MiniCPM4: Ultra-Efficient LLMs on End Devices

This paper introduces MiniCPM4, a highly efficient large language model (LLM) designed explicitly for end-side devices. We achieve this efficiency through systematic innovation in four key dimensions: model architecture, training data, training algorithms, and inference systems. Specifically, in terms of model architecture, we propose InfLLM v2, a trainable sparse attention mechanism that accelerates both prefilling and decoding phases for long-context processing. Regarding training data, we propose UltraClean, an efficient and accurate pre-training data filtering and generation strategy, and UltraChat v2, a comprehensive supervised fine-tuning dataset. These datasets enable satisfactory model performance to be achieved using just 8 trillion training tokens. Regarding training algorithms, we propose ModelTunnel v2 for efficient pre-training strategy search, and improve existing post-training methods by introducing chunk-wise rollout for load-balanced reinforcement learning and data-efficient tenary LLM, BitCPM. Regarding inference systems, we propose CPM.cu that integrates sparse attention, model quantization, and speculative sampling to achieve efficient prefilling and decoding. To meet diverse on-device requirements, MiniCPM4 is available in two versions, with 0.5B and 8B parameters, respectively. Sufficient evaluation results show that MiniCPM4 outperforms open-source models of similar size across multiple benchmarks, highlighting both its efficiency and effectiveness. Notably, MiniCPM4-8B demonstrates significant speed improvements over Qwen3-8B when processing long sequences. Through further adaptation, MiniCPM4 successfully powers diverse applications, including trustworthy survey generation and tool use with model context protocol, clearly showcasing its broad usability.

Real-Time Execution of Action Chunking Flow Policies

Modern AI systems, especially those interacting with the physical world, increasingly require real-time performance. However, the high latency of state-of-the-art generalist models, including recent vision-language action models (VLAs), poses a significant challenge. While action chunking has enabled temporal consistency in high-frequency control tasks, it does not fully address the latency problem, leading to pauses or out-of-distribution jerky movements at chunk boundaries. This paper presents a novel inference-time algorithm that enables smooth asynchronous execution of action chunking policies. Our method, real-time chunking (RTC), is applicable to any diffusion- or flow-based VLA out of the box with no re-training. It generates the next action chunk while executing the current one, "freezing" actions guaranteed to execute and "inpainting" the rest. To test RTC, we introduce a new benchmark of 12 highly dynamic tasks in the Kinetix simulator, as well as evaluate 6 challenging real-world bimanual manipulation tasks. Results demonstrate that RTC is fast, performant, and uniquely robust to inference delay, significantly improving task throughput and enabling high success rates in precise tasks $\unicode{x2013}$ such as lighting a match $\unicode{x2013}$ even in the presence of significant latency. See https://pi.website/research/real_time_chunking for videos.

MoQAE: Mixed-Precision Quantization for Long-Context LLM Inference via Mixture of Quantization-Aware Experts

One of the primary challenges in optimizing large language models (LLMs) for long-context inference lies in the high memory consumption of the Key-Value (KV) cache. Existing approaches, such as quantization, have demonstrated promising results in reducing memory usage. However, current quantization methods cannot take both effectiveness and efficiency into account. In this paper, we propose MoQAE, a novel mixed-precision quantization method via mixture of quantization-aware experts. First, we view different quantization bit-width configurations as experts and use the traditional mixture of experts (MoE) method to select the optimal configuration. To avoid the inefficiency caused by inputting tokens one by one into the router in the traditional MoE method, we input the tokens into the router chunk by chunk. Second, we design a lightweight router-only fine-tuning process to train MoQAE with a comprehensive loss to learn the trade-off between model accuracy and memory usage. Finally, we introduce a routing freezing (RF) and a routing sharing (RS) mechanism to further reduce the inference overhead. Extensive experiments on multiple benchmark datasets demonstrate that our method outperforms state-of-the-art KV cache quantization approaches in both efficiency and effectiveness.

SceneRAG: Scene-level Retrieval-Augmented Generation for Video Understanding

Despite recent advances in retrieval-augmented generation (RAG) for video understanding, effectively understanding long-form video content remains underexplored due to the vast scale and high complexity of video data. Current RAG approaches typically segment videos into fixed-length chunks, which often disrupts the continuity of contextual information and fails to capture authentic scene boundaries. Inspired by the human ability to naturally organize continuous experiences into coherent scenes, we present SceneRAG, a unified framework that leverages large language models to segment videos into narrative-consistent scenes by processing ASR transcripts alongside temporal metadata. SceneRAG further sharpens these initial boundaries through lightweight heuristics and iterative correction. For each scene, the framework fuses information from both visual and textual modalities to extract entity relations and dynamically builds a knowledge graph, enabling robust multi-hop retrieval and generation that account for long-range dependencies. Experiments on the LongerVideos benchmark, featuring over 134 hours of diverse content, confirm that SceneRAG substantially outperforms prior baselines, achieving a win rate of up to 72.5 percent on generation tasks.

BRIGHT+: Upgrading the BRIGHT Benchmark with MARCUS, a Multi-Agent RAG Clean-Up Suite

Retrieval-Augmented Generation (RAG) systems require corpora that are both structurally clean and semantically coherent. BRIGHT is a recent and influential benchmark designed to evaluate complex multi-hop retrieval across diverse, high-reasoning domains. However, its practical effectiveness is limited by common web-crawled artifacts - such as content redundancy and semantic discontinuity - that impair retrieval accuracy and downstream reasoning. Notably, we find that such issues are concentrated in seven StackExchange-derived subdomains, while other domains (e.g., Coding and Theorem-based content) remain relatively clean. In this study, we present MARCUS, a multi-agent pipeline that leverages large language models (LLMs) to systematically clean and re-chunk BRIGHT into a higher-quality corpus: BRIGHT-Plus. MARCUS applies dedicated agents for structural noise removal and semantic segmentation, preserving answer-bearing spans while improving contextual integrity. Experimental evaluations demonstrate that BRIGHT-Plus yields consistent and significant improvements in both retrieval accuracy and multi-hop reasoning across a diverse set of retrievers. We release both the BRIGHT-Plus corpus and the MARCUS pipeline to support future research on robust, reasoning-centric retrieval.

3DFlowAction: Learning Cross-Embodiment Manipulation from 3D Flow World Model

Manipulation has long been a challenging task for robots, while humans can effortlessly perform complex interactions with objects, such as hanging a cup on the mug rack. A key reason is the lack of a large and uniform dataset for teaching robots manipulation skills. Current robot datasets often record robot action in different action spaces within a simple scene. This hinders the robot to learn a unified and robust action representation for different robots within diverse scenes. Observing how humans understand a manipulation task, we find that understanding how the objects should move in the 3D space is a critical clue for guiding actions. This clue is embodiment-agnostic and suitable for both humans and different robots. Motivated by this, we aim to learn a 3D flow world model from both human and robot manipulation data. This model predicts the future movement of the interacting objects in 3D space, guiding action planning for manipulation. Specifically, we synthesize a large-scale 3D optical flow dataset, named ManiFlow-110k, through a moving object auto-detect pipeline. A video diffusion-based world model then learns manipulation physics from these data, generating 3D optical flow trajectories conditioned on language instructions. With the generated 3D object optical flow, we propose a flow-guided rendering mechanism, which renders the predicted final state and leverages GPT-4o to assess whether the predicted flow aligns with the task description. This equips the robot with a closed-loop planning ability. Finally, we consider the predicted 3D optical flow as constraints for an optimization policy to determine a chunk of robot actions for manipulation. Extensive experiments demonstrate strong generalization across diverse robotic manipulation tasks and reliable cross-embodiment adaptation without hardware-specific training.

MIRIAD: Augmenting LLMs with millions of medical query-response pairs

LLMs are bound to transform healthcare with advanced decision support and flexible chat assistants. However, LLMs are prone to generate inaccurate medical content. To ground LLMs in high-quality medical knowledge, LLMs have been equipped with external knowledge via RAG, where unstructured medical knowledge is split into small text chunks that can be selectively retrieved and integrated into the LLMs context. Yet, existing RAG pipelines rely on raw, unstructured medical text, which can be noisy, uncurated and difficult for LLMs to effectively leverage. Systematic approaches to organize medical knowledge to best surface it to LLMs are generally lacking. To address these challenges, we introduce MIRIAD, a large-scale, curated corpus of 5,821,948 medical QA pairs, each rephrased from and grounded in a passage from peer-reviewed medical literature using a semi-automated pipeline combining LLM generation, filtering, grounding, and human annotation. Unlike prior medical corpora, which rely on unstructured text, MIRIAD encapsulates web-scale medical knowledge in an operationalized query-response format, which enables more targeted retrieval. Experiments on challenging medical QA benchmarks show that augmenting LLMs with MIRIAD improves accuracy up to 6.7% compared to unstructured RAG baselines with the same source corpus and with the same amount of retrieved text. Moreover, MIRIAD improved the ability of LLMs to detect medical hallucinations by 22.5 to 37% (increase in F1 score). We further introduce MIRIAD-Atlas, an interactive map of MIRIAD spanning 56 medical disciplines, enabling clinical users to visually explore, search, and refine medical knowledge. MIRIAD promises to unlock a wealth of down-stream applications, including medical information retrievers, enhanced RAG applications, and knowledge-grounded chat interfaces, which ultimately enables more reliable LLM applications in healthcare.

LiPo: A Lightweight Post-optimization Framework for Smoothing Action Chunks Generated by Learned Policies

Recent advances in imitation learning have enabled robots to perform increasingly complex manipulation tasks in unstructured environments. However, most learned policies rely on discrete action chunking, which introduces discontinuities at chunk boundaries. These discontinuities degrade motion quality and are particularly problematic in dynamic tasks such as throwing or lifting heavy objects, where smooth trajectories are critical for momentum transfer and system stability. In this work, we present a lightweight post-optimization framework for smoothing chunked action sequences. Our method combines three key components: (1) inference-aware chunk scheduling to proactively generate overlapping chunks and avoid pauses from inference delays; (2) linear blending in the overlap region to reduce abrupt transitions; and (3) jerk-minimizing trajectory optimization constrained within a bounded perturbation space. The proposed method was validated on a position-controlled robotic arm performing dynamic manipulation tasks. Experimental results demonstrate that our approach significantly reduces vibration and motion jitter, leading to smoother execution and improved mechanical robustness.

Diffusion Transformer-based Universal Dose Denoising for Pencil Beam Scanning Proton Therapy

Purpose: Intensity-modulated proton therapy (IMPT) offers precise tumor coverage while sparing organs at risk (OARs) in head and neck (H&N) cancer. However, its sensitivity to anatomical changes requires frequent adaptation through online adaptive radiation therapy (oART), which depends on fast, accurate dose calculation via Monte Carlo (MC) simulations. Reducing particle count accelerates MC but degrades accuracy. To address this, denoising low-statistics MC dose maps is proposed to enable fast, high-quality dose generation. Methods: We developed a diffusion transformer-based denoising framework. IMPT plans and 3D CT images from 80 H&N patients were used to generate noisy and high-statistics dose maps using MCsquare (1 min and 10 min per plan, respectively). Data were standardized into uniform chunks with zero-padding, normalized, and transformed into quasi-Gaussian distributions. Testing was done on 10 H&N, 10 lung, 10 breast, and 10 prostate cancer cases, preprocessed identically. The model was trained with noisy dose maps and CT images as input and high-statistics dose maps as ground truth, using a combined loss of mean square error (MSE), residual loss, and regional MAE (focusing on top/bottom 10% dose voxels). Performance was assessed via MAE, 3D Gamma passing rate, and DVH indices. Results: The model achieved MAEs of 0.195 (H&N), 0.120 (lung), 0.172 (breast), and 0.376 Gy[RBE] (prostate). 3D Gamma passing rates exceeded 92% (3%/2mm) across all sites. DVH indices for clinical target volumes (CTVs) and OARs closely matched the ground truth. Conclusion: A diffusion transformer-based denoising framework was developed and, though trained only on H&N data, generalizes well across multiple disease sites.

Dynamic Context-Aware Streaming Pretrained Language Model For Inverse Text Normalization

Inverse Text Normalization (ITN) is crucial for converting spoken Automatic Speech Recognition (ASR) outputs into well-formatted written text, enhancing both readability and usability. Despite its importance, the integration of streaming ITN within streaming ASR remains largely unexplored due to challenges in accuracy, efficiency, and adaptability, particularly in low-resource and limited-context scenarios. In this paper, we introduce a streaming pretrained language model for ITN, leveraging pretrained linguistic representations for improved robustness. To address streaming constraints, we propose Dynamic Context-Aware during training and inference, enabling adaptive chunk size adjustments and the integration of right-context information. Experimental results demonstrate that our method achieves accuracy comparable to non-streaming ITN and surpasses existing streaming ITN models on a Vietnamese dataset, all while maintaining low latency, ensuring seamless integration into ASR systems.