Low-Precision Training of Large Language Models: Methods, Challenges, and Opportunities

Large language models (LLMs) have achieved impressive performance across various domains. However, the substantial hardware resources required for their training present a significant barrier to efficiency and scalability. To mitigate this challenge, low-precision training techniques have been widely adopted, leading to notable advancements in training efficiency. Despite these gains, low-precision training involves several components$\unicode{x2013}$such as weights, activations, and gradients$\unicode{x2013}$each of which can be represented in different numerical formats. The resulting diversity has created a fragmented landscape in low-precision training research, making it difficult for researchers to gain a unified overview of the field. This survey provides a comprehensive review of existing low-precision training methods. To systematically organize these approaches, we categorize them into three primary groups based on their underlying numerical formats, which is a key factor influencing hardware compatibility, computational efficiency, and ease of reference for readers. The categories are: (1) fixed-point and integer-based methods, (2) floating-point-based methods, and (3) customized format-based methods. Additionally, we discuss quantization-aware training approaches, which share key similarities with low-precision training during forward propagation. Finally, we highlight several promising research directions to advance this field. A collection of papers discussed in this survey is provided in https://github.com/Hao840/Awesome-Low-Precision-Training.

AnesBench: Multi-Dimensional Evaluation of LLM Reasoning in Anesthesiology

The application of large language models (LLMs) in the medical field has gained significant attention, yet their reasoning capabilities in more specialized domains like anesthesiology remain underexplored. In this paper, we systematically evaluate the reasoning capabilities of LLMs in anesthesiology and analyze key factors influencing their performance. To this end, we introduce AnesBench, a cross-lingual benchmark designed to assess anesthesiology-related reasoning across three levels: factual retrieval (System 1), hybrid reasoning (System 1.x), and complex decision-making (System 2). Through extensive experiments, we first explore how model characteristics, including model scale, Chain of Thought (CoT) length, and language transferability, affect reasoning performance. Then, we further evaluate the effectiveness of different training strategies, leveraging our curated anesthesiology-related dataset, including continuous pre-training (CPT) and supervised fine-tuning (SFT). Additionally, we also investigate how the test-time reasoning techniques, such as Best-of-N sampling and beam search, influence reasoning performance, and assess the impact of reasoning-enhanced model distillation, specifically DeepSeek-R1. We will publicly release AnesBench, along with our CPT and SFT training datasets and evaluation code at https://github.com/MiliLab/AnesBench.

Dual-Splitting Conformal Prediction for Multi-Step Time Series Forecasting

Time series forecasting is crucial for applications like resource scheduling and risk management, where multi-step predictions provide a comprehensive view of future trends. Uncertainty Quantification (UQ) is a mainstream approach for addressing forecasting uncertainties, with Conformal Prediction (CP) gaining attention due to its model-agnostic nature and statistical guarantees. However, most variants of CP are designed for single-step predictions and face challenges in multi-step scenarios, such as reliance on real-time data and limited scalability. This highlights the need for CP methods specifically tailored to multi-step forecasting. We propose the Dual-Splitting Conformal Prediction (DSCP) method, a novel CP approach designed to capture inherent dependencies within time-series data for multi-step forecasting. Experimental results on real-world datasets from four different domains demonstrate that the proposed DSCP significantly outperforms existing CP variants in terms of the Winkler Score, achieving a performance improvement of up to 23.59% compared to state-of-the-art methods. Furthermore, we deployed the DSCP approach for renewable energy generation and IT load forecasting in power management of a real-world trajectory-based application, achieving an 11.25% reduction in carbon emissions through predictive optimization of data center operations and controls.

Retrieval-Augmented Perception: High-Resolution Image Perception Meets Visual RAG

High-resolution (HR) image perception remains a key challenge in multimodal large language models (MLLMs). To overcome the limitations of existing methods, this paper shifts away from prior dedicated heuristic approaches and revisits the most fundamental idea to HR perception by enhancing the long-context capability of MLLMs, driven by recent advances in long-context techniques like retrieval-augmented generation (RAG) for general LLMs. Towards this end, this paper presents the first study exploring the use of RAG to address HR perception challenges. Specifically, we propose Retrieval-Augmented Perception (RAP), a training-free framework that retrieves and fuses relevant image crops while preserving spatial context using the proposed Spatial-Awareness Layout. To accommodate different tasks, the proposed Retrieved-Exploration Search (RE-Search) dynamically selects the optimal number of crops based on model confidence and retrieval scores. Experimental results on HR benchmarks demonstrate the significant effectiveness of RAP, with LLaVA-v1.5-13B achieving a 43% improvement on $V^*$ Bench and 19% on HR-Bench.

ELBA-Bench: An Efficient Learning Backdoor Attacks Benchmark for Large Language Models

Generative large language models are crucial in natural language processing, but they are vulnerable to backdoor attacks, where subtle triggers compromise their behavior. Although backdoor attacks against LLMs are constantly emerging, existing benchmarks remain limited in terms of sufficient coverage of attack, metric system integrity, backdoor attack alignment. And existing pre-trained backdoor attacks are idealized in practice due to resource access constraints. Therefore we establish $\textit{ELBA-Bench}$, a comprehensive and unified framework that allows attackers to inject backdoor through parameter efficient fine-tuning ($\textit{e.g.,}$ LoRA) or without fine-tuning techniques ($\textit{e.g.,}$ In-context-learning). $\textit{ELBA-Bench}$ provides over 1300 experiments encompassing the implementations of 12 attack methods, 18 datasets, and 12 LLMs. Extensive experiments provide new invaluable findings into the strengths and limitations of various attack strategies. For instance, PEFT attack consistently outperform without fine-tuning approaches in classification tasks while showing strong cross-dataset generalization with optimized triggers boosting robustness; Task-relevant backdoor optimization techniques or attack prompts along with clean and adversarial demonstrations can enhance backdoor attack success while preserving model performance on clean samples. Additionally, we introduce a universal toolbox designed for standardized backdoor attack research, with the goal of propelling further progress in this vital area.

Merging Models on the Fly Without Retraining: A Sequential Approach to Scalable Continual Model Merging

Deep model merging represents an emerging research direction that combines multiple fine-tuned models to harness their specialized capabilities across different tasks and domains. Current model merging techniques focus on merging all available models simultaneously, with weight interpolation-based methods being the predominant approaches. However, these conventional approaches are not well-suited for scenarios where models become available sequentially, and they often suffer from high memory requirements and potential interference between tasks. In this study, we propose a training-free projection-based continual merging method that processes models sequentially through orthogonal projections of weight matrices and adaptive scaling mechanisms. Our method operates by projecting new parameter updates onto subspaces orthogonal to existing merged parameter updates while using an adaptive scaling mechanism to maintain stable parameter distances, enabling efficient sequential integration of task-specific knowledge. Our approach maintains constant memory complexity to the number of models, minimizes interference between tasks through orthogonal projections, and retains the performance of previously merged models through adaptive task vector scaling. Extensive experiments on CLIP-ViT models demonstrate that our method achieves a 5-8% average accuracy improvement while maintaining robust performance in different task orderings.

Leveraging Metamemory Mechanisms for Enhanced Data-Free Code Generation in LLMs

Automated code generation using large language models (LLMs) has gained attention due to its efficiency and adaptability. However, real-world coding tasks or benchmarks like HumanEval and StudentEval often lack dedicated training datasets, challenging existing few-shot prompting approaches that rely on reference examples. Inspired by human metamemory-a cognitive process involving recall and evaluation-we present a novel framework (namely M^2WF) for improving LLMs' one-time code generation. This approach enables LLMs to autonomously generate, evaluate, and utilize synthetic examples to enhance reliability and performance. Unlike prior methods, it minimizes dependency on curated data and adapts flexibly to various coding scenarios. Our experiments demonstrate significant improvements in coding benchmarks, offering a scalable and robust solution for data-free environments. The code and framework will be publicly available on GitHub and HuggingFace.

MambaHSI: Spatial-Spectral Mamba for Hyperspectral Image Classification

Transformer has been extensively explored for hyperspectral image (HSI) classification. However, transformer poses challenges in terms of speed and memory usage because of its quadratic computational complexity. Recently, the Mamba model has emerged as a promising approach, which has strong long-distance modeling capabilities while maintaining a linear computational complexity. However, representing the HSI is challenging for the Mamba due to the requirement for an integrated spatial and spectral understanding. To remedy these drawbacks, we propose a novel HSI classification model based on a Mamba model, named MambaHSI, which can simultaneously model long-range interaction of the whole image and integrate spatial and spectral information in an adaptive manner. Specifically, we design a spatial Mamba block (SpaMB) to model the long-range interaction of the whole image at the pixel-level. Then, we propose a spectral Mamba block (SpeMB) to split the spectral vector into multiple groups, mine the relations across different spectral groups, and extract spectral features. Finally, we propose a spatial-spectral fusion module (SSFM) to adaptively integrate spatial and spectral features of a HSI. To our best knowledge, this is the first image-level HSI classification model based on the Mamba. We conduct extensive experiments on four diverse HSI datasets. The results demonstrate the effectiveness and superiority of the proposed model for HSI classification. This reveals the great potential of Mamba to be the next-generation backbone for HSI models. Codes are available at https://github.com/li-yapeng/MambaHSI .

Aligning Few-Step Diffusion Models with Dense Reward Difference Learning

Aligning diffusion models with downstream objectives is essential for their practical applications. However, standard alignment methods often struggle with step generalization when directly applied to few-step diffusion models, leading to inconsistent performance across different denoising step scenarios. To address this, we introduce Stepwise Diffusion Policy Optimization (SDPO), a novel alignment method tailored for few-step diffusion models. Unlike prior approaches that rely on a single sparse reward from only the final step of each denoising trajectory for trajectory-level optimization, SDPO incorporates dense reward feedback at every intermediate step. By learning the differences in dense rewards between paired samples, SDPO facilitates stepwise optimization of few-step diffusion models, ensuring consistent alignment across all denoising steps. To promote stable and efficient training, SDPO introduces an online reinforcement learning framework featuring several novel strategies designed to effectively exploit the stepwise granularity of dense rewards. Experimental results demonstrate that SDPO consistently outperforms prior methods in reward-based alignment across diverse step configurations, underscoring its robust step generalization capabilities. Code is avaliable at https://github.com/ZiyiZhang27/sdpo.

Efficient and Effective Weight-Ensembling Mixture of Experts for Multi-Task Model Merging

Multi-task learning (MTL) leverages a shared model to accomplish multiple tasks and facilitate knowledge transfer. Recent research on task arithmetic-based MTL demonstrates that merging the parameters of independently fine-tuned models can effectively achieve MTL. However, existing merging methods primarily seek a static optimal solution within the original model parameter space, which often results in performance degradation due to the inherent diversity among tasks and potential interferences. To address this challenge, in this paper, we propose a Weight-Ensembling Mixture of Experts (WEMoE) method for multi-task model merging. Specifically, we first identify critical (or sensitive) modules by analyzing parameter variations in core modules of Transformer-based models before and after finetuning. Then, our WEMoE statically merges non-critical modules while transforming critical modules into a mixture-of-experts (MoE) structure. During inference, expert modules in the MoE are dynamically merged based on input samples, enabling a more flexible and adaptive merging approach. Building on WEMoE, we further introduce an efficient-and-effective WEMoE (E-WEMoE) method, whose core mechanism involves eliminating non-essential elements in the critical modules of WEMoE and implementing shared routing across multiple MoE modules, thereby significantly reducing both the trainable parameters, the overall parameter count, and computational overhead of the merged model by WEMoE. Experimental results across various architectures and tasks demonstrate that both WEMoE and E-WEMoE outperform state-of-the-art (SOTA) model merging methods in terms of MTL performance, generalization, and robustness.