This paper introduces a theoretical framework for accelerating the evaluation of Conditional Value-at-Risk (CVaR) value functions in Partially Observable Markov Decision Processes (POMDPs) with formal performance guarantees. It derives novel CVaR bounds for random variables, enabling faster policy evaluation through action elimination using simplified models.

This paper introduces RebuttalAgent, an AI framework that grounds academic rebuttal in Theory of Mind (ToM) to generate strategic and persuasive responses. It proposes a ToM-Strategy-Response (TSR) pipeline, supported by a large-scale synthetic dataset (RebuttalBench) and a specialized evaluation model (Rebuttal-RM), significantly outperforming existing models in automated and human evaluations.

MonoRace is an autonomous drone racing system that utilizes a monocular camera and IMU to achieve champion-level performance, notably winning the A2RL 2025 competition. It features robust state estimation combining neural-network-based gate segmentation with a drone model, an offline optimization procedure, and a neural network for guidance and control.

The paper introduces a comprehensive method and ecosystem (NexAU, NexA4A, NexGAP) to overcome limitations in scaling interactive environments for training agentic Large Language Models (LLMs). This infrastructure enables the systematic generation of diverse, complex, and realistically grounded interaction trajectories for LLMs.

This paper introduces Reward Forcing, a novel framework for efficient streaming video generation that tackles issues like diminished motion dynamics and over-reliance on initial frames. It achieves state-of-the-art performance by combining EMA-Sink for improved long-term context and Rewarded Distribution Matching Distillation (Re-DMD) to enhance motion quality.

This paper introduces Stage-Aware Reinforcement (StARe), a novel module that decomposes long-horizon robotic manipulation tasks into semantically meaningful stages, providing dense, interpretable reinforcement signals. Integrated into the Imitation → Preference → Interaction (IPI) fine-tuning pipeline, StARe significantly improves the performance and robustness of Vision-Language-Action (VLA) models on complex manipulation tasks.

This paper provides a comprehensive tutorial on enhancing Reinforcement Learning (RL) for wireless networks using Large Language Models (LLMs). It proposes a taxonomy for LLM roles in RL (state perceiver, reward designer, decision-maker, generator) and showcases their application in various wireless scenarios to address classical RL's limitations in generalization, interpretability, and sample efficiency.

AdaptVision introduces an efficient VLM paradigm that autonomously determines the minimum number of visual tokens required for each sample by employing a coarse-to-fine visual acquisition strategy, leading to superior performance with significantly reduced computational overhead.

This paper introduces TempR1, a novel temporal-aware multi-task reinforcement learning framework designed to significantly enhance the temporal understanding capabilities of Multimodal Large Language Models (MLLMs). By integrating diverse temporal tasks and tailored reward functions, TempR1 achieves state-of-the-art performance across various video understanding benchmarks and improves generalization.

This paper introduces SpaceTools, a vision-language model trained with Double Interactive Reinforcement Learning (DIRL) to achieve precise spatial reasoning and real-world robot manipulation by effectively coordinating multiple external tools. It demonstrates state-of-the-art performance on various spatial understanding benchmarks.