Research

Everything in this section in one place — newest stories first.

Latesttools

Kimi's Agent Swarm Scales to 300 Sub-Agents, and Moonshot Published the Weights

Moonshot AI's Kimi K2.6 can dispatch up to 300 sub-agents across 4,000 coordinated steps, up from 100 sub-agents and 1,500 steps in K2.5.

LLMsJuly 28, 2026

Why AI Lies to You: It's the Mechanism, Not the Intent, and How to Limit Its Impact

Hallucination is not deliberate deception but an emergent property of model training, which rewards confidence over accuracy or admitting uncertainty.

Editorial image for لماذا يكذب عليك الذكاء الاصطناعي؟ الآلية لا النية، وكيف تحدّ من أثرها

Large Language ModelsJuly 28, 2026

Falcon vs. ALLaM: The Fierce Battle Over Arabic AI in the Gulf Skies

Falcon-H1 Arabic, launched by the UAE's Technology Innovation Institute on January 5, 2026 with a hybrid Mamba-Transformer architecture, scores 75.36% on the OALL benchmark at 34 billion parameters, surpassing according to the institute much larger global models such as Qwen2.5 72B and Llama-3.3 70B.

Editorial image for فالكون أم علّام؟ معركة شرسة في سماء الخليج لتطويع الذكاء الاصطناعي بالعربية

AI ArchitecturesJuly 28, 2026

Why Are Experts Declaring RAG Dead in the Face of the Agentic AI Wave?

The debate is not about the death of RAG as a concept but about the obsolescence of its naive version, a linear pipeline of retrieve-then-stuff-into-context-then-answer without verification.

Editorial image for لماذا يُعلن الخبراء وفاة تقنية RAG أمام طوفان الذكاء الاصطناعي الوكيل؟

AI EconomicsJuly 28, 2026

For CFOs and CTOs: How Tokens Quietly Eat Your Company's Budget

Language model spend is billed by consumption rather than by seat, so it grows with every feature release and never passes through the usual procurement cycle.

شاشة رقمية تعرض عدّاد توكنز متدفق مع ميزان يوازن بين تسعير النماذج وميزانية الشركة

EthicsJuly 26, 2026

Universities Are Quietly Ditching AI Plagiarism Detectors as False Positives Pile Up

Reddit testers found AI detectors flagged the 1776 Declaration of Independence as up to 100% AI-generated, underscoring the tools' unreliability

Developer ToolsJuly 26, 2026

AI Coding Assistants Are Making Code Look the Same, But Not Think the Same

A study of Kaggle contest submissions from 2019 to mid-2026 finds AI coding assistants are driving strong syntactic homogenization, making code more structurally similar across submissions.

AI EthicsJuly 26, 2026

New XAI Framework Brings Auditor-Friendly Explanations to Banking Anomaly Detection

New arXiv paper (2607.13469) presents an explainable AI framework for banking transaction anomaly detection combining Isolation Forest with feature-level explanations

HardwareJuly 23, 2026

Google Is Reportedly Baking Gemini's Architecture Directly Into Silicon With 'Frozen v2'

Google is reportedly developing 'Frozen v2,' an internal server chip that embeds Gemini's model architecture directly into silicon rather than model weights.

AI AgentsJuly 22, 2026

A Water-Filling Rule for Multi-Agent AI: New Paper Tightens the Math Behind Learned Communication Graphs

HIBCG (Heterogeneous Information-Bottleneck Coordination Graphs) is a new method for learning sparse communication graphs in cooperative multi-agent reinforcement learning, posted to arXiv on May 17, 2026 and revised July 10, 2026.

ResearchJuly 22, 2026

New Multi-Agent Framework ARCANA Tackles ARC-AGI-2 with Symbolic Reasoning and a Shared Blackboard

ARCANA is a new multi-agent framework published July 10, 2026, built to solve ARC-AGI-2 tasks under strict time and hardware limits

LLMsJuly 22, 2026

New Framework Tells Researchers Exactly Where to Spend Their Human Survey Budget When LLMs Do the Rest

A new arXiv paper introduces 'rectification difficulty,' a metric quantifying how hard it is to correct LLM errors on specific survey questions using human labels.

LLMsJuly 22, 2026

New Benchmark Exposes a Blind Spot in AI Doctors: They Still Can't Read the Room (or the X-Ray)

MedRealMM is a new benchmark built from 5,620 real multimodal patient-doctor cases across 64 clinical departments at a Chinese internet hospital

ResearchJuly 22, 2026

ALICE: A Single Foundation Model Absorbs Eight Pathology AI Systems Into One Backbone

ALICE is a pathology foundation model built by distilling eight teacher models (vision-only, vision-language, and slide-level) into one unified backbone via multi-stage agglomerative distillation

ResearchJuly 22, 2026

A Unified Geometry for Batch Normalization: LieBN Extends Deep Learning to Nine Curved Data Spaces

LieBN generalizes batch normalization to nine Lie group geometries, including SPD manifolds, rotation matrices, and full-rank correlation matrices.

ResearchJuly 22, 2026

No Retraining Required: TCLA Sharpens Medical Vision-Language Models on the Fly

TCLA is a training-free few-shot adaptation method for Medical Vision-Language Models, detailed in an arXiv paper submitted July 10, 2026

Computer VisionJuly 22, 2026

CtrlVTON Turns Virtual Try-On Into a Precision Editing Task, Not a Guessing Game

CtrlVTON reframes virtual try-on as a controllable image-editing task, using segmentation masks for pixel-level control over garment size, style, and placement.

ResearchJuly 22, 2026

When Medical AI Can't See Who It's Failing, CAPRA Fills the Blind Spot

CAPRA is a new calibrated proxy-axis framework for detecting hidden subgroup performance gaps in medical imaging AI when demographic metadata is missing.

All articles