Tongliang Liu

Yanbei Jiang, Xueqi Ma, Shu Liu et al.

Despite excelling on multimodal benchmarks, vision-language models (VLMs) largely remain a black box. In this paper, we propose a novel interpretability framework to systematically analyze the internal mechanisms of VLMs, focusing on the functional roles of attention heads in multimodal reasoning. To this end, we introduce CogVision, a dataset that decomposes complex multimodal questions into step-by-step subquestions designed to simulate human reasoning through a chain-of-thought paradigm, with each subquestion associated with specific receptive or cognitive functions such as high-level visual reception and inference. Using a probing-based methodology, we identify attention heads that specialize in these functions and characterize them as functional heads. Our analysis across diverse VLM families reveals that these functional heads are universally sparse, vary in number and distribution across functions, and mediate interactions and hierarchical organization. Furthermore, intervention experiments demonstrate their critical role in multimodal reasoning: removing functional heads leads to performance degradation, while emphasizing them enhances accuracy. These findings provide new insights into the cognitive organization of VLMs and suggest promising directions for designing models with more human-aligned perceptual and reasoning abilities.

Ziwen Li, Xin Wang, Hanlue Zhang et al.

The Vision-Language-Action (VLA) models have demonstrated remarkable performance on embodied tasks and shown promising potential for real-world applications. However, current VLAs still struggle to produce consistent and precise target-oriented actions, as they often generate redundant or unstable motions along trajectories, limiting their applicability in time-sensitive scenarios.In this work, we attribute these redundant actions to the spatially uniform perception field of existing VLAs, which causes them to be distracted by target-irrelevant objects, especially in complex environments.To address this issue, we propose an efficient PosA-VLA framework that anchors visual attention via pose-conditioned supervision, consistently guiding the model's perception toward task-relevant regions. The pose-conditioned anchor attention mechanism enables the model to better align instruction semantics with actionable visual cues, thereby improving action generation precision and efficiency. Moreover, our framework adopts a lightweight architecture and requires no auxiliary perception modules (e.g., segmentation or grounding networks), ensuring efficient inference. Extensive experiments verify that our method executes embodied tasks with precise and time-efficient behavior across diverse robotic manipulation benchmarks and shows robust generalization in a variety of challenging environments.

Zhanke Zhou, Chentao Cao, Xiao Feng et al.

We present AlphaApollo, a self-evolving agentic reasoning system that aims to address two bottlenecks in foundation model (FM) reasoning-limited model-intrinsic capacity and unreliable test-time iteration. AlphaApollo orchestrates multiple models with professional tools to enable deliberate, verifiable reasoning. It couples (i) a computation tool (Python with numerical and symbolic libraries) and (ii) a retrieval tool (task-relevant external information) to execute exact calculations and ground decisions. The system further supports multi-round, multi-model solution evolution via a shared state map that records candidates, executable checks, and feedback for iterative refinement. In evaluations on AIME 2024/2025 across multiple models, AlphaApollo delivers consistent gains: +5.15% Average@32 and +23.34% Pass@32 for Qwen2.5-14B-Instruct, and +8.91% Average@32 with +26.67% Pass@32 for Llama-3.3-70B-Instruct. Tool-use analysis shows that more than 80% of tool calls are successfully executed, with consistent outperformance of non-tool baselines, thereby lifting the capability ceiling of FMs. More empirical results and implementation details will be updated at https://github.com/tmlr-group/AlphaApollo.

Gaoge Han, Yongkang Cheng, Zhe Chen et al.

Two-hand reconstruction from monocular images faces persistent challenges due to complex and dynamic hand postures and occlusions, causing significant difficulty in achieving plausible interaction alignment. Existing approaches struggle with such alignment issues, often resulting in misalignment and penetration artifacts. To tackle this, we propose a dual-stage Foundation-to-Diffusion framework that precisely align 2D prior guidance from vision foundation models and diffusion-based generative 3D interaction refinement to achieve occlusion-robust two-hand reconstruction. First, we introduce a lightweight fusion alignment encoder that aligns fused multimodal 2D priors like key points, segmentation maps, and depth cues from vision foundation models during training. This provides robust structured guidance, further enabling efficient inference without heavy foundation model encoders at test time while maintaining high reconstruction accuracy. Second, we implement a two-hand diffusion model explicitly trained to convert interpenetrated 3D poses into plausible, penetration-free counterparts. Through collision gradient-guided denoising, the model rectifies artifacts while preserving natural spatial relationships between hands. Extensive evaluations demonstrate that our method achieves state-of-the-art performance on InterHand2.6M, HIC, and FreiHAND datasets, significantly advancing occlusion handling and interaction robustness. Our code will be publicly released.

Bang Liu, Xinfeng Li, Jiayi Zhang et al. · microsoft-research

The advent of large language models (LLMs) has catalyzed a transformative shift in artificial intelligence, paving the way for advanced intelligent agents capable of sophisticated reasoning, robust perception, and versatile action across diverse domains. As these agents increasingly drive AI research and practical applications, their design, evaluation, and continuous improvement present intricate, multifaceted challenges. This book provides a comprehensive overview, framing intelligent agents within modular, brain-inspired architectures that integrate principles from cognitive science, neuroscience, and computational research. We structure our exploration into four interconnected parts. First, we systematically investigate the modular foundation of intelligent agents, systematically mapping their cognitive, perceptual, and operational modules onto analogous human brain functionalities and elucidating core components such as memory, world modeling, reward processing, goal, and emotion. Second, we discuss self-enhancement and adaptive evolution mechanisms, exploring how agents autonomously refine their capabilities, adapt to dynamic environments, and achieve continual learning through automated optimization paradigms. Third, we examine multi-agent systems, investigating the collective intelligence emerging from agent interactions, cooperation, and societal structures. Finally, we address the critical imperative of building safe and beneficial AI systems, emphasizing intrinsic and extrinsic security threats, ethical alignment, robustness, and practical mitigation strategies necessary for trustworthy real-world deployment. By synthesizing modular AI architectures with insights from different disciplines, this survey identifies key research challenges and opportunities, encouraging innovations that harmonize technological advancement with meaningful societal benefit.

Yexiong Lin, Yu Yao, Tongliang Liu

Flow Matching (FM) is an effective framework for training a model to learn a vector field that transports samples from a source distribution to a target distribution. To train the model, early FM methods use random couplings, which often result in crossing paths and lead the model to learn non-straight trajectories that require many integration steps to generate high-quality samples. To address this, recent methods adopt Optimal Transport (OT) to construct couplings by minimizing geometric distances, which helps reduce path crossings. However, we observe that such geometry-based couplings do not necessarily align with the model's preferred trajectories, making it difficult to learn the vector field induced by these couplings, which prevents the model from learning straight trajectories. Motivated by this, we propose Model-Aligned Coupling (MAC), an effective method that matches training couplings based not only on geometric distance but also on alignment with the model's preferred transport directions based on its prediction error. To avoid the time-costly match process, MAC proposes to select the top-$k$ fraction of couplings with the lowest error for training. Extensive experiments show that MAC significantly improves generation quality and efficiency in few-step settings compared to existing methods. Project page: https://yexionglin.github.io/mac