mlsanityData to 7 Oct 2026

Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition

arXiv 2505.22375 ↗First posted 28 May 2025cs.CLcited by 61

On the map: LLM reasoning → Reinforcement learning for reasoning

Abstract (arXiv)

This work presents Pangu Embedded, an efficient Large Language Model (LLM) reasoner developed on Ascend Neural Processing Units (NPUs), featuring flexible fast and slow thinking capabilities. Pangu Embedded addresses the significant computational costs and inference latency challenges prevalent in existing reasoning-optimized LLMs. We propose a two-stage training framework for its construction. In Stage 1, the model is finetuned via an iterative distillation process, incorporating inter-iteration model merging to effectively aggregate complementary knowledge. This is followed by reinforcement learning on Ascend clusters, optimized by a latency-tolerant scheduler that combines stale synchronous parallelism with prioritized data queues. The RL process is guided by a Multi-source Adaptive Reward System (MARS), which generates dynamic, task-specific reward signals using deterministic metrics and lightweight LLM evaluators for mathematics, coding, and general problem-solving tasks. Stage 2 introduces a dual-system framework, endowing Pangu Embedded with a "fast" mode for routine queries and a deeper "slow" mode for complex inference. This framework offers both manual mode switching for user control and an automatic, complexity-aware mode selection mechanism that dynamically allocates computational resources to balance latency and reasoning depth. Experimental results on benchmarks including AIME 2024, GPQA, and LiveCodeBench demonstrate that Pangu Embedded with 7B parameters, outperforms similar-size models like Qwen3-8B and GLM4-9B. It delivers rapid responses and state-of-the-art reasoning quality within a single, unified model architecture, highlighting a promising direction for developing powerful yet practically deployable LLM reasoners.

Key facts. Cited by 61 papers in the map (data through 7 Oct 2026). Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning. Builds on 47 of its 57 references that are in the map. First posted on arXiv 28 May 2025.

Builds on

47 of 57 references.

Closest ancestors · 6

Same direction, newest first

  1. 2025Fast-Slow Thinking GRPO for Large Vision-Language Model Reasoning25 Apr
  2. 2025SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild24 Mar
  3. 2025DAPO: An Open-Source LLM Reinforcement Learning System at Scale18 Mar
  4. 2025DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning22 Jan
  5. 2023Let's Verify Step by Step31 May
  6. 2017Proximal Policy Optimization Algorithms20 Jul

Other references · 41

  1. 2023GPT-4 Technical Reportcited by 18,132
  2. 2022Training language models to follow instructions with human feedbackcited by 15,712
  3. 2023LLaMA: Open and Efficient Foundation Language Modelscited by 13,958
  4. 2022Chain-of-Thought Prompting Elicits Reasoning in Large Language Modelscited by 12,302
  5. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 8,851
  6. 2025Qwen3 Technical Reportcited by 8,195
  7. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,767
  8. 2022Training a Helpful and Harmless Assistant with Reinforcement Learning from Human Feedbackcited by 3,634
  9. 2023GPQA: A Graduate-Level Google-Proof Q&A Benchmarkcited by 2,748
  10. 2023A Survey on Large Language Model based Autonomous Agentscited by 2,203
  11. 2023Code Llama: Open Foundation Models for Codecited by 2,105
  12. 2024HybridFlow: A Flexible and Efficient RLHF Frameworkcited by 1,980
29 more
  1. 2024LiveCodeBench: Holistic and Contamination Free Evaluation of Large Language Models for Codecited by 1,906
  2. 2025s1: Simple test-time scalingcited by 1,311
  3. 2023LIMA: Less Is More for Alignmentcited by 1,080
  4. 2024Training Large Language Models to Reason in a Continuous Latent Spacecited by 763
  5. 2025LIMO: Less is More for Reasoningcited by 488
  6. 2023What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuningcited by 354
  7. 2023From Quantity to Quality: Boosting LLM Performance with Self-Guided Data Selection for Instruction Tuningcited by 301
  8. 2022Teaching Small Language Models to Reasoncited by 279
  9. 2025O1-Pruner: Length-Harmonizing Fine-Tuning for O1-Like Reasoning Pruningcited by 278
  10. 2024A Survey on Knowledge Distillation of Large Language Modelscited by 276
  11. 2024Are We Done with MMLU?cited by 178
  12. 2024Distilling System 2 into System 1cited by 152
  13. 2025Llama-Nemotron: Efficient Reasoning Modelscited by 151
  14. 2025AdaptThink: Reasoning Models Can Learn When to Thinkcited by 141
  15. 2024Fast Best-of-N Decoding via Speculative Rejectioncited by 112
  16. 2023LawBench: Benchmarking Legal Knowledge of Large Language Modelscited by 111
  17. 2024Forest-of-Thought: Scaling Test-Time Compute for Enhancing LLM Reasoningcited by 110
  18. 2025Small Models Struggle to Learn from Strong Reasonerscited by 109
  19. 2023MoDS: Model-oriented Data Selection for Instruction Tuningcited by 107
  20. 2025Efficient Reasoning Models: A Surveycited by 93
  21. 2024O1 Replication Journey -- Part 2: Surpassing O1-preview through Simple Distillation, Big Progress or Bitter Lesson?cited by 90
  22. 2024LLM Pruning and Distillation in Practice: The Minitron Approachcited by 82
  23. 2025Learning When to Think: Shaping Adaptive Reasoning in R1-Style Models via Multi-Stage RLcited by 59
  24. 2024Survey on Knowledge Distillation for Large Language Models: Methods, Evaluation, and Applicationcited by 55
  25. 2024Learning or Self-aligning? Rethinking Instruction Fine-tuningcited by 43
  26. 2024Entropy Law: The Story Behind Data Compression and LLM Performancecited by 42
  27. 2023PanGu-π: Enhancing Language Model Architectures via Nonlinearity Compensationcited by 17
  28. 2025Marco-o1 v2: Towards Widening The Distillation Bottleneck for Reasoning Modelscited by 13
  29. 2025Pangu Ultra: Pushing the Limits of Dense Large Language Models on Ascend NPUscited by 13

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 14; 2026: 47.

61 papers.

Sort:

By area:

  1. 2025Depth-Breadth Synergy in RLVR: Unlocking LLM Reasoning Gains with Adaptive Explorationcited by 40
  2. 2025Reasoning on a Budget: A Survey of Adaptive and Controllable Test-Time Compute in LLMscited by 31
  3. 2025Towards Concise and Adaptive Thinking in Large Reasoning Models: A Surveycited by 18
  4. 2025ROOT: Robust Orthogonalized Optimizer for Neural Network Trainingcited by 13
  5. 2025Your Models Have Thought Enough: Training Large Reasoning Models to Stop Overthinkingcited by 9
  6. 2025VAR-MATH: Probing True Mathematical Reasoning in LLMS via Symbolic Multi-Instance Benchmarkscited by 8
  7. 2025Fine-tuning Large Language Model for Automated Algorithm Designcited by 6
  8. 2025Revealing the Power of Post-Training for Small Language Models via Knowledge Distillationcited by 6
  9. 2025Efficient Reinforcement Learning with Semantic and Token Entropy for LLM Reasoningcited by 6
  10. 2026Ratio-Variance Regularized Policy Optimization for Efficient LLM Fine-tuningcited by 6
  11. 2026What Matters For Safety Alignment?cited by 6
  12. 2026Streaming-dLLM: Accelerating Diffusion LLMs via Suffix Pruning and Dynamic Decodingcited by 6
38 more
  1. 2025CAMERA: Multi-Matrix Joint Compression for MoE Models via Micro-Expert Redundancy Analysiscited by 5
  2. 2026From Prefix Cache to Fusion RAG Cache: Accelerating LLM Inference in Retrieval-Augmented Generationcited by 5
  3. 2026DLLM Agent: See Farther, Run Fastercited by 5
  4. 2026Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compressioncited by 5
  5. 2025Enhancing the Medical Context-Awareness Ability of LLMs via Multifaceted Self-Refinement Learningcited by 4
  6. 2026Efficient Reasoning with Balanced Thinkingcited by 4
  7. 2025Searching Meta Reasoning Skeleton to Guide LLM Reasoningcited by 3
  8. 2026AGZO: Activation-Guided Zeroth-Order Optimization for LLM Fine-Tuningcited by 3
  9. 2026Taxonomy-Aware Representation Alignment for Hierarchical Visual Recognition with Large Multimodal Modelscited by 3
  10. 2026Near-Policy: Accelerating On-Policy Distillation via Asynchronous Generation and Selective Packingcited by 3
  11. 2025MoRAgent: Parameter Efficient Agent Tuning with Mixture-of-Rolescited by 2
  12. 2026TiFRe: Text-guided Video Frame Reduction for Efficient Video Multi-modal Large Language Modelscited by 2
  13. 2026SpatialText: A Pure-Text Cognitive Benchmark for Spatial Understanding in Large Language Modelscited by 2
  14. 2026OPERA: An Agent for Image Restoration with End-to-End Joint Planning-Execution Optimizationcited by 2
  15. 2026Fairness-Aware and Latency-Controllable Scheduling for Chunked-Prefill LLM Servingcited by 2
  16. 2025Post-Training Quantization of OpenPangu Models for Efficient Deployment on Atlas A2cited by 1
  17. 2026Thinking-Based Non-Thinking: Solving the Reward Hacking Problem in Training Hybrid Reasoning Models via Reinforcement Learningcited by 1
  18. 2026DScheLLM: Enabling Dynamic Scheduling through a Fine-Tuned Dual-System Large language Modelcited by 1
  19. 2026Beyond Static Question Banks: Dynamic Knowledge Expansion via LLM-Automated Graph Construction and Adaptive Generationcited by 1
  20. 2026Traceable Cross-Source RAG for Chinese Tibetan Medicine Question Answeringcited by 1
  21. 2026FMBench: Adaptive Large Language Model Output Formattingcited by 1
  22. 2026Quasar: Quantized Self-Speculative Acceleration for Rapid Inference via Memory-Efficient Verificationcited by 1
  23. 2026Accelerating OpenPangu Inference on NPU via Speculative Decodingcited by 1
  24. 2026Advancing Automated Algorithm Design via Evolutionary Stagewise Design with LLMscited by 1
  25. 2026AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reusecited by 1
  26. 2026Metacognition in LLMs: Foundations, Progress, and Opportunitiescited by 1
  27. 2026AgenticCANN: Automated Ascend C Operator Generation via Knowledge-Augmented Agentic Evolutioncited by 1
  28. 2026A Comprehensive Evaluation of LLM Reasoning: From Single-Model to Multi-Agent Paradigmscited by 0
  29. 2026V-CAGE: Context-Aware Generation and Verification for Scalable Long-Horizon Embodied Taskscited by 0
  30. 2026Adaptive Confidence Gating in Multi-Agent Collaboration for Efficient and Optimized Code Generationcited by 0
  31. 2026Structured Self-Consistency:A Multi-Task Evaluation of LLMs on VirtualHomecited by 0
  32. 2026EvoOpt-LLM: Evolving industrial optimization models with large language modelscited by 0
  33. 2026How Few-shot Demonstrations Affect Prompt-based Defenses Against LLM Jailbreak Attackscited by 0
  34. 2026Differentially Private and Communication Efficient Large Language Model Split Inference via Stochastic Quantization and Soft Promptcited by 0
  35. 2026Silent Inconsistency in Data-Parallel Full Fine-Tuning: Diagnosing Worker-Level Optimization Misalignmentcited by 0
  36. 2026Multi-source Heterogeneous Public Opinion Analysis via Collaborative Reasoning and Adaptive Fusion: A Systematically Integrated Approachcited by 0
  37. 2026Metacognitive Behavioral Tuning of Large Language Models for Multi-Hop Question Answeringcited by 0
  38. 2026VisionPangu: A Compact and Fine-Grained Multimodal Assistant with 1.7B Parameterscited by 0

50 most cited of 61 All 61 (JSON)

Similar, not cited

Close in wording, no citation either way

  1. 2025General-Reasoner: Advancing LLM Reasoning Across All Domains
  2. 2025Reasoning Under 1 Billion: Memory-Augmented Reinforcement Learning for Large Language Models
  3. 2025A Survey of Reinforcement Learning for Large Reasoning Models
  4. 2025Training Language Models to Reason Efficiently
  5. 2025Curriculum Reinforcement Learning from Easy to Hard Tasks Improves LLM Reasoning
  6. 2026Training LLMs for Divide-and-Conquer Reasoning Elevates Test-Time Scalability
  7. 2025SABER: Switchable and Balanced Training for Efficient LLM Reasoning
  8. 2026nabla-Reasoner: LLM Reasoning via Test-Time Gradient Descent in Latent Space