Language model pretraining with next token prediction has proved effective for scaling compute but is limited to the amount of available training data. Scaling reinforcement learning (RL) unlocks a new axis for the continued improvement of artificial intelligence, with the promise that large language models (LLMs) can scale their training data by learning to explore with rewards. However, prior published work has not produced competitive results. In light of this, we report on the training practice of Kimi k1.5, our latest multi-modal LLM trained with RL, including its RL training techniques, multi-modal data recipes, and infrastructure optimization. Long context scaling and improved policy optimization methods are key ingredients of our approach, which establishes a simplistic, effective RL framework without relying on more complex techniques such as Monte Carlo tree search, value functions, and process reward models. Notably, our system achieves state-of-the-art reasoning performance across multiple benchmarks and modalities -- e.g., 77.5 on AIME, 96.2 on MATH 500, 94-th percentile on Codeforces, 74.9 on MathVista -- matching OpenAI's o1. Moreover, we present effective long2short methods that use long-CoT techniques to improve short-CoT models, yielding state-of-the-art short-CoT reasoning results -- e.g., 60.8 on AIME, 94.6 on MATH500, 47.3 on LiveCodeBench -- outperforming existing short-CoT models such as GPT-4o and Claude Sonnet 3.5 by a large margin (up to +550%).

Combining reinforcement learning with language grounding is challenging as the agent needs to explore the environment while simultaneously learning multiple language-conditioned tasks. To address this, we introduce a novel method: the compositionally-enabled reinforcement learning language agent (CERLLA). Our method reduces the sample complexity of tasks specified with language by leveraging compositional policy representations and a semantic parser trained using reinforcement learning and in-context learning. We evaluate our approach in an environment requiring function approximation and demonstrate compositional generalization to novel tasks. Our method significantly outperforms the previous best non-compositional baseline in terms of sample complexity on 162 tasks designed to test compositional generalization. Our model attains a higher success rate and learns in fewer steps than the non-compositional baseline. It reaches a success rate equal to an oracle policy's upper-bound performance of 92%. With the same number of environment steps, the baseline only reaches a success rate of 80%.

Intrinsic reward shaping has emerged as a prevalent approach to solving hard-exploration and sparse-rewards environments in reinforcement learning (RL). While single intrinsic rewards, such as curiosity-driven or novelty-based methods, have shown effectiveness, they often limit the diversity and efficiency of exploration. Moreover, the potential and principle of combining multiple intrinsic rewards remains insufficiently explored. To address this gap, we introduce HIRE (Hybrid Intrinsic REward), a flexible and elegant framework for creating hybrid intrinsic rewards through deliberate fusion strategies. With HIRE, we conduct a systematic analysis of the application of hybrid intrinsic rewards in both general and unsupervised RL across multiple benchmarks. Extensive experiments demonstrate that HIRE can significantly enhance exploration efficiency and diversity, as well as skill acquisition in complex and dynamic settings.

Video encoders optimize compression for human perception by minimizing reconstruction error under bit-rate constraints. In many modern applications such as autonomous driving, an overwhelming majority of videos serve as input for AI systems performing tasks like object recognition or segmentation, rather than being watched by humans. It is therefore useful to optimize the encoder for a downstream task instead of for perceptual image quality. However, a major challenge is how to combine such downstream optimization with existing standard video encoders, which are highly efficient and popular. Here, we address this challenge by controlling the Quantization Parameters (QPs) at the macro-block level to optimize the downstream task. This granular control allows us to prioritize encoding for task-relevant regions within each frame. We formulate this optimization problem as a Reinforcement Learning (RL) task, where the agent learns to balance long-term implications of choosing QPs on both task performance and bit-rate constraints. Notably, our policy does not require the downstream task as an input during inference, making it suitable for streaming applications and edge devices such as vehicles. We demonstrate significant improvements in two tasks, car detection, and ROI (saliency) encoding. Our approach improves task performance for a given bit rate compared to traditional task agnostic encoding methods, paving the way for more efficient task-aware video compression.

Traditional approaches to studying decision-making in neuroscience focus on simplified behavioral tasks where animals perform repetitive, stereotyped actions to receive explicit rewards. While informative, these methods constrain our understanding of decision-making to short timescale behaviors driven by explicit goals. In natural environments, animals exhibit more complex, long-term behaviors driven by intrinsic motivations that are often unobservable. Recent works in time-varying inverse reinforcement learning (IRL) aim to capture shifting motivations in long-term, freely moving behaviors. However, a crucial challenge remains: animals make decisions based on their history, not just their current state. To address this, we introduce SWIRL (SWitching IRL), a novel framework that extends traditional IRL by incorporating time-varying, history-dependent reward functions. SWIRL models long behavioral sequences as transitions between short-term decision-making processes, each governed by a unique reward function. SWIRL incorporates biologically plausible history dependency to capture how past decisions and environmental contexts shape behavior, offering a more accurate description of animal decision-making. We apply SWIRL to simulated and real-world animal behavior datasets and show that it outperforms models lacking history dependency, both quantitatively and qualitatively. This work presents the first IRL model to incorporate history-dependent policies and rewards to advance our understanding of complex, naturalistic decision-making in animals. Historically, decision making in neuroscience has been studied using simplified assays where animals perform repetitive, stereotyped actions (such as licks, nose pokes, or lever presses) in response to sensory stimuli to obtain an explicit reward. While this approach has its advantages, it has limited our understanding of decision making to scenarios where animals are instructed to achieve an explicit goal over brief timescales, usually no more than tens of seconds.

Goal-oriented communications prioritize application-driven objectives over data accuracy, enabling intelligent next-generation wireless systems. Efficient scheduling in multi-device, multi-channel systems poses significant challenges due to high-dimensional state and action spaces. We address these challenges by deriving key structural properties of the optimal solution to the goal-oriented scheduling problem, incorporating Age of Information (AoI) and channel states. Specifically, we establish the monotonicity of the optimal state value function (a measure of long-term system performance) w.r.t. channel states and prove its asymptotic convexity w.r.t. AoI states. Additionally, we derive the monotonicity of the optimal policy w.r.t. channel states, advancing the theoretical framework for optimal scheduling. Leveraging these insights, we propose the structure-guided unified dual on-off policy DRL (SUDO-DRL), a hybrid algorithm that combines the stability of on-policy training with the sample efficiency of off-policy methods. Through a novel structural property evaluation framework, SUDO-DRL enables effective and scalable training, addressing the complexities of large-scale systems. Numerical results show SUDO-DRL improves system performance by up to 45% and reduces convergence time by 40% compared to state-of-the-art methods. It also effectively handles scheduling in much larger systems, where off-policy DRL fails and on-policy benchmarks exhibit significant performance loss, demonstrating its scalability and efficacy in goal-oriented communications.

Multi-Agent Reinforcement Learning (MARL) has gained significant traction for solving complex real-world tasks, but the inherent stochasticity and uncertainty in these environments pose substantial challenges to efficient and robust policy learning. While Distributional Reinforcement Learning has been successfully applied in single-agent settings to address risk and uncertainty, its application in MARL is substantially limited. In this work, we propose a novel approach that integrates distributional learning with a safety-focused loss function to improve convergence in cooperative MARL tasks. Specifically, we introduce a Barrier Function based loss that leverages safety metrics, identified from inherent faults in the system, into the policy learning process. This additional loss term helps mitigate risks and encourages safer exploration during the early stages of training. We evaluate our method in the StarCraft II micromanagement benchmark, where our approach demonstrates improved convergence and outperforms state-of-the-art baselines in terms of both safety and task completion. Our results suggest that incorporating safety considerations can significantly enhance learning performance in complex, multi-agent environments.

Mimicking the real interaction trajectory in the inference of the world model has been shown to improve the sample efficiency of model-based reinforcement learning (MBRL) algorithms. Many methods directly use known state sequences for reasoning. However, this approach fails to enhance the quality of reasoning by capturing the subtle variation between states. Much like how humans infer trends in event development from this variation, in this work, we introduce Global-Local variation Awareness Mamba-based world model (GLAM) that improves reasoning quality by perceiving and predicting variation between states. GLAM comprises two Mambabased parallel reasoning modules, GMamba and LMamba, which focus on perceiving variation from global and local perspectives, respectively, during the reasoning process. GMamba focuses on identifying patterns of variation between states in the input sequence and leverages these patterns to enhance the prediction of future state variation. LMamba emphasizes reasoning about unknown information, such as rewards, termination signals, and visual representations, by perceiving variation in adjacent states. By integrating the strengths of the two modules, GLAM accounts for highervalue variation in environmental changes, providing the agent with more efficient imagination-based training. We demonstrate that our method outperforms existing methods in normalized human scores on the Atari 100k benchmark.

Multi-UAV air combat is a complex task involving multiple autonomous UAVs, an evolving field in both aerospace and artificial intelligence. This paper aims to enhance adversarial performance through collaborative strategies. Previous approaches predominantly discretize the action space into predefined actions, limiting UAV maneuverability and complex strategy implementation. Others simplify the problem to 1v1 combat, neglecting the cooperative dynamics among multiple UAVs. To address the high-dimensional challenges inherent in six-degree-of-freedom space and improve cooperation, we propose a hierarchical framework utilizing the Leader-Follower Multi-Agent Proximal Policy Optimization (LFMAPPO) strategy. Specifically, the framework is structured into three levels. The top level conducts a macro-level assessment of the environment and guides execution policy. The middle level determines the angle of the desired action. The bottom level generates precise action commands for the high-dimensional action space. Moreover, we optimize the state-value functions by assigning distinct roles with the leader-follower strategy to train the top-level policy, followers estimate the leader's utility, promoting effective cooperation among agents. Additionally, the incorporation of a target selector, aligned with the UAVs' posture, assesses the threat level of targets. Finally, simulation experiments validate the effectiveness of our proposed method.

Multimodal sentiment analysis (MSA) identifies individuals' sentiment states in videos by integrating visual, audio, and text modalities. Despite progress in existing methods, the inherent modality heterogeneity limits the effective capture of interactive sentiment features across modalities. In this paper, by introducing a Multi-Modality Collaborative Learning (MMCL) framework, we facilitate cross-modal interactions and capture enhanced and complementary features from modality-common and modality-specific representations, respectively. Specifically, we design a parameter-free decoupling module and separate uni-modality into modality-common and modality-specific components through semantics assessment of cross-modal elements. For modality-specific representations, inspired by the act-reward mechanism in reinforcement learning, we design policy models to adaptively mine complementary sentiment features under the guidance of a joint reward. For modality-common representations, intra-modal attention is employed to highlight crucial components, playing enhanced roles among modalities. Experimental results, including superiority evaluations on four databases, effectiveness verification of each module, and assessment of complementary features, demonstrate that MMCL successfully learns collaborative features across modalities and significantly improves performance. The code can be available at https://github.com/smwanghhh/MMCL.