Full-Stack FP4: Stable LLM Pretraining with Quantized Projections, Optimizers, and Attention

Recent NVFP4 pretraining methods mainly target transformer linear layers, leaving optimizer states, optimizer arithmetic and attention underexplored in 4-bit pipelines. This critical gap blocks stable full-stack 4-bit pretraining, as the three core modules exhibit unique numerical failure patterns: linear layers hit hard quantization noise limits with dimension-propagated error amplification; AdamW second moments are heavy-tailed non-negative values fragile to low-precision denominators; attention carries error-prone computation paths demanding strict forward-backward quantization consistency. We propose Full-Stack FP4, the first complete NVFP4 pretraining framework resolving all three stability bottlenecks via module-wise precision strategies. For linear projections, LoRA-SVD lightweight decomposition suppresses quantization noise and breaks the direct-quantization error ceiling, shrinking the linear-only loss gap from 1.40% to 0.61%. For optimizers, we design AdamW second-moment transformation for robust NVFP4 storage and fully support native NVFP4 Newton-Schulz iterations for the Root (Muon) optimizer. For attention, a mixed-precision scheme quantizes Q/K/V and backward dS while guarding vulnerable paths in BF16, paired with unified tensor reuse to sustain forward-backward alignment. We further analyze fast error accumulation in naive low-bit matrix multiplication and the extreme sensitivity of PV / dOV^T attention branches. All modules are plug-and-play with cumulative stability and efficiency improvements. Our 3B/64B-token pretraining validates near-BF16 performance with merely 1.47% loss gap, verifying feasible stable end-to-end NVFP4 LLM pretraining.

MemTrain: Self-Supervised Context Memory Training

Memory is an indispensable capability for long-horizon LLM agents, enabling them to preserve and utilize information accumulated across extended interactions. Existing memory-agent approaches are typically trained end-to-end with reinforcement learning on downstream tasks. However, collecting high-quality annotated problems for memory-intensive scenarios is costly, and the resulting training data often lack sufficient diversity to cover general memory behaviors. In this work, we propose MemTrain, a self-supervised training framework for generally enhancing the context-memory capability of LLM agents for more effective downstream post-training. MemTrain introduces two coupled proxy tasks over unlabeled Wikipedia corpora: (1) an end-to-end masked reconstruction objective, which requires the model to recover masked entities after multiple rounds of memory updates, thereby encouraging memory maintenance from the final outcome perspective; and (2) an intermediate memory recall objective, which requires the model to reconstruct masked historical information using intermediate memory states, encouraging faithful compression and memory completeness throughout the interaction process. The two objectives are jointly optimized using GRPO. Extensive experiments on long-text QA and search-based QA benchmarks demonstrate that MemTrain consistently improves downstream memory-intensive reasoning performance across different models, achieving gains of up to 17.67 points over direct task-specific post-training.

Trust Region On-Policy Distillation

On-Policy Distillation (OPD) is a fundamental technique for efficient post-training of large language models (LLMs), with broad applications in agent learning, multi-task enhancement, and model compression. However, OPD training becomes unstable when the teacher and student distributions differ substantially, as teacher supervision on student-generated tokens may yield unreliable policy gradients and even cause optimization failure. This work addresses reliable on-policy token-level supervision through credit assignment strategies, and proposes Trust Region On-Policy Distillation, TrOPD. It features the following characteristics: 1) Trust-Region On-Policy Learning: TrOPD performs OPD only in regions where the teacher provides reliable supervision, mitigating the optimization difficulty of the K1 reverse-KL estimator under distribution mismatch. 2) Outlier Estimation: For outlier regions, we explore gradient clipping, masking, and forward-KL estimation to reduce the adverse effects of unreliable supervision. 3) Off-Policy Guidance: The student continues generation from teacher prefixes and uses forward KL to imitate off-policy guidance, encouraging on-policy exploration toward reliable regions. Experiments show that TrOPD consistently outperforms SoTA OPD baselines, including OPD, EOPD, and REOPOLD, across mathematical reasoning, code generation, and general-domain benchmarks.

Position-Aware Depth Decay Decoding ($D^3$): Boosting Large Language Model Inference Efficiency

Due to the large number of parameters, the inference phase of Large Language Models (LLMs) is resource-intensive. Unlike traditional model compression, which needs retraining, recent dynamic computation methods show that not all components are required for inference, enabling a training-free pipeline. In this paper, we focus on the dynamic depth of LLM generation. A token-position aware layer skipping framework is proposed to save 1.5x times operations efficiently while maintaining performance. We first observed that tokens predicted later have lower perplexity and thus require less computation. Then, we propose a training-free algorithm called Position-Aware Depth Decay Decoding ($D^3$), which leverages a power-law decay function, $\left\lfloor L \times (\alpha^i) \right\rfloor$, to determine the number of layers to retain when generating token $T_i$. Remarkably, without any retraining, the $D^3$ achieves success across a wide range of generation tasks for the first time. Experiments on large language models (\ie the Llama) with $7 \sim 70$ billion parameters show that $D^3$ can achieve an average 1.5x speedup compared with the full-inference pipeline while maintaining comparable performance with nearly no performance drop ($<1\%$) on the GSM8K and BBH benchmarks.

EfficientLLM: Scalable Pruning-Aware Pretraining for Architecture-Agnostic Edge Language Models

Modern large language models (LLMs) driven by scaling laws, achieve intelligence emergency in large model sizes. Recently, the increasing concerns about cloud costs, latency, and privacy make it an urgent requirement to develop compact edge language models. Distinguished from direct pretraining that bounded by the scaling law, this work proposes the pruning-aware pretraining, focusing on retaining performance of much larger optimized models. It features following characteristics: 1) Data-scalable: we introduce minimal parameter groups in LLM and continuously optimize structural pruning, extending post-training pruning methods like LLM-Pruner and SparseGPT into the pretraining phase. 2) Architecture-agnostic: the LLM architecture is auto-designed using saliency-driven pruning, which is the first time to exceed SoTA human-designed LLMs in modern pretraining. We reveal that it achieves top-quality edge language models, termed EfficientLLM, by scaling up LLM compression and extending its boundary. EfficientLLM significantly outperforms SoTA baselines with $100M \sim 1B$ parameters, such as MobileLLM, SmolLM, Qwen2.5-0.5B, OLMo-1B, Llama3.2-1B in common sense benchmarks. As the first attempt, EfficientLLM bridges the performance gap between traditional LLM compression and direct pretraining methods, and we will fully open source at https://github.com/Xingrun-Xing2/EfficientLLM.

Enhancing Generalization via Sharpness-Aware Trajectory Matching for Dataset Condensation

Dataset condensation aims to synthesize datasets with a few representative samples that can effectively represent the original datasets. This enables efficient training and produces models with performance close to those trained on the original sets. Most existing dataset condensation methods conduct dataset learning under the bilevel (inner- and outer-loop) based optimization. However, the preceding methods perform with limited dataset generalization due to the notoriously complicated loss landscape and expensive time-space complexity of the inner-loop unrolling of bilevel optimization. These issues deteriorate when the datasets are learned via matching the trajectories of networks trained on the real and synthetic datasets with a long horizon inner-loop. To address these issues, we introduce Sharpness-Aware Trajectory Matching (SATM), which enhances the generalization capability of learned synthetic datasets by optimising the sharpness of the loss landscape and objective simultaneously. Moreover, our approach is coupled with an efficient hypergradient approximation that is mathematically well-supported and straightforward to implement along with controllable computational overhead. Empirical evaluations of SATM demonstrate its effectiveness across various applications, including in-domain benchmarks and out-of-domain settings. Moreover, its easy-to-implement properties afford flexibility, allowing it to integrate with other advanced sharpness-aware minimizers. Our code will be released.

From Pixels to Tokens: Byte-Pair Encoding on Quantized Visual Modalities

Multimodal Large Language Models have made significant strides in integrating visual and textual information, yet they often struggle with effectively aligning these modalities. We introduce a novel image tokenizer that bridges this gap by applying the principle of Byte-Pair Encoding (BPE) to visual data. Unlike conventional approaches that rely on separate visual encoders, our method directly incorporates structural prior information into image tokens, mirroring the successful tokenization strategies used in text-only Large Language Models. This innovative approach enables Transformer models to more effectively learn and reason across modalities. Through theoretical analysis and extensive experiments, we demonstrate that our BPE Image Tokenizer significantly enhances MLLMs' multimodal understanding capabilities, even with limited training data. Our method not only improves performance across various benchmarks but also shows promising scalability, potentially paving the way for more efficient and capable multimodal foundation models.

OmniGen: Unified Image Generation

In this work, we introduce OmniGen, a new diffusion model for unified image generation. Unlike popular diffusion models (e.g., Stable Diffusion), OmniGen no longer requires additional modules such as ControlNet or IP-Adapter to process diverse control conditions. OmniGenis characterized by the following features: 1) Unification: OmniGen not only demonstrates text-to-image generation capabilities but also inherently supports other downstream tasks, such as image editing, subject-driven generation, and visual-conditional generation. Additionally, OmniGen can handle classical computer vision tasks by transforming them into image generation tasks, such as edge detection and human pose recognition. 2) Simplicity: The architecture of OmniGen is highly simplified, eliminating the need for additional text encoders. Moreover, it is more user-friendly compared to existing diffusion models, enabling complex tasks to be accomplished through instructions without the need for extra preprocessing steps (e.g., human pose estimation), thereby significantly simplifying the workflow of image generation. 3) Knowledge Transfer: Through learning in a unified format, OmniGen effectively transfers knowledge across different tasks, manages unseen tasks and domains, and exhibits novel capabilities. We also explore the model's reasoning capabilities and potential applications of chain-of-thought mechanism. This work represents the first attempt at a general-purpose image generation model, and there remain several unresolved issues. We will open-source the related resources at https://github.com/VectorSpaceLab/OmniGen to foster advancements in this field.

SpikeLLM: Scaling up Spiking Neural Network to Large Language Models via Saliency-based Spiking

The recent advancements in large language models (LLMs) with billions of parameters have significantly boosted their performance across various real-world applications. However, the inference processes for these models require substantial energy and computational resources, presenting considerable deployment challenges. In contrast, human brains, which contain approximately 86 billion biological neurons, exhibit significantly greater energy efficiency compared to LLMs with a similar number of parameters. Inspired by this, we redesign 7 to 70 billion parameter LLMs using bio-plausible spiking mechanisms, emulating the efficient behavior of the human brain. We propose the first spiking large language model as recent LLMs termed SpikeLLM. Coupled with the proposed model, a novel spike-driven quantization framework named Optimal Brain Spiking is introduced to reduce the energy cost and accelerate inference speed via two essential approaches: first (second)-order differentiation-based salient channel detection, and per-channel salient outlier expansion with Generalized Integrate-and-Fire neurons. Our proposed spike-driven quantization can plug in main streams of quantization training methods. In the OmniQuant pipeline, SpikeLLM significantly reduces 25.51% WikiText2 perplexity and improves 3.08% average accuracy of 6 zero-shot datasets on a LLAMA2-7B 4A4W model. In the GPTQ pipeline, SpikeLLM realizes a sparse ternary quantization, which achieves additive in all linear layers. Compared with PB-LLM with similar operations, SpikeLLM also exceeds significantly. We will release our code on GitHub.

SpikeLM: Towards General Spike-Driven Language Modeling via Elastic Bi-Spiking Mechanisms

Towards energy-efficient artificial intelligence similar to the human brain, the bio-inspired spiking neural networks (SNNs) have advantages of biological plausibility, event-driven sparsity, and binary activation. Recently, large-scale language models exhibit promising generalization capability, making it a valuable issue to explore more general spike-driven models. However, the binary spikes in existing SNNs fail to encode adequate semantic information, placing technological challenges for generalization. This work proposes the first fully spiking mechanism for general language tasks, including both discriminative and generative ones. Different from previous spikes with {0,1} levels, we propose a more general spike formulation with bi-directional, elastic amplitude, and elastic frequency encoding, while still maintaining the addition nature of SNNs. In a single time step, the spike is enhanced by direction and amplitude information; in spike frequency, a strategy to control spike firing rate is well designed. We plug this elastic bi-spiking mechanism in language modeling, named SpikeLM. It is the first time to handle general language tasks with fully spike-driven models, which achieve much higher accuracy than previously possible. SpikeLM also greatly bridges the performance gap between SNNs and ANNs in language modeling. Our code is available at https://github.com/Xingrun-Xing/SpikeLM.