PaperlayerView exact revision on arXiv

Source-linked AI summary

MiniCPM4: Ultra-Efficient LLMs on End Devices

MiniCPM Team, Chaojun Xiao, Yuxuan Li, Xu Han, Yuzhuo Bai, Jie Cai, Haotian Chen, Wentong Chen, Xin Cong, Ganqu Cui, Ning Ding, Shengda Fan, Yewei Fang, Zixuan Fu, Wenyu Guan, Yitong Guan, Junshao Guo, Yufeng Han, Bingxiang He, Yuxiang Huang, Baoxi Ji, Cunliang Kong, Qiuzuo Li, Siyuan Li, Wenhao Li, Xin Li, Yanghao Li, Yishan Li, Zhen Li, Dan Liu, Biyuan Lin, Yankai Lin, Xiang Long, Quanyu Lu, Yaxi Lu, Peiyan Luo, Hongya Lyu, Litu Ou, Yinxu Pan, Lushi Pu, Zekai Qu, Qundong Shi, Zijun Song, Jiayuan Su, Zhou Su, Ao Sun, Xianghui Sun, Peijun Tang, Fangzheng Wang, Feng Wang, Shuo Wang, Yudong Wang, Zheng Wang, Yesai Wu, Zhenyu Xiao, Jie Xie, Zihao Xie, Xiaoyue Xu, Yukun Yan, Jiarui Yuan, Jinqian Zhang, Kaihuo Zhang, Lei Zhang, Linyue Zhang, Xueren Zhang, Yudi Zhang, Hengyu Zhao, Weilin Zhao, Weilun Zhao, Yuanqian Zhao, Zhi Zheng, Chuyue Zhou, Ge Zhou, Jie Zhou, Wei Zhou, Yanghao Zhou, Zihan Zhou, Zixuan Zhou, Zhiyuan Liu, Guoyang Zeng, Chao Jia, Dahai Li, Maosong Sun

arXiv:2506.07900v2cs.CLcs.AI

TL;DR

MiniCPM4 addresses the high computational cost that has pushed expanding LLMs toward cloud deployment by systematically improving architecture, data, training, and inference for end-side devices. Using these innovations, it reaches comparable performance with existing open-source models using 8 trillion tokens and achieves 5× speedup for long-sequence processing, while remaining subject to open data-construction challenges.

  • Problem

    Expanding LLMs require exponentially more computational resources, leaving them primarily deployed on cloud servers rather than end-side devices.

  • Method

    MiniCPM4 combines innovations in model architecture, training data, training algorithms, and inference systems, including sparse attention, data filtering and generation, supervised fine-tuning data, and efficient inference.

  • Results

    5× speedup for long-sequence processing is achieved while using only 8 trillion tokens to reach comparable performance with existing open-source models.

  • Takeaways & Limitations

    MiniCPM4 provides released model parameters and inference code intended to facilitate open-source development of efficient LLMs.

  • Takeaways & Limitations

    The data pipelines still partially rely on human-designed heuristics, and balancing corpus diversity with task relevance remains challenging.

Abstract

from arXiv · show

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. Furthermore, we construct a hybrid reasoning model, MiniCPM4.1, which can be used in both deep reasoning mode and non-reasoning mode. Evaluation results demonstrate that MiniCPM4 and MiniCPM4.1 outperform similar-sized open-source models across benchmarks, with the 8B variants showing significant speed improvements on long sequence understanding and generation.

1 Introduction

MiniCPM4 targets the resource constraints that push large language models toward cloud deployment by combining innovations in architecture, data, training, and inference. The resulting models are offered in multiple sizes and reasoning modes, with long-context and end-side efficiency as central goals.

  • Motivation: MiniCPM4 addresses the computational and memory demands that have limited large language models primarily to cloud-server deployment.The paper frames end-side deployment as requiring greater efficiency because model scaling increases computational-resource requirements.
  • Systematic efficiency design: MiniCPM4 combines innovations in model architecture, training data, training algorithms, and inference systems.The architecture contribution is InfLLM v2; the data contributions include UltraClean and UltraChat v2; the training and inference contributions include ModelTunnel v2, chunk-wise rollout, BitCPM4, and CPM.cu.
  • Model architecture: InfLLM v2 uses trainable token-level sparse attention with specialized kernels to accelerate both long-context prefilling and decoding.Its specialized training further increases attention sparsity and improves long-context processing capabilities.
  • Training data: UltraClean filters knowledge-intensive data efficiently, while UltraChat v2 generates reasoning-intensive multi-turn dialogues for training and evaluation.UltraClean uses a nearly-trained language model for verification and a classifier trained from selected seed data; UltraChat v2 combines expert models, prompt engineering, and dual-stage filtering.
  • Training algorithms: ModelTunnel v2 improves training-strategy search, chunk-wise rollout reduces idle reinforcement-learning computation, and BitCPM4 lowers ternary-model training costs.The training pipeline also uses multi-token prediction and FP-8 mixed precision to strengthen supervision and exploit GPU computation.
  • Models and training: MiniCPM4 includes 0.5B and 8B variants, extends context from 4K to 128K tokens, and MiniCPM4.1 supports both reasoning and non-reasoning modes.The 8B model is trained on 8.3T high-quality tokens, including 1.3T annealing-phase tokens, before long-context pre-training and supervised fine-tuning.

2 Efficient Architecture and Pre-training

The paper presents efficient sparse attention for long-context processing. InfLLM v2 selects key-value blocks per query group while always retaining initial and local sliding-window tokens.

  • Efficient Architecture and Pre-training: InfLLM v2 enables comparable long-context processing ability to full attention with 81% attention sparsity.The approach is presented as part of an efficient sparse-attention architecture for MiniCPM4.
  • Efficient Architecture and Pre-training: Each query group selects parts of key-value blocks for attention computation, while initial tokens and local sliding-window tokens are always selected.This describes the selection pattern illustrated for InfLLM v2.

2.1 InfLLM v2: Trainable Sparse Attention for Prefilling and Decoding

InfLLM v2 is a trainable sparse attention mechanism that reduces attention computation and memory access for both prefilling and decoding. It selects relevant key-value blocks using semantic relevance scores and efficient Top-K implementation.

  • Design motivation: InfLLM v2 reduces computation and memory access costs for both prefilling and decoding through trainable sparse attention.Unlike several prior sparse-attention designs, it is designed to accelerate both phases without adding attention-output parameters.
  • Efficiency and scope: InfLLM v2 can reduce 60% computational costs compared with NSA through its efficient Top-K context block selection method.When sequences fit within the selected blocks, the mechanism degrades to vanilla dense attention.
  • Overall framework: Each query token attends only to selected key-value blocks rather than the full preceding context.The key-value cache is partitioned into equal-sized blocks, and the highest-relevance blocks are selected before attention is computed over their tokens.
  • Dynamic contextual block selection: InfLLM v2 computes block relevance with fine-grained semantic kernels, avoiding token-level relevance calculations and reducing memory access costs.Query heads within the same group share top-k blocks, further minimizing memory access.
  • Efficient Top-K implementation: Top-K selection is the bottleneck of InfLLM v2 for long-context processing.The proposed LSE approximation uses coarse-grained semantic kernels to reduce the computational and memory access costs of relevance estimation.

2.2 UltraClean: High-Quality Pre-Training Data Filtering and Generation

UltraClean combines efficient verification, classifier-based filtering, and reasoning-intensive data generation to improve training-data quality without increasing training tokens. The resulting data supports comparable or stronger downstream performance, especially on knowledge-intensive and reasoning-intensive tasks, while remaining partly dependent on human-designed heuristics.

  • High-Quality Knowledge-Intensive Data Filtering: High-quality seed selection and balanced positive-negative classifier training enable filtering of web-scale corpora into UltraFineWeb.The pipeline applies verification to seed samples and classifier outputs before large-scale filtering.
  • High-Quality Knowledge-Intensive Data Filtering: UltraClean evaluates candidate data by measuring performance gains from adding corpora during the final training steps of a nearly trained LLM.This replaces costly from-scratch verification with a more efficient quality signal.
  • High-Quality Knowledge-Intensive Data Filtering: FastText filtering reduces large-scale inference cost compared with LLM-based classifiers while retaining competitive performance under certain conditions.Processing 15T tokens takes under 1,000 CPU hours with fastText versus approximately 6,000 GPU hours for an LLM-based classifier.
  • High-Quality Reasoning-Intensive Data Generation: The reasoning-data pipeline uses high-quality seeds, structured curation, and iterative generation to increase knowledge density, logical coherence, and reasoning transferability.It targets general benchmark performance and foundational reasoning generalization.
  • Discussion for Future Training Data: Without increasing, and sometimes reducing, training tokens, optimized data achieves performance comparable to or surpassing full-scale corpora across downstream tasks.Benefits are especially reported for knowledge-intensive and reasoning-intensive tasks, supporting data quality over quantity.
  • Discussion for Future Training Data: The pipelines still rely partly on human-designed heuristics, while balancing corpus diversity and task relevance remains challenging.Future work also targets multilingual, cross-task, and multimodal extensions.

2.3 ModelTunnel v2: Efficient Pre-Training Strategy Search

ModelTunnel v2 reduces the cost of pre-training strategy search by transferring experiments from small models and replacing language-model loss with downstream-relevant ScalingBench signals. The section also describes multi-token prediction and FP8 mixed-precision training as efficiency improvements.

  • Efficient Predictable Scaling: ModelTunnel transfers training-strategy experiments from small models to larger models, addressing the prohibitive cost of full-scale configuration search.A single large-model training experiment can require hundreds of thousands of GPU hours.
  • Efficient Predictable Scaling with Improved Performance Indicator: ScalingBench links its loss to downstream performance through a sigmoid relationship, providing a more suitable indicator than open-source pre-training loss.The relationship is observed across models ranging from 0.36B to 4B parameters with varied data sources and scales.
  • Comparison between µP and Vanilla Architecture: Under the reported conditions, µP and StepLaw show minimal differences in loss and ScalingBench scores, with neither method consistently superior.The comparison is affected by practical hardware constraints, learning-rate scheduling, and data-allocation choices.
  • Pre-Training Engineering: Multi-token prediction adds denser supervision through additional prediction heads, and the final objective combines next-token and multi-token losses.The combined objective is L = LNTP + λLMTP.
  • Pre-Training Engineering: FP8 mixed-precision training applies FP8 computation to selected linear projections while using FP32 accumulation to improve training efficiency.The implementation uses online block-wise quantization and limits FP8 usage to reduce instability.

3 Efficient Post-Training

MiniCPM4’s post-training combines broad supervised fine-tuning data with long chain-of-thought supervision and reinforcement learning. Chunk-wise rollout addresses load imbalance during reinforcement-learning computation.

  • Efficient Post-Training: UltraChat v2 provides diverse supervised fine-tuning instructions intended to activate capabilities acquired during pre-training.The dataset is described as large-scale and comprehensive.
  • Efficient Post-Training: Long chain-of-thought supervision and reinforcement learning are used to enhance deep reasoning ability.These methods extend post-training beyond foundational instruction following.
  • Efficient Post-Training: Chunk-wise rollout is introduced as a load-balanced reinforcement-learning strategy to address inefficient computation from unbalanced rollout workloads.The method targets the rollout process specifically.

3.1 UltraChat v2: Foundational Capability Enhanced SFT Data Generation

UltraChat v2 uses task-oriented synthetic data generation to strengthen foundational capabilities across knowledge, reasoning, instruction following, long-context processing, and tool use. Its construction combines structured knowledge generation, evolving instructions and answers, long-context distractors, function calling, and code-interpreter tasks.

  • Foundational Capability Enhanced SFT Data Generation: The synthetic-data framework organizes post-training examples across five capability tracks: knowledge applications, reasoning, instruction following, long-context processing, and tool use.The tracks provide targeted and structured training signals.
  • Knowledge Applications: Knowledge-driven QA data is built from domain corpora, exam syllabi, and textbooks by extracting knowledge points and generating targeted questions.The process creates an initial structured dataset before later evolution steps.
  • Instruction Following: Instruction evolution and answer-diversity evolution vary prompts and valid responses to improve expressive diversity and robustness.Both strategies are applied to initial practice QA pairs.
  • Reasoning: Math and code reasoning datasets target multi-step inference, mathematical reasoning, procedural thinking, and real-world programming scenarios.The datasets are designed to support more transferable logical skills.
  • Instruction Construction: The pipeline increases instruction complexity progressively and retains outputs that satisfy explicitly verifiable constraints through rule-based filtering.These mechanisms support scalable data generation and controlled task difficulty.
  • Long-Context Processing: Long-context SFT data combines task-oriented queries with related distractor documents and covers context lengths uniformly from 8K to 64K tokens.The design trains models to locate relevant content and reason across mixed-relevance inputs.
  • Tool Use: Function-calling and code-interpreter data combine public and in-house examples, with filtering and sandboxed execution used to improve task relevance and validity.Unsolved code-interpreter examples after 10 attempts are discarded.

3.2 Chunk-wise Rollout: Deep Reasoning with Load-Balanced Reinforcement Learning

The paper introduces chunk-wise rollout to improve reinforcement-learning efficiency for deep reasoning on end-side models. It combines fixed-length trajectory chunks with filtering, reuse, and stabilization techniques to reduce rollout waste while preserving training stability and performance.

  • Training Recipe: Dynamic sampling filters prompts whose responses are uniformly correct or incorrect, ensuring each batch contributes effective gradients.The strategy is intended to reduce gradient variance while maintaining a consistent batch size.
  • Training Recipe: Token-level policy loss weights longer sequences proportionally, encouraging learning of complex reasoning patterns while suppressing verbosity and repetition.Overlong truncated responses are excluded from loss computation to avoid penalizing prematurely cut-off reasoning trajectories.
  • Chunk-wise Rollout: Chunk-wise rollout generates fixed-length trajectory chunks, trains completed trajectories, and reuses incomplete ones across iterations.This design targets throughput degradation from lengthy rollout trajectories by improving computational resource utilization.
  • Stabilized Chunk-wise Rollout: Chunk-level importance sampling corrects for policy changes across chunks, while dual-clip, KL regularization, dynamic reference updates, and garble filtering stabilize optimization.These techniques address distribution shifts, high-variance updates, and corrupted or incoherent reused trajectories.
  • Experimental Results: Chunk-wise rollout reduces training and sampling time per step while maintaining performance, with smaller chunks steadily lowering sampling time.The reported experiments evaluate vanilla rollout against chunk sizes using repeated training runs on DeepSeek-R1-Distill-Qwen-1.5B.

3.3 BitCPM4: Quantization-Aware Training for Ternary LLMs

BitCPM4 is an efficient quantization-aware training approach for adapting high-precision models into ternary LLMs. It reduces QAT training cost while retaining competitive performance, though the smallest model is weaker on challenging mathematics and coding tasks.

  • Method: BitCPM4 uses a two-stage QAT framework initialized from a pre-trained high-precision model instead of training a ternary model from scratch.The method applies ternary quantization to weights rather than activations because activation quantization adds overhead without substantially reducing end-side inference costs.
  • Training Schedule: When QAT tokens exceed 40% of total training tokens, final loss approaches that of training a ternary model from scratch.Experiments on 5M- and 150M-parameter models support continual-training schedules that reproduce the effect of scratch QAT.
  • Results: BitCPM4-0.5B outperforms Qwen3-0.6B on knowledge-related tasks, while BitCPM4-1B performs similarly to competing 2B-parameter models.The comparison is reported in Table 6 against representative models.
  • Efficiency: BitCPM4 uses only 10% of the tokens required by BitNet-2B while delivering competitive results, showing that scratch QAT is unnecessary.The full QAT process trains BitCPM4 variants at 0.5B and 1B parameters.
  • Limitations: The 0.5B model is relatively weaker on challenging mathematical and coding tasks, which the paper attributes to limited reasoning capacity from smaller model size.The authors plan to apply the QAT method to larger models and further consider operators for extremely low-bit models.

4 Efficient Inference and Deployment

CPM.cu targets resource-constrained end-side inference with lightweight CUDA infrastructure, speculative sampling, sparse attention, FR-Spec vocabulary compression, and prefix-aware quantization. These components reduce computation while preserving target-model output correctness and quantized-model performance.

  • CPM.cu: CPM.cu combines static memory management, kernel fusion, speculative sampling, and InfLLM v2 sparse-attention kernels for end-side NVIDIA chips.Speculative sampling uses draft-then-verify generation, with fused kernels tailored to tree-based drafting.
  • FR-Spec: FR-Spec compresses the draft model’s vocabulary to high-frequency tokens while retaining the target model’s full vocabulary.This preserves the target model’s verification process and final output distribution.
  • FR-Spec: 25% of the vocabulary captures 95% of token occurrences and yields a 4× reduction in draft language-model-head overhead.The reduction follows from scaling computation with |Vhigh| instead of |V|.
  • FR-Spec: FR-Spec reduces language-model-head complexity from O(nd|V|) to O(nd|Vhigh|) without retraining or architectural modification.It can replace standard drafting computation in existing speculative-sampling frameworks.
  • Prefix-aware quantization: P-GPTQ addresses initial-token bias in quantization calibration, while S-P-GPTQ achieves the smallest performance degradation among evaluated quantized methods.The evaluation quantizes all linear layers to per-group INT4 using 1,024 calibration sequences.
  • Quantized inference: SpecMQuant adapts speculative sampling to quantized target models by reducing draft-token counts when quantization alleviates target-model memory-access bottlenecks.The method analyzes speculative sampling with W4A16 target models such as GPTQ-quantized models.

4.2 ArkInfer: Cross-Platform Deployment System

ArkInfer is a cross-platform deployment system designed to reduce the engineering burden created by fragmented end-side chips and model formats. It unifies heterogeneous backends through reusable execution, decoding, conversion, and model-access components.

  • System motivation: ArkInfer addresses fragmented end-side hardware by providing a cross-platform compatibility layer with reusable decoding schemes and extensible deployment components.Its design targets diverse platforms and model applications.
  • Architecture: Adapters normalize backend APIs, while unified tensors and KV-cache management provide consistent data handling and token-generation state across platforms.The architecture integrates frameworks for MediaTek, Nvidia, Qualcomm, and Rockchip devices.
  • Architecture: ArkInfer’s abstract executor interface governs neural execution, autoregressive decoding, sampling, and preprocessing through common tensor-based inputs and outputs.This abstraction supports multiple input modalities and model-related runtime processes.
  • Runtime management: Executor-level scheduling and tracing support heterogeneous resource use, debugging, performance analysis, and per-stage precision alignment.These capabilities target adaptation problems across diverse end-side platforms.
  • Decoding: ArkInfer reuses hardware-decoupled speculative and constrained decoding while incorporating BiTA-based speculative decoding without additional draft models.Guidance-based constrained decoding supports structured outputs such as JSON or SQL.
  • Deployment: An extensible frontend and automated model-conversion pipeline streamline access to pre-adapted models across platform-specific formats.The system centers deployment around a model zoo.

5 Evaluations

MiniCPM4 models are evaluated against open-source baselines on knowledge, reasoning, and long-context tasks, while MiniCPM4.1 is tested in deep reasoning and sparse-attention settings. The results report strong benchmark performance, long-context robustness, and substantial inference speed advantages.

  • Standard evaluation: MiniCPM4-0.5B and MiniCPM4-8B are evaluated against similarly sized open-source models across knowledge-intensive and reasoning benchmarks.The benchmark suite includes MMLU, CMMLU, CEval, BBH, GSM8K, MATH500, and AIME.
  • Standard evaluation: MiniCPM4 achieves state-of-the-art performance among similar-sized models and surpasses some open-source models with substantially more parameters.The reported comparisons include MiniCPM4-0.5B versus Llama3.2-1B and Gemma3-1B, and MiniCPM4-8B versus Gemma3-12B and Phi4-14B.
  • Standard evaluation: 8 trillion training tokens deliver performance comparable to Qwen3, which uses 36 trillion tokens, corresponding to 22% of Qwen3’s training-data scale.The comparison is reported as a training-cost advantage alongside comparable performance.
  • Deep reasoning evaluation: MiniCPM4.1-8B reaches a 79.93 overall average and outperforms similar-sized models on deep reasoning tasks.The reported result describes strong performance across varied reasoning tasks.
  • Long-context evaluation: MiniCPM4 achieves 100% accuracy on the needle-in-a-haystack task while attending to only 6K context tokens per token on a 128K context.The model was pretrained at 32K context and achieved 100% accuracy at 4× that length after YaRN extension.
  • Long-context evaluation: Sparse attention changes MiniCPM4.1’s weighted-average RULER score from 88.93% to 85.84%, a 3.09 percentage-point decrease.The reported comparison characterizes sparse attention as retaining comparable performance on many long-sequence tasks.
  • Inference efficiency: MiniCPM4 achieves approximately 7x decoding acceleration over Qwen3-8B on Jetson AGX Orin, with larger efficiency advantages as sequence length increases.The evaluation compares throughput across sequences from 32K to 128K on two end-side chips.

6 Applications

The paper extends MiniCPM4 to survey generation and MCP-based tool use, targeting private, resource-efficient, and interoperable on-device applications. MiniCPM4-Survey combines planning, retrieval, writing, and multi-stage training, while MiniCPM4-MCP adapts the model to tool interaction.

  • Application motivation: On-device survey systems support confidentiality by keeping sensitive resources local and reduce inference costs for token-intensive survey writing.These advantages complement cloud-based survey services.
  • MiniCPM4-Survey: MiniCPM4-Survey generates long-form surveys through a Plan-Retrieve-Write pipeline that structures content before retrieval and writing.The system is built on MiniCPM4-8B and targets trustworthy survey generation.
  • MiniCPM4-Survey: The survey system uses expert-authored surveys and research-paper retrieval data, with supervised fine-tuning followed by section-level and survey-level reinforcement learning.The training pipeline combines curated datasets with multiple optimization stages.
  • MiniCPM4-Survey: Survey training evaluates planning, retrieval, content creation, and citation accuracy through rewards covering coherence, relevance, coverage, novelty, hallucination, and factual accuracy.The reward system addresses the difficulty of assessing survey quality using perplexity alone.
  • MiniCPM4-Survey: MiniCPM4-Survey surpasses open- and closed-source baseline systems on content metrics, matches OpenAI Deep Research competitively, and attains the highest examined factual-metric scores.The method also improves significantly from supervised fine-tuning to reinforcement learning.
  • MiniCPM4-MCP: MiniCPM4-MCP adapts MiniCPM4-8B to interact with tools and data resources through MCP, addressing nonstandardized LLM–tool interactions.The model is evaluated on human-annotated MCP tool-calling data.

7 Conclusion and Future Works

MiniCPM4 achieves comparable performance with efficient pre-training and inference, while future work targets further efficiency improvements across architecture, data, reinforcement learning, and end-side systems.

  • Conclusion: 8 trillion training tokens and a 5× speedup for long-sequence processing summarize MiniCPM4’s reported efficiency.The report attributes these results to efficient pre-training data and infrastructure, plus efficient architecture and inference systems.
  • Conclusion: The authors release MiniCPM4 model parameters and inference code to support the open-source community.The release is presented as part of the report’s contribution to model development and evaluation.
  • Future Works: Future work targets sparse architectures for infinitely long sequences on end-side devices.The planned architecture research focuses on improving sparse model efficiency for increasingly long inputs.
  • Future Works: Future data work will improve corpus quality and synthesize large-scale reasoning-intensive pre-training datasets.The stated goal is to strengthen foundational capabilities through higher-quality and reasoning-focused data.
  • Future Works: The authors will explore reinforcement learning for learning skills from various environments and develop systems for more end-side platforms.These plans aim to broaden future training and inference research around MiniCPM models.

8 Contributions and Acknowledgments

The report credits MiniCPM4 and MiniCPM4.1 to a collective team effort, with separate project design, coordination, supervision, and contributor roles.

  • Acknowledgments: MiniCPM4 and MiniCPM4.1 are credited to the collective efforts of the entire team.
  • Acknowledgments: Project design and coordination are credited to Chaojun Xiao, Yuxuan Li, and Xu Han.
  • Acknowledgments: The report lists a large contributor group in alphabetical order by last name.
  • Acknowledgments: Supervision is credited to Xu Han, Zhiyuan Liu, Guoyang Zeng, Chao Jia, Dahai Li, and Maosong Sun.
Loading 2506.07900v2…