Daily Brief

At WAIC 2026, Arm China emphasizes that edge AI is not a scaled-down version of cloud AI but a new computing market defined by power consumption, real-time capabilities, and reliability. Their Star 300 AIoT platform aims to enable AI capabilities in resource-constrained environments, while the Zhouyi X3-Pro addresses complex inference needs across diverse edge scenarios.

OpenLanguageModel (OLM) is an open-source PyTorch library designed for building and pretraining small language models, emphasizing readability and composability. It achieves 90.6% weak-scaling efficiency on a 348M-parameter workload across four GPUs, making it suitable for educational and research purposes.

SpecLA introduces an efficient speculative decoding runtime for stateful linear-attention models, achieving up to 1.70x speedup over traditional autoregressive decoding on NVIDIA H100 with GDN-1.3B target. It employs topology-aware kernels and confidence pruning to optimize the verification of token candidates, enhancing performance in language tasks.

The paper addresses model collapse in synthetic data learning for iterative instruction tuning, proposing KITE, a two-stage framework that enhances stability in model performance. KITE combines failure-guided data generation with boundary-aware uncertainty curation, outperforming existing synthetic data baselines across various datasets and open-source LLMs.

RIMS introduces a three-stage preference optimization framework for small-scale language models (SLMs) in retrieval-augmented generation, enhancing performance under noisy conditions. It outperforms existing methods like RoseRAG on multi-hop question answering benchmarks, achieving significant improvements in Exact Match and F1 scores. The approach leverages synthetic data generation and a differentiable soft aggregation mechanism to optimize preference selection.

This study develops the QQ equality as an audit criterion for large language models (LLMs), revealing that forced-binary next-token log-probabilities are insufficient for distribution-level QQ audits. The research indicates that 17 out of 18 item pairs were saturated, suggesting a need for pre-specified saturation diagnostics in model evaluations.

MFGNet-Gear is a synthetic 3D dataset for quality inspection in manufacturing, featuring 24,000 paired polygon meshes and point clouds across 12 gear designs and 4 quality classes. This dataset addresses the challenges of class imbalance and data scarcity in defect detection, enabling improved deep learning applications in 3D metrology.

The KyrgyzLLM-Bench benchmark suite evaluates 26 LLMs in Kyrgyz, revealing performance gaps and translation artifacts. Notably, few-shot prompting enhances open-source models in reading comprehension, while proprietary models show inconsistent results. All datasets and evaluation tools are publicly released to advance Kyrgyz NLP research.