Transformer

arXiv 2603.21635v1

cs.CV自动驾驶CV

PGR-Net: Prior-Guided ROI Reasoning Network for Brain Tumor MRI Segmentation

Brain tumor MRI segmentation is essential for clinical diagnosis and treatment planning, enabling accurate lesion detection and radiotherapy target delineation. However, tumor lesions occupy only a small fraction of the volumetric space, resulting in severe spatial sparsity, while existing segmentation networks often overlook clinically observed spatial priors of tumor occurrence, leading to redundant feature computation over extensive background regions. To address this issue, we propose PGR-Net (Prior-Guided ROI Reasoning Network) - an explicit ROI-aware framework that incorporates a data-driven spatial prior set to capture the distribution and scale characteristics of tumor lesions, providing global guidance for more stable segmentation. Leveraging these priors, PGR-Net introduces a hierarchical Top-K ROI decision mechanism that progressively selects the most confident lesion candidate regions across encoder layers to improve localization precision.

PGR-Net Authors

arXiv 2603.21626v1

cs.CVCVTransformer

Efficient Zero-Shot AI-Generated Image Detection

This paper addresses the critical challenge of detecting AI-generated images produced by text-to-image models. The proposed method is training-free and measures representation sensitivity to structured frequency perturbations, enabling detection of subtle manipulations between real and synthetic images. The approach uses only a single Fourier transform for perturbation generation, making it computationally lightweight and achieving one to two orders of magnitude faster inference than existing training-free detectors. Extensive experiments on challenging benchmarks including the OpenFake benchmark demonstrate superior performance over state-of-the-art methods.

Anonymous

arXiv 2603.21619v1

cs.CV自动驾驶端到端

Spatio-Temporal Attention Enhanced Multi-Agent DRL for UAV-Assisted Wireless Networks with Limited Communications

This paper addresses data transmission challenges in UAV-assisted wireless networks where multiple unmanned aerial vehicles serve as relays between ground users and a remote base station. The proposed delay-tolerant multi-agent deep reinforcement learning algorithm jointly optimizes trajectory planning, network formation, and transmission control while incorporating a delay-penalized reward mechanism to encourage inter-UAV information sharing. To handle information loss from unreliable channel conditions, a spatio-temporal attention mechanism predicts and recovers missing network state information, enhancing each UAV's situational awareness for improved collaboration and overall network throughput.

arXiv 2603.21594v1

cs.CV自动驾驶CV

Mind over Space: Can Multimodal Large Language Models Mentally Navigate?

This paper addresses the limitation of Multimodal Large Language Models (MLLMs) in embodied agents, which struggle with spatial reasoning across extensive spatiotemporal scales. The authors introduce Video2Mental, a benchmark that evaluates mental navigation capabilities by requiring models to construct hierarchical cognitive maps from long egocentric videos and generate landmark-based path plans. The research draws inspiration from cognitive science, exploring how biological intelligence uses mental navigation and spatial simulation prior to action. Benchmarking results demonstrate that mental navigation capabilities do not naturally emerge from standard pre-training, with frontier MLLMs showing significant challenges in zero-shot scenarios. Planning accuracy is verified through simulator-based physical interaction, providing a comprehensive evaluation framework for embodied AI systems.

Unknown

arXiv 2603.21577v1

cs.CVCVTransformer

VIGIL: Part-Grounded Structured Reasoning for Generalizable Deepfake Detection

This paper presents VIGIL, a novel part-centric structured forensic framework for deepfake detection using multimodal large language models. The approach employs a plan-then-examine pipeline where the model first plans which facial parts warrant inspection based on global visual cues, then examines each part with independently sourced forensic evidence. A stage-gated injection mechanism delivers part-level forensic evidence only during examination to ensure unbiased part selection. The framework is inspired by expert forensic practice and aims to improve the reliability of deepfake detection by separating evidence generation from manipulation localization, addressing the issue of hallucinated explanations in current MLLM-based methods.

VIGIL Authors

arXiv 2603.21526v1

cs.CL自动驾驶端到端

A transformer architecture alteration to incentivise externalised reasoning

We propose a novel architectural modification and post-training pipeline for enhancing large language model reasoning capabilities by teaching models to truncate forward passes early. Our approach augments the standard transformer architecture with an early-exit mechanism at intermediate layers, enabling the model to exit at shallower layers when tokens can be predicted without deep computation. Through a calibration stage followed by reinforcement learning, we incentivize the model to exit as early as possible while preserving task performance. Preliminary experiments on small reasoning models demonstrate adaptive computation reduction across tokens, suggesting that at appropriate scale, this approach can minimize excess computation for non-myopic planning using internal activations, reserving deep computation only for difficult-to-predict tokens.

Anonymous

arXiv 2603.21376v1

cs.CVCVTransformer

RoboAlign: Learning Test-Time Reasoning for Language-Action Alignment in Vision-Language-Action Models

This paper presents RoboAlign, a systematic framework for improving vision-language-action models (VLAs) by enhancing embodied reasoning capabilities in multimodal large language models (MLLMs). The key innovation involves sampling action tokens through zero-shot natural language reasoning and refining them using reinforcement learning to improve action accuracy. The approach effectively bridges the modality gap between language understanding and low-level robot actions, facilitating knowledge transfer from multimodal LLMs to embodied agents. Experimental validation demonstrates that training VLAs with this framework leads to reliable performance improvements in robotic manipulation tasks.

RoboAlign Authors

arXiv 2603.21341v1

cs.RO端到端Transformer

Evaluating Factor-Wise Auxiliary Dynamics Supervision for Latent Structure and Robustness in Simulated Humanoid Locomotion

This paper evaluates whether factor-wise auxiliary dynamics supervision produces useful latent structure or improved robustness in simulated humanoid locomotion. We introduce DynaMITE, a transformer encoder with a factored 24-dimensional latent space trained using per-factor auxiliary losses during proximal policy optimization. Our method is compared against LSTM, plain Transformer, and MLP baselines on a Unitree G1 humanoid robot across four Isaac Lab tasks. Through comprehensive ablation studies with 10 random seeds, we analyze the contributions of tanh bottlenecks and auxiliary losses to in-distribution reward performance. Results demonstrate that the supervised latent fails to produce decodable or functionally separable factor structure, with probe R-squared near zero and minimal reward changes when subspaces are clamped.

arXiv 2603.21268v1

cs.CV自动驾驶端到端

Enhancing Brain Tumor Classification Using Vision Transformers with Colormap-Based Feature Representation on BRISC2025 Dataset

This study presents a deep learning framework for multi-class brain tumor classification from magnetic resonance imaging scans using Vision Transformers enhanced with colormap-based feature representation. The proposed method leverages transformer architectures to capture long-range dependencies while incorporating color mapping techniques to emphasize structural and intensity variations in MRI scans. Experiments on the BRISC2025 dataset containing glioma, meningioma, pituitary tumor, and non-tumor cases demonstrate superior performance with 98.90% accuracy, significantly outperforming baseline approaches. The framework utilizes standard metrics including accuracy, precision, recall, F1-score, and AUC for comprehensive evaluation.

arXiv 2603.21234v1

cs.CV自动驾驶CV

CounterScene: Counterfactual Causal Reasoning in Generative World Models for Safety-Critical Closed-Loop Evaluation

This paper presents CounterScene, a novel framework that enables structured counterfactual reasoning in generative Bird's Eye View (BEV) world models for safety-critical driving scenario generation. The approach addresses the challenge of creating realistic yet adversarially effective safety scenarios by introducing causal adversarial agent identification to determine critical agents and conflict types. CounterScene develops a conflict-aware interactive world model using causal interaction graphs to explicitly model dynamic inter-agent dependencies. The framework performs minimal interventions through stage-adaptive counterfactual guidance, effectively bridging the gap between realism and adversarial robustness in autonomous driving safety evaluation.

CounterScene Authors

arXiv 2603.21104v1

cs.CVCVTransformer

Cortical Policy: A Dual-Stream View Transformer for Robotic Manipulation

View transformers process multi-view observations to predict actions and have shown impressive performance in robotic manipulation. Existing methods typically extract static visual representations in a view-specific manner, leading to inadequate 3D spatial reasoning ability and a lack of dynamic adaptation. Taking inspiration from how the human brain integrates static and dynamic views, we propose Cortical Policy, a novel dual-stream view transformer for robotic manipulation that jointly reasons from static-view and dynamic-view streams. The static-view stream enhances spatial understanding by aligning features of geometrically consistent keypoints extracted from a pretrained 3D foundation model. The dynamic-view stream achieves adaptive adjustment through position-aware pretraining of an egocentric gaze estimation model, computationally replicating the human cortical dorsal pathway. The complementary view representations enable improved robotic manipulation performance.

Cortical Policy Authors

arXiv 2603.21051v1

cs.CV自动驾驶CV

A Two-stage Transformer Framework for Temporal Localization of Distracted Driver Behaviors

This paper presents a temporal action localization framework specifically designed for driver monitoring systems in autonomous driving applications. The framework employs a two-stage pipeline combining VideoMAE-based feature extraction with an Augmented Self-Mask Attention detector to identify hazardous driving behaviors from in-cabin video streams. A Spatial Pyramid Pooling-Fast module captures multi-scale temporal features for improved localization accuracy. The approach is optimized for transportation safety checkpoints and fleet management assessment systems, demonstrating a trade-off between model capacity and computational efficiency.

arXiv 2603.21048v1

cs.CL大语言模型端到端

Mitigating Selection Bias in Large Language Models via Permutation-Aware GRPO

Large language models often exhibit selection bias in multiple-choice and pairwise evaluation tasks due to non-semantic factors such as option positions and label symbols. Existing inference-time debiasing methods are computationally expensive and may harm reasoning capabilities, while pointwise training approaches ignore the importance of consistent answers across different question permutations. This paper proposes Permutation-Aware Group Relative Policy Optimization (PA-GRPO), which addresses selection bias by enforcing permutation-consistent semantic reasoning through two complementary mechanisms: cross-permutation advantage computation relative to the mean reward over all permutations, and consistency-aware reward that encourages stable decision-making across different permutations. The proposed approach effectively mitigates selection bias while preserving the model's reasoning capabilities.

PA-GRPO Authors

arXiv 2603.21016v1

cs.AI自动驾驶Transformer

AutoMOOSE: An Agentic AI for Autonomous Phase-Field Simulation

This paper introduces AutoMOOSE, an open-source agentic framework designed to orchestrate the complete simulation lifecycle of multiphysics phase-field materials modeling through natural language prompts. The system employs a five-agent pipeline where an Input Writer coordinates six sub-agents, while a Reviewer autonomously diagnoses and corrects runtime failures without human intervention. A modular plugin architecture allows integration of new phase-field formulations, and a Model Context Protocol server exposes ten structured tools for external client interoperability. The framework is validated on a copper grain growth benchmark, demonstrating its ability to generate valid MOOSE simulation input files.

AutoMOOSE Authors

arXiv 2603.20986v1

cs.AI端到端Transformer

Detection of adversarial intent in Human-AI teams using LLMs

Large language models (LLMs) are increasingly deployed in human-AI teams as support agents for complex tasks such as information retrieval, programming, and decision-making assistance. This paper studies the potential role of LLMs as defensive supervisors within mixed human-AI teams to detect malicious behavior. Using a dataset consisting of multi-party conversations and decisions over a 25-round horizon, we formulate the problem of malicious behavior detection from interaction traces. We find that LLMs are capable of identifying malicious behavior effectively, demonstrating their potential as defensive actors in collaborative environments.

Unknown

arXiv 2603.20976v1

cs.CL端到端Transformer

Do LLM-Driven Agents Exhibit Engagement Mechanisms? Controlled Tests of Information Load, Descriptive Norms, and Popularity Cues

Large language models enable increasingly expressive agent-based simulations, but pose methodological challenges regarding behavioral validity. This paper evaluates LLM-driven simulation credibility through a social media test case examining information engagement. Using a Weibo-like environment, the study systematically manipulates information load and descriptive norms while allowing popularity cues to evolve endogenously. The research tests whether simulated user behavior responds systematically to theoretical constructs rather than producing merely plausible outputs. Findings indicate that engagement responds systematically to information load and descriptive norms, with sensitivity to popularity cues varying across contexts. The paper discusses methodological implications for simulation-based communication research, particularly for multi-condition experimental designs involving LLM-driven agents.