth +1.7–2.0 points—at the same trace-aligned QAT budget.

Q‑FIT in depth: stabilizing the KV cache

TAQ alone largely recovers accuracy under W4A4, but performance drops sharply once the KV cache is also quantized (W4A4KV4). The challenge is that RoPE-paired channels can carry asymmetric outliers—so a single shared scaling factor is insufficient—while post-RoPE key magnitudes oscillate across tokens due to RoPE's rotational structure, making a fixed channel-wise shift suboptimal over long decoding sequences.

Q‑FIT jointly calibrates a channel-wise pre-RoPE scaling vector s (folded into projection weights, zero inference overhead) and a post-RoPE shifting vector m (fixed after calibration), selecting both by minimizing the distance between BF16 and KV4 attention outputs. Q‑FIT adapts to each layer's characteristics: where channels show asymmetric outliers with small token-wise variation, it relies mainly on shifting; where key magnitudes oscillate strongly across tokens, it relies mainly on scaling.

Pre/Post-RoPE key cache visualization and Q-FIT calibration results showing asymmetric outliers and token-wise oscillation
Pre-/Post-RoPE key cache visualization and Q‑FIT calibration. RoPE-paired channels exhibit asymmetric outliers (b–c), while post-RoPE magnitudes oscillate across tokens (d–e); Q‑FIT adapts scaling vs. shifting per layer accordingly, recovering most of the W4A4KV4 accuracy gap (a).

Results

AIME‑120 accuracy on R1‑Qwen‑14B as the total fine-tuning budget increases. ReQATT, ReQATTQ, and ReQATTQS denote TAQ only, TAQ+Q‑FIT, and the full method (+SEM).

Bit-PrecisionMethod140M210M280M350M
BF16Baseline56.83
FT63.7064.1765.4664.79
MXFP4
W4A16
Direct PTQ50.37
FT + PTQ54.3857.7156.8756.77
QAT59.8861.3561.0962.29
ReQATT61.1563.6565.0067.29
ReQATTQ61.3665.3266.0466.98
ReQATTQS65.0066.2567.0868.02
MXFP4
W4A4
Direct PTQ43.96
FT + PTQ45.2149.1749.4848.33
QAT54.5955.6754.0658.03
ReQATT56.1559.7959.4859.69
ReQATTQ59.4862.6064.2764.48
ReQATTQS59.6962.8164.4865.94
NVFP4
W4A4KV4
Direct PTQ50.13
FT + PTQ55.0055.8355.2155.73
QAT57.0957.6058.8658.23
ReQATT60.3260.4263.1363.12
ReQATTQ59.7963.4465.9465.21
ReQATTQS59.7964.2864.3765.63

marks the best accuracy within each bit-precision setting. ReQAT improves monotonically as TAQ, Q‑FIT, and SEM are added, and matches or exceeds the best BF16 FT accuracy (65.46%) across every FP4 setting—without increasing the total training budget.

Results on R1‑Llama‑8B under NVFP4 W4A4KV4 (350M-token budget) across benchmarks. Bold = best, underline = second-best.

Bit-PrecisionMethodGSM8KMATH‑500AIME‑120
BF16Baseline88.4990.0036.67
FT91.1592.1848.75
NVFP4
W4A4KV4
Direct PTQ86.4584.6223.13
FT + PTQ88.4288.5334.06
ReQATT89.3889.8038.34
ReQATTQ89.8690.7240.32
ReQATTQS89.8590.5341.85

On the harder AIME benchmark, SEM gives a clear additional gain (40.32→41.85), suggesting that reinforcing low-entropy confidence matters most for the hardest reasoning tasks.

Throughput on NVIDIA Blackwell

We evaluate end-to-end inference throughput with trtllm-bench on two NVIDIA Blackwell platforms—DGX Spark and B200—under bursty, batched serving (1K requests, 512-token prompts, batches up to 256, 8K/16K-token generations). NVFP4 reaches up to 3.93× (DGX Spark) and 3.13× (B200) speedup over BF16; the overhead Q‑FIT adds in ReQAT is small (4–5%), still delivering up to 3.90× and 3.05× speedup.

The gains come from two factors: larger batch sizes enabled by NVFP4's smaller weight and KV-cache memory footprint, and the compute efficiency of 4-bit GEMM. At long output lengths, BF16 cannot sustain large batch sizes due to KV-cache growth, while NVFP4 maintains a batch size of 256 up to 8K tokens—directly translating memory savings into throughput.

End-to-end throughput speedup on DGX Spark and B200, and BF16 vs NVFP4 memory breakdown
End-to-end throughput speedup for (a) DGX Spark and (b) B200, comparing native NVFP4 and ReQAT-trained NVFP4 against BF16. (c) Memory breakdown at batch size 256: BF16's KV cache growth caps the achievable batch size at long output lengths, while NVFP4's smaller footprint sustains batch 256 up to 8K tokens.

BibTeX

@inproceedings{lee2026reqat,
  title     = {{ReQAT}: Achieving Full-Precision Reasoning Accuracy with 4-bit Floating-Point Quantization-Aware Training},
  author    = {Lee, Janghwan and Lee, Sihwa and Kim, Jinseok and Kim, Yongjik and Lim, Jieun and Oh, Jinwook and Choi, Jungwook},
  booktitle = {Proceedings of the 42nd International Conference on Machine Learning (ICML)},
  year      = {2026},
  eprint    = {2606.15682},
  archivePrefix = {arXiv}
}