The Mixture of Experts (MoE) architecture is a cornerstone of modern state-of-the-art (SOTA) large language models (LLMs). MoE models facilitate scalability by enabling sparse parameter activation. However, traditional MoE architecture uses homogeneous experts of a uniform size, activating a fixed number of parameters irrespective of input complexity and thus limiting computational efficiency. To overcome this limitation, we introduce Grove MoE, a novel architecture incorporating experts of varying sizes, inspired by the heterogeneous big.LITTLE CPU architecture. This architecture features novel adjugate experts with a dynamic activation mechanism, enabling model capacity expansion while maintaining manageable computational overhead. Building on this architecture, we present GroveMoE-Base and GroveMoE-Inst, 33B-parameter LLMs developed by applying an upcycling strategy to the Qwen3-30B-A3B-Base model during mid-training and post-training. GroveMoE models dynamically activate 3.14 - 3.28B parameters based on token complexity and achieve performance comparable to SOTA open-source models of similar or even larger size.

Diffusion large language models (dLLMs) accelerate language generation by predicting multiple masks in a single forward pass.However, existing dLLMs can suffer from unreliable predictions in early denoising stages under aggressive parallelism strategies, leading to errors that can propagate to later stages. To tackle this issue, we present Consistency Forcing (CForce) for dLLMs, a distillation method to force the mask predictions of early stages to align with those of later stages. CForce trains the model on pre-collected self-rollout trajectories, thereby improving training-inference alignment.We introduce Confidence Adaptive KL Divergence as a distillation objective to conjoin the merits of forward and reverse KL.We further provide a theoretical analysis for the consistency objective to explain why CForce can approximately minimize the prediction error of early stages.Critically, the same formulation applies to both mask-to-token decoding and edit-capable decoding; in the edit-capable case, later token-to-token refinements provide additional supervision for earlier masked-state predictions. Experiments on non-edit and edit-capable LLaDA models show improved speed-quality trade-offs, especially under high-parallelism decoding budgets.

Diffusion large language models (dLLMs) generate text through iterative denoising. In commonly adopted parallel decoding schemes, each step confirms only high-confidence positions while remasking the others. By analyzing dLLM denoising traces, we uncover a key inefficiency: models often predict the correct target token several steps before its confidence becomes high enough to be decoded. This gap between early prediction and late decoding forces repeated remasking of already-correct tokens, causing redundant iterations and limiting acceleration. To exploit this temporal redundancy, we introduce Trace Credit to quantify a token’s decoding potential by accumulating historical evidence. Building on this, we propose CreditDecoding, a training-free parallel decoding method that fuses Trace Credit with current logits to boost the confidence of correct but underconfident tokens, thereby accelerating denoising and improving robustness. On eight benchmarks, CreditDecoding achieves up to 5.48 times speedup with +0.48 accuracy on LLaDA-8B and consistently improves performance across diverse dLLM architectures and parameter scales. It further scales to long contexts and remains orthogonal to mainstream inference optimizations, making it a practical and widely applicable solution.

This work explores efficient ultra-long context modeling. We posit that an effective solution requires three fundamental properties: sparsity, random-access flexibility, and length generalization. To achieve this, we leverage Hierarchical Sparse Attention (HSA), a novel attention mechanism that satisfies all three properties. We integrate HSA into the Transformer architecture to develop HSA-UltraLong, an 8B-parameter Mixture-of-Experts (MoE) model trained on over 8 trillion tokens. We rigorously evaluate the model across tasks with both in-domain and out-of-domain context lengths to validate its capabilities. Our model demonstrates comparable performance to full-attention baselines on in-domain sequence lengths. Crucially, it achieves over 90% accuracy on most in-context retrieval tasks with contexts up to 512 times the pre-training context length. This work reports our findings and remaining issues throughout the experiments, offering insights for future research in ultra-long context modeling.

Processing long contexts is increasingly important for Large Language Models (LLMs) in tasks like multi-turn dialogues, code generation, and document summarization. This paper addresses the challenges of achieving high long-context performance, low computational complexity, and compatibility with pretrained models – collectively termed the “impossible triangle”. We introduce E2LLM (Encoder Elongated Large Language Models), a novel approach that effectively navigates this paradox. E2LLM divides long contexts into chunks, compresses each into soft prompts using a pretrained text encoder, and aligns these representations with a decoder-only LLM via an adapter. To enhance the LLM’s reasoning with these soft prompts, we employ two training objectives: encoder output reconstruction and long-context instruction fine-tuning. Extensive experiments reveal that E2LLM not only outperforms 8 state-of-the-art (SOTA) methods in effectiveness and efficiency for document summarization and question answering, but also achieves the best performance on LongBench v2 among models of comparable size. The source code is available at https://github.com/codefuse-ai/E2LLM.

Programming languages possess rich semantic information - such as data flow - that is represented by graphs and not available from the surface form of source code. Recent code language models have scaled to billions of parameters, but model source code solely as text tokens while ignoring any other structural information. Conversely, models that do encode structural information of code make modifications to the Transformer architecture, limiting their scale and compatibility with pretrained LLMs. In this work, we take the best of both worlds with GALLa - Graph Aligned Large Language Models. GALLa utilizes graph neural networks and cross-modal alignment technologies to inject the structural information of code into LLMs as an auxiliary task during finetuning. This framework is both model-agnostic and task-agnostic, as it can be applied to any code LLM for any code downstream task, and requires the structural graph data only at training time from a corpus unrelated to the finetuning data, while incurring no cost at inference time over the baseline LLM. Experiments on five code tasks with six different baseline LLMs ranging in size from 350M to 14B validate the effectiveness of GALLa, demonstrating consistent improvement over the baseline, even for powerful models such as LLaMA3 and Qwen2.5-Coder.

Multi-task learning (MTL) benefits the fine-tuning of large language models (LLMs) by providing a single model with improved performance and generalization ability across tasks, presenting a resource-efficient alternative to developing separate models for each task. Yet, existing MTL strategies for LLMs often fall short by either being computationally intensive or failing to ensure simultaneous task convergence. This paper presents CoBa, a new MTL approach designed to effectively manage task convergence balance with minimal computational overhead. Utilizing Relative Convergence Scores (RCS), Absolute Convergence Scores (ACS), and a Divergence Factor (DF), CoBa dynamically adjusts task weights during the training process, ensuring that the validation loss of all tasks progress towards convergence at an even pace while mitigating the issue of individual task divergence. The results of our experiments involving three disparate datasets underscore that this approach not only fosters equilibrium in task improvement but enhances the LLMs’ performance by up to 13% relative to the second-best baselines. Code is open-sourced at https://github.com/codefuse-ai/MFTCoder.

The key challenge in semantic search is to create models that are both accurate and efficient in pinpointing relevant sentences for queries. While BERT-style bi-encoders excel in efficiency with pre-computed embeddings, they often miss subtle nuances in search tasks. Conversely, GPT-style LLMs with cross-encoder designs capture these nuances but are computationally intensive, hindering real-time applications. In this paper, we present D2LLMs—Decomposed and Distilled LLMs for semantic search—that combines the best of both worlds. We decompose a cross-encoder into an efficient bi-encoder integrated with Pooling by Multihead Attention and an Interaction Emulation Module, achieving nuanced understanding and pre-computability. Knowledge from the LLM is distilled into this model using contrastive, rank, and feature imitation techniques. Our experiments show that D2LLM surpasses five leading baselines in terms of all metrics across three tasks, particularly improving NLI task performance by at least 6.45%

Self-attention and position embedding are two crucial modules in transformer-based Large Language Models (LLMs). However, the potential relationship between them is far from well studied, especially for long context window extending. In fact, anomalous behaviors that hinder long context extrapolation exist between Rotary Position Embedding (RoPE) and vanilla self-attention.Incorrect initial angles between Q and K can cause misestimation in modeling rotary position embedding of the closest tokens.To address this issue, we propose Collinear Constrained Attention mechanism, namely CoCA. Specifically, we enforce a collinear constraint between Q and K to seamlessly integrate RoPE and self-attention.While only adding minimal computational and spatial complexity, this integration significantly enhances long context window extrapolation ability. We provide an optimized implementation, making it a drop-in replacement for any existing transformer-based models.Extensive experiments demonstrate that CoCA excels in extending context windows. A CoCA-based GPT model, trained with a context length of 512, can extend the context window up to 32K (60×) without any fine-tuning.Additionally, incorporating CoCA into LLaMA-7B achieves extrapolation up to 32K within a training length of only 2K.Our code is publicly available at: https://github.com/codefuse-ai/Collinear-Constrained-Attention

Many text classification tasks require handling unseen domains with plenty of unlabeled data, thus giving rise to the self-adaption or the so-called transductive zero-shot learning (TZSL) problem. However, current methods based solely on encoders or decoders overlook the possibility that these two modules may promote each other. As a first effort to bridge this gap, we propose an autoencoder named ZeroAE. Specifically, the text is encoded with two separate BERT-based encoders into two disentangled spaces, i.e., label-relevant (for classification) and label-irrelevant respectively. The two latent spaces are then decoded by prompting GPT-2 to recover the text as well as to further generate text with labels in the unseen domains to train the encoder in turn. To better exploit the unlabeled data, a novel indirect uncertainty-aware sampling (IUAS) approach is proposed to train ZeroAE. Extensive experiments show that ZeroAE largely surpasses the SOTA methods by 15.93% and 8.70% on average respectively in the label-partially-unseen and label-fully-unseen scenario. Notably, the label-fully-unseen ZeroAE even possesses superior performance to the label-partially-unseen SOTA methods.