efficiency — AI Digest

31 июл TurboVLA: Real-Time Vision-Language-Action Model at 32 Hz on an RTX 4090 with <1 GB VRAM research
12 мая Qwen-Image-2.0: Unified Image Generation and Editing at 2K Resolution, Top-1 on AI Arena Alibaba research
13 мая Baidu Releases ERNIE 5.1 at 6% of Industry Pre-Training Cost, Enters Global Top-10 Search Baidu models-llm
26 июн JetSpec: Breaking the Scaling Ceiling of Speculative Decoding with Parallel Tree Drafting Hao AI Lab, UC San Diego research
2 июл Dockerless: Environment-Free Program Verifier for Coding Agents ByteDance research
2 июл DOPD: Dual On-Policy Distillation with Advantage-Aware Token Routing research
4 июл A Systematic Analysis of Hybrid Linear Attention: 72-Model Study ByteDance Seed research
6 июл Program-as-Weights: Compiling Task Specs into LoRA Adapters for On-Device Inference University of Waterloo / Harvard research
6 июл MrFlow: Training-Free 10-25x Speedup for Flow-Matching Text-to-Image Models Beihang University research
31 июл Mage-VL: An Efficient Codec-Native Streaming Multimodal Foundation Model Microsoft research
15 сен Tencent Hunyuan open-sources SAS sparse-attention routers trained end-to-end on Qwen3 Tencent Hunyuan research
11 июн Kwai Keye-VL-2.0: Open-Source 30B MoE Multimodal Model with 256K Context for Long Video Kwai research
14 июн MiniMax Sparse Attention: 28× Compute Reduction at 1M-Token Context with No Quality Loss MiniMax research
21 июн Moebius: 0.2B Lightweight Image Inpainting Framework Matches 11.9B FLUX Model Huazhong University of Science and Technology research
12 авг BDH-CQ: Recurrent Latent Reasoning Model Breaks ARC-AGI-1 Cost-Efficiency Frontier Pathway research
9 сен Uno: lossless LLM speedups via diffusion-augmented autoregression MBZUAI research
12 сен WMRL: scaling automatic research agents via world models University of Illinois / NVIDIA research
12 сен NCP-ArchPreview: latent-space language models at 8.9B via Next Concept Prediction Shanghai AI Lab research
16 сен Vidu S2: real-time interactive, editable, and spatial video generation ShengShu Intelligence research
16 сен ZGCM-1: a fully open, extremely efficient 7B foundation model for math and agentic search ZGCAGI research
16 сен Dream-RSI: recursive self-improvement through evolving worlds Google research
16 мая Orthrus: 7.8x Inference Speedup for Qwen3 via Autoregressive-Diffusion KV Sharing research
16 мая SANA-WM: Minute-Scale 720p World Modeling on a Single GPU NVIDIA research
4 июн ThoughtFold: Introspective Preference Learning Cuts Reasoning Tokens by 56% Without Accuracy Loss research
16 июн FastContext: Specialized Exploration Subagent Cuts Coding Agent Token Usage by 60% Microsoft / Shanghai Jiao Tong University research
25 июн Quantized Reasoning Models Think They Need to Think Longer, but They Do Not Meta research
2 июл vLLM v0.24.0: Model Runner V2 Default, Rust Frontend, SM90 FP8 Speedups vLLM tools
3 июл Program-as-Weights: Compile-Once Adapter Paradigm Matches 32B Models at 1/50 the Memory research
10 июл Super Weights in LLMs: Why High-Salience Parameters Fail as Fine-Tuning Targets Amazon research
1 сен Qwen team details the Qwen3.8-Next architecture: hybrid attention, n-gram embeddings, Muon Qwen (Alibaba) research
2 сен Google DeepMind ships agentic video understanding in Gemini Google DeepMind tools
7 сен LatentPress: compressing context into continuous memory tokens a frozen LLM reads directly research
18 сен DeepSeek-V4.1-Flash: Pushing the Limits of KV Cache Compression DeepSeek research
25 июн Are We Ready For an Agent-Native Memory System? SJTU Benchmarks 12 Architectures research
2 июл BlockPilot: Instance-Adaptive Block Size for Diffusion-Based Speculative Decoding research
13 июл Jet-Long: Efficient Long-Context Extension with Dynamic Bifocal RoPE MIT / NVIDIA research
13 июл Scalable Visual Pretraining for Language Intelligence research
2 сен SMELT: looped MoE transformers match baseline scaling at matched compute ByteDance Seed research
15 сен COBRA-Skills: contextual bandit-guided evolution for agent skills Chinese University of Hong Kong, Shenzhen research
9 июн On the Geometry of On-Policy Distillation: A Training Paradigm Distinct from SFT and RLVR Hong Kong University of Science and Technology research
24 июн SHERLOC: Structured Diagnostic Localization Cuts Code Repair Token Usage by 36.7% research
26 июн OPRD: On-Policy Representation Distillation for Post-Training LLMs research
3 июл ELDR: Expert-Locality-Aware Routing Cuts MoE Serving Latency by up to 14% Microsoft Research research
3 июл FlashMorph: Data-Driven Hybrid Attention Layer Placement via Learnable Gates ByteDance Seed research
23 июл Mage-Flow: An Efficient Native-Resolution Foundation Model for Image Generation and Editing Microsoft research
14 авг LycheeMemory V2: Efficient Long-Term Memory for LLM Agents via Semantic Segment-Level Consolidation research
1 сен Does on-policy distillation really distill? Teacher-free OPSA beats it on AIME24 Purdue University research
8 сен Language Models Can Control Their Own Attention research
8 сен Don't Drop Dropout: Optimizing Layer Sparsity for Efficient LLM Training and Inference Cerebras research
21 сен SoL-Pi: Recursively Scaling Auto-Research Loops for Efficient Agent Harness research
23 окт vLLM 0.32.0 widens low-precision KV cache and drops Python 3.10 vLLM tools