Tagged articles
Model Efficiency
35 articles · Page 1 of 1
The AREX-2 model, a 27B parameter model trained with long-horizon reflection on machine learning and algorithmic programming tasks, outperforms much larger models on benchmarks like MLE-bench and BrowseComp, demonstrating that reflective self-improvement can transfer across domains and provide a verifiable path to recursive self-improvement (RSI).
AI researchAREX-2BrowseComp
0 likes · 12 min read

With the 5‑hour usage cap temporarily lifted for Plus, Business, and Pro plans and the efficiency‑focused GPT 5.6 Sol update, developers must rethink model quotas, shifting from simple time windows to task‑budget planning that considers tokens, concurrency, tool calls, and model tiers to maintain productivity in high‑frequency AI workspaces.
AI workflowChatGPTCodex
0 likes · 8 min read

This article surveys the latest on‑policy distillation (OPD) research, categorizing each work as either a reinterpretation of an existing problem or a modification of a different module, and highlights the experimental findings, design choices, and trade‑offs reported across the papers.
LLMModel EfficiencyOPD
0 likes · 31 min read

The article breaks down the three core factors—tokens, electricity, and data—that together determine AI model cost, speed, capability, and industry competition, illustrating the trade‑offs with concrete numbers, examples, and future outlooks.
AIAI industryModel Efficiency
0 likes · 10 min read


SuanNiMay 28, 2026 · Artificial Intelligence
Microsoft’s Lens team shows that a 3.8 B‑parameter image‑generation model can match or surpass 6 B‑plus models while consuming only about 19 % of the GPU compute, thanks to aggressive model compression, dense captioning, mixed‑resolution training, optimized VAE and language encoders, and targeted RL fine‑tuning.
Image GenerationModel Efficiencybenchmarking
0 likes · 14 min read

The ICLR 2026 paper introduces Next‑ToBE, a training‑objective modification that replaces the one‑hot next‑token label with a soft distribution over a future token window, unlocking latent foresight in LLMs, improving future‑token hit rate, downstream reasoning performance, and reducing training memory and time.
Future Token PredictionModel EfficiencyNext-ToBE
0 likes · 12 min read

The paper introduces Next‑ToBE, a training‑objective modification that replaces the one‑hot next‑token label with a soft distribution covering a future token window, thereby activating latent anticipatory capacity in large language models and yielding significant gains in token‑hit rates, reasoning accuracy, and training efficiency.
Anticipatory CapacityModel EfficiencyNext-ToBE
0 likes · 11 min read

An analysis of eight frontier coding agents shows that token consumption in agentic coding tasks is highly variable, often orders of magnitude higher than simple code reasoning, and that spending more tokens does not reliably improve accuracy, with significant differences across models and limited predictability of costs.
AI agentsModel Efficiencycoding agents
0 likes · 11 min read

Tokens are the smallest language building blocks that AI models process, representing characters, words, subwords, punctuation or emojis; they determine context window size and generation speed, so tokenization directly impacts model understanding accuracy and efficiency, as explained in the 2026 Token Report.
AI FundamentalsModel Efficiencycontext window
0 likes · 4 min read

The paper presents a systematic empirical study that derives a power‑law scaling formula for reinforcement‑learning‑after‑training of large language models, demonstrating accurate inter‑ and intra‑model performance prediction, learning‑efficiency saturation, data‑reuse benefits, and cross‑architecture validity.
Data ReuseLlama-3Model Efficiency
0 likes · 11 min read

Google’s newly unveiled “Banana” visual Transformer introduces dynamic sparse attention that cuts inference cost 3‑5×, reduces memory by 70%, and improves ImageNet accuracy, while demonstrating real‑world gains in autonomous driving, medical imaging, and satellite analysis.
Computer VisionDynamic Sparse AttentionGoogle
0 likes · 6 min read

The article reviews the Kimi team's Attention Residuals approach, which substitutes traditional ResNet additive shortcuts with learned attention‑based weighting, explains the theoretical motivation linking depth to time, details full‑attention and block‑wise implementations, presents experimental results showing up to 1.25× compute efficiency and improved performance on reasoning and knowledge tasks.
Attention MechanismModel EfficiencyResidual Networks
0 likes · 11 min read

The STEM architecture replaces the Transformer feed‑forward network with a static token‑indexed embedding table, enabling lookup‑based memory that decouples capacity from compute, improves training stability, expands addressable memory, and delivers consistent performance gains on long‑context and knowledge‑intensive tasks.
Lookup MemoryModel EfficiencySTEM Architecture
0 likes · 8 min read


AntTechMar 4, 2026 · Artificial Intelligence
A new Region‑to‑Image Distillation (R2I) approach lets multimodal large language models perceive tiny visual details in a single forward pass, eliminating costly tool calls while achieving state‑of‑the‑art accuracy on the ZoomBench fine‑grained benchmark.
Model EfficiencyZoomBenchfine-grained perception
0 likes · 11 min read

Alec Radford’s latest Anthropic model, backed by a $1 billion funding round, claims significant performance gains through more efficient algorithms, challenging OpenAI and Google while pushing the AI field toward safer, more controllable large‑scale models.
AI industryAI safetyAlec Radford
0 likes · 5 min read


SuanNiMar 2, 2026 · Artificial Intelligence
A recent study from top universities examines how the Skills framework enables small language models to reduce memory usage, improve accuracy, and handle complex industrial tasks, revealing performance gaps across model sizes, dataset challenges, and code‑specialized variants while highlighting cost‑effective deployment strategies.
AIModel EfficiencySkills framework
0 likes · 8 min read


SuanNiFeb 26, 2026 · Artificial Intelligence
BitDance’s new multimodal AI model achieves an 8.7‑fold inference acceleration using only 2.6 billion parameters, surpasses 14‑billion‑parameter state‑of‑the‑art architectures in image generation quality, and introduces binary visual tokens, a binary diffusion head, and next‑block diffusion for efficient parallel autoregressive prediction.
AIBinary TokenizationModel Efficiency
0 likes · 11 min read

The article presents STEM, a method that transforms dense and MoE transformer architectures by converting the expert routing step into a static table‑lookup operation, achieving higher parameter efficiency, lower communication overhead, and improved interpretability while maintaining or boosting downstream task performance.
Embedding LookupInterpretabilityModel Efficiency
0 likes · 6 min read

An analysis of the 5,984 papers accepted at NeurIPS 2025 shows a decisive move from ever‑larger models toward agents, reasoning‑focused LLMs, efficiency engineering, AI for Science, and trustworthy AI, signaling the transition from a research‑toy era to an engineering‑driven AI ecosystem.
AI TrendsAI for ScienceAgents
0 likes · 7 min read


AntTechNov 8, 2025 · Artificial Intelligence
Ant Group’s President Luo Ji outlined how the AntBaiLing suite, featuring trillion‑parameter open‑source models, three efficiency breakthroughs, and a domestic compute cluster, is advancing AGI research and inclusive applications, especially in healthcare, while emphasizing ethical, trustworthy AI.
AGIModel Efficiencylarge language models
0 likes · 5 min read

Thinking Machines Lab introduces On-Policy Distillation, a post‑training technique that matches reinforcement‑learning performance while reducing compute cost by up to tenfold, and demonstrates its effectiveness through extensive experiments on reasoning, personalization, and catastrophic‑forgetting mitigation.
Model EfficiencyOn-Policy Distillationknowledge distillation
0 likes · 15 min read

The paper presents a systematic scaling‑law study of the linear‑time xLSTM architecture versus quadratic‑time Transformers, evaluating parameter‑data loss surfaces, optimal model size under equal FLOP budgets, and inference latency components, and shows that xLSTM consistently offers better cost‑effectiveness across diverse contexts and budgets.
Inference OptimizationLinear Time ComplexityModel Efficiency
0 likes · 11 min read

This article analyzes the rStar2‑Agent paper, revealing how Agentic Reinforcement Learning, the GRPO‑RoC algorithm, a high‑throughput code‑execution service, and a three‑stage training recipe let a modest 14‑billion‑parameter model surpass much larger LLMs on challenging math benchmarks.
AI researchLLMModel Efficiency
0 likes · 18 min read

The article introduces KAT‑V1 AutoThink, a dual‑mode large language model that automatically switches between thinking and non‑thinking modes based on problem difficulty, details its novel training paradigm, reinforcement‑learning enhancements, performance benchmarks against leading open‑source models, and provides open‑source resources for further research.
Model Efficiencyauto-thinkknowledge distillation
0 likes · 14 min read


AIWalkerJul 15, 2025 · Artificial Intelligence
This article presents Dynamic Vision Mamba (DyVM), a method that tackles token and block redundancy in Mamba‑based visual models through a novel re‑ordering pruning strategy and dynamic block selection, achieving a 35.2% FLOPs reduction with only a 1.7% accuracy loss while demonstrating strong generalization across tasks and architectures.
Computer VisionDynamic Block SelectionFLOPs Reduction
0 likes · 22 min read


AIWalkerApr 16, 2025 · Artificial Intelligence
This article reviews recent multi‑scale attention breakthroughs—including EMA, MSDA, VWA, and related modules—showing how they improve accuracy, cut FLOPs by up to 70%, and can be inserted into existing models with minimal effort, backed by code and paper links.
Computer VisionModel EfficiencyPlug and Play
0 likes · 10 min read


AIWalkerMar 14, 2025 · Artificial Intelligence
Researchers He Kaiming, Yann LeCun and colleagues propose a 9‑line Dynamic Tanh (DyT) layer that replaces LayerNorm/RMSNorm in Transformers, showing comparable or superior accuracy across vision, language, speech and DNA tasks while also reducing inference latency on modern GPUs.
AI researchDynamic TanhModel Efficiency
0 likes · 18 min read

DeepSeek R1, an open‑source large language model, leverages rule‑based, large‑scale reinforcement learning and mixed supervised‑fine‑tuning data to achieve deep reasoning comparable to OpenAI o1, illustrating China’s rapid AI progress, the importance of efficiency, and the democratizing impact of open AI research.
DeepSeekModel Efficiencyopen-source AI
0 likes · 11 min read

This issue reviews China's first open‑source short‑film model SkyReels‑V1, DeepSeek's Native Sparse Attention breakthrough, xAI's massive Grok‑3 deployment on 200k H100 GPUs, and a knowledge‑graph‑guided RAG framework, highlighting their performance gains, architectural innovations, and industry impact.
AIModel EfficiencyRAG
0 likes · 15 min read


ArchitectFeb 19, 2025 · Artificial Intelligence
The article critically examines whether the pre‑training Scaling Law still applies to Grok 3, compares its compute usage and model size with DeepSeek and OpenAI models, evaluates the cost‑effectiveness of pre‑training, RL and test‑time scaling, and explores how these insights shape future large‑language‑model development strategies.
Grok 3Model EfficiencyRL scaling
0 likes · 11 min read

This article derives DeepSeek V3's training Model FLOPs Utilization (MFU) using publicly available data, showing an MFU of roughly 37%—about a 60% improvement over V2—and provides detailed formulas, parameter settings, and a reproducible Python script.
AI performanceDeepSeekMFU
0 likes · 8 min read

In a keynote at the 2024 Beijing Zhiyuan Conference, Baidu’s CTO Wang Haifeng explained how large‑model universality and comprehensive capabilities are driving artificial general intelligence forward, highlighting scale laws, multimodal advances, agent technologies, and the industrial‑scale production of AI.
AI TrendsAI industrializationGeneral AI
0 likes · 7 min read

The article discusses how large AI models are moving toward a unified architecture that reduces task‑algorithm coupling, outlines the multi‑layer efficiency challenges—from model sparsity and quantization to software and infrastructure optimization—and highlights recent NVIDIA GTC 2024 and China AI Day events with registration details.
China AI DayModel EfficiencyNVIDIA GTC
0 likes · 12 min read

Adaptive Parameter Generation Network (APG) dynamically creates sample‑specific model parameters for click‑through‑rate prediction using low‑rank factorization, parameter sharing, and over‑parameterization, achieving up to 0.2% AUC improvement, 3% CTR lift, and up to 96.6% storage reduction with faster inference.
CTR predictionModel Efficiencyadaptive parameter generation
0 likes · 14 min read

This article explains how the Fine‑ and Coarse‑Granularity Hybrid Self‑Attention (FCA) mechanism reduces BERT’s computational cost by over 50% while keeping accuracy loss under 1%, detailing the method, experimental results, and its significance for efficient large‑scale language models.
BERTFCAModel Efficiency
0 likes · 8 min read
