Skip to content
gentic.news — AI News Intelligence Platform
Connecting to the Living Graph…
gentic.news

AI Research

Breaking AI research news: latest papers from arXiv, NeurIPS, ICML, and top labs. Track transformer architecture advances, reasoning breakthroughs, and scientific discoveries in machine learning and artificial intelligence.

AI ResearchFunding & BusinessProducts & LaunchesBig TechStartupsOpen SourcePolicy & EthicsOpinion & Analysis
A bar chart comparing human and AI performance on the ActiveVision benchmark shows a tall blue bar at 96.1% for…
AI Research
85

ActiveVision Benchmark: Humans 96.1%, Best AI 10.6%

ActiveVision benchmark: humans 96.1%, best AI 10.6%. The 85.5-point gap reveals fundamental limits in iterative visual reasoning for current models.

x.com/Jul 23, 2026/3 min read
reasoningbenchmarksevaluation
Paul Iusztin's course diagram showing coding agent harness architecture with LangChain and terminal connections…
AI Research
89

Open-Source Course Shows Harness, Not Model, Lifts Coding Agent 25 Places

Open-source course shows harness engineering, not model swap, moved a coding agent from ~30th to top 5 on Terminal-Bench. Course builds Decode from scratch.

x.com/Jul 23, 2026/3 min read
open-sourcecoding agentsai engineering
AI system on a computer screen displays complex mathematical equations and a highlighted counterexample, with a…
AI Research
85

AI Disproves 87-Year-Old Conjecture, Finds Counterexample Humans Missed

AI disproves 87-year-old math conjecture, finding a counterexample humans missed, per @rohanpaul_ai.

x.com/Jul 23, 2026/3 min read
discoverymathematicsai
An open laptop screen displays a web browser interface with a glowing AI agent icon hovering over a task completion…
AI Research
90

Microsoft Fara1.5-27B Open-Source Agent Scores 72.3% on Web Tasks

Microsoft released Fara1.5-27B, a vision-only web browsing agent scoring 72.3% on Online-Mind2Web, open-source on Hugging Face.

x.com/Jul 22, 2026/3 min read
web-agentsopen-sourcecomputer-use
A first-person view inside a vibrant generative world, with a floating text prompt bar at the bottom, showing a lush…
AI Research
85

AlayaRenderer-Flash Hits 31.54 FPS, Enables Playable Generative Worlds

AlayaRenderer-Flash accelerates real-time scene synthesis from 0.56 to 31.54 FPS, a 56x speedup enabling playable generative worlds.

x.com/Jul 22, 2026/3 min read
real-time renderingworld modelsai
Security researcher at a computer terminal displaying a breached network diagram with red alerts, indicating an AI…
AI Research
88

AI Breached Real Production Systems, Not Just Sandboxes

AI breached real production systems, not just test environments. First documented case of operational security breach.

x.com/Jul 21, 2026/3 min read
ai securityai agentscybersecurity
A colorful bar chart showing token savings, with a 39% highlighted bar labeled SWE-Pruner Pro, next to a diagram of…
AI Research
87

SWE-Pruner Pro Saves 39% Tokens by Reading LLM Hidden States

SWE-Pruner Pro saves up to 39% tokens on coder LLMs by reading keep-or-prune signals from hidden states, maintaining task quality without external heuristics.

x.com/Jul 21, 2026/3 min read
efficiencyllm inferenceai research
A desktop screenshot showing a macOS-style interface generated by AI, with a dock, menu bar, window controls, and…
AI Research
92

Qwen 3.8 Max Generates macOS Clone in Single HTML Pass

Qwen 3.8 Max generated a full macOS clone in HTML/CSS/JS in a single pass, demonstrating unusually high UI fidelity for a code generation model.

x.com/Jul 20, 2026/3 min read/Multi-Source
code generationui developmentai
A system architecture diagram showing KV cache blocks being selectively reused across multiple document queries…
AI Research
92

CacheBlend: 2-4x Faster KV Cache for Multi-Doc Queries

CacheBlend reuses per-document KV caches by recomputing only boundary tokens, achieving 2-4x speedups on multi-document queries. Alibaba data shows 10% of blocks serve 77% of hits.

x.com/Jul 20, 2026/3 min read
cachingai infrastructureinference optimization
Google engineer holding a server rack with a glowing blue chip labeled 'Frozen v2' in a bright data center, rows of…
AI Research
97

Google’s Frozen v2 chip: 6–10× tokens/W for Gemini, 2028 target

Google is developing Frozen v2, a chip freezing Gemini architecture into silicon for 6–10× tokens/W, deployment as early as 2028, driven by compute shortage.

x.com/Jul 20, 2026/3 min read
geminiinference optimizationai hardware
A glowing digital brain icon on a dark background surrounded by branching lines and nodes, suggesting AI neural…
AI Research
85

239-Paper Survey Maps How AI Agents Self-Improve via Scaffold Updates

A survey of 239 papers shows 68% of AI agent self-improvement methods focus on scaffold updates rather than model retraining, raising evaluation quality concerns.

x.com/Jul 19, 2026/3 min read
agentic systemsmachine learningai research
Stanford researchers survey over 100 papers on LLMs' metacognition, covering confidence calibration and…
AI Research
75

100+ Papers Surveyed: LLMs' Metacognition Gap

A systematic survey of 100+ papers reveals gaps in LLM metacognition, including 10-30% miscalibration in top models like GPT-4 and Claude 3.

x.com/Jul 19, 2026/3 min read
ai safetyllmsbenchmarks
Engineer monitors a robot arm on a workbench while a screen displays GigaWorld-Policy-0.5 inference metrics on an…
AI Research
85

GigaWorld-Policy-0.5 Hits 85ms on RTX 4090 for Robot Control

GigaWorld-Policy-0.5 runs robot control at 85ms on an RTX 4090, using a Mixture-of-Transformers architecture for real-time local deployment.

x.com/Jul 19, 2026/3 min read
real-time controledge aitransformer architecture
A gaming monitor displays a detailed fantasy scene from Black Myth: Wukong, showing a character with a staff in a…
AI Research
78

90 Hours of Black Myth: Wukong Fuel New World Model Benchmark

A new survey and benchmark rethinks interactive world models as game engines, with a data engine collecting over 90 hours of Black Myth: Wukong gameplay.

x.com/Jul 19, 2026/3 min read
world modelsresearchbenchmark
Diagram comparing low, medium, and high reasoning effort modes in LLMs, with arrows showing dynamic switching at…
AI Research
75

LLMs Learn to Switch Reasoning Effort at Inference Time

@rasbt explains how LLMs switch reasoning effort using inference-time methods and training, potentially cutting token usage by 30–50% on simple queries.

x.com/Jul 18, 2026/3 min read
reasoningllminference
A split screen shows four video frames with a runner and a person on a scooter, illustrating keyframe generation quality
AI Research
80

KeyFrame-Compass Benchmark Targets Keyframe Video Generation Gaps

KeyFrame-Compass is the first benchmark for keyframe-conditioned video generation, with 386 samples and six metrics.

x.com/Jul 18, 2026/3 min read
benchmarksvideo generationevaluation
A hierarchical knowledge graph with interconnected nodes and color-coded clusters, illustrating graph-traversal…
AI ResearchBreakthrough
82

HG-RAG Beats Flat Retrieval on Graph Queries Across 800-Node Worlds

HG-RAG uses graph-traversal over knowledge graphs for RAG, beating flat retrieval on hierarchical and multi-hop queries across worlds up to 800 nodes.

arxiv.org/Jul 18, 2026/3 min read
knowledge graphsarxivrag
Diagram showing LongStraw's branch replay method on 8 H20 GPUs reaching 2.1M token positions for RL post-training
AI Research
87

LongStraw Reaches 2.1M Tokens on 8 H20 GPUs via Branch Replay

LongStraw reaches 2.1M token positions for RL post-training on 8 H20 GPUs by replaying short response branches, cutting compute 8-16x vs prior art.

x.com/Jul 17, 2026/3 min read
efficient traininglong contextreinforcement learning
A futuristic tech interface shows a glowing AI chip labeled 'Kimi K3' with data streams and numbers like 2.8T and…
AI Research
100

Moonshot AI's Kimi K3: 2.8T params, 1M token window, $3/M input

Moonshot AI released Kimi K3, a 2.8T-parameter mixture-of-experts model with 1M token context window and $3/M input pricing, claiming autonomous chip design and research capabilities.

x.com/Jul 16, 2026/3 min read/Widely Reported
long contextmoonshot aiai models
NVIDIA's #1 RTEB Embedding Model Skips Token Generation Entirely
AI Research
83

NVIDIA's #1 RTEB Embedding Model Skips Token Generation Entirely

NVIDIA's RTEB hit #1 on MTEB by skipping token generation. The cheapest reasoning token is the unused one.

x.com/Jul 16, 2026/3 min read/Multi-Source
ai-modelsbenchmarksnvidia
A large neural network model diagram with 1T parameters, showing reinforcement learning pipeline steps and emergent…
AI Research
99

Ring-Zero Trains 1T-Parameter Model via Reinforcement Learning

Ring-Zero scales RL with verifiable rewards to 1T parameters, revealing emergent reasoning like self-verification and context anxiety.

x.com/Jul 16, 2026/3 min read
reinforcement learninglarge language modelsai research
A bar chart comparing SWE-Bench scores for 7B to 14B models, showing a +5.4 improvement with function-aware FIM…
AI Research
85

Function-Aware Fill-in-the-Middle Boosts SWE-Bench by +5.4 on 14B Models

Function-aware FIM mid-training boosts SWE-Bench by +2.8 to +5.4 on 7B-14B models, preserving general abilities. Six checkpoints and 400K dataset open-sourced.

x.com/Jul 16, 2026/3 min read
coding agentsbenchmarksmodel training
A glowing red robotic hand reaches toward a blue digital brain, symbolizing AI security testing, with data streams…
AI ResearchBreakthrough
91

GPT-Red: OpenAI's LLM Super-Hacker Finds 84% of Attacks, Humans 13%

OpenAI's GPT-Red LLM finds 84% of attacks vs 13% for humans, hardening GPT-5.6 Sol. Automated red-teaming shifts safety paradigm.

technologyreview.com/Jul 15, 2026/3 min read
llm securityai safetyred-teaming
Researchers examine a synthetic document benchmark on a screen, analyzing long-context visual understanding and…
AI Research
85

ServiceNow's SynthDocBench Teases Apart VLM Long-Context Failure Modes

ServiceNow releases SynthDocBench, a controlled synthetic benchmark for long-context visual document understanding that varies length, layout, modality, and reasoning to diagnose VLM failures.

x.com/Jul 15, 2026/3 min read
ai benchmarksservicenowvision-language models
A person holds a smartphone displaying a vivid AI-generated landscape, with text overlay reading 'SpectraReward' and…
AI Research
85

ByteDance SpectraReward: Training-Free Reward Reads Prompt Back From Image

ByteDance Seed releases SpectraReward, a training-free reward that reads a prompt back from a generated image using prompt log-likelihood. No training or preference labels needed.

x.com/Jul 15, 2026/3 min read
ai alignmentbytedancetext-to-image
Two men in a modern office, one gesturing toward a whiteboard with AI agent diagrams, the other smiling while…
AI ResearchBreakthrough
100

Rich Sutton Launches Oak Lab to Build Self-Learning AI Agents

Rich Sutton founded Oak Lab to build self-learning AI agents. He rejects static datasets for real-time reinforcement learning with a trillion-parameter goal at 20W.

the-decoder.com/Jul 13, 2026/3 min read/Multi-Source
startupsreinforcement learningai agents
A laptop with 25GB RAM running a 744-billion-parameter AI model, no GPU, displaying inference output on a dark screen
AI Research
85

Colibri Runs 744B-Parameter Model on 25GB RAM, No GPU

Colibri claims to run a 744B-parameter model on 25GB RAM without GPU, but lacks evidence. If true, it could democratize large-model inference.

x.com/Jul 13, 2026/3 min read
quantizationmodel compressioninference
Professor Yuji Tachikawa at a whiteboard, with Claude Fable 5 displayed on a laptop, highlighting a calculation…
AI Research
100

Claude Fable 5 Solves String Theory Problem Stalled for Six Months

Claude Fable 5 solved a string theory problem stalled for six months. Professor Yuji Tachikawa says the model made a non-trivial observation and used SymPy for verification.

x.com/Jul 13, 2026/3 min read/Widely Reported
claudeanthropicphysics
GenCeption: Video Diffusion Backbone Beats Specialists on 5 Vision Tasks
AI ResearchBreakthrough
82

GenCeption: Video Diffusion Backbone Beats Specialists on 5 Vision Tasks

GenCeption uses video diffusion as a vision backbone, matching specialists with 7-500x less data and generalizing from synthetic to real footage.

arxiv.org/Jul 13, 2026/3 min read
foundation modelscomputer visionai research
HuggingFace researcher at a desk with a laptop displaying a video captioning interface, a diagram of parallel…
AI Research
85

PadCaptioner: 3B video caption model beats 7B rivals with parallel decoding

PadCaptioner, a 3B model, beats 7B rivals in dense video captioning via lossless parallel autoregressive decoding, challenging scaling orthodoxy.

x.com/Jul 12, 2026/3 min read
aivideomodels