2026

10,670 articles · 135 days · 10 categories · avg 79 per day

GLEAM: Spectral Losses for Turbulence Forecasting on Unstructured Meshes
Research·arXiv cs.LG

GLEAM: Spectral Losses for Turbulence Forecasting on Unstructured Meshes

Jul 25, 2026·2 min
LAARA vs. LoRA: Layer-Wise Adaptive Rank for Efficient Fine-Tuning
Creative

LAARA vs. LoRA: Layer-Wise Adaptive Rank for Efficient Fine-Tuning

Jul 25, 2026·2 min
SAE Interpretability Evaluation: The Measurement Pipeline Matters More Than the Model Architecture
LLM

SAE Interpretability Evaluation: The Measurement Pipeline Matters More Than the Model Architecture

Jul 25, 2026·2 min
FNO and UNO Neural Operators for Calculating Two-Dimensional Neutron Flux
Research

FNO and UNO Neural Operators for Calculating Two-Dimensional Neutron Flux

Jul 25, 2026·2 min
SUM: A Server-Side Fix for Forgetting in Federated Continual Learning, Up to 22% Gain
Research·arXiv cs.LG

SUM: A Server-Side Fix for Forgetting in Federated Continual Learning, Up to 22% Gain

SUM corrects "adaptation vectors" on the server and delivers up to 22% gains over earlier federated continual learning methods, with no extra load on clients.

Jul 25, 2026·2 min
STN-TGAT: A Graph Neural Network for Ranking S&P 500 Stocks and Building Portfolios
Research·arXiv cs.LG

STN-TGAT: A Graph Neural Network for Ranking S&P 500 Stocks and Building Portfolios

Researchers have introduced STN-TGAT, a model that combines a Transformer with a graph attention network to pick 5 stocks from the S&P 500 top 50 and build a…

Jul 25, 2026·2 min
How to Explain Continual Learning in Neural Networks: Attention Rollout and Grad-CAM on Time Series
Research·arXiv cs.LG

How to Explain Continual Learning in Neural Networks: Attention Rollout and Grad-CAM on Time Series

Researchers used attention rollout and Grad-CAM to explain how neural networks adapt on nonstationary time series, using groundwater monitoring data as a test case.

Jul 25, 2026·2 min
Air Quality Arena: A Benchmark for Forecasting Air Pollution with Foundation Models
Medicine·arXiv cs.LG

Air Quality Arena: A Benchmark for Forecasting Air Pollution with Foundation Models

Researchers have released Air Quality Arena, an open dataset of 14,000+ series covering 6 pollutants in 7 countries, plus a benchmark of 11 time-series foundation…

Jul 25, 2026·2 min
Top themes
01CruiseBench: A New Benchmark for Evaluating Neural Networks on Aircraft Engine Lifetime PredictionJul 25, 2026·2 min02HyenaND: A Subquadratic Operator That Replaces Attention on Multidimensional DataJul 25, 2026·2 min03ECoM Reasoning: Compressed Reasoning Speeds Up Voice Models on Spoken MathJul 25, 2026·2 min04Can Transformers Select a Model? An arXiv Bayesian Test at 2.8M ParametersJul 25, 2026·2 min05emb-diversity: An Embedding-Based Tool for Measuring Data DiversityJul 25, 2026·2 min
FinMMEval 2026 Task 2: A Five-Language Financial QA Benchmark, With the Top 4 Systems Within 1%
LLM·arXiv cs.CL

FinMMEval 2026 Task 2: A Five-Language Financial QA Benchmark, With the Top 4 Systems Within 1%

Jul 25, 2026·2 min
D2VBench: A 10,000-Dilemma Benchmark Tests LLM Value Alignment
LLM·arXiv cs.CL

D2VBench: A 10,000-Dilemma Benchmark Tests LLM Value Alignment

Jul 25, 2026·2 min
Sentence Splitter: How a T5 Model Finds the Hidden Factual Structure of Sentences
Research·arXiv cs.CL

Sentence Splitter: How a T5 Model Finds the Hidden Factual Structure of Sentences

Jul 25, 2026·2 min
VizRAG: Visualizing Hypergraphs to Improve RAG Retrieval for Multimodal Models
LLM·arXiv cs.CL

VizRAG: Visualizing Hypergraphs to Improve RAG Retrieval for Multimodal Models

Jul 25, 2026·2 min
TriAgent: A Committee of Small Models Instead of an Expensive LLM Saves $9.3M a Year
LLM·arXiv cs.CL

TriAgent: A Committee of Small Models Instead of an Expensive LLM Saves $9.3M a Year

Researchers unveiled TriAgent, a multi-agent financial sentiment system where cheap models replace a cloud LLM, saving $9.3M a year at 10M users.

Jul 25, 2026·2 min
Rubric4Setwise: Evaluating Whole Document Sets for RAG Instead of Ranking by Relevance
Creative·arXiv cs.CL

Rubric4Setwise: Evaluating Whole Document Sets for RAG Instead of Ranking by Relevance

Researchers introduce SetwiseEvalKit, a benchmark of ~28,000 criteria for judging the quality of an entire document set, and Rubric4Setwise, a method that generates…

Jul 25, 2026·2 min
Neural Network Pinpoints a Speaker's Origin from Their Arabic Dialect to Within 481 km
Research·arXiv cs.CL

Neural Network Pinpoints a Speaker's Origin from Their Arabic Dialect to Within 481 km

A new neural network predicts where an Arabic speaker comes from as latitude and longitude, with a median error of 481.2 km, without sorting speakers into countries.

Jul 25, 2026·2 min
HIPE-2026: How to Extract Person-Place Relations from Historical Newspapers Without Neural Networks
LLM·arXiv cs.CL

HIPE-2026: How to Extract Person-Place Relations from Historical Newspapers Without Neural Networks

Jul 25, 2026·2 min
SLPO: A Method That Brings Verifiable-Reward RL to Latent Reasoning Models
LLM·arXiv cs.CL

SLPO: A Method That Brings Verifiable-Reward RL to Latent Reasoning Models

Jul 25, 2026·2 min
MultiMDM: A Diffusion Language Model That Generates Text in a Few Steps
Creative·arXiv cs.CL

MultiMDM: A Diffusion Language Model That Generates Text in a Few Steps

Jul 25, 2026·2 min
NLI Hypergraph vs. LLM Judge: A New Way to Audit AI Reasoning Without a Reference Answer
Medicine·arXiv cs.CL

NLI Hypergraph vs. LLM Judge: A New Way to Audit AI Reasoning Without a Reference Answer

Jul 25, 2026·2 min
Adaptive Capitulation in LLMs: How Chatbots Give In to Vulnerable Users (Study)
Creative·arXiv cs.CL

Adaptive Capitulation in LLMs: How Chatbots Give In to Vulnerable Users (Study)

Jul 25, 2026·2 min
Small Qwen Language Models Ignore Instructions but Look Accurate on Benchmarks
Creative

Small Qwen Language Models Ignore Instructions but Look Accurate on Benchmarks

Jul 25, 2026·2 min
Why Pure Transformers Can't Learn Structural Generalization: A Proof
Autonomous

Why Pure Transformers Can't Learn Structural Generalization: A Proof

Jul 25, 2026·2 min
Hypernetworks Instead of LoRA: Scaling Laws for Knowledge Injection into LLMs
Creative

Hypernetworks Instead of LoRA: Scaling Laws for Knowledge Injection into LLMs

Jul 25, 2026·2 min
SFT Fine-Tuning Narrows AI Behavior: arXiv Study on Diversity Collapse
LLM·arXiv cs.CL

SFT Fine-Tuning Narrows AI Behavior: arXiv Study on Diversity Collapse

Standard SFT fine-tuning makes language models more accurate at games but quietly strips away the variety of their moves. A new arXiv study explains why and…

Jul 25, 2026·2 min
CRA-Bench and CRA-Net: Catching Harm That Builds Up Over the Course of an LLM Conversation
LLM·arXiv cs.CL

CRA-Bench and CRA-Net: Catching Harm That Builds Up Over the Course of an LLM Conversation

A new framework, CRA, and a benchmark, CRA-Bench, aim to catch harm that builds up gradually from harmless messages in an LLM conversation rather than in a single…

Jul 25, 2026·2 min
Spectral-LSH: Up to 16× Prompt Compression Without Fine-Tuning to Speed Up LLM Inference
LLM·arXiv cs.AI

Spectral-LSH: Up to 16× Prompt Compression Without Fine-Tuning to Speed Up LLM Inference

The new Spectral-LSH method compresses long prompts 8–16× with no model fine-tuning: it hashes similar tokens into groups in attention's spectral space and…

Jul 25, 2026·2 min
AI Tutor "Scaffolding Collapse": New Method Keeps Socratic Dialogue Intact on Qwen3-8B
LLM·arXiv cs.AI

AI Tutor "Scaffolding Collapse": New Method Keeps Socratic Dialogue Intact on Qwen3-8B

Scaffold-Preserving Representation Alignment targets "scaffolding collapse" in AI tutors: on Qwen3-8B, the collapse rate fell to 32% and dialogues hold for more…

Jul 25, 2026·2 min
Top themes
01AI Data Centers and Heat Overload US Power Grids: PJM Capacity Costs Jump 60%Jul 25, 2026·2 min02China vs. the US: How Each Protects Children From AI Companions, and Why Beijing's Approach Is StricterJul 25, 2026·2 min03Why a Successful AI Pilot Fails to Pay Off at Scale: The Topic of INFOSTART CIO CAMP 2026Jul 25, 2026·2 min04Timeweb Cloud: one in three documentation requests comes from AI agents, not humansJul 25, 2026·2 min05Vibe Coding All the Way to Production: A Google Play App Where AI Wrote 99.9% of the CodeJul 25, 2026·2 min
Mobileye: Founder Amnon Shashua to Step Down as CEO as Board Searches for Successor
Business·36Kr (36氪)

Mobileye: Founder Amnon Shashua to Step Down as CEO as Board Searches for Successor

Jul 25, 2026·2 min
GridAI: Drones from a Chinese Startup Fly Under the Forest Canopy Without GNSS or Heavy Compute
Autonomous·36Kr (36氪)

GridAI: Drones from a Chinese Startup Fly Under the Forest Canopy Without GNSS or Heavy Compute

Jul 25, 2026·3 min
Chinese Tech Companies Ride the Global AI Boom: Chip Exports Jump 96% to $177 Billion
Hardware·SCMP Tech

Chinese Tech Companies Ride the Global AI Boom: Chip Exports Jump 96% to $177 Billion

Jul 25, 2026·2 min
Google starts training Gemini 4 as Alphabet raises capex forecast to $205 billion
LLM·The Decoder

Google starts training Gemini 4 as Alphabet raises capex forecast to $205 billion

Jul 25, 2026·2 min
DeepSeek Raised Its Registered Capital 65% in 2026, and Its Founder Says AGI Is the Only Goal
LLM·36Kr (36氪)

DeepSeek Raised Its Registered Capital 65% in 2026, and Its Founder Says AGI Is the Only Goal

DeepSeek's registered capital rose 65% in 2026 after two capital increases, and founder Liang Wenfeng told investors in a closed-door meeting that the path to AGI…

Jul 25, 2026·2 min
ProGraph: A Two-Layer Memory for LLM Agents Beats Mem0, RAG and HippoRAG
LLM·arXiv cs.AI

ProGraph: A Two-Layer Memory for LLM Agents Beats Mem0, RAG and HippoRAG

ProGraph, a two-layer memory for LLM agents, scored 80.1% on the new multi-hop MemHop benchmark and beat Mem0, A-Mem, HippoRAG and RAG without building an explicit…

Jul 25, 2026·2 min
LISA: Sparse Attention Speeds Up Long LLM Reasoning by 50% Without Retraining
Creative·arXiv cs.AI

LISA: Sparse Attention Speeds Up Long LLM Reasoning by 50% Without Retraining

The LISA method replaces the standard attention mechanism in language models and speeds up long reasoning by 50% at a 16K-token context, with no training from scratch.

Jul 25, 2026·2 min
NEXUS: Runtime Safety Control for Tool-Using LLM Agents — F1 0.949
LLM·arXiv cs.AI

NEXUS: Runtime Safety Control for Tool-Using LLM Agents — F1 0.949

Jul 25, 2026·2 min
Learn2Discern Benchmark: LLMs Barely Tell Reliable Sources from Popular Ones
LLM·arXiv cs.AI

Learn2Discern Benchmark: LLMs Barely Tell Reliable Sources from Popular Ones

Jul 25, 2026·2 min
FormulaSPIN: A Neural Network That Writes Spreadsheet Formulas From Descriptions and Beats Proprietary Models
Creative·arXiv cs.AI

FormulaSPIN: A Neural Network That Writes Spreadsheet Formulas From Descriptions and Beats Proprietary Models

Jul 25, 2026·3 min
Confidential Inference on NVIDIA H100 with Intel TDX: Privacy Costs 17–21% of Speed
Hardware·arXiv cs.AI

Confidential Inference on NVIDIA H100 with Intel TDX: Privacy Costs 17–21% of Speed

Jul 25, 2026·2 min
OpenEvoShield: Defending Multi-Agent LLM Systems Against Evolving Injection Attacks
LLM·arXiv cs.AI

OpenEvoShield: Defending Multi-Agent LLM Systems Against Evolving Injection Attacks

Jul 25, 2026·2 min
It's Not the Model, It's the Architecture: Breaking Down 3 High-Profile AI Failures (Deloitte, AI Cameras, an Agent)
LLM

It's Not the Model, It's the Architecture: Breaking Down 3 High-Profile AI Failures (Deloitte, AI Cameras, an Agent)

Jul 25, 2026·2 min
FineServe: A Dataset of Real-World LLM Serving Workloads From a Global Marketplace
LLM

FineServe: A Dataset of Real-World LLM Serving Workloads From a Global Marketplace

Jul 25, 2026·2 min
STMicroelectronics Forecasts Revenue Growth as AI Data Centers Drive Chip Demand
Hardware

STMicroelectronics Forecasts Revenue Growth as AI Data Centers Drive Chip Demand

Jul 25, 2026·2 min
Elon Musk: Tesla's AI6 Chip Will Be the World's Best for Edge AI Computing
Robotics·3DNews AI

Elon Musk: Tesla's AI6 Chip Will Be the World's Best for Edge AI Computing

Elon Musk says Tesla's upcoming AI6 chip will be the world's best for edge AI computing, serving as the foundation for robotaxis, Optimus robots and Autopilot.

Jul 25, 2026·2 min
Apple is preparing an AI shopping assistant: a section about it turned up in the Apple Store privacy policy
Creative·3DNews AI

Apple is preparing an AI shopping assistant: a section about it turned up in the Apple Store privacy policy

Apple added a section on an unreleased AI shopping assistant to the Apple Store app's privacy policy, the first official mention of the assistant.

Jul 25, 2026·2 min
Tencent's Marvis: System-Level AI Agent for PCs Hits 300,000 DAU in Two Days
LLM·36Kr (36氪)

Tencent's Marvis: System-Level AI Agent for PCs Hits 300,000 DAU in Two Days

Tencent unveiled Marvis, an AI agent that controls a PC's files, hardware and browser at the system level rather than acting as a general-purpose chatbot. It drew…

Jul 25, 2026·3 min
Google Isn't Falling Behind in the AI Race, Pichai Tells Skeptics, Pointing to 82% Cloud Growth
LLM·TNW

Google Isn't Falling Behind in the AI Race, Pichai Tells Skeptics, Pointing to 82% Cloud Growth

On the July 22, 2026 earnings call, Sundar Pichai rejected the claim that Google is lagging in AI and cited 82% cloud growth as his main counterargument.

Jul 25, 2026·2 min
Poolside's Laguna S: a 118B MoE That Beats Thinky's Trillion-Parameter Open-Weights Model
LLM·Latent Space

Poolside's Laguna S: a 118B MoE That Beats Thinky's Trillion-Parameter Open-Weights Model

Jul 25, 2026·2 min
Laguna S 2.1 vs DeepSeek V4: neolab promises cheaper than Flash and stronger than Pro
LLM·Latent Space

Laguna S 2.1 vs DeepSeek V4: neolab promises cheaper than Flash and stronger than Pro

Jul 25, 2026·2 min
Nunchaku Lite: 4-bit Image Generation in Diffusers — Up to 50% Less VRAM
Creative·Hugging Face Blog

Nunchaku Lite: 4-bit Image Generation in Diffusers — Up to 50% Less VRAM

Jul 25, 2026·2 min
Claude Sonnet 4.5 Knows When It's Being Tested: Why a Green Safety Benchmark Doesn't Mean Green in Production
LLM·Habr AI

Claude Sonnet 4.5 Knows When It's Being Tested: Why a Green Safety Benchmark Doesn't Mean Green in Production

Jul 25, 2026·2 min