论文检索

Sang-Hoon Lee, Ha-Yeong Choi

Despite recent progress in diffusion and conditional flow matching (CFM) models for low-resolution domains such as latent representations, their application to high-resolution data like raw waveform signals remains underexplored. Generative adversarial networks (GANs) have been the dominant approach in neural vocoder and neural audio codecs for realistic waveform generation. However, under low-bitrate conditions, these models suffer from degraded performance due to information loss caused by heavy compression and quantization, often resulting in mispronunciations. To address the aforementioned problem, we first leverage CFM to iteratively generate raw waveform in an extremely low-bitrate scenario. We then introduce hierarchical representation alignment learning (REPA-H) to enable efficient and robust CFM training. Furthermore, we propose dense vector quantization (DVQ), a novel factorized quantization method using a single quantizer. Our model, FlowTokenizer, outperforms state-of-the-art neural audio codecs in audio quality and semantic intelligibility under low-bitrate conditions, using only 25 tokens per second for 24 kHz waveform generation.

Kwok-Ho Ng, Tingting Song, Yongdong WU, Zhihua Xia

Advanced speech synthesis technologies have enabled highly realistic speech generation, posing security risks that motivate research into audio deepfake detection (ADD). While state space models (SSMs) offer linear complexity, pure causal SSMs architectures often struggle with the content-based retrieval required to capture global frequency-domain artifacts. To address this, we explore the scaling properties of hybrid architectures by proposing XLSR-MamBo, a modular framework integrating an XLSR front-end with synergistic Mamba-Attention backbones. We systematically evaluate four topological designs using advanced SSM variants, Mamba, Mamba2, Hydra, and Gated DeltaNet. Experimental results demonstrate that the MamBo-3-Hydra-N3 configuration achieves competitive performance compared to other state-of-the-art systems on the ASVspoof 2021 LA, DF, and In-the-Wild benchmarks. This performance benefits from Hydra’s native bidirectional modeling, which captures holistic temporal dependencies more efficiently than the heuristic dual-branch strategies employed in prior works. Furthermore, evaluations on the DFADD dataset demonstrate robust generalization to unseen diffusion- and flow-matching-based synthesis methods. Crucially, our analysis reveals that scaling backbone depth effectively mitigates the performance variance and instability observed in shallower models. These results demonstrate the hybrid framework’s ability to capture artifacts in spoofed speech signals, providing an effective method for ADD. Codes are publicly available at https://github.com/saki-ciallo/XLSR-MamBo.

Haoyang Chen, Yi Liu, Jianzhi Shao, Tao Zhang, Chengfu Huo, Wei Hu

Thinking LLMs produce reasoning traces before answering. Prior activation steering work mainly targets on shaping these traces. It remains less understood how answer tokens actually read and integrate the reasoning to produce reliable outcomes. Focusing on quantitative reasoning, we analyze the answer-to-reasoning attention and observe a benign self-reading pattern aligned with correctness, characterized by a forward drift of the reading focus along the reasoning trace and a persistent concentration on key semantic anchors, whereas incorrect solutions exhibit diffuse and irregular attention pattern. We interpret this as internal certainty during answer decoding, where the model commits to a viable solution branch and integrates key evidence. Following this, we propose a training-free steering method driven by Self-Reading Quality (SRQ) scores combining geometric metrics for process control with semantic metrics for content monitoring. SRQ selects data to build steering vectors that guide inference toward benign self-reading and away from uncertain and disorganized reading. Experiments show that our method yields consistent accuracy gains.

Zhenqin Li, ShengYong Ding, Shuangyin Li

Question-Answer Generation (QAG) is essential for alleviating the cold-start problem in domain-specific large language model (LLM) post-training, where high-quality data is severely scarce.Effective training samples include rich semantic diversity and rigorous factual consistency.Thus, it is necessary to consider the inherent tension between semantic breadth and factual fidelity.However, it is extremely challenging to trade off semantic diversity against factual consistency, in that generalization across the semantic space must be achieved effectively and reliably, and factual integrity must be ensured as well.To address this issue, we propose an effective framework, namely DiFRa, that integrates continuous concept diffusion with discrete knowledge graph constraints to balance semantic diversity and factual consistency.Specifically, the proposed DiFRa models discrete concepts as a continuous latent distribution to sample embeddings that capture rich semantic variations, and constructs a refined knowledge graph as explicit factual constraints.Then, a diversity and consistency aware mechanism is designed to dynamically integrate both embeddings and the knowledge graph for QA pairs generation.Furthermore, we introduce SeFa, which harmonizes semantic entropy and consistency scores to quantify the trade-off between diversity and correctness.Extensive experiments demonstrate that DiFRa consistently outperforms the baseline models, validating its efficacy in reconciling the tension to generate semantically diverse and factually consistent QA pairs. The source code is publicly available.

Jingfan Yang, Rui Zhang, Liang Hong, Ke Yuan

Multimodal learning aims to learn unified multimodal representations from heterogeneous modalities and supports many natural language processing tasks. However, multimodal models often exhibit modality laziness: over-relying on a dominant modality and under-exploiting complementary signals. Existing approaches typically strengthen unimodal training or rebalance modality contributions, but they may still emphasize shared semantics and overlook modality-specific cues. To address this, we propose SCOPE, a unified framework for learning complete multimodal representations, achieving Shared-and-COmplementary cue PrEservation. Firstly, SCOPE uses a mutual information-guided disentanglement module to separate shared semantics from modality-specific cues and mitigate representation collapse. Secondly, SCOPE aligns modalities by enforcing structural consistency between modality-wise semantic graphs, avoiding brittle point-wise matching. Finally, SCOPE performs balanced fusion via structure-aware diffusion attention to integrate shared and complementary cues without feature homogenization. Experiments on four benchmark datasets show that SCOPE consistently outperforms SOTA baselines, achieving up to 27.10% accuracy improvement.

Shenyang Chen, Liuwan Zhu

Standard evaluations of backdoor attacks on text-to-image (T2I) models primarily measure trigger activation and visual fidelity. We challenge this paradigm, demonstrating that encoder-side poisoning induces persistent, trigger-free semantic corruption that fundamentally reshapes the representation manifold. We trace this vulnerability to a geometric mechanism: a Jacobian-based analysis reveals that backdoors act as low-rank, target-centered deformations that amplify local sensitivity, causing distortion to propagate coherently across semantic neighborhoods. To rigorously quantify this structural degradation, we introduce SEMAD (Semantic Alignment and Drift), a diagnostic framework that measures both internal embedding drift and downstream functional misalignment. Our findings, validated across diffusion and contrastive paradigms, expose the deep structural risks of encoder poisoning and highlight the necessity of geometric audits beyond simple attack success rates.

Huaisheng Zhu, MingYu Liu, Junze Liu, Zhen Ge, Tian Wang, Jiri Gesi, Dakuo Wang, Weiqi Zhang, Houyu Zhang, Yufan Guo 等

Diffusion Large Language Models (DLLMs) have recently achieved strong performance, e.g., masked diffusion models (MDMs) can surpass autoregressive models (ARMs) in various tasks. However, DLLMs often struggle with inaccurate early-stage predictions due to limited context, which hinders both the model’s inference efficiency and the output’s overall quality. We propose Calibrated On-Policy Self-Distillation (COPSD) for DLLMs, a simple and efficient method to calibrate early token predictions without requiring demonstration data. COPSD distills an unnormalized target distribution derived from later decoding steps into the original model, enabling more accurate early predictions during inference. Experiments on math, planning, and RLHF tasks show that COPSD improves both effectiveness and efficiency, and further enhances performance when combined with supervised fine-tuning.

Maisha Maliha, Dean F. Hougen

Text-to-image diffusion models achieve remarkable generation quality, yet their internal mechanisms for grounding prompt semantics into visual structure remain poorly understood. We present a novel mechanistic interpretability framework for Stable Diffusion that probes how individual prompt tokens are represented and utilized during the denoising process. Given a prompt, we record cross-attention activations throughout UNet denoising and convert them into token-level spatial grounding maps that indicate where each token contributes signal during image synthesis. To establish causal faithfulness, we perform controlled prompt interventions by removing a single word at a time while keeping the sampling seed fixed, producing counterfactual generations. To quantify mechanistic sensitivity, we introduce a head-resolved spike score based on divergence between per-head token contribution distributions before and after intervention, enabling module-wise and head-wise attribution of semantic changes. Experiments on compositional prompts and challenging relational descriptions reveal systematic patterns of token grounding, semantic drift, and head specialization across denoising timesteps. Our results provide a practical and reproducible toolkit for analyzing how diffusion models encode and apply semantic information, supporting deeper transparency in text-to-image generation.

Yangryeol Park, Kunhui Lee, Hanback Choi, Cheoneum Park, Donghyeon Jeon, Inho Kang, Seung-Hoon Na

Masked diffusion language models (MDLMs) enable efficient parallel decoding but are limited by a monotonic unmasking policy, where committed tokens cannot be revised. While remasking-based methods mitigate early errors, they mainly intervene during generation. In this work, we study post-hoc refinement of a completed draft and find that naive correction often fails because of contextual lock-in, a phenomenon in which local error patterns become self-reinforcing. To address this, we propose PURE (Post-hoc Unlocking and REfinement), a training-free inference algorithm for two-phase decoding. PURE profiles confidence dynamics during drafting to identify unstable regions via an instability score (\Delta_i), then unlocks them through deterministic window masking and stochastic leftward relaxation. On reasoning benchmarks, PURE substantially improves accuracy when applied to LLaDA-8B-Instruct, including a gain of +12.9 points over the baseline on GSM8K. These gains require only a small refinement budget, yielding a favorable compute-quality trade-off for discrete diffusion decoding.

Kewei Chen, Yayu Long, Shuai Li, Mingsheng Shang

Diffusion policies have demonstrated exceptional performance in embodied AI. However, their iterative denoising process results in high latency, and existing acceleration methods often sacrifice physical consistency. To address this, we propose ElasticFlow, a distillation-free, physics-consistent one-step policy framework. We reconstruct the Mean Field Theory by directly modeling the average velocity field, enabling a direct single-step mapping from noise to action. Addressing the Temporal Heterogeneity of robotic tasks, we introduce the Elastic Time Horizons mechanism. This mechanism effectively overcomes Spectral Bias by explicitly encoding control granularity, achieving efficient alignment between semantic instructions and physical execution horizons. Experiments on benchmarks such as LIBERO, CALVIN, and RoboTwin demonstrate that ElasticFlow achieves efficient 1-NFE inference (approximately 71Hz). Furthermore, it outperforms state-of-the-art methods, including OpenVLA and \pi_0, on long-horizon tasks, highlighting its potential for efficient, robust, and semantically aligned control.

Zheng He, Yiwei Wang, Hongxing Wang, Yujun Cai

Large Vision-Language Models (LVLMs) confront an escalating threat from sophisticated multimodal jailbreak attacks. However, existing defense strategies suffer from three critical limitations: (1) the neglect of visual threats; (2) a lack of fine-grained specificity regarding specific attack semantics; and (3) the absence of a dedicated jailbreak detection mechanism, which leads to unnecessary defensive measures against benign inputs. To address these limitations, we propose ReCon, a novel black-box defense framework. ReCon integrates a diffusion-based image purifier to neutralize visual perturbations and an autoencoder-based detector for anomaly filtration. At its core, it employs a Reverse Safety Concept Injection module that maps detected unsafe concepts to fine-grained, constructive Safe Concepts, generating targeted prompts to precisely rectify attack semantics. Extensive experiments demonstrate that ReCon significantly enhances the robustness of LVLMs against jailbreak attacks while preserving performance on benign tasks. Disclaimer: Samples in this paper may be harmful and cause discomfort.

Yanru Huo, Ziyue Jiang, Zuoli Tang, Qingyang Hong, Zhou Zhao

While Diffusion Transformers (DiT) have advanced non-autoregressive (NAR) speech synthesis, their high computational demands remain an obvious limitation. Existing DiT-based text-to-speech (TTS) model acceleration approaches predominantly focus on reducing sampling steps through distillation techniques, yet they remain constrained by training costs. We introduce DiTReducio, a training-free acceleration framework that compresses computations in DiT-based TTS models through a progressive calibration process. We propose two compression methods, Temporal Skipping and Branch Skipping, to eliminate redundant computations during inference. Moreover, based on two characteristic attention patterns identified within DiT layers, we devise a pattern-guided strategy to selectively apply the compression methods. Our method allows flexible modulation between generation quality and computational efficiency through adjustable compression thresholds. Experimental evaluations conducted on F5-TTS and MegaTTS 3 demonstrate that DiTReducio achieves a 75.4% reduction in FLOPs and improves the Real-Time Factor (RTF) by 37.1%, while preserving generation quality. The code is available at https://github.com/MM-Speech/DiTReducio.

Qinglin Zeng, Jusheng Zhang, Jing Yang, Ningyuan Liu, Keze Wang

Masked Discrete Diffusion Models (MDMs) enable parallel generation via iterative refinement. However, we identify a critical decisional mismatch. The MDM architecture is inherently dynamic and capable of sensing context shifts. In contrast, prevailing decoding paradigms remain static and myopic. They treat each denoising step as an isolated snapshot, effectively discarding valuable temporal feedback that signals logical conflicts. To bridge this gap, we propose Regret-Aware Confidence Calibration (RACC). This training-free framework aligns decoding decisions with the model’s latent self-correction capabilities. RACC introduces a momentum anchor to track confidence trajectories. When a token’s probability drops abruptly below its historical trend, the system triggers a "regret" signal. Unlike expensive re-masking or lookahead search, RACC utilizes this signal to proactively demote unstable candidates. Extensive experiments on reasoning benchmarks, such as HumanEval and GSM8K, demonstrate that RACC significantly improves generation consistency. Crucially, RACC achieves these gains with zero additional inference overhead, effectively balancing decoding quality and efficiency.

Shuang Cheng, Yihan Bian, Dawei Liu, Yuhua Jiang, Yihao Liu, Linfeng Zhang, Qian Yao, Zhongbo Tian, Wenhai Wang, Qipeng Guo 等

Autoregressive (AR) language modeling remains the dominant paradigm due to its dense supervision signal and highly optimized serving infrastructure, but its strictly causal, token-by-token decoding limits parallelism and non-causal modeling. While masked diffusion offers a promising path toward parallel generation, it faces two critical bottlenecks: training inefficiency stemming from sparse masked objectives, and high latency caused by iterative whole-sequence denoising. We present a systematic study of blockwise discrete diffusion, a pragmatic middle ground that preserves AR-compatible serving while enabling parallel intra-block generation. Our study proceeds in four steps: (i) a controlled, compute- and scale-matched comparison revealing that AR is a more effective backbone for blockwise hybrids than masked diffusion objectives; (ii) a scalable conversion recipe, SDAR, validating that AR models spanning 1.7B to 30B parameters can be adapted into block diffusion models with minimal compute while preserving backbone capabilities; and (iii) a systematic characterization of decoding dynamics, which reveals a virtuous cycle where larger models enable more aggressive parallel decoding, achieving theoretical speedups over 5\times and wall-clock speedups of 2.3\times on H200 GPUs in latency-critical regimes; and (iv) an investigation of local non-causal modeling capabilities, showing that SDAR’s local bidirectional attention overcomes causal bottlenecks in scientific domains (e.g., chemistry) and enables robust test-time scaling. We release the full model suite, the training framework, and our inference engines for further innovation in non-autoregressive generative paradigms.

Zijing Wang, YongKang Liu, Mingyang Wang, Ercong Nie, Deyuan Chen, Zhengjie Zhao, Shi Feng, Daling Wang, Xiaocui Yang, Yifei Zhang 等

Multimodal Large Language Models (MLLMs) rely on strong linguistic reasoning inherited from their base language models. However, multimodal instruction fine-tuning paradoxically degrades this text’s reasoning capability, undermining multimodal performance. To address this issue, we propose a training-free framework to mitigate this degradation. Through layer-wise vision token masking, we reveal a common three-stage pattern in multimodal large language models: early-modal separation, mid-modal alignment, and late-modal degradation. By analyzing the behavior of MLLMs at different stages, we propose a plateau-guided model merging method that selectively injects base language model parameters into MLLMs. Experimental results based on five MLLMs on nine benchmarks demonstrate the effectiveness of our method. Attention-based analysis further reveals that merging shifts attention from diffuse, scattered patterns to focused localization on task-relevant visual regions.Our repository is on https://github.com/wzj1718/PlaM .

Shaokai He, Kaiwen Wei, Xinyi Zeng, Xiang Chen, Xue Yang, Zhenyang Li, Jiang Zhong, Yu Tian

The "reversal curse" refers to the phenomenon where large language models (LLMs) exhibit predominantly unidirectional behavior when processing logically bidirectional relationships. Prior work attributed this to autoregressive training—predicting the next token inherently favors left-to-right information flow over genuine bidirectional knowledge associations. However, we observe that Diffusion LLMs (DLLMs), despite being trained bidirectionally, also suffer from the reversal curse. To investigate the root causes, we conduct systematic experiments on DLLMs and identify three key reasons: 1) entity fragmentation during training, 2) data asymmetry, and 3) missing entity relations. Motivated by the analysis of these reasons, we propose Diffusion Entity-Relation Modeling (DiffER), which addresses the reversal curse through entity-aware training and balanced data construction. Specifically, DiffER introduces whole-entity masking, which mitigates entity fragmentation by predicting complete entities in a single step. DiffER further employs distribution-symmetric and relation-enhanced data construction strategies to alleviate data asymmetry and missing relations. Extensive experiments demonstrate that DiffER effectively alleviates the reversal curse in Diffusion LLMs, offering new perspectives for future research. The code is available at https://github.com/CQU-MM-Intelligent-Lab/DiffER.

Guanghao Li, Zhihui Fu, Min Fang, Qibin Zhao, Ming Tang, Chun Yuan, Jun Wang

Autoregressive (AR) decoding in large language models (LLMs) is latency-bounded by strictly sequential token generation.Speculative decoding mitigates this bottleneck by letting a fast drafter propose multi-token candidates that are then verified in parallel by the target model; yet most existing systems still rely on AR drafters, limiting wall-clock gains.We present **DiffuSpec**, which repurposes a *diffusion language model* (DLM) as a *parallel* drafter to generate multi-token proposals in a single forward pass while remaining compatible with standard AR verifiers.However, DLM drafting presents unique challenges: 1) bidirectional conditioning produces a token lattice where locally optimal tokens may fail to form a valid causal sequence; 2) the mechanism requires tuning the draft length, which induces a speed–quality trade-off. To address these issues, we introduce (i) *Causal-consistency Path Search* (CPS) to extract verifier-aligned causal paths from the lattice, and (ii) an *Adaptive Draft-Length* (ADL) controller that adjusts proposal lengths using online acceptance feedback.Across benchmarks, DiffuSpec achieves up to 3\times wall-clock speedup and consistently outperforms strong baselines, demonstrating diffusion-based drafting as a competitive alternative to AR drafters for speculative decoding.

Ruixuan Xu, Jiexi Xu, Qiyan Zhao, Xiaofeng Zhang

Recent advances in diffusion-based Multimodal Large Language Models (dMLLMs) offer a compelling alternative to autoregressive counterparts; however, they remain prone to hallucinations. Through information flow analysis on LLaDA-V, we identify two intertwined factors contributing to this issue. First, although the special tokens serve as semantic anchors for aggregating visual information, they simultaneously induce severe attention sinks, excessively consuming the model’s attention budget. Second, the long-range decay inherent in Rotary Position Embedding (RoPE) leads to semantic blind spots, preventing these anchors from uniformly perceiving the entire visual input. Accordingly, our objective is to moderately alleviate the attention sink effect on semantic anchors while enhancing their ability to aggregate global visual information, thereby eliminating semantic blind spots. To this end, we propose Extrinsic Distance-Aware Regularization (EDAR), a training-free decoding strategy that augments the attention key space with a static, distance-aware matrix. This matrix jointly redistributes excessive attention away from anchors and injects absolute positional bias to ensure uniform visual coverage. Experiments on LLaDA-V demonstrate that EDAR effectively eliminates semantic blind spots and achieves state-of-the-art performance on both hallucination-specific and general multimodal benchmarks.

Chenkai Xu, Xu Wang, Zhenyi Liao, Yishun Li, TianQi Hou, Zhijie Deng

Consistency models (CMs) have shown promise in the efficient generation of both image and text. This raises the natural question of whether we can learn a unified CM for efficient multimodal generation (e.g., text-to-image) and understanding (e.g., image-to-text). Intuitively, such a model could be acquired by applying the consistency distillation (CD) to existing unified multimodal models. However, the key challenge is establishing a unified denoising perspective for both image and text generation, which is essential for establishing the consistency mapping. To tackle this, at the representation level, we advocate for discrete tokens for both modalities to best preserve language modeling capabilities. Critically, instead of defining the text denoising trajectory via recent discrete diffusion language modeling principles, we specify it using the parallel decoding trace of an autoregressive language model, benefiting from the latter’s superior performance in general text generation tasks. The denoising trajectory of image tokens adheres to standard discrete diffusion. We train our unified consistency models (UniCMs) on these combined multimodal trajectories simultaneously with a unified objective. We introduce a trajectory segmentation strategy to improve the training convergence. Empirically, in text-to-image generation, UniCMs outperform SD3 on GenEval and Image Reward, while requiring only approximately {1}/{8} of the sampling time. Meanwhile, in image-to-text generation, UniCMs surpass Show-o on the MMMU benchmark while being 1.5 \times faster at long-sequence generating speed.

Zhengnan Guo, Fei Tan

While Diffusion Large Language Models (dLLMs) have emerged as a promising non-autoregressive paradigm comparable to auto-regressive (AR) models, their faithfulness, specifically regarding hallucination, remains largely underexplored. To bridge this gap, we present the first controlled comparative study to evaluate hallucination patterns in dLLMs. Our results demonstrate that current dLLMs exhibit a higher propensity for hallucination than AR counterparts controlled for architecture, scale, and pre-training weights. Furthermore, an analysis of inference-time compute reveals divergent dynamics: while quasi-autoregressive generation suffers from early saturation, non-sequential decoding unlocks potential for continuous refinement. Finally, we identify distinct failure modes unique to the diffusion process, including premature termination, incomplete denoising, and context intrusion. Our findings underscore that although dLLMs have narrowed the performance gap on general tasks, their distinct hallucination mechanisms pose a critical challenge to model reliability.