Pengju Ren

Haiduo Huang, Fuwei Yang, Zhenhua Liu et al.

Speculative decoding (SD) accelerates large language model inference by using a smaller draft model to predict multiple tokens, which are then verified in parallel by the larger target model. However, the limited capacity of the draft model often necessitates tree-based sampling to improve prediction accuracy, where multiple candidates are generated at each step. We identify a key limitation in this approach: the candidates at the same step are derived from the same representation, limiting diversity and reducing overall effectiveness. To address this, we propose Jakiro, leveraging Mixture of Experts (MoE), where independent experts generate diverse predictions, effectively decoupling correlations among candidates. Furthermore, we introduce a hybrid inference strategy, combining autoregressive decoding for initial tokens with parallel decoding for subsequent stages, and enhance the latter with contrastive mechanism in features to improve accuracy. Our method significantly boosts prediction accuracy and achieves higher inference speedups. Extensive experiments across diverse models validate the effectiveness and robustness of our approach, establishing a new SOTA in speculative decoding. Our codes are available at https://github.com/haiduo/Jakiro.

Haiduo Huang, Fuwei Yang, Zhenhua Liu et al.

Speculative decoding is a powerful way to accelerate autoregressive large language models (LLMs), but directly porting it to vision-language models (VLMs) faces unique systems constraints: the prefill stage is dominated by visual tokens whose count scales with image resolution and video length, inflating both compute and memory, especially the key-value (KV) cache. We study speculative decoding for VLMs and introduce SpecVLM, a practical system that (1) establishes a strong EAGLE-2-style baseline, EagleVLM, delivering 1.5--2.3x end-to-end speedups over full autoregressive inference, and (2) further accelerates VLM inference with an elastic visual compressor that adaptively selects among pruning, pooling, convolution, and resampler primitives to balance FLOPs/parameters and accuracy per input. To avoid costly offline distillation corpora, we propose an online-logit distillation protocol that trains the draft model with on-the-fly teacher logits and penultimate features using a combined cross-entropy and Smooth L1 objective, eliminating storage and preprocessing while remaining compute-efficient. This protocol reveals a training-time scaling effect: longer online training monotonically increases the draft model's average accepted length, improving speculative efficiency. Empirically, SpecVLM achieves additional acceleration, culminating in 2.5--2.9x end-to-end speedups within 5 epochs across LLaVA and MMMU, consistently over resolutions and task difficulties, while preserving the target model's output distribution (lossless decoding). Our code is available at https://github.com/haiduo/SpecVLM.

Haiduo Huang, Jiangcheng Song, Yadong Zhang et al.

Recent advances in knowledge distillation have emphasized the importance of decoupling different knowledge components. While existing methods utilize momentum mechanisms to separate task-oriented and distillation gradients, they overlook the inherent conflict between target-class and non-target-class knowledge flows. Furthermore, low-confidence dark knowledge in non-target classes introduces noisy signals that hinder effective knowledge transfer. To address these limitations, we propose DeepKD, a novel training framework that integrates dual-level decoupling with adaptive denoising. First, through theoretical analysis of gradient signal-to-noise ratio (GSNR) characteristics in task-oriented and non-task-oriented knowledge distillation, we design independent momentum updaters for each component to prevent mutual interference. We observe that the optimal momentum coefficients for task-oriented gradient (TOG), target-class gradient (TCG), and non-target-class gradient (NCG) should be positively related to their GSNR. Second, we introduce a dynamic top-k mask (DTM) mechanism that gradually increases K from a small initial value to incorporate more non-target classes as training progresses, following curriculum learning principles. The DTM jointly filters low-confidence logits from both teacher and student models, effectively purifying dark knowledge during early training. Extensive experiments on CIFAR-100, ImageNet, and MS-COCO demonstrate DeepKD's effectiveness. Our code is available at https://github.com/haiduo/DeepKD.

Jin Cui, Jiaqi Guo, Jiepeng Zhou et al.

Chain-of-Thought (CoT) reasoning empowers Large Language Models (LLMs) with remarkable capabilities but typically requires prohibitive parameter scales. CoT distillation has emerged as a promising paradigm to transfer reasoning prowess into compact Student Models (SLMs), but existing approaches often rely on a solitary teacher, capping the student's potential since individual LLMs often exhibit distinct capability biases and may suffer from catastrophic forgetting. While leveraging diverse teachers seems appealing, effectively fusing their supervisions remains challenging: teacher-student incompatibility risks amplifying hallucinations, and passive supervision fails to ensure genuine logic internalization. To address this, we introduce COMPACT, a framework that adaptively fuses supervisions from different teachers by dynamically weighting teacher gradients based on the student's real-time compatibility evaluated by a multi-dimensional metric: (1) Graph-based Consensus to filter misleading rationales by identifying mainstream reasoning paths; (2) Mutual-Information-based Adaptability to detect "epiphany moments" for genuinely understanding the reasoning process rather than merely imitating; and (3) Loss-based Difficulty to assess student receptivity to the teacher's guidance and prevent negative transfer. Extensive experiments and latent space analysis demonstrate that COMPACT effectively integrates diverse reasoning capabilities without damaging the model's original knowledge structure, achieving state-of-the-art performance on various benchmarks while mitigating catastrophic forgetting.

Jin Cui, Jiaqi Guo, Jiepeng Zhou et al.

While Large Language Models (LLMs) have emerged with remarkable capabilities in complex tasks through Chain-of-Thought reasoning, practical resource constraints have sparked interest in transferring these abilities to smaller models. However, achieving both domain performance and cross-domain generalization remains challenging. Existing approaches typically restrict students to following a single golden rationale and treat different reasoning paths independently. Due to distinct inductive biases and intrinsic preferences, alongside the student's evolving capacity and reasoning preferences during training, a teacher's "optimal" rationale could act as out-of-distribution noise. This misalignment leads to a degeneration of the student's latent reasoning distribution, causing suboptimal performance. To bridge this gap, we propose MIND, a capability-adaptive framework that transitions distillation from passive mimicry to active cognitive construction. We synthesize diverse teacher perspectives through a novel "Teaching Assistant" network. By employing a Feedback-Driven Inertia Calibration mechanism, this network utilizes inertia-filtered training loss to align supervision with the student's current adaptability, effectively enhancing performance while mitigating catastrophic forgetting. Extensive experiments demonstrate that MIND achieves state-of-the-art performance on both in-distribution and out-of-distribution benchmarks, and our sophisticated latent space analysis further confirms the mechanism of reasoning ability internalization.