We study the problem of causal function estimation in the Proxy Causal Learning (PCL) framework, where confounders are not observed but proxies for the confounders are available. Two main approaches have been proposed: outcome bridge-based and treatment bridge-based methods. In this work, we propose two kernel-based doubly robust estimators that combine the strengths of both approaches, and naturally handle continuous and high-dimensional variables. Our identification strategy builds on a recent density ratio-free method for treatment bridge-based PCL; furthermore, in contrast to previous approaches, it does not require indicator functions or kernel smoothing over the treatment variable. These properties make it especially well-suited for continuous or high-dimensional treatments. By using kernel mean embeddings, we propose the first density-ratio free doubly robust estimators for proxy causal learning, which have closed form solutions and strong uniform consistency guarantees. Our estimators outperform existing methods on PCL benchmarks, including a prior doubly robust method that requires both kernel smoothing and density ratio estimation.

While real-world applications increasingly demand intricate scene manipulation, existing instruction-guided image editing benchmarks often oversimplify task complexity and lack comprehensive, fine-grained instructions. To bridge this gap, we introduce CompBench, a large-scale benchmark specifically designed for complex instruction-guided image editing. CompBench features challenging editing scenarios that incorporate fine-grained instruction following, spatial and contextual reasoning, thereby enabling comprehensive evaluation of image editing models' precise manipulation capabilities. To construct CompBench, we propose an MLLM-human collaborative framework with tailored task pipelines. Furthermore, we propose an instruction decoupling strategy that disentangles editing intents into four key dimensions: location, appearance, dynamics, and objects, ensuring closer alignment between instructions and complex editing requirements. Extensive evaluations reveal that CompBench exposes fundamental limitations of current image editing models and provides critical insights for the development of next-generation instruction-guided image editing systems. Our project page is available at https://comp-bench.github.io/.

Improving energy efficiency in residential buildings is critical to combating climate change and reducing greenhouse gas emissions. Retrofitting existing buildings, which contribute a significant share of energy use, is therefore a key priority, especially in regions with outdated building stock. Artificial Intelligence (AI) and Machine Learning (ML) can automate retrofit decision-making and find retrofit strategies. However, their use faces challenges of data availability, model transparency, and compliance with national and EU AI regulations including the AI act, ethics guidelines and the ALTAI. This paper presents a trustworthy-by-design ML-based decision support framework that recommends energy efficiency strategies for residential buildings using minimal user-accessible inputs. The framework merges Conditional Tabular Generative Adversarial Networks (CTGAN) to augment limited and imbalanced data with a neural network-based multi-label classifier that predicts potential combinations of retrofit actions. To support explanation and trustworthiness, an Explainable AI (XAI) layer using SHapley Additive exPlanations (SHAP) clarifies the rationale behind recommendations and guides feature engineering. Two case studies validate performance and generalization: the first leveraging a well-established, large EPC dataset for England and Wales; the second using a small, imbalanced post-retrofit dataset from Latvia (RETROFIT-LAT). Results show that the framework can handle diverse data conditions and improve performance up to 53% compared to the baseline. Overall, the proposed framework provides a feasible, interpretable, and trustworthy AI system for building retrofit decision support through assured performance, usability, and transparency to aid stakeholders in prioritizing effective energy investments and support regulation-compliant, data-driven innovation in sustainable energy transition.

This paper proposes a novel knowledge-Base (KB) assisted semantic communication framework for image transmission. At the receiver, a Facebook AI Similarity Search (FAISS) based vector database is constructed by extracting semantic embeddings from images using the Contrastive Language-Image Pre-Training (CLIP) model. During transmission, the transmitter first extracts a 512-dimensional semantic feature using the CLIP model, then compresses it with a lightweight neural network for transmission. After receiving the signal, the receiver reconstructs the feature back to 512 dimensions and performs similarity matching from the KB to retrieve the most semantically similar image. Semantic transmission success is determined by category consistency between the transmitted and retrieved images, rather than traditional metrics like Peak Signal-to-Noise Ratio (PSNR). The proposed system prioritizes semantic accuracy, offering a new evaluation paradigm for semantic-aware communication systems. Experimental validation on CIFAR100 demonstrates the effectiveness of the framework in achieving semantic image transmission.

We address the setting of Proxy Causal Learning (PCL), which has the goal of estimating causal effects from observed data in the presence of hidden confounding. Proxy methods accomplish this task using two proxy variables related to the latent confounder: a treatment proxy (related to the treatment) and an outcome proxy (related to the outcome). Two approaches have been proposed to perform causal effect estimation given proxy variables; however only one of these has found mainstream acceptance, since the other was understood to require density ratio estimation - a challenging task in high dimensions. In the present work, we propose a practical and effective implementation of the second approach, which bypasses explicit density ratio estimation and is suitable for continuous and high-dimensional treatments. We employ kernel ridge regression to derive estimators, resulting in simple closed-form solutions for dose-response and conditional dose-response curves, along with consistency guarantees. Our methods empirically demonstrate superior or comparable performance to existing frameworks on synthetic and real-world datasets.

Societal and personal values are transmitted to younger generations through interaction and exposure. Traditionally, children and adolescents learned values from parents, educators, or peers. Nowadays, social platforms serve as a significant channel through which youth (and adults) consume information, as the main medium of entertainment, and possibly the medium through which they learn different values. In this paper we extract implicit values from TikTok movies uploaded by online influencers targeting children and adolescents. We curated a dataset of hundreds of TikTok movies and annotated them according to the well established Schwartz Theory of Personal Values. We then experimented with an array of language models, investigating their utility in value identification. Specifically, we considered two pipelines: direct extraction of values from video and a 2-step approach in which videos are first converted to elaborated scripts and values are extracted from the textual scripts. We find that the 2-step approach performs significantly better than the direct approach and that using a few-shot application of a Large Language Model in both stages outperformed the use of a fine-tuned Masked Language Model in the second stage. We further discuss the impact of continuous pretraining and fine-tuning and compare the performance of the different models on identification of values endorsed or confronted in the TikTok. Finally, we share the first values-annotated dataset of TikTok videos. To the best of our knowledge, this is the first attempt to extract values from TikTok specifically, and visual social media in general. Our results pave the way to future research on value transmission in video-based social platforms.

The convergence of drone delivery systems, virtual worlds, and blockchain has transformed logistics and supply chain management, providing a fast, and environmentally friendly alternative to traditional ground transportation methods;Provide users with a real-world experience, virtual service providers need to collect up-to-the-minute delivery information from edge devices. To address this challenge, 1) a reinforcement learning approach is introduced to enable drones with fast training capabilities and the ability to autonomously adapt to new virtual scenarios for effective resource allocation.2) A semantic communication framework for meta-universes is proposed, which utilizes the extraction of semantic information to reduce the communication cost and incentivize the transmission of information for meta-universe services.3) In order to ensure that user information security, a lightweight authentication and key agreement scheme is designed between the drone and the user by introducing blockchain technology. In our experiments, the drone adaptation performance is improved by about 35\%, and the local offloading rate can reach 90\% with the increase of the number of base stations. The semantic communication system proposed in this paper is compared with the Cross Entropy baseline model. Introducing blockchain technology the throughput of the transaction is maintained at a stable value with different number of drones.

Multi-Objective Reinforcement Learning (MORL) aims to learn a set of policies that optimize trade-offs between multiple, often conflicting objectives. MORL is computationally more complex than single-objective RL, particularly as the number of objectives increases. Additionally, when objectives involve the preferences of agents or groups, incorporating fairness becomes both important and socially desirable. This paper introduces a principled algorithm that incorporates fairness into MORL while improving scalability to many-objective problems. We propose using Lorenz dominance to identify policies with equitable reward distributions and introduce lambda-Lorenz dominance to enable flexible fairness preferences. We release a new, large-scale real-world transport planning environment and demonstrate that our method encourages the discovery of fair policies, showing improved scalability in two large cities (Xi'an and Amsterdam). Our methods outperform common multi-objective approaches, particularly in high-dimensional objective spaces.

Recent research has generated hope that inference scaling, such as resampling solutions until they pass verifiers like unit tests, could allow weaker models to match stronger ones. Beyond inference, this approach also enables training reasoning models, where data is curated using rejection sampling against a verifier. However, we show that this approach is fundamentally limited when verifiers are imperfect and have a non-zero probability of producing false positives. Resampling cannot decrease this probability, so it imposes an upper bound to the accuracy of resampling-based inference scaling, regardless of compute budget. Our analysis shows that there is a strong correlation between the model's single-sample accuracy and its false positive rate on HumanEval and MBPP, whose unit tests have limited coverage. Therefore, no amount of inference scaling of weaker models can enable them to match the single-sample accuracy of a sufficiently strong model. Empirical results show that optimal sampling attempts are often fewer than 10, as the negative utility of false positives outweighs benefits, bending inference scaling curves downward. Finally, false positives may have other undesirable qualities, like poor adherence to coding style conventions.

Recent advancements in pre-trained vision-language models like CLIP have enabled the task of open-vocabulary segmentation. CLIP demonstrates impressive zero-shot capabilities in various downstream tasks that require holistic image understanding. However, due to the image-level contrastive learning and fully global feature interaction, ViT-based CLIP struggles to capture local details, resulting in poor performance in segmentation tasks. Our analysis of ViT-based CLIP reveals that anomaly tokens emerge during the forward process, attracting disproportionate attention from normal patch tokens and thereby diminishing spatial awareness. To address this issue, we propose Self-Calibrated CLIP (SC-CLIP), a training-free method that calibrates CLIP to generate finer representations while preserving its original generalization ability-without introducing new parameters or relying on additional backbones. Specifically, we mitigate the negative impact of anomaly tokens from two complementary perspectives. First, we explicitly identify the anomaly tokens and replace them based on local context. Second, we reduce their influence on normal tokens by enhancing feature discriminability and attention correlation, leveraging the inherent semantic consistency within CLIP's mid-level features. In addition, we introduce a two-pass strategy that effectively integrates multi-level features to enrich local details under the training-free setting. Together, these strategies enhance CLIP's feature representations with improved granularity and semantic coherence. Experimental results demonstrate the effectiveness of SC-CLIP, achieving state-of-the-art results across all datasets and surpassing previous methods by 9.5%. Notably, SC-CLIP boosts the performance of vanilla CLIP ViT-L/14 by 6.8 times. Our source code is available at https://github.com/SuleBai/SC-CLIP.

Artificial General Intelligence (AGI) Agents and Robots must be able to cope with everchanging environments and tasks. They must be able to actively construct new internal causal models of their interactions with the environment when new structural changes take place in the environment. Thus, we claim that active causal structure learning with latent variables (ACSLWL) is a necessary component to build AGI agents and robots. This paper describes how a complex planning and expectation-based detour behavior can be learned by ACSLWL when, unexpectedly, and for the first time, the simulated robot encounters a sort of transparent barrier in its pathway towards its target. ACSWL consists of acting in the environment, discovering new causal relations, constructing new causal models, exploiting the causal models to maximize its expected utility, detecting possible latent variables when unexpected observations occur, and constructing new structures-internal causal models and optimal estimation of the associated parameters, to be able to cope efficiently with the new encountered situations. That is, the agent must be able to construct new causal internal models that transform a previously unexpected and inefficient (sub-optimal) situation, into a predictable situation with an optimal operating plan.

With the broader adoption and highly successful development of Large Language Models (LLMs), there has been growing interest and demand for applying LLMs to autonomous driving technology. Driven by their natural language understanding and reasoning capabilities, LLMs have the potential to enhance various aspects of autonomous driving systems, from perception and scene understanding to interactive decision-making. This paper first introduces the novel concept of designing Large Language Models for Autonomous Driving (LLM4AD), followed by a review of existing LLM4AD studies. Then, a comprehensive benchmark is proposed for evaluating the instruction-following and reasoning abilities of LLM4AD systems, which includes LaMPilot-Bench, CARLA Leaderboard 1.0 Benchmark in simulation and NuPlanQA for multi-view visual question answering. Furthermore, extensive real-world experiments are conducted on autonomous vehicle platforms, examining both on-cloud and on-edge LLM deployment for personalized decision-making and motion control. Next, the future trends of integrating language diffusion models into autonomous driving are explored, exemplified by the proposed ViLaD (Vision-Language Diffusion) framework. Finally, the main challenges of LLM4AD are discussed, including latency, deployment, security and privacy, safety, trust and transparency, and personalization.

Generating realistic and controllable traffic scenes from natural language can greatly enhance the development and evaluation of autonomous driving systems. However, this task poses unique challenges: (1) grounding free-form text into spatially valid and semantically coherent layouts, (2) composing scenarios without predefined locations, and (3) planning multi-agent behaviors and selecting roads that respect agents' configurations. To address these, we propose a modular framework, TTSG, comprising prompt analysis, road retrieval, agent planning, and a novel plan-aware road ranking algorithm to solve these challenges. While large language models (LLMs) are used as general planners, our design integrates them into a tightly controlled pipeline that enforces structure, feasibility, and scene diversity. Notably, our ranking strategy ensures consistency between agent actions and road geometry, enabling scene generation without predefined routes or spawn points. The framework supports both routine and safety-critical scenarios, as well as multi-stage event composition. Experiments on SafeBench demonstrate that our method achieves the lowest average collision rate (3.5\%) across three critical scenarios. Moreover, driving captioning models trained on our generated scenes improve action reasoning by over 30 CIDEr points. These results underscore our proposed framework for flexible, interpretable, and safety-oriented simulation.

This paper presents CodeRefine, a novel framework for automatically transforming research paper methodologies into functional code using Large Language Models (LLMs). Our multi-step approach first extracts and summarizes key text chunks from papers, analyzes their code relevance, and creates a knowledge graph using a predefined ontology. Code is then generated from this structured representation and enhanced through a proposed retrospective retrieval-augmented generation approach. CodeRefine addresses the challenge of bridging theoretical research and practical implementation, offering a more accurate alternative to LLM zero-shot prompting. Evaluations on diverse scientific papers demonstrate CodeRefine's ability to improve code implementation from the paper, potentially accelerating the adoption of cutting-edge algorithms in real-world applications.

In drug discovery, predicting the absorption, distribution, metabolism, excretion, and toxicity (ADMET) properties of small-molecule drugs is critical for ensuring safety and efficacy. However, the process of accurately predicting these properties is often resource-intensive and requires extensive experimental data. To address this challenge, we propose SMILES-Mamba, a two-stage model that leverages both unlabeled and labeled data through a combination of self-supervised pretraining and fine-tuning strategies. The model first pre-trains on a large corpus of unlabeled SMILES strings to capture the underlying chemical structure and relationships, before being fine-tuned on smaller, labeled datasets specific to ADMET tasks. Our results demonstrate that SMILES-Mamba exhibits competitive performance across 22 ADMET datasets, achieving the highest score in 14 tasks, highlighting the potential of self-supervised learning in improving molecular property prediction. This approach not only enhances prediction accuracy but also reduces the dependence on large, labeled datasets, offering a promising direction for future research in drug discovery.

Multiple benchmarks have been developed to assess the alignment between deep neural networks (DNNs) and human vision. In almost all cases these benchmarks are observational in the sense they are composed of behavioural and brain responses to naturalistic images that have not been manipulated to test hypotheses regarding how DNNs or humans perceive and identify objects. Here we introduce the toolbox \textit{MindSet: Vision}, consisting of a collection of image datasets and related scripts designed to test DNNs on 30 psychological findings. In all experimental conditions, the stimuli are systematically manipulated to test specific hypotheses regarding human visual perception and object recognition. In addition to providing pre-generated datasets of images, we provide code to regenerate these datasets, offering many configurable parameters which greatly extend the dataset versatility for different research contexts, and code to facilitate the testing of DNNs on these image datasets using three different methods (similarity judgments, out-of-distribution classification, and decoder method), accessible via https://github.com/MindSetVision/MindSetVision. To illustrate the challenges these datasets pose for developing better DNN models of human vision, we test several models on range of datasets included in the toolbox.

The application of supervised learning techniques in combination with model predictive control (MPC) has recently generated significant interest, particularly in the area of approximate explicit MPC, where function approximators like deep neural networks are used to learn the MPC policy via optimal state-action pairs generated offline. While the aim of approximate explicit MPC is to closely replicate the MPC policy, substituting online optimization with a trained neural network, the performance guarantees that come with solving the online optimization problem are typically lost. This paper considers an alternative strategy, where supervised learning is used to learn the optimal value function offline instead of learning the optimal policy. This can then be used as the cost-to-go function in a myopic MPC with a very short prediction horizon, such that the online computation burden reduces significantly without affecting the controller performance. This approach differs from existing work on value function approximations in the sense that it learns the cost-to-go function by using offline-collected state-value pairs, rather than closed-loop performance data. The cost of generating the state-value pairs used for training is addressed using a sensitivity-based data augmentation scheme.

Language-conditioned robot manipulation is an emerging field aimed at enabling seamless communication and cooperation between humans and robotic agents by teaching robots to comprehend and execute instructions conveyed in natural language. This interdisciplinary area integrates scene understanding, language processing, and policy learning to bridge the gap between human instructions and robot actions. In this comprehensive survey, we systematically explore recent advancements in language-conditioned robot manipulation. We categorize existing methods based on the primary ways language is integrated into the robot system, namely language for state evaluation, language as a policy condition, language for cognitive planning and reasoning, and language in unified vision-language-action models. Specifically, we further analyze state-of-the-art techniques from five axes of action granularity, data and supervision regimes, system cost and latency, environments and evaluations, and cross-modal task specification. Additionally, we highlight the key debates in the field. Finally, we discuss open challenges and future research directions, focusing on potentially enhancing generalization capabilities and addressing safety issues in language-conditioned robot manipulators.

Score-based generative models, commonly referred to as diffusion models, have proven to be successful at generating text and image data. However, their adaptation to mixed-type tabular data remains underexplored. In this work, we propose CDTD, a Continuous Diffusion model for mixed-type Tabular Data. CDTD is based on a novel combination of score matching and score interpolation to enforce a unified continuous noise distribution for both continuous and categorical features. We explicitly acknowledge the necessity of homogenizing distinct data types by relying on model-specific loss calibration and initialization schemes. To further address the high heterogeneity in mixed-type tabular data, we introduce adaptive feature- or type-specific noise schedules. These ensure balanced generative performance across features and optimize the allocation of model capacity across features and diffusion time. Our experimental results show that CDTD consistently outperforms state-of-the-art benchmark models, captures feature correlations exceptionally well, and that heterogeneity in the noise schedule design boosts sample quality. Replication code is available at https://github.com/muellermarkus/cdtd.

Image-Text Retrieval (ITR) is challenging in bridging visual and lingual modalities. Contrastive learning has been adopted by most prior arts. Except for limited amount of negative image-text pairs, the capability of constrastive learning is restricted by manually weighting negative pairs as well as unawareness of external knowledge. In this paper, we propose our novel Coupled Diversity-Sensitive Momentum Constrastive Learning (CODER) for improving cross-modal representation. Firstly, a novel diversity-sensitive contrastive learning (DCL) architecture is invented. We introduce dynamic dictionaries for both modalities to enlarge the scale of image-text pairs, and diversity-sensitiveness is achieved by adaptive negative pair weighting. Furthermore, two branches are designed in CODER. One learns instance-level embeddings from image/text, and it also generates pseudo online clustering labels for its input image/text based on their embeddings. Meanwhile, the other branch learns to query from commonsense knowledge graph to form concept-level descriptors for both modalities. Afterwards, both branches leverage DCL to align the cross-modal embedding spaces while an extra pseudo clustering label prediction loss is utilized to promote concept-level representation learning for the second branch. Extensive experiments conducted on two popular benchmarks, i.e. MSCOCO and Flicker30K, validate CODER remarkably outperforms the state-of-the-art approaches. Our code is available at: https://github.com/BruceW91/CODER.

Dynamic game arises as a powerful paradigm for multi-robot planning, for which safety constraint satisfaction is crucial. Constrained stochastic games are of particular interest, as real-world robots need to operate and satisfy constraints under uncertainty. Existing methods for solving stochastic games handle chance constraints using exponential penalties with hand-tuned weights. However, finding a suitable penalty weight is nontrivial and requires trial and error. In this paper, we propose the chance-constrained iterative linear-quadratic stochastic games (CCILQGames) algorithm. CCILQGames solves chance-constrained stochastic games using the augmented Lagrangian method. We evaluate our algorithm in three autonomous driving scenarios, including merge, intersection, and roundabout. Experimental results and Monte Carlo tests show that CCILQGames can generate safe and interactive strategies in stochastic environments.

Byzantine attacks hinder the deployment of federated learning algorithms. Although we know that the benign gradients and Byzantine attacked gradients are distributed differently, to detect the malicious gradients is challenging due to (1) the gradient is high-dimensional and each dimension has its unique distribution and (2) the benign gradients and the attacked gradients are always mixed (two-sample test methods cannot apply directly). To address the above, for the first time, we propose MANDERA which is theoretically guaranteed to efficiently detect all malicious gradients under Byzantine attacks with no prior knowledge or history about the number of attacked nodes. More specifically, we transfer the original updating gradient space into a ranking matrix. By such an operation, the scales of different dimensions of the gradients in the ranking space become identical. The high-dimensional benign gradients and the malicious gradients can be easily separated. The effectiveness of MANDERA is further confirmed by experimentation on four Byzantine attack implementations (Gaussian, Zero Gradient, Sign Flipping, Shifted Mean), comparing with state-of-the-art defenses. The experiments cover both IID and Non-IID datasets.

Large Language Model based agents increasingly operate in high stakes, multi turn settings where factual grounding is critical, yet their memory systems typically rely on flat key value stores or plain vector retrieval with no mechanism to track the provenance or trustworthiness of stored knowledge. We present ElephantBroker, an open source cognitive runtime that unifies a Neo4j knowledge graph with a Qdrant vector store through the Cognee SDK to provide durable, verifiable agent memory. The system implements a complete cognitive loop (store, retrieve, score, compose, protect, learn) comprising a hybrid five source retrieval pipeline, an eleven dimension competitive scoring engine for budget constrained context assembly, a four state evidence verification model, a five stage context lifecycle with goal aware assembly and continuous compaction, a six layer cheap first guard pipeline for safety enforcement, an AI firewall providing enforceable tool call interception and multi tier safety scanning, a nine stage consolidation engine that strengthens useful patterns while decaying noise, and a numeric authority model governing multi organization identity with hierarchical access control. Architectural validation through a comprehensive test suite of over 2,200 tests spanning unit, integration, and end to end levels confirms subsystem correctness. The modular design supports three deployment tiers, five profile presets with inheritance, multi gateway isolation, and a management dashboard for human oversight, enabling configurations from lightweight memory only agents to full cognitive runtimes with enterprise grade safety and auditability.

Machine learning models can achieve high predictive accuracy in hydrological applications but often lack physical interpretability. The Mass-Conserving Perceptron (MCP) provides a physics-aware artificial intelligence (AI) framework that enforces conservation principles while allowing hydrological process relationships to be learned from data. In this study, we investigate how progressively embedding physically meaningful representations of hydrological processes within a single MCP storage unit improves predictive skill and interpretability in rainfall-runoff modeling. Starting from a minimal MCP formulation, we sequentially introduce bounded soil storage, state-dependent conductivity, variable porosity, infiltration capacity, surface ponding, vertical drainage, and nonlinear water-table dynamics. The resulting hierarchy of process-aware MCP models is evaluated across 15 catchments spanning five hydroclimatic regions of the continental United States using daily streamflow prediction as the target. Results show that progressively augmenting the internal physical structure of the MCP unit generally improves predictive performance. The influence of these process representations is strongly hydroclimate dependent: vertical drainage substantially improves model skill in arid and snow-dominated basins but reduces performance in rainfall-dominated regions, while surface ponding has comparatively small effects. The best-performing MCP configurations approach the predictive skill of a Long Short-Term Memory benchmark while maintaining explicit physical interpretability. These results demonstrate that embedding hydrological process constraints within AI architectures provides a promising pathway toward interpretable and process-aware rainfall-runoff modeling.

Most vision-language systems are static observers: they describe pixels, do not act, and cannot safely improve under shift. This passivity limits generalizable, physically grounded visual intelligence. Learning through action, not static description, is essential beyond curated data. We present Pixelis, a pixel-space agent that operates directly on images and videos via a compact set of executable operations (zoom/crop, segment, track, OCR, temporal localization) and learns from its consequences. Pixelis trains in three phases: (1) Supervised Fine-Tuning learns a pixel-tool grammar from Chain-of-Thought-Action traces with a masked imitation loss that upweights operation/argument tokens and auxiliary heads to stabilize pixel-grounded arguments; (2) Curiosity-Coherence Reward Fine-Tuning optimizes a dual-drive objective marrying prediction-error curiosity with adjacent-step coherence and a mild efficiency prior under a KL anchor, yielding short, valid, structured toolchains; (3) Pixel Test-Time RL performs label-free adaptation by retrieving neighbors, voting over complete trajectories rather than answers, and updating toward short, high-fidelity exemplars while constraining drift with a KL-to-EMA safety control. Across six public image and video benchmarks, Pixelis yields consistent improvements: the average relative gain is +4.08% over the same 8B baseline (peaking at +6.03% on VSI-Bench), computed as (ours-baseline)/baseline, while producing shorter, auditable toolchains and maintaining in-corridor KL during test-time learning. Acting within pixels, rather than abstract tokens, grounds multimodal perception in the physical world, linking visual reasoning with actionable outcomes, and enables embodied adaptation without external feedback.

We present THEMIS, a novel multi-task benchmark designed to comprehensively evaluate multimodal large language models (MLLMs) on visual fraud reasoning within real-world academic scenarios. Compared to existing benchmarks, THEMIS introduces three major advances. (1) Real-World Scenarios and Complexity: Our benchmark comprises over 4,000 questions spanning seven scenarios, derived from authentic retracted-paper cases and carefully curated multimodal synthetic data. With 60.47% complex-texture images, THEMIS bridges the critical gap between existing benchmarks and the complexity of real-world academic fraud. (2) Fraud-Type Diversity and Granularity: THEMIS systematically covers five challenging fraud types and introduces 16 fine-grained manipulation operations. On average, each sample undergoes multiple stacked manipulation operations, with the diversity and difficulty of these manipulations demanding a high level of visual fraud reasoning from the models. (3) Multi-Dimensional Capability Evaluation: We establish a mapping from fraud types to five core visual fraud reasoning capabilities, thereby enabling an evaluation that reveals the distinct strengths and specific weaknesses of different models across these core capabilities. Experiments on 16 leading MLLMs show that even the best-performing model, GPT-5, achieves an overall performance of only 56.15%, demonstrating that our benchmark presents a stringent test. We expect THEMIS to advance the development of MLLMs for complex, real-world fraud reasoning tasks.

Multimodal Large Language Models often suffer from object hallucination. While existing research utilizes attention enhancement and visual retracing, we find these works lack sufficient interpretability regarding attention drift in final model stages. In this paper, we investigate the layer wise evolution of visual features and discover that hallucination stems from deep layer attention regressing toward initial visual noise from early layers. We observe that output reliability depends on acquiring visual anchors at intermediate layers rather than final layers. Based on these insights, we propose CLVA, which stands for Cross-Layer Visual Anchors, a training free method that reinforces critical mid layer features while suppressing regressive noise. This approach effectively pulls deep layer attention back to correct visual regions by utilizing essential anchors captured from attention dynamics. We evaluate our method across diverse architectures and benchmarks, demonstrating outstanding performance without significant increase in computational time and GPU memory.

Out-of-Distribution (OOD) generalization has become a primary metric for evaluating image analysis systems. Since deep learning models tend to capture domain-specific context, they often develop shortcut dependencies on these non-causal features, leading to inconsistent performance across different data sources. Current techniques, such as invariance learning, attempt to mitigate this. However, they struggle to isolate highly mixed features within deep latent spaces. This limitation prevents them from fully resolving the shortcut learning problem.In this paper, we propose Hierarchical Causal Dropout (HCD), a method that uses channel-level causal masks to enforce feature sparsity. This approach allows the model to separate causal features from spurious ones, effectively performing a causal intervention at the representation level. The training is guided by a Matrix-based Mutual Information (MMI) objective to minimize the mutual information between latent features and domain labels, while simultaneously maximizing the information shared with class labels.To ensure stability, we incorporate a StyleMix-driven VICReg module, which prevents the masks from accidentally filtering out essential causal data. Experimental results on OOD benchmarks show that HCD performs better than existing top-tier methods.

Extending Reinforcement Learning with Verifiable Rewards (RLVR) to multimodal large language models (MLLMs) faces a fundamental challenge: their responses inherently interleave perception-related tokens, which ground visual content, with reasoning-related tokens, which construct reasoning chains. These token types instantiate distinct yet interdependent capacities -- visual grounding and symbolic reasoning -- making isolated optimization insufficient. Through token-level empirical analysis, we demonstrate that optimizing either perception- or reasoning-only tokens consistently underperforms full optimization, underscoring their inherent coupling. To address this, we propose a plug-and-play Token-Reweighting (ToR) strategy that explicitly models this interdependence by identifying critical tokens of both types and dynamically reweighting them during RLVR training. Applied on top of existing methods (e.g., GRPO and DAPO), ToR delivers consistent performance gains across multiple multi-modal reasoning benchmarks, achieving state-of-the-art performance with both accurate visual grounding and coherent reasoning.

Sparse autoencoders (SAEs) improve interpretability in multimodal models, but it remains unclear whether SAE features form modular, composable units for reasoning-an assumption underlying many intervention-based steering methods. We test this modularity hypothesis and find it often fails: intervening on a task-selective feature set can modestly improve reasoning accuracy, while intervening on the union of two such sets reliably induces output drift (large unintended changes in predictions) and degrades accuracy, even under norm-matched perturbations. This non modular circuit interference is consistent with shared internal pathways where feature unions amplify activation shifts. We develop a reproducible causal pipeline to localize and test these sparse visual thought circuits in Qwen3-VL-8B. On a controlled synthetic benchmark with seven task types and three difficulty levels, linear probes identify a mid decoder locus for task type information. We train SAEs at this layer, construct task-selective sets via an explicit rule, and perform inference time scaling and ablation while quantifying accuracy and drift. Our findings-validated with bootstrapped subsamples and permutation controls, and replicated across multiple VLM families and five diverse datasets clarify the boundaries of SAE feature composability and provide a rigorous diagnostic framework for more reliable VLM control.