efficiency

Skyfall AI unveiled SKY-001, an engineering world model that reconstructs a Jaguar car with thousands of 3D parts, claiming CAD accuracy on par with GPT-6 Astra while running 51× faster and costing 150× less.

本文提出结构保持的专家剪枝框架，通过 PEEP 与 DCEP 两种方法利用专家输出的结构冗余性对稀疏 MoE 大模型进行剪枝，在 DeepSeek-V2-Lite 和 Qwen3-30B-A3B 上超越现有基线。

Flide is an efficient and exact inference engine for LLM decision models that decouples logical tree structures from a single flattened token stream via ForkAttention, answering queries up to 9.5x faster than vLLM and SGLang while staying closest to a float32 reference.

A Zyphra Research paper introduces Zyphra Hierarchical Attention (ZHA), a subquadratic sparse attention method using a hierarchical tree-structured indexer over the KV cache, which can be adapted to pretrained models via short KL distillation and scales dramatically better beyond 1M context.

A research paper formulating long-context diffusion costs as dual budgets—attention graph edges and activated experts—presenting the DualBudget sparse-expert configuration with graph/compute bounds and evaluations showing gains through 16K context and a 20% cost reduction over DiffuSeq.

The paper compares continuous diffusion language models (LangFlow) and speculative drafters (DSpark) as standalone generators under matched training budgets, finding LangFlow superior on perplexity and MAUVE at longer sequence lengths while DSpark-RNN leads at length 16, highlighting iterative full-sequence refinement as a key advantage.

SkillMem introduces a compact external latent memory for reusable skills in latent multi-agent reasoning, achieving 38× storage reduction and up to 4.23× inference speedup across benchmarks with Qwen3 models from 4B to 14B.

Sam Altman argues that local inference is an inefficient allocation of resources in a constrained environment: equipping everyone's computer with large amounts of scarce DRAM that is only used about 5% of the time is far less effective than keeping cloud resources at high utilization.

DecisionRank 使用 item-space on-policy distillation 将生成式教师模型的排序知识迁移到高效决策模型，在 TREC DL、BEIR、BRIGHT 等检索基准上达到最先进性能，同时推理速度显著提升（最高快 24.5 倍），1.7B 参数即可超越 27B 生成式教师。

At EMNLP 2026, T-Tech researchers presented LIMe, a lightweight attention extension that reuses key-value buffers from earlier layers to mitigate representation collapse, reducing LLM training compute by ~15% at equal quality versus baseline LLaMA on ~1B-parameter models.

IDM: Intrinsic Distribution Matching 提出让生成器自身同时提供分数监督和分布匹配表示的方法，在保留完整 DMD 校正的基础上，对生成器 EMA 特征的均值和逐坐标标准差添加显式损失。在 2 NFE 预算下，基于 Z-Image 的 IDM 在 LongText-CN 和 LongText-EN 上取得最佳分数。

Cartesia 的 From Tokens to Bytes 系列第一篇探讨 Hierarchical Networks (H-Nets) 的扩展规律，展示学习式 tokenization 带来 3-5 倍计算效率提升，2 阶段 H-Net 以 10 倍更少算力和数据超越匹配的开源模型，并呈现 in-context tokenization 等新行为以及在生物与视觉等跨模态任务上的潜力。

Albert Gu announces a new version of hierarchical networks (H-Nets) that claim 3-5x efficiency gains in scaling law experiments, matching an open-source model trained on 11T tokens with roughly 10x less data and compute, as a step toward end-to-end learning from raw bytes.

A successor to the Tesseract OCR engine has been released as indra by datalab-to, notable for being just 1M parameters while showing strong gains in both quality and speed on benchmarks.

据 SkalskiP 的测试，GPT-6.1 Terra 在文本、目标检测和计数等视觉基准上超越 Astra，同时速度快约 47%、成本低约 92%，结果令人难以置信。

发布了 Spark-X2.6-1.7B 小模型，参数量不到 4B 版本的一半，但在 SWE Verified (59.8 vs 62.8) 和 τ²-bench (85.3 vs 86.9) 两项基准上性能接近。

MORE 是一个自进化训练框架，通过两阶段范式与 MATE-PPO 联合学习推理模式选择和计算预算分配，使智能体在编码、数学和通用推理基准上以更短的响应长度和更低的推理时间取得一致的性能提升。

SLICE-Mem 提出一种轻量级 LLM Agent 长期记忆框架，通过显著性引导的记忆准入与基于谓词的语义抽取实现免 LLM 调用的在线记忆构建，在保持接近 MemOS 基线精度的同时将 token 消耗降低 92.4%–99.5%。

ConsMSA introduces semantic distribution consistency learning across intra- and inter-modalities for multimodal sentiment analysis, enabling token reweighting and pruning while achieving state-of-the-art results on CMU-MOSI, CMU-MOSEI, and CH-SIMS with only 10% of tokens retained.

荷兰政府宣布计划通过AI驱动的效率改革裁减约25,000个公务员岗位（约占16%），目标是到2035年每年节省35亿欧元，并将于明年初公布AI部署方案。