Hanlin Xu

Jianuo Huang, Yaojie Zhang, Qituan Zhang et al.

Speculative decoding accelerates LLM inference by drafting multiple tokens and verifying them in parallel with the target model. However, its practical speedup is constrained by the trade-off between draft quality and drafting cost: autoregressive drafters model causal dependencies among draft tokens but incur sequential overhead, while parallel drafters reduce drafting cost but weaken intra-block dependency modeling. In this paper, we propose Domino, a speculative decoding framework that decouples causal dependency modeling from expensive autoregressive draft execution. Domino first uses a parallel draft backbone to produce preliminary draft distributions for the entire block, and then applies a lightweight Domino head to refine them with prefix-dependent causal information. To stabilize teacher-forced causal encoding, we further introduce a base-anchored training curriculum that first strengthens the parallel backbone and then gradually shifts optimization toward the causally corrected final distribution. Experiments on Qwen3 models show that Domino achieves up to \(5.49\times\) end-to-end speedup under the Transformers backend and up to \(5.8\times\) throughput speedup under SGLang serving.

Yuhang Han, Yuyang Wu, Zhengbo Jiao et al.

Reinforcement Learning from Verifiable Rewards (RLVR) has substantially enhanced the reasoning capabilities of large language models in abstract reasoning tasks. However, its application to Large Vision-Language Models (LVLMs) remains constrained by a structural representational bottleneck. Existing approaches generally lack explicit modeling and effective utilization of visual information, preventing visual representations from being tightly coupled with the reinforcement learning optimization process and thereby limiting further improvements in multimodal reasoning performance. To address this limitation, we propose KAWHI (Key-Region Aligned Weighted Harmonic Incentive), a plug-and-play reward reweighting mechanism that explicitly incorporates structured visual information into uniform reward policy optimization methods (e.g., GRPO and GSPO). The method adaptively localizes semantically salient regions through hierarchical geometric aggregation, identifies vision-critical attention heads via structured attribution, and performs paragraph-level credit reallocation to align spatial visual evidence with semantically decisive reasoning steps. Extensive empirical evaluations on diverse reasoning benchmarks substantiate KAWHI as a general-purpose enhancement module, consistently improving the performance of various uniform reward optimization methods. Project page: KAWHI (https://kawhiiiileo.github.io/KAWHI_PAGE/)

Yifeng Gao, Ziang Ji, Yuxuan Wang et al.

Diffusion-based Large Language Models (dLLMs) have emerged as a competitive alternative to autoregressive models, offering unique advantages through bidirectional attention and parallel generation paradigms. However, the generation results of current parallel decoding methods deviate from stepwise decoding, introducing potential performance degradation, which limits their practical deployment. To address this problem, we propose \textbf{S}elf \textbf{S}peculative \textbf{D}ecoding (SSD), a lossless inference acceleration method that leverages the dLLM itself as both speculative decoding drafter and verifier without auxiliary modules. SSD introduces a self-drafting mechanism where the model generates predictions for multiple positions, then verifies them through hierarchical verification trees in a single forward pass. Unlike traditional speculative decoding that requires separate draft models, SSD eliminates model redundancy and memory overhead by exploiting the dLLM's inherent parallel prediction capability for multiple positions. This self-speculative approach allows the model to progressively verify and accept multiple tokens in a single forward pass. Our experiments demonstrate that SSD achieves up to 3.46$\times$ speedup while keeping the output identical to stepwise decoding on open source models such as LLaDA and Dream. Code will be made publicly available on GitHub.

Xinyu Zhan, Lixin Yang, Yifei Zhao et al.

We present OAKINK2, a dataset of bimanual object manipulation tasks for complex daily activities. In pursuit of constructing the complex tasks into a structured representation, OAKINK2 introduces three level of abstraction to organize the manipulation tasks: Affordance, Primitive Task, and Complex Task. OAKINK2 features on an object-centric perspective for decoding the complex tasks, treating them as a sequence of object affordance fulfillment. The first level, Affordance, outlines the functionalities that objects in the scene can afford, the second level, Primitive Task, describes the minimal interaction units that humans interact with the object to achieve its affordance, and the third level, Complex Task, illustrates how Primitive Tasks are composed and interdependent. OAKINK2 dataset provides multi-view image streams and precise pose annotations for the human body, hands and various interacting objects. This extensive collection supports applications such as interaction reconstruction and motion synthesis. Based on the 3-level abstraction of OAKINK2, we explore a task-oriented framework for Complex Task Completion (CTC). CTC aims to generate a sequence of bimanual manipulation to achieve task objectives. Within the CTC framework, we employ Large Language Models (LLMs) to decompose the complex task objectives into sequences of Primitive Tasks and have developed a Motion Fulfillment Model that generates bimanual hand motion for each Primitive Task. OAKINK2 datasets and models are available at https://oakink.net/v2.

Amirmohsen Sattarifard, Sepehr Lavasani, Ehsan Imani et al.

Deploying Large Language Models (LLMs) on edge hardware demands aggressive, prompt-aware dynamic pruning to reduce computation without degrading quality. Static or predictor-based schemes either lock in a single sparsity pattern or incur extra runtime overhead, and recent zero-shot methods that rely on statistics from a single prompt fail on short prompt and/or long generation scenarios. We introduce A/I-GLASS: Activation- and Impact-based Global-Local neural importance Aggregation for feed-forward network SparSification, two training-free methods that dynamically select FFN units using a rank-aggregation of prompt local and model-intrinsic global neuron statistics. Empirical results across multiple LLMs and benchmarks demonstrate that GLASS significantly outperforms prior training-free methods, particularly in challenging long-form generation scenarios, without relying on auxiliary predictors or adding any inference overhead.

Yicun Yang, Cong Wang, Shaobo Wang et al.

Diffusion-based large language models (dLLMs) have exhibited substantial potential for parallel text generation, which may enable more efficient generation compared to autoregressive models. However, current dLLMs suffer from fixed generation lengths, which indicates the generation lengths of dLLMs have to be determined before decoding as a hyper-parameter, leading to issues in efficiency and flexibility. To solve these problems, in this work, we propose to train a diffusion LLM with native variable generation lengths, abbreviated as dLLM-Var. Concretely, we aim to train a model to accurately predict the [EOS] token in the generated text, which makes a dLLM be able to natively infer in a block diffusion manner, while still maintaining the ability of global bi-directional (full) attention and high parallelism. Experiments on standard benchmarks demonstrate that our method achieves a 30.1x speedup over traditional dLLM inference paradigms and a 2.4x speedup relative to autoregressive models such as Qwen and Llama. Our method achieves higher accuracy and faster inference, elevating dLLMs beyond mere academic novelty and supporting their practical use in real-world applications. Codes and models have been released.

Qingwen Zeng, Hanlin Xu, Nanjun Xu et al.

Carbon emissions significantly contribute to climate change, and carbon credits have emerged as a key tool for mitigating environmental damage and helping organizations manage their carbon footprint. Despite their growing importance across sectors, fully leveraging carbon credits remains challenging. This study explores engineering practices and fintech solutions to enhance carbon emission management. We first review the negative impacts of carbon emission non-disclosure, revealing its adverse effects on financial stability and market value. Organizations are encouraged to actively manage emissions and disclose relevant data to mitigate risks. Next, we analyze factors influencing carbon prices and review advanced prediction algorithms that optimize carbon credit purchasing strategies, reducing costs and improving efficiency. Additionally, we examine corporate carbon emission prediction models, which offer accurate performance assessments and aid in planning future carbon credit needs. By integrating carbon price and emission predictions, we propose research directions, including corporate carbon management cost forecasting. This study provides a foundation for future quantitative research on the financial and market impacts of carbon management practices and is the first systematic review focusing on computing solutions and engineering practices for carbon credits.