REVIEW 17 cited by
Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition
Not yet reviewed by Pith; the record is open.
This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.
SPECIMEN: schema-true, not a live event
T0 review · schema-true
One-sentence machine reading of the paper's core claim.
pith:XXXXXXXX · record.json · timestamp
Pangu Embedded: An Efficient Dual-system LLM Reasoner with Metacognition
read the original abstract
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.
Forward citations
Cited by 17 Pith papers
-
COSM: A Cooperative Scheduling Framework for Concurrent PIM and CPU Execution on Mobile Devices
COSM is a cooperative scheduling framework for concurrent PIM and CPU execution on mobile devices that hides PIM latency and overlaps execution with data transfer, achieving up to 2.8x PIM throughput with less than 2%...
-
TypePro: Boosting LLM-Based Type Inference via Inter-Procedural Slicing
TypePro reaches 88.9% and 86.6% Top-1 exact match on Python and TypeScript type-inference datasets by feeding LLMs inter-procedural slices plus structurally derived candidate types.
-
Selective Left-Shift: Turning Test-Time Compute and Difficulty-based Curation into Training Data for Low-Resource Code Generation
Left-shifting iterative compiler/test refinement into verified SFT data, then GRPO on difficulty-curated IO rewards, lifts Qwen3-8B Julia pass@1 past prior SOTA at 1/3 data and 1/6 cost, and bootstraps Ballerina.
-
COSM: A Cooperative Scheduling Framework for Concurrent PIM and CPU Execution on Mobile Devices
COSM enables concurrent PIM and CPU execution on mobiles via low-interference control and idleness-aware scheduling, delivering up to 2.8x PIM throughput with under 2% CPU slowdown.
-
From Rigid to Dynamic: Entropy-Guided Adaptive Inference for Long-Context LLMs
EntropyInfer adaptively allocates inference compute using per-head attention entropy for rigid/dynamic classification during prefilling and compresses KV cache with generated tokens, achieving up to 2.39x speedup on l...
-
Focus-then-Context: Subject-Centric Progressive Visual Token Reduction for Vision-Language Models
SPpruner reduces visual tokens in VLMs via focus identification followed by context-aware scanning, retaining 22.2% tokens for 2.53x speedup on Qwen2.5-VL with negligible accuracy loss.
-
AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reuse
AdapShot adaptively tunes shot count via entropy probes and reuses semantically-matched KV caches with position decoupling to deliver ~10% accuracy gains and 4.64x speedup over fixed-shot baselines.
-
AdapShot: Adaptive Many-Shot In-Context Learning with Semantic-Aware KV Cache Reuse
AdapShot adaptively optimizes shot counts via probe entropy and semantic KV cache reuse with decoupling, reporting ~10% gain and 4.64x speedup over DBSA.
-
Towards Efficient Large Language Reasoning Models via Extreme-Ratio Chain-of-Thought Compression
Extra-CoT trains a semantic compressor on math CoT data, applies mixed-ratio SFT, and uses CHRPO reinforcement learning to achieve over 73% token reduction on MATH-500 with 0.6% accuracy gain on Qwen3-1.7B.
-
AGZO: Activation-Guided Zeroth-Order Optimization for LLM Fine-Tuning
AGZO restricts ZO perturbations to an activation-derived low-rank subspace, claiming higher gradient cosine similarity and better benchmark performance than isotropic ZO baselines on Qwen3 and Pangu models.
-
PrunePath: Towards Highly Structured Sparse Language Models
PrunePath introduces budget-adaptive structured sparsification for FFN layers via softmax routing and cumulative-mass thresholds on top of MoEfication, with Triton kernels for inference speedups.
-
Ratio-Variance Regularized Policy Optimization
R²VPO uses ratio-variance regularization as a distributional soft brake on policy updates, claiming better performance than PPO on math reasoning and robotic control without hard clipping.
-
OPERA: An Agent for Image Restoration with End-to-End Joint Planning-Execution Optimization
OPERA jointly optimizes restoration planning via RL over tool compositions and execution via agent-guided co-training of tools, claiming consistent gains over all-in-one models and prior agent methods on multi-degrada...
-
Near-Policy: Accelerating On-Policy Distillation via Asynchronous Generation and Selective Packing
NPD accelerates on-policy distillation 8.1 times faster than baselines by using asynchronous SFT with Δ-IFD filtering, outperforming standard SFT and enabling a 1B model to achieve 68.73% SOTA score.
-
A-IO: Adaptive Inference Orchestration for Memory-Bound NPUs
A-IO adaptively orchestrates LLM inference on NPUs to address memory bottlenecks, model scaling paradoxes, and synchronization costs in speculative decoding.
-
Fairness-Aware and Latency-Controllable Scheduling for Chunked-Prefill LLM Serving
The paper introduces an aging-based scheduler with LPRS and APC for chunked-prefill LLM engines that cuts mean end-to-end latency by over 10% and lowers P99 tail latency versus FCFS on real hardware.
-
Pangu-ACE: Adaptive Cascaded Experts for Educational Response Generation on EduBench
Pangu-ACE improves educational response quality on EduBench from 0.457 to 0.538 and format validity from 0.707 to 0.866 by routing 19.7% of samples to a 1B model while escalating the rest to 7B.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.