Kimi K2: Open Agentic Intelligence

We introduce Kimi K2, a Mixture-of-Experts (MoE) large language model with 32 billion activated parameters and 1 trillion total parameters. We propose the MuonClip optimizer, which improves upon Muon with a novel QK-clip technique to address training instability while enjoying the advanced token efficiency of Muon. Based on MuonClip, K2 was pre-trained on 15.5 trillion tokens with zero loss spike. During post-training, K2 undergoes a multi-stage post-training process, highlighted by a large-scale agentic data synthesis pipeline and a joint reinforcement learning (RL) stage, where the model improves its capabilities through interactions with real and synthetic environments. Kimi K2 achieves state-of-the-art performance among open-source non-thinking models, with strengths in agentic capabilities. Notably, K2 obtains 66.1 on Tau2-Bench, 76.5 on ACEBench (En), 65.8 on SWE-Bench Verified, and 47.3 on SWE-Bench Multilingual -- surpassing most open and closed-sourced baselines in non-thinking settings. It also exhibits strong capabilities in coding, mathematics, and reasoning tasks, with a score of 53.7 on LiveCodeBench v6, 49.5 on AIME 2025, 75.1 on GPQA-Diamond, and 27.1 on OJBench, all without extended thinking. These results position Kimi K2 as one of the most capable open-source large language models to date, particularly in software engineering and agentic tasks. We release our base and post-trained model checkpoints to facilitate future research and applications of agentic intelligence.

Test-Time Scaling with Reflective Generative Model

We introduce our first reflective generative model MetaStone-S1, which obtains OpenAI o3's performance via the self-supervised process reward model (SPRM). Through sharing the backbone network and using task-specific heads for next token prediction and process scoring respectively, SPRM successfully integrates the policy model and process reward model(PRM) into a unified interface without extra process annotation, reducing over 99% PRM parameters for efficient reasoning. Equipped with SPRM, MetaStone-S1 is naturally suitable for test time scaling (TTS), and we provide three reasoning effort modes (low, medium, and high), based on the controllable thinking length. Moreover, we empirically establish a scaling law that reveals the relationship between total thinking computation and TTS performance. Experiments demonstrate that our MetaStone-S1 achieves comparable performance to OpenAI-o3-mini's series with only 32B parameter size. To support the research community, we have open-sourced MetaStone-S1 at https://github.com/MetaStone-AI/MetaStone-S1.

Probing Association Biases in LLM Moderation Over-Sensitivity

Large Language Models are widely used for content moderation but often misclassify benign comments as toxic, leading to over-sensitivity. While previous research attributes this issue primarily to the presence of offensive terms, we reveal a potential cause beyond token level: LLMs exhibit systematic topic biases in their implicit associations. Inspired by cognitive psychology's implicit association tests, we introduce Topic Association Analysis, a semantic-level approach to quantify how LLMs associate certain topics with toxicity. By prompting LLMs to generate free-form scenario imagination for misclassified benign comments and analyzing their topic amplification levels, we find that more advanced models (e.g., GPT-4 Turbo) demonstrate stronger topic stereotype despite lower overall false positive rates. These biases suggest that LLMs do not merely react to explicit, offensive language but rely on learned topic associations, shaping their moderation decisions. Our findings highlight the need for refinement beyond keyword-based filtering, providing insights into the underlying mechanisms driving LLM over-sensitivity.

R3-RAG: Learning Step-by-Step Reasoning and Retrieval for LLMs via Reinforcement Learning

Retrieval-Augmented Generation (RAG) integrates external knowledge with Large Language Models (LLMs) to enhance factual correctness and mitigate hallucination. However, dense retrievers often become the bottleneck of RAG systems due to their limited parameters compared to LLMs and their inability to perform step-by-step reasoning. While prompt-based iterative RAG attempts to address these limitations, it is constrained by human-designed workflows. To address these limitations, we propose $\textbf{R3-RAG}$, which uses $\textbf{R}$einforcement learning to make the LLM learn how to $\textbf{R}$eason and $\textbf{R}$etrieve step by step, thus retrieving comprehensive external knowledge and leading to correct answers. R3-RAG is divided into two stages. We first use cold start to make the model learn the manner of iteratively interleaving reasoning and retrieval. Then we use reinforcement learning to further harness its ability to better explore the external retrieval environment. Specifically, we propose two rewards for R3-RAG: 1) answer correctness for outcome reward, which judges whether the trajectory leads to a correct answer; 2) relevance-based document verification for process reward, encouraging the model to retrieve documents that are relevant to the user question, through which we can let the model learn how to iteratively reason and retrieve relevant documents to get the correct answer. Experimental results show that R3-RAG significantly outperforms baselines and can transfer well to different retrievers. We release R3-RAG at https://github.com/Yuan-Li-FNLP/R3-RAG.

FamilyTool: A Multi-hop Personalized Tool Use Benchmark

The integration of tool learning with Large Language Models (LLMs) has expanded their capabilities in handling complex tasks by leveraging external tools. However, existing benchmarks for tool learning inadequately address critical real-world personalized scenarios, particularly those requiring multi-hop reasoning and inductive knowledge adaptation in dynamic environments. To bridge this gap, we introduce FamilyTool, a novel benchmark grounded in a family-based knowledge graph (KG) that simulates personalized, multi-hop tool use scenarios. FamilyTool challenges LLMs with queries spanning 1 to 3 relational hops (e.g., inferring familial connections and preferences) and incorporates an inductive KG setting where models must adapt to unseen user preferences and relationships without re-training, a common limitation in prior approaches that compromises generalization. We further propose KGETool: a simple KG-augmented evaluation pipeline to systematically assess LLMs' tool use ability in these settings. Experiments reveal significant performance gaps in state-of-the-art LLMs, with accuracy dropping sharply as hop complexity increases and inductive scenarios exposing severe generalization deficits. These findings underscore the limitations of current LLMs in handling personalized, evolving real-world contexts and highlight the urgent need for advancements in tool-learning frameworks. FamilyTool serves as a critical resource for evaluating and advancing LLM agents' reasoning, adaptability, and scalability in complex, dynamic environments. Code and dataset are available at Github.

Prior-Fitted Networks Scale to Larger Datasets When Treated as Weak Learners

Prior-Fitted Networks (PFNs) have recently been proposed to efficiently perform tabular classification tasks. Although they achieve good performance on small datasets, they encounter limitations with larger datasets. These limitations include significant memory consumption and increased computational complexity, primarily due to the impracticality of incorporating all training samples as inputs within these networks. To address these challenges, we investigate the fitting assumption for PFNs and input samples. Building on this understanding, we propose \textit{BoostPFN} designed to enhance the performance of these networks, especially for large-scale datasets. We also theoretically validate the convergence of BoostPFN and our empirical results demonstrate that the BoostPFN method can outperform standard PFNs with the same size of training samples in large datasets and achieve a significant acceleration in training times compared to other established baselines in the field, including widely-used Gradient Boosting Decision Trees (GBDTs), deep learning methods and AutoML systems. High performance is maintained for up to 50x of the pre-training size of PFNs, substantially extending the limit of training samples. Through this work, we address the challenges of efficiently handling large datasets via PFN-based models, paving the way for faster and more effective tabular data classification training and prediction process. Code is available at Github.

Category-Selective Neurons in Deep Networks: Comparing Purely Visual and Visual-Language Models

Category-selective regions in the human brain, such as the fusiform face area (FFA), extrastriate body area (EBA), parahippocampal place area (PPA), and visual word form area (VWFA), play a crucial role in high-level visual processing. Here, we investigate whether artificial neural networks (ANNs) exhibit similar category-selective neurons and how these neurons vary across model layers and between purely visual and vision-language models. Inspired by fMRI functional localizer experiments, we presented images from different categories (faces, bodies, scenes, words, scrambled scenes, and scrambled words) to deep networks and identified category-selective neurons using statistical criteria. Comparing ResNet and the structurally controlled ResNet-based CLIP model, we found that both models contain category-selective neurons, with their proportion increasing across layers, mirroring category selectivity in higher-level visual brain regions. However, CLIP exhibited a higher proportion but lower specificity of category-selective neurons compared to ResNet. Additionally, CLIP's category-selective neurons were more evenly distributed across feature maps and demonstrated greater representational consistency across layers. These findings suggest that language learning increases the number of category-selective neurons while reducing their selectivity strength, reshaping visual representations in deep networks. Our study provides insights into how ANNs mirror biological vision and how multimodal learning influences category-selective representations.

F3D-Gaus: Feed-forward 3D-aware Generation on ImageNet with Cycle-Consistent Gaussian Splatting

This paper tackles the problem of generalizable 3D-aware generation from monocular datasets, e.g., ImageNet. The key challenge of this task is learning a robust 3D-aware representation without multi-view or dynamic data, while ensuring consistent texture and geometry across different viewpoints. Although some baseline methods are capable of 3D-aware generation, the quality of the generated images still lags behind state-of-the-art 2D generation approaches, which excel in producing high-quality, detailed images. To address this severe limitation, we propose a novel feed-forward pipeline based on pixel-aligned Gaussian Splatting, coined as F3D-Gaus, which can produce more realistic and reliable 3D renderings from monocular inputs. In addition, we introduce a self-supervised cycle-consistent constraint to enforce cross-view consistency in the learned 3D representation. This training strategy naturally allows aggregation of multiple aligned Gaussian primitives and significantly alleviates the interpolation limitations inherent in single-view pixel-aligned Gaussian Splatting. Furthermore, we incorporate video model priors to perform geometry-aware refinement, enhancing the generation of fine details in wide-viewpoint scenarios and improving the model's capability to capture intricate 3D textures. Extensive experiments demonstrate that our approach not only achieves high-quality, multi-view consistent 3D-aware generation from monocular datasets, but also significantly improves training and inference efficiency.

Constructing Confidence Intervals for Average Treatment Effects from Multiple Datasets

Constructing confidence intervals (CIs) for the average treatment effect (ATE) from patient records is crucial to assess the effectiveness and safety of drugs. However, patient records typically come from different hospitals, thus raising the question of how multiple observational datasets can be effectively combined for this purpose. In our paper, we propose a new method that estimates the ATE from multiple observational datasets and provides valid CIs. Our method makes little assumptions about the observational datasets and is thus widely applicable in medical practice. The key idea of our method is that we leverage prediction-powered inferences and thereby essentially `shrink' the CIs so that we offer more precise uncertainty quantification as compared to na\"ive approaches. We further prove the unbiasedness of our method and the validity of our CIs. We confirm our theoretical results through various numerical experiments. Finally, we provide an extension of our method for constructing CIs from combinations of experimental and observational datasets.

A Graph-Based Synthetic Data Pipeline for Scaling High-Quality Reasoning Instructions

Synthesizing high-quality reasoning data for continual training has been proven to be effective in enhancing the performance of Large Language Models (LLMs). However, previous synthetic approaches struggle to easily scale up data and incur high costs in the pursuit of high quality. In this paper, we propose the Graph-based Synthetic Data Pipeline (GSDP), an economical and scalable framework for high-quality reasoning data synthesis. Inspired by knowledge graphs, we extracted knowledge points from seed data and constructed a knowledge point relationships graph to explore their interconnections. By exploring the implicit relationships among knowledge, our method achieves $\times$255 data expansion. Furthermore, GSDP led by open-source models, achieves synthesis quality comparable to GPT-4-0613 while maintaining $\times$100 lower costs. To tackle the most challenging mathematical reasoning task, we present the GSDP-MATH dataset comprising over 1.91 million pairs of math problems and answers. After fine-tuning on GSDP-MATH, GSDP-7B based on Mistral-7B achieves 37.7% accuracy on MATH and 78.4% on GSM8K, demonstrating the effectiveness of our method. The dataset and models trained in this paper will be available.