iGEN
Visit IGEN World Explore IGEN Expo
EXPLORE UPGRADE PLANS
BREAKING
Werner Enterprises Posts Highest Revenue Per Truck Growth in One-Way Segment in a Decade CMA CGM and Stonepeak Launch United Ports LLC in $2.4 Billion Terminal Joint Venture UPS shift away from Amazon shows bigger payoff Lanesurf: 62% of Loads Get Vetted Carrier Offers Before Brokers Arrive India-China Border Trade Via Lipulekh Resumes Aug 1; China Permits 20 Traders Geopolitics Drives CMA CGM Q2 Profit Surge of 42% as Volumes and Rates Climb Benchmark Diesel Price Rises Third Week as Futures Plunge; Spread Hits Record Indian Government Limits Sugar Dealers to 400 Tonnes Stock Until November to Curb Hoarding Tenants signing longer leases for larger warehouses as 3PLs lock in capacity US stock market flat as S&P 500 and Dow barely move, Nasdaq slides over 1% on chip rout Werner Enterprises Posts Highest Revenue Per Truck Growth in One-Way Segment in a Decade CMA CGM and Stonepeak Launch United Ports LLC in $2.4 Billion Terminal Joint Venture UPS shift away from Amazon shows bigger payoff Lanesurf: 62% of Loads Get Vetted Carrier Offers Before Brokers Arrive India-China Border Trade Via Lipulekh Resumes Aug 1; China Permits 20 Traders Geopolitics Drives CMA CGM Q2 Profit Surge of 42% as Volumes and Rates Climb Benchmark Diesel Price Rises Third Week as Futures Plunge; Spread Hits Record Indian Government Limits Sugar Dealers to 400 Tonnes Stock Until November to Curb Hoarding Tenants signing longer leases for larger warehouses as 3PLs lock in capacity US stock market flat as S&P 500 and Dow barely move, Nasdaq slides over 1% on chip rout
Home ›› Technology ›› Ai ›› Llms ›› Researchers Identify Shrinkage Bias in LLM FP4 Pretraining, Propose UFP4 Recipe for Stability

Researchers Identify Shrinkage Bias in LLM FP4 Pretraining, Propose UFP4 Recipe for Stability

A new study from researchers on arXiv identifies 'Shrinkage Bias' in E2M1-based FP4 pretraining for large language models, a systematic error that accumulates across layers. The proposed UFP4 recipe, using uniform grids like E1M2/INT4, demonstrates lower BF16-relative loss degradation on models up to 124B parameters, urging hardware support for uniform 4-bit formats.

iG
iGEN Editorial
June 20, 2026
Researchers Identify Shrinkage Bias in LLM FP4 Pretraining, Propose UFP4 Recipe for Stability

FP4 training promises substantial reductions in memory and computation cost for LLM pretraining, yet current hardware paths and recipes—including NVIDIA Blackwell/Rubin-class systems and AMD MI350-series GPUs—remain centered on E2M1 data elements. A new study from researchers on arXiv identifies a fundamental limitation of that choice: non-uniform formats such as E2M1 inherently suffer from Shrinkage Bias, a systematic negative rounding error caused by the geometric asymmetry of their representable bins.

The Origin of Shrinkage Bias

The researchers show that Shrinkage Bias accumulates multiplicatively across layers and is amplified by the Random Hadamard Transform (RHT), providing a unified explanation for the training instability observed in existing E2M1-based FP4 recipes. In contrast, uniform grids (E1M2/INT4) bypass this grid-geometry error and better convert the improved bucket utilization from RHT into higher quantization quality.

UFP4: A Uniform 4-bit Recipe

Based on this finding, the researchers propose UFP4, a uniform 4-bit training recipe that applies RHT to all three training GEMMs while restricting stochastic rounding to dY alone. This approach leverages the benefits of uniform quantization while minimizing additional complexity.

Experimental Validation

The experiments covered Dense 1.5B, MoE 7.9B, and MoE 124B long-run pretraining. According to the study, UFP4 consistently achieves lower BF16-relative loss degradation than strong E2M1-based baselines, supported by scaling-law analysis and ablation studies. The following table summarizes the key results:

Model E2M1 Baseline Loss Degradation UFP4 Loss Degradation Improvement
Dense 1.5B Higher (not specified) Lower (not specified) Significant
MoE 7.9B Higher Lower Significant
MoE 124B Higher Lower Significant

Note: Exact numbers are not provided in the abstract.

Implications for Hardware Design

The paper suggests that future accelerators should support E1M2/INT4-style uniform 4-bit grids as first-class training primitives alongside E2M1. This recommendation aligns with the observed benefits of uniform formats in mitigating Shrinkage Bias and improving training stability.

For enterprise technology leaders, this research highlights a critical consideration when adopting low-precision training for large models. The choice of quantization format—E2M1 versus E1M2/INT4—directly impacts training stability and final model quality. As LLMs become integral to supply chain and logistics applications (e.g., demand forecasting, route optimization), ensuring efficient and stable pretraining will reduce operational costs and accelerate deployment timelines. Hardware vendors and cloud providers may need to adapt their roadmaps to include uniform 4-bit support, as advocated by the UFP4 recipe.


Sources:

Keep Reading

Recommended Stories

New Research Shows Pretraining Data Composition Can Engineer Neural Scaling Laws for Particle Physics Technology

New Research Shows Pretraining Data Composition Can Engineer Neural Scaling Laws for Particle Physics

A new arXiv paper demonstrates that neural scaling laws in particle physics can be engineered by adjusting pretraining data composition. The study shows that including more diverse and task-aligned synthetic data can shift scaling behavior to require more data rather than larger models, offering insights for efficient AI training.

July 8, 2026
RaBiT: Residual-Aware Binarization Training for Accurate and Efficient Large Language Models Technology

RaBiT: Residual-Aware Binarization Training for Accurate and Efficient Large Language Models

Researchers propose RaBiT, a quantization framework that resolves pathological feature co-adaptation in residual binarized LLMs. RaBiT delivers state-of-the-art 2-bit accuracy and 4.49x inference speed-up on an RTX 4090, rivaling hardware-intensive Vector Quantization methods.

June 16, 2026
SPRI: SVD-Partitioned Residual Initialization Boosts Data-Constrained MoE Upcycling for Multilingual Translation Technology

SPRI: SVD-Partitioned Residual Initialization Boosts Data-Constrained MoE Upcycling for Multilingual Translation

Researchers propose SPRI, a method that initializes Mixture-of-Experts (MoE) models from pretrained dense models using SVD-partitioned residuals. Evaluated on multilingual speech-to-text translation, SPRI achieves gains of 2.58 BLEU and 3.32 COMET over fine-tuned dense models, and outperforms prior MoE upcycling baselines by 3.39 BLEU and 4.34 COMET points.

June 16, 2026
New Research Reveals How Visual Tokens Evolve Inside Vision-Language Models Technology

New Research Reveals How Visual Tokens Evolve Inside Vision-Language Models

A new computer vision paper from arXiv investigates how visual tokens are integrated into large language models (LLMs) under two paradigms: in-context prompting and layer-wise injection. The authors find that visual tokens enter the LLM as 'disguised visual context' lacking linguistic structure, then evolve differently depending on the integration architecture. They show that attention allocation alone is insufficient, and performance depends on the quality of visual representations at each layer.

July 8, 2026