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.

Hallucination in large language models (LLMs) continues to be a significant issue, particularly in tasks like question answering, where models often generate plausible yet incorrect or irrelevant information. Although various methods have been proposed to mitigate hallucinations, the relationship between attention patterns and hallucinations has not been fully explored. In this paper, we analyze the distribution of attention scores across each layer and attention head of LLMs, revealing a common and intriguing phenomenon: Shallow layers of LLMs primarily rely on uniform attention patterns, where the model distributes its attention evenly across the entire sequence. This uniform attention pattern can lead to hallucinations, as the model fails to focus on the most relevant information. To mitigate this issue, we propose a training-free method called Attention Replacement Technique (ART), which replaces these uniform attention patterns in the shallow layers with local attention patterns. This change directs the model to focus more on the relevant contexts, thus reducing hallucinations. Through extensive experiments, ART demonstrates significant reductions in hallucinations across multiple LLM architectures, proving its effectiveness and generalizability without requiring fine-tuning or additional training data.

Large reasoning models (LRMs) that generate long chains of thought now perform well on multi-step math, science, and coding tasks. However, their behavior is still unstable and hard to interpret, and existing analysis tools struggle with such long, structured reasoning traces. We introduce Step-Saliency, which pools attention–gradient scores into step-to-step maps along the question–thinking–summary trajectory. Across several models, Step-Saliency reveals two recurring information-flow failures: Shallow Lock-in, where shallow layers over-focus on the current step and barely use earlier context, and Deep Decay, where deep layers gradually lose saliency on the thinking segment and the summary increasingly attends to itself and the last few steps. Motivated by these patterns, we propose StepFlow, a saliency-inspired test-time intervention that adjusts shallow saliency patterns measured by Step-Saliency via Odds-Equal Bridge and adds a small step-level residual in deep layers via Step Momentum Injection. StepFlow improves accuracy on math, science, and coding tasks across multiple LRMs without retraining, indicating that repairing information flow can recover part of their missing reasoning performance.

Aligning Large Vision-Language Models (LVLMs) to mitigate hallucinations typically relies on high-quality preference data. However, in self-supervised settings, standard binary preference optimization (e.g., DPO) suffers from noisy supervision and semantic ambiguity, as automatically generated chosen responses are not guaranteed to be superior to rejected ones. In this work, we propose Trident, a fully self-supervised framework that ensures robust alignment via a structured triplet paradigm. Trident autonomously constructs reliable preference triplets—comprising semantically enriched (chosen), degraded (rejected), and neutral (anchor) responses—through automated visual perturbations and self-summarization. We further introduce Trident Preference Regularization (TPR), a novel objective that utilizes an adaptive margin to enforce semantic separation between the triplet components while preventing deviation from the pretrained distribution. Despite requiring no human annotations or external reward models, Trident consistently outperforms state-of-the-art RLHF and RLAIF baselines. For instance, on LLaVA-1.5-7B, it reduces the hallucination rate on AMBER to 11.3% and achieves 95.70% precision on POPE using only 4k self-generated triplets and a single epoch. This validates structured triplet supervision as a scalable paradigm for robust self-supervised alignment.

Multimodal large language models (MLLMs) are increasingly deployed in Web-scale applications—such as image search, social media captioning, and e-commerce product description generation—where factual consistency is critical for user trust and content reliability. However, we observe that MLLMs frequently hallucinate in these settings due to two relevant phenomena: the massive activation phenomenon and positional information decay. The former refers to the tendency of attention mechanisms to concentrate on a small set of tokens with extreme activation values in query and key projections, which play indispensable roles in contextual understanding. In our investigation, we discover that perturbing these tokens leads to significant performance drops, highlighting their utmost importance. As for positional information decay, it occurs due to the common rotary position encoding strategy, where the attention to early visual tokens diminishes over time, especially in long-sequence generation tasks, such as image caption. To address these challenges, we propose TokenTruth, a token-level intervention strategy that dynamically suppresses irrelevant visual tokens while preserving key contextual signals. Our method is grounded in an in-depth analysis of massive activations and attention sink behaviors, and introduces a targeted token penalty mechanism that reallocates attention more faithfully toward informative visual regions. Extensive experiments demonstrate that TokenTruth significantly improves factual consistency across various MLLMs on standard image understanding benchmarks.

While diffusion Multimodal Large Language Models (dMLLMs) have recently achieved remarkable strides in multimodal generation, the development of interpretability mechanisms has lagged behind their architectural evolution. Unlike traditional autoregressive models that produce sequential activations, diffusion-based architectures generate tokens via parallel denoising, resulting in smooth, distributed activation patterns across the entire sequence. Consequently, existing Class Activation Mapping (CAM) methods, which are tailored for local, sequential dependencies, are ill-suited for interpreting these non-autoregressive behaviors. To bridge this gap, we propose Diffusion-CAM, the first interpretability method specifically tailored for dMLLMs. We derive raw activation maps by differentiably probing intermediate representations in the transformer backbone, accordingly capturing both latent features and their class-specific gradients. To address the inherent stochasticity of these raw signals, we incorporate four key modules to resolve spatial ambiguity and mitigate intra-image confounders and redundant token correlations. Extensive experiments demonstrate that Diffusion-CAM significantly outperforms SoTA methods in both localization accuracy and visual fidelity, establishing a new standard for understanding the parallel generation process of diffusion multimodal systems.

Latent reasoning enables Large Language Models (LLMs) to perform multi-step inference within continuous hidden states, offering efficiency gains over explicit Chain-of-Thought (CoT). However, the opacity of these continuous thought vectors hinders their reliability and controllability. This paper bridges the gap between mechanistic interpretability and actionable control. We first present a systematic analysis using structural, causal, and geometric probes, revealing that latent vectors encode compressed, faithful representations of reasoning steps, with early vectors acting as critical causal hubs. Building on this, we operationalize these interpretability insights into a suite of training-free, decode-time interventions that refine the latent reasoning process by imposing the identified geometric and semantic priors. Extensive experiments across multiple model scales and diverse task domains demonstrate that our approaches consistently improve reasoning accuracy. Our interpretability-guided interventions consistently unlock latent capabilities and improve reasoning accuracy without any parameter updates.

Chain-of-Thought (CoT) prompting enhances the math reasoning capability of large language models (LLMs) to a large margin. However, the mechanism underlying such improvements remains unexplored. In this paper, we present SalaMAnder (Shapley-based Mathematical Expression Attribution and Metric), a theoretically grounded methodology as well as a mathematically rigorous evaluation metric for quantifying component-level contributions in few-shot CoT reasoning. Concretely, we leverage the Shapley value for mathematical expression attribution and develop an efficient stratified sampling algorithm that significantly reduces the computational complexity. Besides, we develop the CoSP (Cardinality of Shapley Positives) metric through covariance analysis. Comprehensive validation across popular LLM models and diverse mathematical benchmarks demonstrates that the CoSP metric within our SalaMAnder framework exhibits a robust monotonic correlation with model performance, not only providing theoretical explanations for the empirical success of existing few-shot CoT but also establishing mathematically rigorous principles for prompt construction optimization. Furthermore, we verify the reliability of the explanation, based on which we unify the insights of previous work.

Large Vision–Language Models (LVLMs) have garnered substantial interest owing to their impressive ability to interpret visual inputs and converse with users.Nevertheless, LVLMs still suffer from object hallucination – generating descriptions for objects that are absent from the image, which undermines reliability and hinders real-world deployment. We propose DAPE-BR, a positional-alignment scheme that (i) preserves the pretrained weight order while globally—- visual–text distances, (ii) embeds an isotropic fused patch-distance metric, and (iii) applies a patch-distance causal mask to enforce spatial causality. Extensive experiments on POPE, MMStar and SQA show that DAPE-BR consistently reduces hallucinations and boosts.

Multimodal large language models (MLLMs) demonstrate excellent abilities for understanding visual information, while the hallucination remains. Albeit image tokens constitute the majority of the MLLMs input, the relation between image tokens and hallucinations is still unexplored. In this paper, we analyze the attention score distribution of image tokens across layers and attention heads in models, revealing an intriguing but common phenomenon: most hallucinations are closely linked to the attention sink patterns of image tokens attention matrix, where shallow layers exhibit dense sinks and deep layers exhibit the sparse. We further explore the attention heads of different layers, finding: heads with high-density attention sink of the image part act positively in mitigating hallucinations. Inspired by these findings, we propose a training-free approach called Enhancing Vision Attention Sinks (EVAS) to facilitate the convergence of the image token attention sink within shallow layers. Specifically, EVAS identifies the attention heads that emerge as the densest visual sink in shallow layers and extracts its attention matrix, which is then broadcast to other heads of the same layer, thereby strengthing the layer’s focus on the image itself. Extensive empirical results of various MLLMs illustrate the superior performance of the proposed EVAS, demonstrating its effectiveness and generality.

The increasing ubiquity of text-to-image (T2I) models as tools for visual content generation raises concerns about their ability to accurately represent diverse cultural contexts - where missed cues can stereotype communities and undermine usability. In this work, we present the first study to systematically quantify the alignment of T2I models and evaluation metrics with respect to both explicit (stated) as well as implicit (unstated, implied by the prompt’s cultural context) cultural expectations. To this end, we introduce CulturalFrames, a novel benchmark designed for rigorous human evaluation of cultural representation in visual generations. Spanning 10 countries and 5 socio-cultural domains, CulturalFrames comprises 983 prompts, 3637 corresponding images generated by 4 state-of-the-art T2I models, and over 10k detailed human annotations. We find that across models and countries, cultural expectations are missed an average of 44% of the time. Among these failures, explicit expectations are missed at a surprisingly high average rate of 68%, while implicit expectation failures are also significant, averaging 49%. Furthermore, we show that existing T2I evaluation metrics correlate poorly with human judgments of cultural alignment, irrespective of their internal reasoning. Collectively, our findings expose critical gaps, provide a concrete testbed, and outline actionable directions for developing culturally informed T2I models and metrics that improve global usability.

Large Vision Language Models (LVLMs) achieve great performance on visual-language reasoning tasks, however, the black-box nature of LVLMs hinders in-depth research on the reasoning mechanism. As all images need to be converted into image tokens to fit the input format of large language models (LLMs) along with natural language prompts, sequential visual representation is essential to the performance of LVLMs, and the information flow analysis approach can be an effective tool for determining interactions between these representations. In this paper, we propose integrating attention analysis with LLaVA-CAM, concretely, attention scores highlight relevant regions during forward propagation, while LLaVA-CAM captures gradient changes through backward propagation, revealing key image features. By exploring the information flow from the perspective of visual representation contribution, we observe that it tends to converge in shallow layers but diversify in deeper layers. To validate our analysis, we conduct comprehensive experiments with truncation strategies across various LVLMs for visual question answering and image captioning tasks, and experimental results not only verify our hypothesis but also reveal a consistent pattern of information flow convergence in the corresponding layers, and the information flow cliff layer will be different due to different contexts.

Transferability estimation has been attached to great attention in the computer vision fields. Researchers try to estimate with low computational cost the performance of a model when transferred from a source task to a given target task. Considering the effectiveness of such estimations, the communities of natural language processing also began to study similar problems for the selection of pre-trained language models. However, there is a lack of a comprehensive comparison between these estimation methods yet. Also, the differences between vision and language scenarios make it doubtful whether previous conclusions can be established across fields. In this paper, we first conduct a thorough survey of existing transferability estimation methods being able to find the most suitable model, then we conduct a detailed empirical study for the surveyed methods based on the GLUE benchmark. From qualitative and quantitative analyses, we demonstrate the strengths and weaknesses of existing methods and show that H-Score generally performs well with superiorities in effectiveness and efficiency. We also outline the difficulties of consideration of training details, applicability to text generation, and consistency to certain metrics which shed light on future directions.

Mixture-of-Experts (MoE) networks have been proposed as an efficient way to scale up model capacity and implement conditional computing. However, the study of MoE components mostly focused on the feedforward layer in Transformer architecture. This paper proposes the Mixture of Attention Heads (MoA), a new architecture that combines multi-head attention with the MoE mechanism. MoA includes a set of attention heads that each has its own set of parameters. Given an input, a router dynamically selects a subset of k attention heads per token. This conditional computation schema allows MoA to achieve stronger performance than the standard multi-head attention layer. Furthermore, the sparsely gated MoA can easily scale up the number of attention heads and the number of parameters while preserving computational efficiency. Despite performance improvements, MoA also automatically differentiates heads’ utilities, providing a new perspective to discuss the model’s interpretability. We conducted experiments on several important tasks, including Machine Translation and Masked Language Modeling. Experiments have shown promising results on several tasks against strong baselines that involve large and very deep models.

Few-shot Text Classification predicts the semantic label of a given text with a handful of supporting instances. Current meta-learning methods have achieved satisfying results in various few-shot situations. Still, they often require a large amount of data to construct many few-shot tasks for meta-training, which is not practical in real-world few-shot scenarios. Prompt-tuning has recently proved to be another effective few-shot learner by bridging the gap between pre-train and downstream tasks. In this work, we closely combine the two promising few-shot learning methodologies in structure and propose a Prompt-Based Meta-Learning (PBML) model to overcome the above meta-learning problem by adding the prompting mechanism. PBML assigns label word learning to base-learners and template learning to meta-learner, respectively. Experimental results show state-of-the-art performance on four text classification datasets under few-shot settings, with higher accuracy and good robustness. We demonstrate through low-resource experiments that our method alleviates the shortcoming that meta-learning requires too much data for meta-training. In the end, we use the visualization to interpret and verify that the meta-learning framework can help the prompting method converge better. We release our code to reproduce our experiments.

Generating long text conditionally depending on the short input text has recently attracted more and more research efforts. Most existing approaches focus more on introducing extra knowledge to supplement the short input text, but ignore the coherence issue of the generated texts. To address aforementioned research issue, this paper proposes a novel two-stage approach to generate coherent long text. Particularly, we first build a document-level path for each output text with each sentence embedding as its node, and a revised self-organising map (SOM) is proposed to cluster similar nodes of a family of document-level paths to construct the directed semantic graph. Then, three subgraph alignment methods are proposed to extract the maximum matching paths or subgraphs. These directed subgraphs are considered to well preserve extra but relevant content to the short input text, and then they are decoded by the employed pre-trained model to generate coherent long text. Extensive experiments have been performed on three real-world datasets, and the promising results demonstrate that the proposed approach is superior to the state-of-the-art approaches w.r.t. a number of evaluation criteria.