arXiv:2603.02228v1 Announce Type: cross Abstract: The proof that Large Language Models (LLMs) augmented with external read-write memory constitute a computationally universal system has established the theoretical foundation for general-purpose agents. However, existing implementations face a critical bottleneck: the finite and costly Context Window, which functions not as infinite memory but as a scarce semantic cache. In this work, we introduce \textit{Neural Paging}, a hierarchical architecture that decouples symbolic reasoning from information resource management. We formulate the \textit{Context Paging Problem (CPP)} and propose a lightweight, differentiable \textit{Page Controller} designed to approximate ``Semantic Belady's Optimality'' -- retaining tokens with high future utility under explicit assumptions on access patterns. We provide theoretical analysis showing that, under bounded context window size~$K$, Neural Paging reduces the asymptotic complexity of long-horizon reasoning from quadratic $O(N^2)$ to $O(N \cdot K^2)$, and we derive a robustness bound (Theorem~4) that quantifies competitive-ratio degradation under policy-dependent access with bounded sensitivity. We validate these bounds on synthetic paging traces, confirming that the theoretical guarantees hold and identifying significant slack that motivates learned policies.

arXiv:2506.07218v3 Announce Type: replace-cross Abstract: Enhancing the multimodal reasoning capabilities of Multimodal Large Language Models (MLLMs) is a challenging task that has attracted increasing attention in the community. Recently, several studies have applied Reinforcement Learning with Verifiable Rewards (RLVR) to the multimodal domain in order to enhance the reasoning abilities of MLLMs. However, these works largely overlook the enhancement of multimodal perception capabilities in MLLMs, which serve as a core prerequisite and foundational component of complex multimodal reasoning. Through McNemar's test, we find that existing RLVR method fails to effectively enhance the multimodal perception capabilities of MLLMs, thereby limiting their further improvement in multimodal reasoning. To address this limitation, we propose Perception-R1, which introduces a novel visual perception reward that explicitly encourages MLLMs to perceive the visual content accurately, thereby can effectively incentivizing both their multimodal perception and reasoning capabilities. Specifically, we first collect textual visual annotations from the CoT trajectories of multimodal problems, which will serve as visual references for reward assignment. During RLVR training, we employ a judging LLM to assess the consistency between the visual annotations and the responses generated by MLLM, and assign the visual perception reward based on these consistency judgments. Extensive experiments on several multimodal reasoning benchmarks demonstrate the effectiveness of our Perception-R1, which achieves state-of-the-art performance on most benchmarks using only 1,442 training data. Our code and dataset will be available at https://github.com/tongxiao2002/Perception-R1.

arXiv:2602.07801v2 Announce Type: replace-cross Abstract: In long-video understanding, conventional uniform frame sampling often fails to capture key visual evidence, leading to degraded performance and increased hallucinations. To address this, recent agentic thinking-with-videos paradigms have emerged, adopting a localize-clip-answer pipeline in which the model actively identifies relevant video segments, performs dense sampling within those clips, and then produces answers. However, existing methods remain inefficient, suffer from weak localization, and adhere to rigid workflows. To solve these issues, we propose VideoTemp-o3, a unified agentic thinking-with-videos framework that jointly models video grounding and question answering. VideoTemp-o3 exhibits strong localization capability, supports on-demand clipping, and can refine inaccurate localizations. Specifically, in the supervised fine-tuning stage, we design a unified masking mechanism that encourages exploration while preventing noise. For reinforcement learning, we introduce dedicated rewards to mitigate reward hacking. Besides, from the data perspective, we develop an effective pipeline to construct high-quality long video grounded QA data, along with a corresponding benchmark for systematic evaluation across various video durations. Experimental results demonstrate that our method achieves remarkable performance on both long video understanding and grounding.