loss-function

该论文提出 LILIE（Language-Informed and Lossless Image Encoder），利用可逆归一化流编码器为文本到图像生成提供语言监督损失，在 GPIC 数据集上将 JiT 模型的 gFD 从 66.84 降至 53.01，相比纯像素损失显著提升保真度。

Researchers from Allora Foundation and Liverpool John Moores University introduce the CZAR (Composite Zero-Agnostic Return) loss, a convex, asymmetric, piecewise-quadratic objective tailored for financial log-return prediction that prevents models from collapsing to zero-forecast behavior, and show improved directional accuracy on BTC intraday returns with LightGBM.

This paper proposes Adaptive Margin Ordinal Loss (AMOL) to address center-class hedging in ordinal classification tasks, achieving improved performance over existing methods.

The paper proposes UASPL, a method that integrates predictive reliability into sample selection for self-paced learning using evidential neural networks, improving classification performance and interpretability.

The paper introduces AUF (Accept-Until-Fail), a simple modification to the cross-entropy loss for masked block drafters in speculative decoding that restricts supervision to the prefix up to the first predicted failure, improving average emitted length across benchmarks without changing inference.

Introduces Smooth Maximum Mean Discrepancy (SMMD), a loss function that aligns predicted numeric distributions with targets using kernel matching and graph-based smoothness, improving numerical prediction accuracy in LLMs across multiple tasks.

This paper proposes a confusion matrix-based graph construction method and a hybrid loss function for Graph Neural Networks to improve multi-site pollution prediction accuracy and interpretability, evaluated on real-world air pollution data.

This blog post explains the math behind gradient descent, the fundamental optimization algorithm used to train machine learning models, with a step-by-step numeric example and intuition.

This paper introduces Digit Entropy Loss (DEL), a novel loss function for numerical learning in large language models that reformulates entropy optimization to improve digit-level prediction accuracy and handle floating-point numbers, consistently outperforming existing methods on mathematical reasoning benchmarks.

This paper introduces D-PACE, a dynamic position-aware cross-entropy loss for training speculative decoding drafters that adaptively weights positions to improve acceptance length and inference speed, achieving consistent wall-clock speedups across benchmarks with minimal overhead.