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%).

With the growing success of multi-modal learning, research on the robustness of multi-modal models, especially when facing situations with missing modalities, is receiving increased attention. Nevertheless, previous studies in this domain exhibit certain limitations, as they often lack theoretical insights or their methodologies are tied to specific network architectures or modalities. We model the scenarios of multi-modal models encountering missing modalities from an information-theoretic perspective and illustrate that the performance ceiling in such scenarios can be approached by efficiently utilizing the information inherent in non-missing modalities. In practice, there are two key aspects: (1) The encoder should be able to extract sufficiently good features from the non-missing modality; (2) The extracted features should be robust enough not to be influenced by noise during the fusion process across modalities. To this end, we introduce Uni-Modal Ensemble with Missing Modality Adaptation (UME-MMA). UME-MMA employs uni-modal pre-trained weights for the multi-modal model to enhance feature extraction and utilizes missing modality data augmentation techniques to better adapt to situations with missing modalities. Apart from that, UME-MMA, built on a late-fusion learning framework, allows for the plug-and-play use of various encoders, making it suitable for a wide range of modalities and enabling seamless integration of large-scale pre-trained encoders to further enhance performance. And we demonstrate UME-MMA's effectiveness in audio-visual datasets~(e.g., AV-MNIST, Kinetics-Sound, AVE) and vision-language datasets~(e.g., MM-IMDB, UPMC Food101).

Large language models (LLMs) with memory are computationally universal. However, mainstream LLMs are not taking full advantage of memory, and the designs are heavily influenced by biological brains. Due to their approximate nature and proneness to the accumulation of errors, conventional neural memory mechanisms cannot support LLMs to simulate complex reasoning. In this paper, we seek inspiration from modern computer architectures to augment LLMs with symbolic memory for complex multi-hop reasoning. Such a symbolic memory framework is instantiated as an LLM and a set of SQL databases, where the LLM generates SQL instructions to manipulate the SQL databases. We validate the effectiveness of the proposed memory framework on a synthetic dataset requiring complex reasoning. The project website is available at https://chatdatabase.github.io/ .

In vision-based reinforcement learning (RL) tasks, it is prevalent to assign the auxiliary task with a surrogate self-supervised loss so as to obtain more semantic representations and improve sample efficiency. However, abundant information in self-supervised auxiliary tasks has been disregarded, since the representation learning part and the decision-making part are separated. To sufficiently utilize information in the auxiliary task, we present a simple yet effective idea to employ self-supervised loss as an intrinsic reward, called Intrinsically Motivated Self-Supervised learning in Reinforcement learning (IM-SSR). We formally show that the self-supervised loss can be decomposed as exploration for novel states and robustness improvement from nuisance elimination. IM-SSR can be effortlessly plugged into any reinforcement learning with self-supervised auxiliary objectives with nearly no additional cost. Combined with IM-SSR, the previous underlying algorithms achieve salient improvements on both sample efficiency and generalization in various vision-based robotics tasks from the DeepMind Control Suite, especially when the reward signal is sparse.

The world provides us with data of multiple modalities. Intuitively, models fusingdata from different modalities outperform unimodal models, since more informationis aggregated. Recently, joining the success of deep learning, there is an influentialline of work on deep multimodal learning, which has remarkable empirical resultson various applications. However, theoretical justifications in this field are notablylacking.Can multimodal provably perform better than unimodal? In this paper, we answer this question under a most popular multimodal learningframework, which firstly encodes features from different modalities into a commonlatent space and seamlessly maps the latent representations into the task space. Weprove that learning with multiple modalities achieves a smaller population risk thanonly using its subset of modalities. The main intuition is that the former has moreaccurate estimate of the latent space representation. To the best of our knowledge,this is the first theoretical treatment to capture important qualitative phenomenaobserved in real multimodal applications. Combining with experiment results, weshow that multimodal learning does possess an appealing formal guarantee.

We consider a variant of this problem, where instead of requiring the data to be completely private so that no one gets In the classical multi-party computation setting, any information about it, we only require data to be partially multiple parties jointly compute a function without private. That is, no one can efficiently recover the original revealing their own input data. We consider a data, but users can extract other useful information from the variant of this problem, where the input data can encrypted data. Although being different, our requirement be shared for machine learning training purposes, has the flavor of differential privacy (Dwork et al., 2006), but the data are also encrypted so that they cannot e.g., users can obtain the average salary of all employees, be recovered by other parties. We present a but cannot figure out the salary of each individual.