1
全国唯一！商汤大装置临港AIDC获“算效+算电”双5A认证
业界首个5A算电协同认证
2
AdaR: A Framework for Equipping LLMs with Adaptive Reasoning
让大模型按需调整推理深度，AdaR框架为复杂问题动态分配计算资源，兼顾效率与准确。
arXiv:2510.04617v3 Announce Type: replace Abstract: Mathematical reasoning is a primary indicator of large language models (LLMs) intelligence. Howeve…
3
TELLME: Test-Enhanced Learning for Language Model Enrichment
用测试增强学习替代海量语料，为语言模型持续预训练省下高额计算成本，稳准快。
arXiv:2608.11788v1 Announce Type: cross Abstract: Continual pre-training (CPT) has been widely adopted as a method for domain adaptation in large lang…
4
Statistical Early Stopping for Reasoning Models
提出用统计方法为推理模型设计早停策略，显著提升计算效率而不损失准确度。
arXiv:2602.13935v2 Announce Type: replace-cross Abstract: While LLMs have seen substantial improvement in reasoning capabilities, they also sometimes …
5
In-Place Tokenizer Expansion for Pre-trained LLMs
提出一种原地扩展分词器的方案，解决预训练大模型因后期语言加入导致的 token 效率低下问题，降低延迟与能耗。
arXiv:2607.15232v1 Announce Type: cross Abstract: A tokenizer fixed at the start of pre-training allocates vocabulary in proportion to the pre-trainin…
6
SigLIP-HD by Fine-to-Coarse Supervision
从粗到细监督实现高效高分辨率视觉表示，SigLIP-HD在处理多模态大模型时兼顾计算效率与细粒度特征。
arXiv:2607.09488v1 Announce Type: new Abstract: High-quality visual representation is a long-standing pursuit in computer vision. In the context of mu…
7
HBM 之父金正浩：AI 的本质是内存，GPU 真正工作的时间只有 10%-30%
HBM之父一语道破：AI计算的真正瓶颈不在GPU，而在内存，多数时间GPU都在“等待”。
IT之家 7 月 6 日消息，被誉为“HBM 之父”的韩国科学技术院（KAIST）电气系教授金正浩近日接受《东亚日报》采访时表示，AI 的核心竞争力正在从 GPU 转向内存。 金正浩认为 AI 的本质是内存，GPU 在 AI 推理中的利用率远低于理论水平。AI 每次输出结果，都必须先从 HBM 读取…
8
Compress the Easy, Explore the Hard: Difficulty-Aware Entropy Regularization for Efficient LLM Reasoning
难度感知熵正则化，让大模型推理时自动压缩简单问题算力，集中资源攻克难题，效率与精度兼得。
arXiv:2602.22642v2 Announce Type: replace Abstract: Chain-of-Thought (CoT) has substantially empowered Large Language Models (LLMs) to tackle complex …
9
AI hit the memory wall — now it needs a new context tier
GPU算力不再是瓶颈，AI的下一场硬仗是“记忆墙”——上下文管理成2026关键战场。
Presented by Solidigm As inference workloads evolve from discrete question-and-answer exchanges into persistent, multi-step agentic systems, GPU avail…
10
New AI optimization framework beats Claude Code and Codex by 2.5x on the same compute budget
新AI优化框架Arbor在相同计算预算下比Claude Code和Codex性能提升2.5倍，为AI编码工具树立新标杆。
Imagine your engineering team just deployed an AI agent to search through internal company documents and answer employee questions. It works perfectly…
11
Do We Really Need Diffusion? A Fast U-Net for Paired Medical Image Translation
挑战扩散模型主流地位，快速U-Net实现配对医学图像翻译，兼顾速度与质量。
arXiv:2606.17675v1 Announce Type: new Abstract: Magnetic resonance imaging-signal fat fraction (MRI-SFF) quantifies tissue fat and serves as an establ…
12
CacheRL:Multi-Turn Tool-Calling Agents via Cached Rollouts and Hybrid Reward
小模型多步工具调用准确率达92%，逼近GPT-5且算力需求骤降100倍，极限效率训练新范式。
arXiv:2606.14179v1 Announce Type: new Abstract: We present CacheRL, a system for training small agent foundation models that achieves 92 percent proce…
13
Can Deep Neural Networks Improve Compression of Very Large Scientific Data?
深度神经网络能否突破极限，为超大规模科学数据压缩带来革命性提升？这篇arXiv新研究直击核心问题。
arXiv:2606.14353v1 Announce Type: new Abstract: Error-bounded lossy compression is a fundamental technique for managing the rapidly growing volumes of…
14
FOCUS: DLLMs Know How to Tame Their Compute Bound
DLLM解码计算并行化存在严重低效，只有少数token可解码，此文精准定位瓶颈并提出优化方向。
arXiv:2601.23278v2 Announce Type: replace Abstract: Diffusion Large Language Models (DLLMs) offer a compelling alternative to Auto-Regressive models, …
15
A Riemannian Approach to Low-Rank Optimal Transport
将低秩最优传输问题转化为黎曼流形上的优化，显著提升计算效率与可扩展性，理论突破值得关注。
arXiv:2606.12120v1 Announce Type: new Abstract: Low-rank optimal transport (OT) mitigates the quadratic scaling of classical solvers, yet existing app…
16
Unifying Data, Memory, and Compute Efficiency in LLM training: A Survey
从数据、内存、计算三方面统一梳理LLM训练效率优化策略，系统总结前沿方法
arXiv:2606.10706v1 Announce Type: cross Abstract: Resource constraints increasingly determine what can be trained, fine-tuned, and deployed in large l…
17
Threshold-Based Exclusive Batching for LLM Inference
挑战LLM推理中的混合批处理惯例，全新阈值独占批处理策略或改写效率规则。
arXiv:2606.00516v1 Announce Type: new Abstract: Mixed batching (MB)--interleaving prefill and decode in a single batch--has become the standard schedu…
18
Video Reasoning without Training
无需昂贵训练与冗长思维链，探索视频推理的高效新路径。
arXiv:2510.17045v2 Announce Type: replace-cross Abstract: Video reasoning using Large Multimodal Models (LMMs) relies on costly reinforcement learning…
19
LookWise: Knowing When and Where to Look for Fine-Grained Visual Reasoning in Multimodal Large Language Models
无需训练的轻量级视觉推理方法LookWise，让多模态大模型更高效地“细看”图像细节
arXiv:2603.00171v3 Announce Type: replace-cross Abstract: Multimodal Large Language Models (MLLMs) are shifting towards "Thinking with Images" by acti…
20
MoBiQuant: Mixture-of-Bits Quantization for Token-Adaptive Any-Precision LLM
MoBiQuant通过混合位数量化实现token自适应任意精度，大幅降低大模型内存与计算开销，推理更高效。
arXiv:2602.20191v2 Announce Type: replace-cross Abstract: Dynamic runtime latency and memory constraints necessitate flexible large language model (LL…