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
Mira Murati presents a chart comparing Thinking Machines' error rate and inference cost against frontier models…
AI Research
92

Mira Murati's Thinking Machines beats frontier models by 29.8% with Bridgewater's expert judgments

Thinking Machines beat frontier models by 29.8% fewer errors using Bridgewater's expert judgments, at 13.8x lower inference cost.

x.com/10h ago/3 min read
financeenterprise-aistartups
AI Security Inst Shows Test-Time Compute Skews Frontier Evaluations
AI Research
86

AI Security Inst Shows Test-Time Compute Skews Frontier Evaluations

AISecInst research shows test-time compute budgets skew frontier model evaluations, challenging standard practices.

x.com/11h ago/3 min read
ai safetyai researchmodel evaluation
A robotic arm in a lab setting grasps a transparent plastic cup on a table, with a monitor displaying code and…
AI Research
85

DART: One-Shot Robot Adaptation via Weight Space Arithmetic

DART from Seoul National University adapts robot policies with one demonstration using weight space arithmetic, achieving 73% success on unseen domain shifts.

x.com/15h ago/3 min read
domain adaptationweight space arithmeticrobot learning
A diagram showing ELDR routing decode requests to GPU experts using prefill signatures, with arrows indicating…
AI Research
85

ELDR: Expert-Locality Decode Routing Cuts MoE TPOT by 13.9%

ELDR uses prefill expert signatures to route decode requests, cutting median TPOT by 5.9–13.9% in vLLM at scale.

x.com/23h ago/3 min read
ai infrastructurellm servingmoe
3D scene decomposition showing token groups representing distinct objects, with reconstruction and segmentation…
AI Research
87

Feed-Forward Model Decomposes 3D Scenes as Objects Without 3D Labels

A feed-forward model decomposes 3D scenes into objects from unposed images without 3D annotations, enabling one-pass reconstruction, segmentation, and manipulation.

x.com/1d ago/3 min read
computer visionscene understanding3d vision
Seven LLM logos arranged in a grid above a bar chart comparing posterior match rates and downstream prediction…
AI Research
87

BayesBench: LLMs Match Bayesian Posteriors But Fail Downstream Prediction

BayesBench tests 7 LLMs on multi-turn Bayesian reasoning. Scaling improves latent inference but not prediction, exposing a critical gap for agentic deployment.

arxiv.org/2d ago/3 min read/Multi-Source
llm evaluationbayesian inferenceai research
A colorful heatmap overlay on a digital brain scan highlights distinct neural regions labeled for language, math…
AI Research
85

LLMs Spontaneously Develop Human-Like Brain Regions for Language, Math

LLMs spontaneously develop human-like brain regions for language, math, physics, and social reasoning, per @LiorOnAI. Two optimization processes converged on the same solution.

x.com/2d ago/3 min read
emergent behaviorneuroscienceai research
Server racks with domestic chips powering a 1.6 trillion-parameter AI model, LongCat-2.0, open-sourced by Meituan in…
AI ResearchBreakthrough
100

Meituan Open-Sources 1.6T-Parameter LongCat-2.0 Trained on Domestic Chips

Meituan open-sourced 1.6T-parameter LongCat-2.0 trained on 50,000 domestic ASICs, claiming China's first full-process domestic-chip trillion-parameter model.

scmp.com/3d ago/3 min read/Widely Reported
chinaopen-sourcehardware
A diagram showing SingGuard processing text and image inputs through fast and slow reasoning modules to evaluate…
AI Research
85

SingGuard: Runtime Guardrails for Multimodal AI Treat Safety as Input

SingGuard treats safety rules as runtime inputs for multimodal AI, achieving SOTA across 6 families and 35 datasets via fast/slow reasoning.

x.com/3d ago/3 min read
guardrailsai safetymultimodal ai
A diagram showing multiple hash IDs replacing traditional token embeddings in a Transformer architecture, with…
AI Research
85

MultiHashFormer Brings Hash-Based Autoregression to Causal LMs

MultiHashFormer brings hash-based autoregression to causal LMs, slashing embedding memory and outperforming standard Transformers from 100M to 3B parameters.

x.com/3d ago/3 min read
efficiencylanguage modelsai research
Open textbook on mathematical foundations of reinforcement learning with grid-world examples, 16.2K GitHub stars…
AI Research
81

Free RL Textbook 'Math Foundations' Hits 16.2K GitHub Stars

Free RL textbook by Shiyu Zhao hits 16.2K GitHub stars and 2.1M video views, filling a gap in RL education with rigorous math and a unified grid-world example.

x.com/4d ago/3 min read
open-sourcereinforcement-learningmachine-learning
A human hand in a blue glove demonstrates a task while a robot arm mirrors the motion, with a green overlay showing…
AI Research
82

ByteDance Seed Turns Cheap Human Videos Into Robot Skills

ByteDance Seed replaces noisy 6DoF hand poses with relative wrist translation, creating a shared action space for humans and bi-manual robots that scales with cheap data and outperforms full-pose baselines.

x.com/4d ago/3 min read
roboticsbytedanceimitation learning
Two computer monitors side by side display lines of code in a dimly lit office, suggesting a software bug-hunting or…
AI ResearchBreakthrough
85

Zhipu GLM-5.2 beats Anthropic's Mythos on bug-hunt benchmark

Zhipu AI's GLM-5.2 beat Anthropic's Claude Opus 4.8 on a cybersecurity bug-hunting benchmark, then matched it with extra instructions, marking another 'DeepSeek moment'.

scmp.com/4d ago/3 min read/Multi-Source
anthropiczhipu aichinese ai
Bar chart showing GPT-5.4 performance on PlanBench-XL dropping from 51.90% to 11.36% on hardest tool-use tasks with…
AI Research
90

PlanBench-XL: GPT-5.4 Scores 11.36% on Hard Tool-Use Tasks

PlanBench-XL shows GPT-5.4 drops from 51.90% to 11.36% accuracy on long-horizon tool-use tasks with 1,665 tools, revealing a fundamental planning weakness.

x.com/4d ago/3 min read
planningbenchmarksllm-agents
Alibaba's Qwen-AgentWorld open-source model interface on Hugging Face with code and streaming inference tools
AI Research
82

Alibaba Open-Sources Qwen-AgentWorld for Generalist Agent Training

Alibaba open-sourced Qwen-AgentWorld and Wan-Streamer v0.1 on Hugging Face, targeting generalist agent training and real-time streaming. The releases include 8 additional papers on agent benchmarks and architectures.

x.com/5d ago/3 min read
open-sourceagentic aiworld models
A large neural network diagram overlays molecular structures, protein chains, and text tokens, illustrating…
AI Research
95

BioMatrix: A single decoder reads proteins, molecules, language on 304B tokens

BioMatrix, a decoder-only biological foundation model, achieves SOTA on 77 of 80 tasks after training on 304B tokens of sequences, structures, and language.

x.com/5d ago/3 min read
foundation modelsprotein designmolecular generation
Two stacked line charts compare attention cost and prefill speed between standard and Grouped Query Experts methods…
AI Research
85

Grouped Query Experts cuts long-context attention cost 44%

GQE speeds long-context attention prefill 1.7–1.8× by routing tokens to 9 of 16 query heads, matching baseline accuracy at 56.04.

x.com/5d ago/3 min read
efficiencyresearchattention
A diagram shows EvoEmbedding's latent memory queue processing a long text passage, generating dynamic embeddings…
AI Research
85

EvoEmbedding Beats Static Embedders 3× Larger via Latent Memory Queue

EvoEmbedding uses a latent memory queue to beat static embedders 3× its size on long-context retrieval, per @HuggingPapers.

x.com/5d ago/3 min read
embedding modelsresearchretrieval
A terminal window displays command-line output with benchmark results, showing a 33.4% score, while a bar chart…
AI Research
87

CLI-Universe: Qwen3-32B fine-tuned on 6K trajectories beats models 10x larger on Terminal-Bench 2.0

CLI-Universe synthesizes terminal-agent tasks; Qwen3-32B fine-tuned on 6K trajectories hits 33.4% on Terminal-Bench 2.0, beating models 10x larger.

x.com/5d ago/3 min read
agentic aifine-tuningbenchmarks
Robot with a new limb configuration adapting its movement on a lab floor, surrounded by sensors and a computer…
AI Research
85

ICWM Lets Robots Adapt to Unseen Morphologies in Seconds

ICWM learns world dynamics from seconds of self-generated interaction, enabling zero-shot generalization to unseen cameras and morphologies without fine-tuning.

x.com/6d ago/3 min read
roboticsresearchai
Two researchers point at a large monitor displaying a chart comparing iLLaDA and Qwen2.5 benchmark scores, with the…
AI ResearchBreakthrough
93

ByteDance iLLaDA: 8B Diffusion LM Matches Qwen2.5 Base, Lags on Instruct

ByteDance iLLaDA, an 8B diffusion LM trained on 12T tokens, matches Qwen2.5 7B on base benchmarks (63.9 vs 63.3) but trails 10 points after instruction tuning, revealing the alignment gap for diffusion models.

the-decoder.com/6d ago/3 min read/Multi-Source
llm benchmarksdiffusion modelsbytedance
A taxonomy diagram with branching nodes connecting world models, video generation, and vision-language-action…
AI Research
87

World Action Models Survey Unifies 100+ Methods Under One Taxonomy

A survey reviews 100+ world action models, unifying world models, video generation, and VLA policies under one taxonomy.

x.com/6d ago/3 min read
world modelssurveyembodied ai
Agent traces overlaid on a grid map, with numbered steps showing hierarchical skill decomposition from hindsight, no…
AI Research
82

OPID: Agents Learn From Hindsight Without External Memory

OPID lets agents learn hierarchical skills from hindsight, improving sample efficiency on ALFWorld, WebShop, Search QA without external memory at inference.

x.com/6d ago/3 min read
hierarchical-reinforcement-learningagent-learningsample-efficiency
NVIDIA logo displayed on a dark background with abstract geometric patterns suggesting depth and stereo vision…
AI Research
85

NVIDIA Drops Fast-FoundationStereo: 10× Faster Depth Estimation

NVIDIA released Fast-FoundationStereo, a real-time foundation model for zero-shot stereo depth estimation that is 10× faster than FoundationStereo with matching accuracy.

x.com/6d ago/3 min read
roboticscomputer visionedge ai
Alibaba Qwen-Image-Agent agentic framework diagram showing context-aware image generation pipeline with planning…
AI Research
87

Qwen-Image-Agent: Alibaba's Agentic Framework for Context-Aware Image Gen

Alibaba's Qwen-Image-Agent uses planning, reasoning, search, and memory to build context for text-to-image models, bridging the context gap in real-world generation.

x.com/Jun 26, 2026/3 min read
image generationagentic aialibaba
A graph network diagram with red nodes and lines representing dynamic red-teaming connections overlaid on a digital…
AI Research
85

RIFT-Bench Tests 45 Agentic Systems With Dynamic Red-Teaming

RIFT-Bench evaluates 45 agentic AI systems via a graph-driven two-phase pipeline, enabling unified security comparison across heterogeneous architectures.

arxiv.org/Jun 24, 2026/3 min read/Widely Reported
ai safetyagentic aibenchmarks
ReMMD Agent Hits 41.8% Accuracy on Multilingual Misinformation, Cuts Cost 79.9%
AI Research
72

ReMMD Agent Hits 41.8% Accuracy on Multilingual Misinformation, Cuts Cost 79.9%

ReMMD-Agent achieves 41.8% accuracy on multilingual misinformation detection with 79.9% cost reduction, using a persistent memory approach.

arxiv.org/Jun 24, 2026/3 min read
agentic systemsmisinformationai research
A screenshot of a Twitter post by Tencent announcing the open-source release of TencentDB Agent Memory, showing…
AI Research
100

Tencent Open-Sources Agent Memory System Cutting Token Use 61%

Tencent open-sourced TencentDB Agent Memory, cutting token usage by 61.38% and boosting task success by 51.52% on WideSearch, running fully local.

x.com/Jun 23, 2026/3 min read/Multi-Source
open-sourcememory-systemsai-agents
Two bar charts comparing OpenAI GPT-5.5-Cyber and Anthropic Mythos scores on CyberGym, ExploitGym, and SEC-bench…
AI ResearchBreakthrough
100

OpenAI GPT-5.5-Cyber Beats Anthropic Mythos on Security Benchmarks

OpenAI's GPT-5.5-Cyber beats Anthropic's Mythos on security benchmarks. Updated Codex plugin auto-patches after scanning 30M commits.

the-decoder.com/Jun 23, 2026/3 min read/Widely Reported
anthropicbenchmarksai models
A person in a dark sweater and glasses looks at a smartphone, their face lit by the screen, with a small robot…
AI Research
75

Pew: Only 16% of Americans Expect AI to Help Society in 2026

Pew report: 16% of Americans expect AI to help society, down from 37% in 2024 — a 21-point drop in two years.

x.com/Jun 23, 2026/3 min read
industry trendsai safetyai policy