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 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
A dynamic dashboard with interconnected nodes representing multiple LLMs, coordinated by Sakana AI's Fugu…
AI Research
85

Sakana AI's Fugu Orchestrator Matches Anthropic Fable 5 Without Using It

Sakana AI's Fugu orchestrator matches Anthropic's top models on benchmarks without using them, offering a hedge against vendor lock-in amid export controls.

the-decoder.com/Jun 22, 2026/3 min read/Widely Reported
startupsbenchmarksai models
A 3D spatial tree diagram with branching nodes and arrows illustrating hierarchical spatial reasoning, with…
AI ResearchBreakthrough
100

ByteDance Seed's SpatialTree Redefines MLLM Spatial Reasoning at CVPR 2026

ByteDance Seed's SpatialTree achieves 79.8% on SEAL-Bench, 12.4 points above GPT-4V, using hierarchical spatial decomposition. Open-sourced at CVPR 2026.

pandaily.com/Jun 22, 2026/3 min read/Widely Reported
bytedancecomputer visionai research
A laptop screen displays the HuggingFace dataset page for LOCUS-v1, showing 2.2M US laws, with code and data tables…
AI Research
89

LOCUS-v1: 2.2M US Laws Hit HuggingFace via AI Pipeline

LOCUS-v1, a dataset of 2.2M US laws built via AI pipeline, released on HuggingFace. First comprehensive legal database of its kind, but quality and validation metrics remain undisclosed.

x.com/Jun 21, 2026/3 min read
hugging-facelegal-techdatasets
A Miami startup's LLM inference dashboard shows 12 million tokens processed for $8, compared to $2,600 on Claude…
AI ResearchBreakthrough
90

Miami Startup Claims 12M-Token LLM Inference at $8 vs. $2,600 on Claude

Miami startup claims 12M-token LLM inference for $8 vs. $2,600 on Claude Opus 4.6. No paper or benchmarks released yet.

pub.towardsai.net/Jun 21, 2026/3 min read
ai startupsllm inferenceanthropic
A diagram shows multiple robot agents connected by arrows, with a central meta-skill node labeled 'orchestration'…
AI Research
80

Meta-skill evolution lets multi-agent systems self-improve without retraining

Multi-agent systems can improve orchestration by evolving a meta-skill via RL on interactions, without retraining agents. Demonstrated on a simulated benchmark.

x.com/Jun 20, 2026/3 min read
multi-agentmeta-learningreinforcement learning
A bar chart comparing Zhipu GLM 5.2 and Claude Fable 5 scores on web design benchmarks, with GLM 5.2 leading in…
AI Research
92

Zhipu's GLM 5.2 claims Design Arena's top HTML spot with Elo 1,360 — edging a hobbled Claude Fable 5

Zhipu AI's 753-billion-parameter open-weight model GLM 5.2 topped the Design Arena HTML benchmark with an Elo score of 1,360, edging Anthropic's Claude Fable 5 (1,350). The win coincides with a Commerce Department export-control order that pulled Fable 5 from non-US users, and GLM 5.2's API pricing

pandaily.com/Jun 20, 2026/3 min read/Widely Reported
anthropicchinese aibenchmarks
A person using a laptop with ChatGPT interface open, surrounded by colorful AI-related graphics and charts…
AI ResearchBreakthrough
95

OpenAI shows small doses of beneficial-trait RL improve 44 of 53 safety benchmarks — and the gains generalize

OpenAI researchers Jagadeesh, Saab, Singhal et al. published findings on June 18 showing RL training on traits like honesty and corrigibility improved 44 of 53 safety benchmarks. Gains generalized across domains not used in training, and the model resisted harmful fine-tuning better than the baselin

the-decoder.com/Jun 19, 2026/3 min read/Widely Reported
alignmentai safetyreinforcement learning
AI Generates Chest X-Rays Clinicians Cannot Tell Apart From Real Ones
AI Research
85

AI Generates Chest X-Rays Clinicians Cannot Tell Apart From Real Ones

RadiT XL, a 1.3B-parameter rectified flow transformer trained on 1.2 million chest radiographs, produces synthetic images that clinical experts cannot reliably distinguish from real ones — a milestone that could break the data bottleneck limiting medical AI fairness and generalization.

arxiv.org/Jun 19, 2026/3 min read/Widely Reported
medical imagingai modelsgenerative ai
A large language model interface displays Qwen 2.5 7B with a near-constant confidence score of 0.856, while…
AI Research
92

Qwen 2.5 7B Expresses Near-Constant Confidence Whether It Is Right or Wrong, Study Finds

A June 2026 arXiv preprint from University of Minnesota researchers tested Qwen 2.5 7B on structured clinical prediction data and found its verbalized confidence scores are essentially uninformative -- clustering between 0.856 and 0.937 no matter how well or badly the model performs. Combining SHAP-

arxiv.org/Jun 19, 2026/3 min read/Widely Reported
researchsafetytabular data
OpenAI Can Predict Model Failures via Past Chat Replay
AI Research
100

OpenAI Can Predict Model Failures via Past Chat Replay

OpenAI can estimate model failures by replaying past chats, enabling proactive error detection without new labeled data. No benchmark numbers disclosed.

x.com/Jun 18, 2026/3 min read/Multi-Source
ai safetyresearch

BeliefDiffusion Uses Diffusion Models for Robot Navigatio…

AI Research
69

BeliefDiffusion Uses Diffusion Models for Robot Navigation in Partially

BeliefDiffusion combines diffusion models with MPC for robot navigation in partially observable environments, outperforming model-free RL and generative baselines in synthetic maps.

arxiv.org/Jun 18, 2026/3 min read
research-papersroboticsreinforcement-learning
A diagram of a science safety risk management framework with colored nodes for risk dimensions like ethics…
AI Research
68

SciRisk-Bench Tests 10 Risk Dimensions Across 7 Science Disciplines

SciRisk-Bench evaluates LLMs across 10 risk dimensions and 7 disciplines. Safety omission and lab safety show highest vulnerability.

arxiv.org/Jun 18, 2026/3 min read
ai safetybenchmarkslarge language models
OpenAI researchers analyzing data on a computer screen, showing graphs of GPT-5 error predictions with 92% accuracy…
AI Research
90

OpenAI DeploymentSim predicts GPT-5 errors 92% of the time pre-launch

OpenAI's Deployment Simulation predicted GPT-5 errors with 92% accuracy using 1.3M real conversations, outperforming standard safety tests.

the-decoder.com/Jun 17, 2026/3 min read/Widely Reported
gpt-5ai safetyopenai
A person with a concerned expression sits before a glowing computer screen displaying lines of code and a warning…
AI Research
80

Alignment Pretraining Could Backfire, LessWrong Post Warns

LessWrong post warns synthetic alignment pretraining data could backfire in capable LLMs, leading to rebel personas.

lesswrong.com/Jun 17, 2026/3 min read/Multi-Source
anthropicai safetyalignment research
Side-by-side comparison of images generated by vanilla LoRA and Pareto LoRA, with the Pareto LoRA output showing…
AI Research
90

Pareto LoRA Boosts Image Quality 44.9% vs Vanilla LoRA on Emu2

Pareto LoRA reformulates multimodal instruction tuning as bi-objective optimization, achieving up to 44.9% image quality gains on Emu2 while maintaining text performance.

arxiv.org/Jun 17, 2026/3 min read/Widely Reported
nlpmultimodal modelscomputer vision
A stylized abstract illustration of a glowing brain network overlaid on a world map, with red and blue data streams…
AI Research
72

Estonian Institute: Claude Tops Russian Propaganda Benchmark, Mistral Trails

Estonian Language Institute benchmark tests 60 AI models vs Russian propaganda. Claude tops, Mistral trails with 36.67% misinformation rate.

the-decoder.com/Jun 16, 2026/3 min read
anthropicai safetybenchmark
Visual-Seeker: Active Visual Reasoning Beats Proprietary MLLMs on 5 Benchmarks
AI Research
72

Visual-Seeker: Active Visual Reasoning Beats Proprietary MLLMs on 5 Benchmarks

Visual-Seeker achieves SOTA on five multimodal search benchmarks, surpassing proprietary models by actively harvesting visual evidence during search.

arxiv.org/Jun 16, 2026/3 min read
agentsresearchmultimodal
Two researchers in a lab analyzing a chart showing cost reduction, with a laptop displaying a graph of annotation…
AI Research
70

Metric Match Cuts LLM Judge Annotation Cost 32.5% via Subset Selection

MIT and Stanford researchers developed Metric Match, a subset selection method that reduces LLM judge annotation costs by 32.5% and estimation error by 18.7%, achieving a 0.838 win-rate against random selection.

arxiv.org/Jun 16, 2026/3 min read
paperresearchllm
Researchers analyze fusion strategies on a computer dashboard displaying patient data and survival curves for PE…
AI Research
70

No single fusion strategy wins

Zhang et al. test 4 fusion strategies on 7K+ patients, finding no universal best. Contrastive alignment with CLMBR wins for PE mortality; cross-attention and co-attention split for CVD.

arxiv.org/Jun 16, 2026/3 min read
healthcare aimultimodal learningai research
Close-up of a Final Cut Pro timeline on a computer screen, showing multiple video tracks with colored clips and a…
AI Research
65

AI editor matches pro on 84% of video cuts in blind test

AI editor matched pro on 84% of video cuts in blind test of 4-hour project. Suggests editorial judgment is partially automatable.

x.com/Jun 15, 2026/3 min read
ai videovideo editingbenchmark
Smartphone displaying LLaDA-8B inference interface with latency reduction metrics, NPU chip schematic overlay
AI Research
84

llada.cpp Cuts LLaDA-8B Latency 17-42x on Mobile NPU

llada.cpp, the first NPU-aware dLLM inference framework, cuts LLaDA-8B latency 17-42x on smartphones, enabling real-time on-device generation.

arxiv.org/Jun 15, 2026/3 min read/Multi-Source
ai inferencemobile hardwarediffusion models
Bar chart comparing AI model scores on MA-ProofBench, with GPT-5.5 reaching 16% on undergraduate and 5% on PhD…
AI Research
82

MA-ProofBench: GPT-5.5 Hits 16% on Math Analysis, Most Models Near 0%

MA-ProofBench, a new theorem-proving benchmark for mathematical analysis, shows GPT-5.5 achieving 16% on undergraduate problems and 5% on PhD-level, with most models near 0% on the harder set.

arxiv.org/Jun 15, 2026/3 min read/Multi-Source
mathematicstheorem provingbenchmarks
Mirage Probes Paper Reveals Two Distinct VLM Failure Modes
AI Research
90

Mirage Probes Paper Reveals Two Distinct VLM Failure Modes

Mirage Probes paper reveals VLMs have two distinct failure modes—textual biases and spurious images—requiring different mitigations. Text cleaning only fixes one; the other needs representational interventions.

arxiv.org/Jun 15, 2026/3 min read/Multi-Source
ai safetycomputer visionresearch
A sleek robotic arm on a lab bench precisely assembles a small electronic circuit board, with glowing blue lights…
AI Research
84

WorkBench Revisited: Claude Opus 4.8 Hits 89% Task Completion

Claude Opus 4.8 completes 89% of WorkBench tasks with 2.5% harm rate, up from GPT-4's 43% and 26% in 2024, showing capability and safety align.

arxiv.org/Jun 15, 2026/3 min read/Multi-Source
anthropicagent safetybenchmarks
A line graph titled SWE-Explore showing low coverage rates around 14-19% for Claude Code, Codex 5.3, and OpenHands…
AI Research
92

SWE-Explore: AI coding agents find files but miss 81-86% of critical lines

SWE-Explore benchmark shows Claude Code, Codex cover only 14-19% of critical lines despite finding the right file. Model strength doesn't fix the structural weakness.

the-decoder.com/Jun 14, 2026/3 min read/Widely Reported
code generationresearchai agents
A futuristic digital rendering of a robotic arm manipulating glowing tokens on a simulated grid, with Nvidia…
AI Research
87

Nvidia Cosmos 3 Unifies Physical AI — Action as Token

Nvidia's Cosmos 3 unifies physical AI perception, simulation, and action in one model via action-as-token. No benchmark data disclosed yet.

x.com/Jun 14, 2026/3 min read
physical-airoboticsworld-models
A person typing on a laptop with code on the screen, likely representing the AI model Rio 3.5 Open 397B from Rio de…
AI Research
85

Rio 3.5 Open 397B: City Gov't Model Tops Open-Source Leaderboard

Rio 3.5 Open 397B claims SOTA open-source status but lacks any evidence. A single tweet is the only source.

x.com/Jun 13, 2026/3 min read
municipal-aiopen-source-llmbenchmark-claims
Bar chart comparing BIRD benchmark scores shows Gemini-SQL2 at 80.04% and GPT-5.5-xhigh around 73%, with Gemini 3.1…
AI Research
95

Google Gemini-SQL2 Hits 80.04% on BIRD, Beating GPT-5.5 by 7 Points

Google's Gemini-SQL2 hits 80.04% on BIRD, beating GPT-5.5 by 7 points and Claude Opus 4.6 by 9 points, with no public release or paper yet.

the-decoder.com/Jun 13, 2026/3 min read/Widely Reported
nl-to-sqlfoundation-modelsbenchmarks
A doctor in a white coat studies a tablet displaying AI diagnostic data, surrounded by medical charts and a…
AI Research
83

General LLMs Beat Clinical AI Tools in Doctor Study

Frontier LLMs beat clinical AI tools like OpenEvidence in all evaluations, matching Google Search AI Overview.

x.com/Jun 12, 2026/3 min read/Multi-Source
llmbenchmarkai