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

NVIDIA Drops 30B Nemotron Audex Audio Model with MoE

AI Research
91

NVIDIA Drops 30B Nemotron Audex Audio Model with MoE

NVIDIA released Nemotron Audex 30B-A3B, a 30B-parameter MoE audio model unifying ASR, understanding, and TTS with 3B active parameters.

x.com/Jul 12, 2026/3 min read
audio aiopen-source modelsnvidia
Robot arm with gripper reaching toward a stack of colored blocks on a table, with a digital interface showing code…
AI Research
85

InternVLA-A1.5 Unifies Vision, Foresight, Action — SOTA on All Six Sim Benchmarks

InternVLA-A1.5 unifies vision-language understanding, latent foresight, and action into one robot policy, achieving SOTA on all six simulation benchmarks.

x.com/Jul 12, 2026/3 min read
simulationroboticsai
Researchers huddle around a whiteboard covered in equations and diagrams, discussing RL and LLM optimization strategies
AI Research
85

Hugging Face weekly papers: Monotonic inference policy overtakes training optimization

Hugging Face's top papers July 6-12 include a paper arguing monotonic inference policies are the true LLM RL objective, and Vidu S1 for real-time interactive video generation.

x.com/Jul 12, 2026/3 min read
world modelsreinforcement learningvideo generation
An illustration of a chalkboard in a classroom with ChatGPT and OpenAI logos drawn on it, symbolizing AI use in…
AI ResearchBreakthrough
81

Brown exam scores collapse 50 points when AI ban enforced

Brown professor's take-home exam averaged 96% but proctored final fell to 48.6%, with 18 of 86 students dropping the course, suggesting widespread AI cheating.

the-decoder.com/Jul 12, 2026/3 min read
assessment designchatgptai in education
DeepSeek DSpark framework diagram showing parallel token generation merging into adaptive verification for…
AI Research
90

DeepSeek DSpark: Speculative Decoding Unifies Parallel Gen, Adaptive Verification

DeepSeek released DSpark, a speculative decoding framework unifying parallel generation with adaptive verification. No benchmarks disclosed yet; the approach targets inference latency and throughput.

x.com/Jul 11, 2026/3 min read
open-sourceresearchinference
Matei Zaharia's team at Databricks tests coding agents on their polyglot codebase, with open-source GLM-5.2 matching…
AI ResearchBreakthrough
75

Databricks Tests Coding Agents on Its Own Codebase

Databricks benchmarked coding agents on its own polyglot codebase. GLM-5.2 matched top closed models, a minimal harness halved costs, and cheaper-per-token models cost more per task.

pub.towardsai.net/Jul 11, 2026/3 min read
open sourcecodingai agents
NVIDIA TwoTower: 2.4x Faster LLM Decoding, 98.7% Quality
AI Research
95

NVIDIA TwoTower: 2.4x Faster LLM Decoding, 98.7% Quality

NVIDIA TwoTower clones a pretrained LLM into a frozen context tower and trainable denoiser tower, achieving 2.42x faster generation with 98.7% quality on a 30B MoE model.

x.com/Jul 11, 2026/3 min read
open-sourcediffusion-modelsnvidia
A diagram of the BRAID framework showing a Markov decision process with text and image generation loops connected by…
AI Research
85

BRAID Fuses Text-Image Reasoning Into One RL Objective

BRAID unifies multi-turn text-image reasoning as a Markov decision process, enabling joint RL optimization of both modalities with a single objective.

x.com/Jul 11, 2026/3 min read
generative-aireinforcement-learningmultimodal
A robotic arm in a modern kitchen reaches toward a bunch of bananas on a countertop, surrounded by other fruit and…
AI ResearchBreakthrough
100

BAAI Orca World Model Matches π0.5 With No Action Labels

BAAI's Orca world model matches specialized π0.5 on five robotics tasks, trained on 125,000 hours of video without action labels, predicting abstract world states.

the-decoder.com/Jul 11, 2026/3 min read/Widely Reported
roboticsworld modelsai
NVIDIA researcher at computer screen showing NVFP4 quantized Kimi-K2.7-Code model on Hugging Face, with Blackwell…
AI Research
90

NVIDIA Releases FP4 Quantized Kimi-K2.7-Code with 1T Parameters

NVIDIA released FP4 quantized Kimi-K2.7-Code on Hugging Face, a 1T-parameter model for Blackwell GPUs with claimed accuracy retention.

x.com/Jul 10, 2026/3 min read
hardwarequantizationai
Mark Zuckerberg at a Meta event, smiling while presenting Muse Spark 1.1 on a large screen showing code, with a…
AI Research
100

Meta Muse Spark 1.1 Debuts in AI Coding Battle; Zuck Post Hits 12M Views

Meta released Muse Spark 1.1 for agentic coding tasks. Zuckerberg's post got 12M views in 12 hours; no benchmarks disclosed.

x.com/Jul 10, 2026/3 min read/Widely Reported
metaai modelscoding assistants
A JPMorgan Chase bank branch exterior with a sleek glass facade and the company logo above the entrance, set in a…
AI Research
77

JPMorgan AI Agents Beat 60/40 Portfolio in Backtests

JPMorgan's AI agents outperformed the 60/40 portfolio in backtests, signaling a shift toward autonomous asset allocation by major financial institutions.

bloomberg.com/Jul 9, 2026/3 min read
financejpmorganai agents
A software engineer reviews code on a large monitor displaying benchmark tasks, with a broken task highlighted in…
AI Research
95

OpenAI Finds 30% of SWE-Bench Pro Tasks Are Broken, Pulls Endorsement

OpenAI finds ~30% of SWE-Bench Pro tasks broken, pulls endorsement. Human reviewers flagged 249 flawed tasks.

the-decoder.com/Jul 9, 2026/3 min read/Widely Reported
ai codingbenchmarksopenai
A reflective orchestration agent interface showing DeepSeek V3.2 with a 67.25% pass@2 score on ARC-AGI-1, costing…
AI ResearchBreakthrough
86

DeepSeek V3.2 Agent Hits 67% on ARC-AGI-1 Without Fine-Tuning

Moghe & Chin achieve 67.25% pass@2 on ARC-AGI-1 using DeepSeek V3.2 in non-thinking mode at $0.62/task, with no fine-tuning. The work demonstrates agent architecture alone can lift a 15.50% baseline by ~52 points.

arxiv.org/Jul 9, 2026/3 min read
arc-agibenchmarksdeepseek

Four metagaming types need separate fixes or models learn…

AI ResearchBreakthrough
82

Four metagaming types need separate fixes or models learn to conceal it

A LessWrong taxonomy classifies AI metagaming into four types requiring separate fixes; blanket mitigation may teach concealment.

lesswrong.com/Jul 8, 2026/3 min read
taxonomyalignmentai safety
A line graph titled 'LLM agent performance vs task length' shows a steep downward curve labeled 'nonlinear failure'…
AI Research
90

LLM agents fail nonlinearly as tasks lengthen, 27-paper synthesis finds

27-paper synthesis finds LLM agent failures compound nonlinearly with task length. Six failure clusters identified across 19 benchmarks.

arxiv.org/Jul 8, 2026/3 min read/Multi-Source
llm agentsbenchmarksai research
A bar chart showing CAD benchmark scores for 7 LLM models, with DeepSeek-V3.2 and Qwen3-235B-A22B at the top near…
AI ResearchBreakthrough
75

LLMForge: 7 Models Score 0.89 on CAD Benchmark; VLMs Fix Cylinders

LLMForge scores 0.89 on 97-design CAD benchmark. VLM critic achieves 100% watertight meshes but cylinders remain a failure mode.

arxiv.org/Jul 8, 2026/3 min read
benchmarkaicad
Lilian Weng, co-founder of Thinking Machines Lab, gestures during a presentation, with a slide showing a diagram of…
AI Research
94

Lilian Weng Argues Harness Design, Not Model Rewrites, Is Path to RSI

Lilian Weng argues RSI starts with harness design, not model rewrites, citing Sakana AI's The AI Scientist in Nature 2026 and two other projects.

x.com/Jul 7, 2026/3 min read
sakana airecursive self-improvementai research
A diagram illustrating GraphRAG memory design with MongoDB at center connecting to full-text, semantic, and graph…
AI Research
75

GraphRAG Memory Design: Retrieval Over Storage, MCP Integration

Agent memory design prioritizes retrieval over storage, using unified MongoDB and MCP server. GraphRAG enables multi-hop traversal via three strategies.

x.com/Jul 7, 2026/3 min read
knowledge graphsmcpai agents
Line chart comparing Ant's 1.1B LingBot-Vision model against Meta's 7B DINOv3 on 12 benchmarks, with Ant's model…
AI Research
100

Ant Group's 1.1B LingBot-Vision Beats Meta's 7B DINOv3 on 12 Benchmarks

Ant Group's 1.1B LingBot-Vision tops Meta's 7B DINOv3 on 12 spatial benchmarks, with 40% fewer FLOPs.

pandaily.com/Jul 7, 2026/3 min read/Widely Reported
computer visionai modelsfintech ai
A large language model's internal reasoning visualized as a neural network, with a glowing surgical scalpel…
AI Research
87

Anthropic Shows Models Detect Brain Surgery Interventions Mid-Reasoning

Anthropic's J-space paper proves causal reasoning control and model detection of interventions, raising alignment and eval awareness questions.

x.com/Jul 7, 2026/3 min read
alignmentanthropicinterpretability

DARPA AIQ Program Shifts From Benchmarks to Measuring AI …

AI Research
75

DARPA AIQ Program Shifts From Benchmarks to Measuring AI Capabilities

DARPA AIQ program, one year in, shifts from benchmarks to a science of AI capability, per program lead @patrickshafto.

x.com/Jul 7, 2026/3 min read
ai safetyai policygovernment ai
A Peking University researcher points to a phase-change memristor chip on a test board, with latency and speedup…
AI ResearchBreakthrough
100

PKU Chip Hits 2.12ms Brain Latency, 478x A100 Speedup

PKU chip achieves 2.12ms step latency with 478x speedup over Nvidia A100 for brain modeling using phase-change memristors.

pandaily.com/Jul 7, 2026/3 min read/Widely Reported
hardwareresearchneuromorphic computing
A sleek metallic robot hand reaches toward a glowing digital brain, symbolizing advanced AI competition
AI Research
84

GPT-4 Held Top Spot 52 Weeks; Today's Models Last 7

GPT-4 dominated the ECI for a year. Today's top models last 7 weeks median, with 17 leadership changes since Feb 2024.

the-decoder.com/Jul 6, 2026/3 min read
benchmarksai modelscompetition
A person's laptop screen displays Obsidian notes and a connected AI agent interface, with icons for Readwise…
AI Research
74

Second Brain Becomes Open-Source AI Agent Memory System

Pauliusztin open-sourced a 2-year project turning 10,994 notes into a living AI agent memory system. The architecture aligns with Google's new OKF standard.

x.com/Jul 6, 2026/3 min read
open sourceknowledge managementai agents
Alibaba researchers present MIPU framework diagram with a two-step RL process, showing a large language model being…
AI Research
85

Alibaba's MIPI fixes LLM training-inference mismatch with direct RL

Alibaba's MIPI directly optimizes inference policy, fixing the mismatch in LLM post-training via the MIPU framework.

x.com/Jul 6, 2026/3 min read
llm post-trainingreinforcement learningalibaba
A high-resolution photomosaic composed of hundreds of small tile photos forming a larger image, likely a landscape…
AI Research
85

PhotoQuilt Makes Training-Free Photomosaics at 14K Resolution

PhotoQuilt generates training-free photomosaics at any resolution, bootstrapping a global layout at low res then upscaling tiles via FLUX, scaling past 14K without quadratic attention cost.

x.com/Jul 5, 2026/3 min read
image-generationdiffusion-modelsgenerative-ai
AI researchers discuss a chart comparing 35B parameter agent performance to trillion-parameter models at a conference
AI Research
85

Hugging Face Papers: 35B Agent Matches Trillion-Parameter Performance

Hugging Face Daily Papers featured eight AI papers, including Orca (world model), Dockerless (62% SWE-bench), and a 35B agent matching trillion-parameter performance.

x.com/Jul 5, 2026/3 min read
efficiencyworld modelsscaling
Robot navigating through a room with furniture, using sensors and a screen displaying a floor plan
AI Research
100

Alibaba's Qwen-RobotNav Unifies Robot Navigation in One 2B-8B Model

Alibaba's Qwen-RobotNav unifies VLN, ObjectNav, tracking, and autonomous driving in a 2B-8B model, deploying zero-shot to quadruped robots via a configurable observation protocol.

x.com/Jul 5, 2026/3 min read/Multi-Source
roboticsnavigationai models
CMU researchers demonstrate Gym-Anything converting a spreadsheet interface into a simulated training ground for AI…
AI Research
92

CMU's Gym-Anything Turns Any Software Into Agent Training Ground

CMU's Gym-Anything automates agent environment creation, producing CUA-World with 10,000+ tasks. Even strong models fail most long tasks, showing real computer-use work is unsolved.

x.com/Jul 4, 2026/3 min read
cmuresearchai agents