arXiv subjects

Linfeng Zhang

Publications and source records attributed to Linfeng Zhang.

At least 163 records · Page 9Linked to original sources

arXiv2025

SDAR: A Synergistic Diffusion-AutoRegression Paradigm for Scalable Sequence Generation

We propose SDAR, a Synergistic Diffusion-Autoregression paradigm that unifies the training efficiency of autoregressive models with the parallel inference capability of diffusion. Instead of costly end-to-end diffusion training, SDAR performs a lightweight paradigm conversion that transforms a well-trained autoregressive (AR) model into a blockwise diffusion model through brief, data-efficient adaptation. During inference, SDAR generates sequences autoregressively across blocks for global coherence while decoding all tokens within each block in parallel via a discrete diffusion process. Extensive experiments show that AR models remain substantially more compute-efficient than masked diffusion models, providing a strong foundation for adaptation. Building on this insight, SDAR achieves efficient AR-to-diffusion conversion with minimal cost, preserving AR-level performance while enabling parallel generation. Scaling studies across dense and Mixture-of-Experts architectures confirm that SDAR scales without compromise: larger models exhibit stronger robustness to block size and decoding thresholds, yielding greater speedups without accuracy loss. Beyond efficiency, SDAR demonstrates enhanced reasoning and domain adaptability. Our 30B MoE model surpasses its AR counterpart on challenging scientific reasoning benchmarks such as GPQA and ChemBench, and gains further improvements under test-time scaling methods like majority voting and pass@k. Together, these results establish SDAR as a practical paradigm that combines the strengths of autoregression and diffusion for scalable, high-throughput reasoning.

cs.LG↗

arXiv2025

Reasoning-Enhanced Large Language Models for Molecular Property Prediction

Molecular property prediction is crucial for drug discovery and materials science, yet existing approaches suffer from limited interpretability, poor cross-task generalization, and lack of chemical reasoning capabilities. Traditional machine learning models struggle with task transferability, while specialized molecular language models provide little insight into their decision-making processes. To address these limitations, we propose \textbf{MPPReasoner}, a multimodal large language model that incorporates chemical reasoning for molecular property prediction. Our approach, built upon Qwen2.5-VL-7B-Instruct, integrates molecular images with SMILES strings to enable comprehensive molecular understanding. We develop a two-stage training strategy: supervised fine-tuning (SFT) using 16,000 high-quality reasoning trajectories generated through expert knowledge and multiple teacher models, followed by Reinforcement Learning from Principle-Guided Rewards (RLPGR). RLPGR employs verifiable, rule-based rewards that systematically evaluate chemical principle application, molecular structure analysis, and logical consistency through computational verification. Extensive experiments across 8 datasets demonstrate significant performance improvements, with MPPReasoner outperforming the best baselines by 7.91\% and 4.53\% on in-distribution and out-of-distribution tasks respectively. MPPReasoner exhibits exceptional cross-task generalization and generates chemically sound reasoning paths that provide valuable insights into molecular property analysis, substantially enhancing both interpretability and practical utility for chemists. Code is available at https://anonymous.4open.science/r/MPPReasoner-12687.

cs.LG↗

arXiv2025

Innovator: Scientific Continued Pretraining with Fine-grained MoE Upcycling

A large language model (LLM) with knowledge in both scientific and general tasks is the foundation of science general intelligence. However, directly continued pretraining an LLM using science data usually leads to catastrophic forgetting, which indicates severe degradation in general ability. In this report, we present Innovator, which solves this problem by upcycling a pre-trained dense LLM into a fine-grained Mixtures-of-Experts model during continued pretraining, where different experts are expected to learn science knowledge in different disciplines, and a shared expert is utilized for general tasks. Innovator introduces a four-stage upcycle training paradigm: (1) Scientific Expert Induction on discipline-specific data, (2) Fine-grained Expert Splitting via FFN dimension decomposition, (3) Science-Aware Routing warmup, and (4) Generalist-Scientist Integration training on hybrid datasets. Such a paradigm enables knowledge in the general domain, and different scientific disciplines can be decoupled, avoiding the negative influence among knowledge in different domains. With 53.3B total parameters and 13.3B activated, Innovator extends Qwen2.5-7B using a shared general expert and 64 specialized scientific experts with 8 activated. Trained on 300B tokens with tri-level quality-controlled data, Innovator achieves 25% average improvement across 30 scientific tasks with a win rate as 70%, while retaining 99% performance in general tasks. Furthermore, Innovator-Reason, which is post-trained from Innovator for reasoning boosting, exhibits excellent reasoning performance in solving complex scientific problems with improvements over 30%.

cs.LG↗

arXiv2025

AI for Service: Proactive Assistance with AI Glasses

In an era where AI is evolving from a passive tool into an active and adaptive companion, we introduce AI for Service (AI4Service), a new paradigm that enables proactive and real-time assistance in daily life. Existing AI services remain largely reactive, responding only to explicit user commands. We argue that a truly intelligent and helpful assistant should be capable of anticipating user needs and taking actions proactively when appropriate. To realize this vision, we propose Alpha-Service, a unified framework that addresses two fundamental challenges: Know When to intervene by detecting service opportunities from egocentric video streams, and Know How to provide both generalized and personalized services. Inspired by the von Neumann computer architecture and based on AI glasses, Alpha-Service consists of five key components: an Input Unit for perception, a Central Processing Unit for task scheduling, an Arithmetic Logic Unit for tool utilization, a Memory Unit for long-term personalization, and an Output Unit for natural human interaction. As an initial exploration, we implement Alpha-Service through a multi-agent system deployed on AI glasses. Case studies, including a real-time Blackjack advisor, a museum tour guide, and a shopping fit assistant, demonstrate its ability to seamlessly perceive the environment, infer user intent, and provide timely and useful assistance without explicit prompts.

cs.AI↗

arXiv2025

Shifting AI Efficiency From Model-Centric to Data-Centric Compression

The advancement of large language models (LLMs) and multi-modal LLMs (MLLMs) has historically relied on scaling model parameters. However, as hardware limits constrain further model growth, the primary computational bottleneck has shifted to the quadratic cost of self-attention over increasingly long sequences by ultra-long text contexts, high-resolution images, and extended videos. In this position paper, \textbf{we argue that the focus of research for efficient artificial intelligence (AI) is shifting from model-centric compression to data-centric compression}. We position data-centric compression as the emerging paradigm, which improves AI efficiency by directly compressing the volume of data processed during model training or inference. To formalize this shift, we establish a unified framework for existing efficiency strategies and demonstrate why it constitutes a crucial paradigm change for long-context AI. We then systematically review the landscape of data-centric compression methods, analyzing their benefits across diverse scenarios. Finally, we outline key challenges and promising future research directions. Our work aims to provide a novel perspective on AI efficiency, synthesize existing efforts, and catalyze innovation to address the challenges posed by ever-increasing context lengths.

cs.CL↗

arXiv2025

Rethinking LLM Evaluation: Can We Evaluate LLMs with 200x Less Data?

As the demand for comprehensive evaluations of diverse model capabilities steadily increases, benchmark suites have correspondingly grown significantly in scale. Despite notable advances in redundancy reduction and subset-level performance prediction, a systematic framework that effectively integrates these methods to ensure both prediction accuracy and ranking consistency is still largely elusive. In this paper, we first perform a sample-level analysis of benchmark redundancy and identify several highly similar samples that can be eliminated. Besides, we frame benchmark compression as an optimization problem with the aim of score reconstruction. Building on these, we then propose EssenceBench, a coarse-to-fine framework utilizing an iterative Genetic Algorithm (GA), which takes the advantages of fitness-based subset search and attribution-based sample search. Compared to previous methods, our approach yields superior compression results with lower reconstruction error and markedly higher efficiency. In particular, on the HellaSwag benchmark (10K samples), our method preserves the ranking of all models shifting within 5% using 25x fewer samples, and achieves 95% ranking preservation shifting within 5% using only 200x fewer samples.

cs.CL↗

arXiv2025

OmniSAM: Omnidirectional Segment Anything Model for UDA in Panoramic Semantic Segmentation

Segment Anything Model 2 (SAM2) has emerged as a strong base model in various pinhole imaging segmentation tasks. However, when applying it to $360^\circ$ domain, the significant field-of-view (FoV) gap between pinhole ($70^\circ \times 70^\circ$) and panoramic images ($180^\circ \times 360^\circ$) poses unique challenges. Two major concerns for this application includes 1) inevitable distortion and object deformation brought by the large FoV disparity between domains; 2) the lack of pixel-level semantic understanding that the original SAM2 cannot provide. To address these issues, we propose a novel OmniSAM framework, which makes the first attempt to apply SAM2 for panoramic semantic segmentation. Specifically, to bridge the first gap, OmniSAM first divides the panorama into sequences of patches. These patches are then treated as image sequences in similar manners as in video segmentation tasks. We then leverage the SAM2's memory mechanism to extract cross-patch correspondences that embeds the cross-FoV dependencies, improving feature continuity and the prediction consistency along mask boundaries. For the second gap, OmniSAM fine-tunes the pretrained image encoder and reutilize the mask decoder for semantic prediction. An FoV-based prototypical adaptation module with dynamic pseudo label update mechanism is also introduced to facilitate the alignment of memory and backbone features, thereby improving model generalization ability across different sizes of source models. Extensive experimental results demonstrate that OmniSAM outperforms the state-of-the-art methods by large margins, e.g., 79.06% (+10.22%) on SPin8-to-SPan8, 62.46% (+6.58%) on CS13-to-DP13.

cs.CV↗

arXiv2025

Thinking Inside the Mask: In-Place Prompting in Diffusion LLMs

Despite large language models (LLMs) have achieved remarkable success, their prefix-only prompting paradigm and sequential generation process offer limited flexibility for bidirectional information. Diffusion large language models (dLLMs) present new opportunities through their bidirectional attention mechanisms and iterative refinement processes, enabling more flexible in-place prompting strategies. We introduce ICE (In-Place Chain-of-Thought Prompting with Early Exit), a novel framework that transforms prefix-only prompting into in-place prompting specifically designed for dLLMs. ICE integrates in-place prompts directly within masked token positions during iterative refinement and employs a confidence-aware early exit mechanism to significantly reduce computational overhead. Extensive experiments demonstrate ICE's effectiveness, achieving up to 17.29% accuracy improvement with 4.12$\times$ speedup on GSM8K, and up to 276.67$\times$ acceleration on MMLU while maintaining competitive performance.

cs.CL↗

arXiv2025

Don't Just Chase "Highlighted Tokens" in MLLMs: Revisiting Visual Holistic Context Retention

Despite their powerful capabilities, Multimodal Large Language Models (MLLMs) suffer from considerable computational overhead due to their reliance on massive visual tokens. Recent studies have explored token pruning to alleviate this problem, which typically uses text-vision cross-attention or [\texttt{CLS}] attention to assess and discard redundant visual tokens. In this work, we identify a critical limitation of such attention-first pruning approaches, i.e., they tend to preserve semantically similar tokens, resulting in pronounced performance drops under high pruning ratios. To this end, we propose {HoloV}, a simple yet effective, plug-and-play visual token pruning framework for efficient inference. Distinct from previous attention-first schemes, HoloV rethinks token retention from a holistic perspective. By adaptively distributing the pruning budget across different spatial crops, HoloV ensures that the retained tokens capture the global visual context rather than isolated salient features. This strategy minimizes representational collapse and maintains task-relevant information even under aggressive pruning. Experimental results demonstrate that our HoloV achieves superior performance across various tasks, MLLM architectures, and pruning ratios compared to SOTA methods. For instance, LLaVA1.5 equipped with HoloV preserves 95.8\% of the original performance after pruning 88.9\% of visual tokens, achieving superior efficiency-accuracy trade-offs.

cs.CV↗

arXiv2025

Mask Tokens as Prophet: Fine-Grained Cache Eviction for Efficient dLLM Inference

Diffusion large language models (dLLMs) present a promising alternative to dominant autoregressive models (ARMs) by the ability of parallel decoding at the expense of substantial computation and memory costs. Specifically, the cache mechanism for bidirectional attention in dLLMs demands large memory footprint, restricting their ability to handle long contexts under resource-limited settings. Existing cache eviction strategies are designed for ARMs and ignore the unique characteristics of dLLMs, thus leading to unsatisfactory performance. To address these challenges, we introduce MaskKV, a training-free cache eviction framework tailored to dLLMs, focusing on the effect of mask tokens in dLLMs. MaskKV is built on two key innovations: (1) a mask-query guided scoring mechanism that leverages attention weights to identify and evict less critical prompt tokens for each head; (2) an adaptive cache budgeting strategy that improves efficiency by reducing allocation in intermediate layers and concentrating resources on prompt-preferring heads. On LLaDA with MaskKV, compressing the KV cache to only 256 pairs (less than 5% of tokens) retains 94% of the full-cache performance on LongBench and achieves up to 31x acceleration at 32k prompt length. The code is publicly available at: https://github.com/jianuo-huang/MaskKV

cs.CL↗

arXiv2025

Unified Cross-Scale 3D Generation and Understanding via Autoregressive Modeling

3D structure modeling is essential across scales, enabling applications from fluid simulation and 3D reconstruction to protein folding and molecular docking. Yet, despite shared 3D spatial patterns, current approaches remain fragmented, with models narrowly specialized for specific domains and unable to generalize across tasks or scales. We propose Uni-3DAR, a unified autoregressive framework for cross-scale 3D generation and understanding. At its core is a coarse-to-fine tokenizer based on octree data structures, which compresses diverse 3D structures into compact 1D token sequences. We further propose a two-level subtree compression strategy, which reduces the octree token sequence by up to 8x. To address the challenge of dynamically varying token positions introduced by compression, we introduce a masked next-token prediction strategy that ensures accurate positional modeling, significantly boosting model performance. Extensive experiments across multiple 3D generation and understanding tasks, including small molecules, proteins, polymers, crystals, and macroscopic 3D objects, validate its effectiveness and versatility. Notably, Uni-3DAR surpasses previous state-of-the-art diffusion models by a substantial margin, achieving up to 256\% relative improvement while delivering inference speeds up to 21.8x faster.

cs.LG↗

arXiv2025

FreqCa: Accelerating Diffusion Models via Frequency-Aware Caching

The application of diffusion transformers is suffering from their significant inference costs. Recently, feature caching has been proposed to solve this problem by reusing features from previous timesteps, thereby skipping computation in future timesteps. However, previous feature caching assumes that features in adjacent timesteps are similar or continuous, which does not always hold in all settings. To investigate this, this paper begins with an analysis from the frequency domain, which reveal that different frequency bands in the features of diffusion models exhibit different dynamics across timesteps. Concretely, low-frequency components, which decide the structure of images, exhibit higher similarity but poor continuity. In contrast, the high-frequency bands, which decode the details of images, show significant continuity but poor similarity. These interesting observations motivate us to propose Frequency-aware Caching (FreqCa) which directly reuses features of low-frequency components based on their similarity, while using a second-order Hermite interpolator to predict the volatile high-frequency ones based on its continuity. Besides, we further propose to cache Cumulative Residual Feature (CRF) instead of the features in all the layers, which reduces the memory footprint of feature caching by 99%. Extensive experiments on FLUX.1-dev, FLUX.1-Kontext-dev, Qwen-Image, and Qwen-Image-Edit demonstrate its effectiveness in both generation and editing. Codes are available in the supplementary materials and will be released on GitHub.

cs.LG↗

arXiv2025

AudioMarathon: A Comprehensive Benchmark for Long-Context Audio Understanding and Efficiency in Audio LLMs

Processing long-form audio is a major challenge for Large Audio Language models (LALMs). These models struggle with the quadratic cost of attention ($O(N^2)$) and with modeling long-range temporal dependencies. Existing audio benchmarks are built mostly from short clips and do not evaluate models in realistic long context settings. To address this gap, we introduce AudioMarathon, a benchmark designed to evaluate both understanding and inference efficiency on long-form audio. AudioMarathon provides a diverse set of tasks built upon three pillars: long-context audio inputs with durations ranging from 90.0 to 300.0 seconds, which correspond to encoded sequences of 2,250 to 7,500 audio tokens, respectively, full domain coverage across speech, sound, and music, and complex reasoning that requires multi-hop inference. We evaluate state-of-the-art LALMs and observe clear performance drops as audio length grows. We also study acceleration techniques and analyze the trade-offs of token pruning and KV cache eviction. The results show large gaps across current LALMs and highlight the need for better temporal reasoning and memory-efficient architectures. We believe AudioMarathon will drive the audio and multimodal research community to develop more advanced audio understanding models capable of solving complex audio tasks.

cs.SD↗

arXiv2025

LMM-Incentive: Large Multimodal Model-based Incentive Design for User-Generated Content in Web 3.0

Web 3.0 represents the next generation of the Internet, which is widely recognized as a decentralized ecosystem that focuses on value expression and data ownership. By leveraging blockchain and artificial intelligence technologies, Web 3.0 offers unprecedented opportunities for users to create, own, and monetize their content, thereby enabling User-Generated Content (UGC) to an entirely new level. However, some self-interested users may exploit the limitations of content curation mechanisms and generate low-quality content with less effort, obtaining platform rewards under information asymmetry. Such behavior can undermine Web 3.0 performance. To this end, we propose \textit{LMM-Incentive}, a novel Large Multimodal Model (LMM)-based incentive mechanism for UGC in Web 3.0. Specifically, we propose an LMM-based contract-theoretic model to motivate users to generate high-quality UGC, thereby mitigating the adverse selection problem from information asymmetry. To alleviate potential moral hazards after contract selection, we leverage LMM agents to evaluate UGC quality, which is the primary component of the contract, utilizing prompt engineering techniques to improve the evaluation performance of LMM agents. Recognizing that traditional contract design methods cannot effectively adapt to the dynamic environment of Web 3.0, we develop an improved Mixture of Experts (MoE)-based Proximal Policy Optimization (PPO) algorithm for optimal contract design. Simulation results demonstrate the superiority of the proposed MoE-based PPO algorithm over representative benchmarks in the context of contract design. Finally, we deploy the designed contract within an Ethereum smart contract framework, further validating the effectiveness of the proposed scheme.

cs.AI↗

arXiv2025

Self Speculative Decoding for Diffusion Large Language Models

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.

cs.CL↗

arXiv2025

Let Features Decide Their Own Solvers: Hybrid Feature Caching for Diffusion Transformers

Diffusion Transformers offer state-of-the-art fidelity in image and video synthesis, but their iterative sampling process remains a major bottleneck due to the high cost of transformer forward passes at each timestep. To mitigate this, feature caching has emerged as a training-free acceleration technique that reuses or forecasts hidden representations. However, existing methods often apply a uniform caching strategy across all feature dimensions, ignoring their heterogeneous dynamic behaviors. Therefore, we adopt a new perspective by modeling hidden feature evolution as a mixture of ODEs across dimensions, and introduce HyCa, a Hybrid ODE solver inspired caching framework that applies dimension-wise caching strategies. HyCa achieves near-lossless acceleration across diverse domains and models, including 5.55 times speedup on FLUX, 5.56 times speedup on HunyuanVideo, 6.24 times speedup on Qwen-Image and Qwen-Image-Edit without retraining.

cs.CV↗

arXiv2025

MASC: Boosting Autoregressive Image Generation with a Manifold-Aligned Semantic Clustering

Autoregressive (AR) models have shown great promise in image generation, yet they face a fundamental inefficiency stemming from their core component: a vast, unstructured vocabulary of visual tokens. This conventional approach treats tokens as a flat vocabulary, disregarding the intrinsic structure of the token embedding space where proximity often correlates with semantic similarity. This oversight results in a highly complex prediction task, which hinders training efficiency and limits final generation quality. To resolve this, we propose Manifold-Aligned Semantic Clustering (MASC), a principled framework that constructs a hierarchical semantic tree directly from the codebook's intrinsic structure. MASC employs a novel geometry-aware distance metric and a density-driven agglomerative construction to model the underlying manifold of the token embeddings. By transforming the flat, high-dimensional prediction task into a structured, hierarchical one, MASC introduces a beneficial inductive bias that significantly simplifies the learning problem for the AR model. MASC is designed as a plug-and-play module, and our extensive experiments validate its effectiveness: it accelerates training by up to 57% and significantly improves generation quality, reducing the FID of LlamaGen-XL from 2.87 to 2.58. MASC elevates existing AR frameworks to be highly competitive with state-of-the-art methods, establishing that structuring the prediction space is as crucial as architectural innovation for scalable generative modeling.

cs.CV↗

arXiv2025

Efficient Multi-modal Large Language Models via Progressive Consistency Distillation

Visual tokens consume substantial computational resources in multi-modal large models (MLLMs), significantly compromising their efficiency. Recent works have attempted to improve efficiency by compressing visual tokens during training, either through modifications to model components or by introducing additional parameters. However, they often overlook the increased learning difficulty caused by such compression, as the model's parameter space struggles to quickly adapt to the substantial perturbations in the feature space induced by token compression. In this work, we propose to develop Efficient MLLMs via Progressive Consistency Distillation (EPIC), a progressive learning framework. Specifically, by decomposing the feature space perturbations introduced by token compression along the token-wise and layer-wise dimensions, we introduce token consistency distillation and layer consistency distillation, respectively, aiming to reduce the training difficulty by leveraging guidance from a teacher model and following a progressive learning trajectory. Extensive experiments demonstrate the superior effectiveness, robustness, and generalization capabilities of our proposed framework.

cs.CV↗