Recent advances in multimodal language models (MLLMs) have made thinking with images a dominant paradigm for multimodal reasoning. However, existing methods still fail to ensure evidence–answer consistency, where correct answers must be supported by correct visual evidence. To address this issue, we propose DeFacto, a counterfactual reasoning framework that explicitly aligns visual evidence with final answers by jointly optimizing for task correctness and evidence–answer consistency. Our approach integrates three complementary training paradigms: (i) positive, (ii) counterfactual, and (iii) random-masking. We further develop a language-guided evidence construction pipeline that automatically localizes question-relevant regions and generates counterfactual variants, resulting in DeFacto-100K, a dataset of about 100k images. Building on this dataset, we train MLLMs with GRPO-based reinforcement learning and design three complementary rewards to promote correct answering, structured reasoning, and consistent evidence selection. Moreover, we introduce DeFacto-1.5K, a human-annotated benchmark for systematically evaluating evidence-grounded consistency beyond answer accuracy. Experiments on diverse benchmarks demonstrate that DeFacto substantially improves both answer accuracy and evidence–answer consistency over strong baselines. The code and datasets will be released upon acceptance.

Multimodal latent-space reasoning aims to replace explicit “thinking with images” by performing visual reasoning directly in a compact latent space. However, existing approaches largely rely on visual supervision and produce latent representations that lack sufficient semantic richness, limiting their ability to support diverse region-level reasoning tasks. In this work, we introduce Semantic-Enriched Latent Visual Reasoning (SLVR), a two-stage learning framework that enriches latent representations with attribute-level visual semantics and aligns them with diverse reasoning objectives. In the first stage, SLVR learns semantically enriched region-centric latents under fine-grained attribute supervision. In the second stage, we design Multi-query Group Relative Policy Optimization (M-GRPO) to align latent representations across multiple queries grounded in the same region, encouraging consistent utilization of semantic information under varying reasoning demands. To support this framework, we construct SLV-Set, comprising approximately 400K region-level attribute annotations and 800K multi-query question answering samples, and introduce SV-QA, a benchmark that evaluates latent reasoning under semantic variation. Experiments demonstrate that SLVR improves the robustness and semantic consistency of latent visual reasoning compared to existing baselines.

In robotic perception, cross-granularity object detection is essential for identifying and localizing targets at varying levels of detail. Traditional detection methods often struggle to bridge the gap between coarse object detection and fine-grained component localization, limiting their ability to associate parts, such as a cup and its handle. Vision-language models (VLMs), while effective in spatial reasoning, face challenges in fine-grained detection due to the scarcity of annotated datasets. To address these issues, we first propose the chain-of-detection (CoD) framework, which focuses on guiding detection in a step-by-step manner from coarse recognition to fine-grained localization. During this process, we observe that existing detectors still lack sufficient capability in recognizing fine-grained components. To overcome this limitation, we further combine the CoD framework with Monte Carlo tree search (MCTS) to automatically generate fine-grained datasets, eliminating the need for manual labeling and significantly improving detector performance. Experiments show that our approach achieves an average improvement of 17.31% in robotic manipulation success rates for common objects, 51.39% for larger object operations, and about 50% in simulated environments. These results demonstrate the effectiveness of CoD in advancing cross-granularity detection and enhancing precise robotic manipulation. The implementation is publicly available at https://github.com/tinnel123666888/CoD and the CoD dataset is released at https://huggingface.co/datasets/tinnel123/CoD_dataset.

In-context learning (ICL) is a valuable capability exhibited by Transformers pretrained on diverse sequence tasks. However, previous studies have observed that ICL often conflicts with the model's inherent in-weight learning (IWL) ability. By examining the representation space learned by a toy model in synthetic experiments, we identify the shared encoding space for context and samples in Transformers as a potential source of this conflict. To address this, we modify the model architecture to separately encode the context and samples into two distinct spaces: a task representation space and a sample representation space. We model these two spaces under a simple yet principled framework, assuming a linear representational structure and treating them as a pair of dual spaces. Both theoretical analysis and empirical results demonstrate the effectiveness of our proposed architecture, CoQE, in the single-value answer setting. It not only enhances ICL performance through improved representation learning, but also successfully reconciles ICL and IWL capabilities across synthetic few-shot classification and a newly designed pseudo-arithmetic task. Code: https://github.com/McGuinnessChen/dual-representation-space-encoding

Deep networks often exhibit a preference for "simple" solutions, and such a simplicity bias is widely believed to play a key role in generalization. Yet a broadly applicable, quantitative measure of simplicity remains elusive. We introduce as a distribution-aware, low-dimensional surrogate for neural functions: we approximate a network’s predictive behavior along data-dependent interpolation paths using orthogonal polynomial bases, yielding a compact functional representation. We show that the of this representation serves as a practical simplicity metric that is predictive of generalization across tasks and architectures, and consistently outperforms existing generalization proxies such as sharpness. Finally, polynomial representations naturally yield a simplicity regularizer, which consistently improves generalization in image and text classification, fine-tuning contrastive vision–language models, and reinforcement learning.

Vision-Language-Action (VLA) models have recently demonstrated strong capabilities in mapping multimodal inputs to robotic control. However, a critical limitation persists: reasoning and action remain decoupled. In existing frameworks, language is used to generate auxiliary textual explanations—external commentaries on intended behavior—while actions are produced by a separate, implicitly aligned module. This architectural separation creates a semantic gap between high-level planning and low-level control, limiting fine-grained grounding and robustness in complex, cluttered environments. To bridge this gap, we propose Chain-of-Boxes Reasoning VLA (Box-Chain VLA), a framework that unifies reasoning and action within a shared latent space. Instead of treating language as a passive bystander, our model embeds structured, reinforcement-optimized reasoning tokens directly into the action generation process. These tokens—encoding task decomposition, spatial references, and subgoal structure—are not merely for interpretability, but serve as active guidance signals for trajectory prediction. By integrating reasoning and action into a single generative stream, Box-Chain VLA enables reasoning to directly guide action, where linguistic concepts act as inductive biases in motor control. Experiments in simulation and real-world settings show that our method consistently outperforms prior VLAs, achieving more reliable and precise manipulation in long-horizon and fine-grained tasks—without requiring additional reasoning supervision.

Spiking neural networks (SNNs) possess great potential for energy-efficient computation; however, their practical deployment is often constrained by the high computational cost associated with training across multiple time steps. Unlike previous approaches that simply reduce the number of time steps, this work investigates temporal feature redundancy and reveals the feature overlapping phenomenon in SNNs-showing that substantial computational redundancy exists across temporal dimensions. To address this issue, we propose our core contribution, temporal differential decoupling (TDD), which transforms network computation into the differential domain to disentangle static and dynamic feature components. This decoupling enables focused processing of informative signals and significantly reduces redundant computation without compromising model accuracy. Building upon this framework, we further design the TDD-based differential domain low-sparsity approximation (TDD-DDLA) algorithm, which is based on the gradient sensitivity criterion, as an implementation strategy to quantify each temporal feature's contribution to gradient updates and achieve efficient energy optimization. In contrast to prior studies that only adjusted time steps without explicit feature-level analysis, our framework provides a structured and theoretically grounded analysis of temporal feature evolution. Experimental results demonstrate that our method achieves up to 80.9% fewer spikes per time step and 57.8% fewer total spikes, with no degradation in classification performance, offering a promising pathway toward scalable, low-cost, and high-accuracy SNN deployment.

Spiking neural networks (SNNs) inspired by biological neurons offer energy-efficient and interpretable computation but is limited by the simplistic structure of point neurons. We introduce a multi-compartment spiking neuron model (MCN) with trainable cross-compartment connections that simulate soma-dendrite interactions. Theoretically, we prove that these connections act as spatiotemporal momentum, guiding learning dynamics toward global optima. To leverage this, we propose a multi-compartment spatiotemporal backpropagation (MCST-BP) algorithm that enhances gradient flow stability. Experimental results for multiple benchmark datasets, including S-MNIST, CIFAR-10, Spiking Heidelberg Digits (SHD), and ECG, show that MC-SNNs outperform traditional SNNs in both convergence speed and accuracy. Our work bridges neurobiological structure and computational modeling, providing a theoretical and practical foundation for high-performance brain-inspired learning systems.

Model-based reinforcement learning (MBRL) leverages an agent's experience to construct a dynamics model, thereby improving the data efficiency of policy learning. While MBRL methods have achieved impressive results on fully observable (Markovian) tasks, their effectiveness on partially observable tasks remains limited. Among existing methods, Dreamer-built upon a recurrent variational autoencoder to model latent dynamics and perform trajectory imagination-has shown the most promise for such tasks. However, Dreamer struggles in environments that require long-term memory reasoning due to the large history space, which renders dynamics prediction and policy learning inefficient. To address this issue, we propose Causal Dreamer, which introduces a self-supervised task called history counterfactual inference to identify and discard uninformative observations from the agent's history. By retaining only the observations that are critical for accurate dynamics prediction, Causal Dreamer improves multi-step transition modeling and enhances policy performance. Experiments on long-horizon partially observable tasks, including Maze, BabyAI, and robotic navigation environments, demonstrate that Causal Dreamer significantly outperforms prior methods in both sample efficiency and final performance.

Spiking Neural Networks (SNNs) often rely on rate coding, where high-precision inference depends on long time-steps, leading to significant latency and energy cost—especially for ANN-to-SNN conversions. To address this, we propose Adaptive Fission, a post-training encoding technique that selectively splits high-sensitivity neurons into groups with varying scales and weights. This enables neuron-specific, on-demand precision and threshold allocation while introducing minimal spatial overhead. As a generalized form of population coding, it seamlessly applies to a wide range of pretrained SNN architectures without requiring additional training or fine-tuning. Experiments on neuromorphic hardware demonstrate up to 80\% reductions in latency and power consumption without degrading accuracy.

Recent advances in multimodal language models (MLLMs) have achieved remarkable progress in vision-language reasoning, especially with the emergence of "thinking with images," which integrates explicit visual steps into the reasoning process. While this paradigm strengthens image-based reasoning, a significant challenge remains: models may arrive at correct answers by relying on irrelevant or spurious regions, driven by prior knowledge or dataset biases. Even when the answer is correct, flawed reasoning indicates that the model has not truly understood the image, highlighting the critical importance of reasoning fidelity in multimodal tasks. To address this issue, we propose DeFacto, a counterfactual reasoning framework that jointly enforces accurate answering and faithful reasoning. A key component of our approach is the design of three complementary training paradigms: (i) positive, (ii) counterfactual, and (iii) random-masking. To enable these paradigms, we develop a pipeline that automatically localizes question-relevant evidence and constructs positive, counterfactual, and random variants, resulting in a dataset of about 100k images. Building on this framework, we train multimodal language models with GRPO-based reinforcement learning, where we design three complementary rewards to guide the model toward accurate answering and evidence-grounded reasoning. Experiments on diverse benchmarks demonstrate that DeFacto substantially improves both answer accuracy and reasoning faithfulness, establishing a stronger foundation for interpretable multimodal reasoning. The code is available on GitHub and the dataset is released on HuggingFace.

This paper introduces FALCON , a novel F ast A utonomous exp L oration framework using CO verage path guida N ce, which aims at setting a new performance benchmark in the field of autonomous aerial exploration. Despite recent advancements in the domain, existing exploration planners often suffer from inefficiencies such as frequent revisitations of previously explored regions. FALCON effectively harnesses the full potential of online generated coverage paths in enhancing exploration efficiency. The framework begins with an incremental connectivity-aware space decomposition and connectivity graph construction, which facilitate efficient coverage path planning. Subsequently, a hierarchical planner generates a coverage path spanning the entire unexplored space, serving as a global guidance. Then, a local planner optimizes the frontier visitation order, minimizing traversal time while consciously incorporating the intention of the global guidance. Finally, minimum-time smooth and safe trajectories are produced to visit the frontier viewpoints. For fair and comprehensive benchmark experiments, we introduce a lightweight exploration planner evaluation environment that allows for comparing exploration planners across a variety of testing scenarios using an identical quadrotor simulator. Additionally, an in-depth analysis and evaluation is conducted to highlight the significant performance advantages of FALCON in comparison with the state-of-the-art exploration planners based on objective criteria. Extensive ablation studies demonstrate the effectiveness of each component in the proposed framework. Real-world experiments conducted fully onboard further validate FALCON 's practical capability in complex and challenging environments. The source code of both the exploration planner FALCON and the exploration planner evaluation environment has been released to benefit the community

Underwater image enhancement (UIE) is critical in various applications, including marine biology research, underwater archaeology, and autonomous underwater vehicle (AUV) navigation. The unpredictable nature of underwater environments frequently leads to degradation in contrast, color, and perceptual visual quality. Previous methods using the single receptive field to extract features are not capable of handling varying light conditions, which hinders detail preservation, color correction, and image quality improvement. To address these challenges, we propose Multi Core Token Mixer (MCTM) by introducing a distinctive multi-core mechanism. This mechanism is adept at extracting varied receptive fields, thereby enabling the model to capture the degradation at different scales caused by inhomogeneous underwater conditions. We performed experiments on three datasets (UIEB, EUVP, and UFO-120), and MCTM consistently outperforms existing models in image enhancement, color correction, and perceptual visual quality. Our work sets a new standard in the field and emphasizes the promise held by task-specific architectures that harness the power of Transformer models to tackle domain-specific challenges, particularly in UIE.

Domain generalization seeks to train models that can generalize to unseen out-of-distribution (OOD) test domains. A typical strategy is to learn "invariant'' features that are assumed to be identically distributed across all domains, in order to extract in-distribution features for downstream tasks. However, this hypothesis often fails in real-world scenarios, leading to feature shifts that degrade performance. To address this, we first theoretically decompose the generalization error in the feature space and prove that ensuring in-distribution features can provide a tighter error bound. We then propose a novel solution, Iterative In-Distribution Compression, which models the exaction of in-distribution features as a convergent iterative process. It adaptively measures the feature space to select the most similar domain for each test sample, and through iterative contraction mappings, it progressively reduces the feature distance, aligning the features with known domains. Extensive experiments on five benchmarks demonstrate the effectiveness of our approach, highlighting the advantages of in-distribution mapping for enhancing generalization performance.

Multimodal large models have made significant progress, yet fine-grained understanding of complex scenes remains a challenge. High-quality, large-scale vision-language datasets are essential for addressing this issue. However, existing methods often rely on labor-intensive manual annotations or closed-source models with optimal performance, making large-scale data collection costly. To overcome these limitations, we propose a self-bootstrapped training pipeline that leverages the model's own multimodal capabilities to recursively refine its understanding. By decomposing existing multimodal data into localized sub-regions and generating hierarchical scene descriptions and multi-faceted question-answer pairs, we construct a dataset based on 1.4M image-task instances. We further utilize this dataset to train the base model, significantly enhancing its ability to interpret complex visual scenes and perform various vision-related tasks. Our OURO model, fine-tuned on Qwen2-VL-7B-Instruct using LoRA, achieves substantial improvements over both the base model and similarly-sized counterparts across multiple multimodal benchmarks. Our self-bootstrapped training pipeline offers a novel paradigm for the continuous improvement of multimodal models. Code and datasets are available at https://github.com/tinnel123666888/OURO.git.

Humans develop world models that capture the underlying generation process of data. Whether neural networks can learn similar world models remains an open problem. In this work, we present the first theoretical results for this problem, showing that in a multi-task setting, models with a low-degree bias provably recover latent data-generating variables under mild assumptions--even if proxy tasks involve complex, non-linear functions of the latents. However, such recovery is sensitive to model architecture. Our analysis leverages Boolean models of task solutions via the Fourier-Walsh transform and introduces new techniques for analyzing invertible Boolean transforms, which may be of independent interest. We illustrate the algorithmic implications of our results and connect them to related research areas, including self-supervised learning, out-of-distribution generalization, and the linear representation hypothesis in large language models.

Large language models (LLMs) have been able to perform various forms of reasoning tasks ina wide range of scenarios, but are they truly engaging in task abstraction and rule-based reasoning beyond mere memorization? To answer this question, we propose a novel experimentalapproach, Misleading Fine-Tuning (MisFT), to examine whether LLMs perform abstract reasoning by altering their original understanding of fundamental rules. In particular, by constructing datasets with math expressions or logical formulas that contradict correct principles, we fine-tune the model to learn those contradictory rules and assess its generalization ability on unseen test domains. Through a series of experiments, we find that current LLMs are capable of applying contradictory rules to solve practical math word problems and natural language reasoning tasks, implying the presence of an internal mechanism in LLMs that abstracts before reasoning.

The efficiency of large vision-language models (LVLMs) is constrained by the computational bottleneck of the attention mechanism during the prefill phase and the memory bottle-neck of fetching the key-value (KV) cache in the decoding phase, particularly in scenarios involving high-resolution images or videos. Visual content often exhibits substantial redundancy, resulting in highly sparse attention maps within LVLMs. This sparsity can be leveraged to accelerate attention computation or compress the KV cache through various approaches. However, most studies focus on addressing only one of these bottlenecks and do not adequately support dynamic adjustment of sparsity concerning distinct layers or tasks. In this paper, we present ZipVL, an efficient inference framework designed for LVLMs through a dynamic ratio allocation strategy of important tokens. This ratio is adaptively determined based on the layer-specific distribution of attention scores, rather than fixed hyper-parameters, thereby improving efficiency for less complex tasks while maintaining high performance for more challenging ones. Then we select important tokens based on their normalized attention scores and perform sparse attention mechanism solely on those important tokens, reducing the latency in the prefill phase. Tokens deemed less important will be discarded to reduce KV cache size, alleviating the memory bottleneck in the decoding phase. Our experiments demonstrate that ZipVL can accelerate the prefill phase by $2.3 \times$ and improve decoding throughput by $2.8 \times$, with a minimal accuracy reduction of only 0.5% on VQAv2 benchmark over LLaVA-Next-13B model, effectively enhancing the generation efficiency of LVLMs.

We present RoboGen, a generative robotic agent that automatically learns diverse robotic skills at scale via generative simulation. RoboGen leverages the latest advancements in foundation and generative models. Instead of directly using or adapting these models to produce policies or low-level actions, we advocate for a generative scheme, which uses these models to automatically generate diversified tasks, scenes, and training supervisions, thereby scaling up robotic skill learning with minimal human supervision. Our approach equips a robotic agent with a self-guided propose-generate-learn cycle: the agent first proposes interesting tasks and skills to develop, and then generates corresponding simulation environments by populating pertinent objects and assets with proper spatial configurations. Afterwards, the agent decomposes the proposed high-level task into sub-tasks, selects the optimal learning approach (reinforcement learning, motion planning, or trajectory optimization), generates required training supervision, and then learns policies to acquire the proposed skill. Our work attempts to extract the extensive and versatile knowledge embedded in large-scale models and transfer them to the field of robotics. Our fully generative pipeline can be queried repeatedly, producing an endless stream of skill demonstrations associated with diverse tasks and environments.