We propose a novel method that leverages sparse autoencoders (SAEs) and clustering techniques to analyze the internal token representations of large language models (LLMs) and guide generations in mathematical reasoning tasks. Our approach first trains an SAE to generate sparse vector representations for training tokens, then applies k-means clustering to construct a graph where vertices represent token clusters and weighted edges capture sequential token transitions. Using this graph, we define an edge-weight based reward function to quantify adherence to established reasoning traces, thereby identifying exploitative reasoning trajectories. Additionally, we measure generation diversity from clustering to assess the extent of exploration. Our findings indicate that balancing both exploitation and exploration is crucial for achieving high accuracy in mathematical reasoning tasks. During generation, the SAE can serve as a scalable reward model to guide generations, ensuring a balanced trade-off between exploitation and exploration. This prevents extreme behaviors in either direction, ultimately fostering a higher-quality reasoning process in LLMs.

Although current large Vision-Language Models (VLMs) have advanced in multimodal understanding and reasoning, their fundamental perceptual and reasoning abilities remain limited. Specifically, even on simple jigsaw tasks, existing VLMs perform near randomly, revealing deficiencies in core perception and reasoning capabilities. While high-quality vision-language data can enhance these capabilities, its scarcity and limited scalability impose significant constraints. To address this, we propose AGILE, an Agentic jiGsaw Interaction Learning for Enhancing visual perception and reasoning in VLMs. AGILE formulates jigsaw solving as an interactive process, enabling the model to progressively engage with the environment. At each step, the model generates executable code to perform an action based on the current state, while the environment provides fine-grained visual feedback to guide task completion. Through this iterative cycle of observation and interaction, the model incrementally improves its perceptual and reasoning capabilities via exploration and feedback. Experimental results show that AGILE not only substantially boosts performance on jigsaw tasks of varying complexity (e.g., increasing accuracy from 9.5% to 82.8% under the 2 $\times$ 2 setting) but also demonstrates strong generalization across 9 general vision tasks, achieving an average improvement of 3.1%. These results indicate notable enhancements in both perceptual and reasoning abilities. This work opens a new avenue for advancing reasoning and generalization in multimodal models and provides an efficient, scalable solution to the scarcity of multimodal reinforcement learning data. The code and datasets is available at https://github.com/yuzeng0-0/AGILE .

Reasoning has emerged as a pivotal capability in Large Language Models (LLMs). Through Reinforcement Learning (RL), typically Group Relative Policy Optimization (GRPO), these models are able to solve complex tasks such as mathematics and code generation. Building on these advances, recent research has sought to extend reasoning to Vision-Language Models (VLMs), yielding promising results across diverse visual tasks. Despite this progress, our study uncovers the dual nature of multimodal reasoning: while it substantially enhances logical inference and facilitates performance on challenging problems, it may gradually impair perceptual grounding, leading to recognition failures on otherwise basic visual questions. Through further analysis, we attribute this phenomenon to visual forgetting, wherein prolonged reasoning causes the model to increasingly disregard visual input. To address this, we propose Vision-Anchored Policy Optimization (VAPO), a simple yet effective method that explicitly steers the reasoning process toward visually grounded trajectories. Our result model, VAPO-Thinker-7B, significantly strengthens the model's reliance on visual information and achieves new state-of-the-art results on a wide range of established benchmarks. Project page: https://xytian1008.github.io/VAPO/

Work done during internship at BIGAI.

In absence of reliable domain expertise, one might resort to inferring the latent graph structure, which is often difficult due to the vast search space of possible graphs.

Complex numerical time series analysis often demands multi-step reasoning capabilities beyond current models' reach. Tasks like medical diagnosis and weather forecasting require sequential reasoning processes -- including counterfactual analysis, logical deduction, knowledge application, and multi-modal contextual integration -- that existing time series models cannot explicitly perform. While recent research has shown large language models (LLMs) can achieve sophisticated Chain-of-Thought (CoT) reasoning through reinforcement learning (RL), these advances have primarily focused on mathematical and coding domains, with LLMs still demonstrating poor performance on time series tasks. We introduce Chain Of thought for Understanding Numerical Time Series (COUNTS), the first framework that trains LLMs to perform CoT reasoning across diverse time series tasks using RL with verifiable rewards. Our approach employs a Residual Vector-Quantized VAE to create high-fidelity discrete tokens that seamlessly integrate into a pre-trained LLM's vocabulary. COUNTS undergoes a two-stage training process: first, supervised fine-tuning on time series analysis tasks to master our novel representations, followed by Group Relative Policy Optimization training on verifiable problems using prompting strategies that encourage explicit reasoning steps before producing final answers. Our experiments demonstrate that this RL-driven approach with intermediate CoT reasoning significantly enhances LLM performance across various time series analysis tasks, opening new possibilities for complex temporal data reasoning.

Despite rapid progress in autonomous robotics, executing complex or long-horizon tasks remains a fundamental challenge. Most current approaches follow an open-loop paradigm with limited reasoning and no feedback, resulting in poor robustness to environmental changes and severe error accumulation. We present RoboPilot, a dual-thinking closed-loop framework for robotic manipulation that supports adaptive reasoning for complex tasks in real-world dynamic environments. RoboPilot leverages primitive actions for structured task planning and flexible action generation, while introducing feedback to enable replanning from dynamic changes and execution errors. Chain-of-Thought reasoning further enhances high-level task planning and guides low-level action generation. The system dynamically switches between fast and slow thinking to balance efficiency and accuracy. To systematically evaluate the robustness of RoboPilot in diverse robot manipulation scenarios, we introduce RoboPilot-Bench, a benchmark spanning 21 tasks across 10 categories, including infeasible-task recognition and failure recovery. Experiments show that RoboPilot outperforms state-of-the-art baselines by 25.9\% in task success rate, and the real-world deployment on an industrial robot further demonstrates its robustness in real-world settings.

Foundation models have reshaped AI by unifying fragmented architectures into scalable backbones with multimodal reasoning and contextual adaptation. In parallel, the long-standing notion of AI agents, defined by the sensing-decision-action loop, is entering a new paradigm: with FMs as their cognitive core, agents transcend rule-based behaviors to achieve autonomy, generalization, and self-reflection. This dual shift is reinforced by real-world demands such as autonomous driving, robotics, virtual assistants, and GUI agents, as well as ecosystem advances in embedded hardware, edge computing, mobile deployment platforms, and communication protocols that together enable large-scale deployment. Yet this convergence collides with reality: while applications demand long-term adaptability and real-time interaction, mobile and edge deployments remain constrained by memory, energy, bandwidth, and latency. This creates a fundamental tension between the growing complexity of FMs and the limited resources of deployment environments. This survey provides the first systematic characterization of adaptive, resource-efficient agentic AI systems. We summarize enabling techniques into elastic inference, test-time adaptation, dynamic multimodal integration, and agentic AI applications, and identify open challenges in balancing accuracy-latency-communication trade-offs and sustaining robustness under distribution shifts. We further highlight future opportunities in algorithm-system co-design, cognitive adaptation, and collaborative edge deployment. By mapping FM structures, cognition, and hardware resources, this work establishes a unified perspective toward scalable, adaptive, and resource-efficient agentic AI. We believe this survey can help readers to understand the connections between enabling technologies while promoting further discussions on the fusion of agentic intelligence and intelligent agents.

Traditional recommender systems rely on passive feedback mechanisms that limit users to simple choices such as like and dislike. However, these coarse-grained signals fail to capture users' nuanced behavior motivations and intentions. In turn, current systems cannot also distinguish which specific item attributes drive user satisfaction or dissatisfaction, resulting in inaccurate preference modeling. These fundamental limitations create a persistent gap between user intentions and system interpretations, ultimately undermining user satisfaction and harming system effectiveness. To address these limitations, we introduce the Interactive Recommendation Feed (IRF), a pioneering paradigm that enables natural language commands within mainstream recommendation feeds. Unlike traditional systems that confine users to passive implicit behavioral influence, IRF empowers active explicit control over recommendation policies through real-time linguistic commands. To support this paradigm, we develop RecBot, a dual-agent architecture where a Parser Agent transforms linguistic expressions into structured preferences and a Planner Agent dynamically orchestrates adaptive tool chains for on-the-fly policy adjustment. To enable practical deployment, we employ simulation-augmented knowledge distillation to achieve efficient performance while maintaining strong reasoning capabilities. Through extensive offline and long-term online experiments, RecBot shows significant improvements in both user satisfaction and business outcomes.