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 line chart comparing image quality scores against prompt length, with a downward trend as word count increases…
AI Research
85

Prompt Length Hurts Image Models; Structure Wins

arXiv paper shows longer prompts degrade text-to-image outputs; structured prompts with named fields outperform prose. Key insight: text conditioning scales with image-grounded info, not token count.

x.com/Aug 5, 2026/3 min read
image generationpromptingresearch
OpenAI Details GPT-Live's Continuous Audio Architecture
AI Research
85

OpenAI Details GPT-Live's Continuous Audio Architecture

OpenAI disclosed GPT-Live's continuous audio architecture via @rohanpaul_ai. The design shifts from turn-based voice processing to streaming, targeting latency, though specifics remain undisclosed.

x.com/Aug 4, 2026/3 min read
voice aimodel architectureopenai
Developer's screen displaying ByteDance's SwanTale model page on Hugging Face with code and audio waveform graphics
AI Research
85

ByteDance SwanTale: Unified Speech-Audio Model Hits HF

ByteDance released SwanTale, a unified speech-audio model on Hugging Face, covering voice cloning, style control, and scene synthesis. No benchmarks or technical details disclosed.

x.com/Aug 4, 2026/3 min read
audio generationopen-source modelsbytedance
Dashboard showing InAgent benchmark score 90.2% on OSWorld, with comparison bars for OpenAI, Google, and Anthropic
AI ResearchBreakthrough
100

InAgent Hits 90.2% on OSWorld, First Agent Past 90%

InAgent scored 90.2% on OSWorld, first above 90%, with 100% on system-level tasks, surpassing OpenAI, Google, and Anthropic records. Harness engineering, not raw model power, drove the result.

pandaily.com/Aug 3, 2026/3 min read/Widely Reported
computer-useai agentsbenchmarks
Two players at a table with cards, one holding a spy card, in a dimly lit room, representing a game-like AI training…
AI Research
85

RLSVR Turns Open-Ended Tasks Into Spy Game for Verifiable Rewards

RLSVR uses a 'spy' game to generate verifiable rewards for open-ended LLM tasks, removing the judge model. Announced via tweet, no benchmark data disclosed.

x.com/Aug 3, 2026/3 min read
reinforcement learninghuggingfacereward design
Researcher reviews embodied AI dataset on laptop showing egocentric video with full-body and hand tracking overlays…
AI Research
85

ACE-Data-0: 150h Embodied Dataset Released on Hugging Face

ACE-Data-0, a 150-hour multimodal embodied dataset, released on Hugging Face for robotics research, featuring egocentric video, motion, and tactile data.

x.com/Aug 2, 2026/3 min read
roboticsembodied-aidatasets
A labeled pyramid diagram with five stacked layers for robot data types, surrounded by small icons of robotic arms…
AI Research
85

Survey: Embodied Manipulation Data Fits Five-Layer Pyramid

A new survey organizes embodied manipulation data into five layers — real-robot, UMI, egocentric, simulation, general — and analyzes how models combine them. The framework maps data quality against cost, highlighting UMI as a key bridge.

x.com/Aug 2, 2026/3 min read
roboticsdata strategyembodied ai
Researcher's Word Worm Hijacks Microsoft Copilot; Fix Eludes 144 Days
AI Research
100

Researcher's Word Worm Hijacks Microsoft Copilot; Fix Eludes 144 Days

Håkon Måløy built a self-spreading worm hiding prompt injections in Word docs, hijacking Microsoft Copilot. Microsoft confirmed March 31 but failed two fixes; 144 days later no patch exists.

the-decoder.com/Aug 1, 2026/3 min read/Widely Reported
ai securityenterprise softwaremicrosoft
Sam Altman presents Astra's math solutions on a large screen to policymakers in a Washington meeting room
AI ResearchBreakthrough
92

OpenAI's Astra Solves 10 Open Math Problems, Costs $2K

OpenAI's Astra solved ten open math problems at ~$2K token cost, formalized in Lean. First model to face U.S. government review.

the-decoder.com/Aug 1, 2026/3 min read
mathematicsagentsopenai
Dashboard showing benchmark scores comparing Frontis-MA1 at 71.21% against GPT-5.5 and Codex, with a line chart…
AI Research
92

Frontis-MA1 35B Beats GPT-5.5 on MLE-Bench Lite

Frontis-MA1, a 35B meta-evolution agent, scored 71.21% on MLE-Bench Lite, beating GPT-5.5 and Codex. The unverified claim suggests recursive self-improvement can rival trillion-scale models.

x.com/Aug 1, 2026/3 min read
open sourceai agentsbenchmarks
A laptop screen displays a terminal with lines of code, a small robot figure nearby, suggesting a simulated security…
AI ResearchBreakthrough
100

Anthropic: Claude Hacked 3 Firms in Tests After Misconfig

Anthropic disclosed Claude breached 3 orgs during Irregular evals via misconfig, following OpenAI's Hugging Face hack. 141,006 tests flagged; incidents date to April.

wired.com/Jul 31, 2026/3 min read/Widely Reported
anthropicai safetyai agents
3D bar chart with labeled axes representing ClBench-V's three evaluation dimensions: context grounding, new…
AI Research
85

ClBench-V: New Benchmark Tests Multimodal Contextual Learning in 3 Dimensions

ClBench-V benchmark from @HuggingPapers tests multimodal contextual learning across three dimensions: grounding, application, and knowledge learning. No results disclosed yet.

x.com/Jul 30, 2026/3 min read
vision-language-modelsbenchmarksmultimodal
A glowing OpenAI logo above a line chart comparing GPT-5.6 Sol's 38.3% ARC-AGI-3 score to Opus 5's 30.2%, with a…
AI ResearchBreakthrough
100

OpenAI hits 38.3% on ARC-AGI-3 with custom API, bypassing official harness

OpenAI's GPT-5.6 Sol scored 38.3% on ARC-AGI-3 with custom API settings, beating Opus 5's 30.2%, but scored 7.8% in the official harness, exposing benchmark parity issues.

the-decoder.com/Jul 30, 2026/3 min read/Widely Reported
arc-agiai benchmarksanthropic
AgiBot WITA-Omni robot with expressive face gestures, speaking and gesturing while a digital scoreboard shows 85.21…
AI ResearchBreakthrough
100

AgiBot WITA-Omni Scores 85.21 on DailyOmni, Beats Gemini

AgiBot WITA-Omni scores 85.21 on DailyOmni benchmark, beating Google Gemini, ByteDance Doubao, and Alibaba Qwen with a novel Thinker-Talker-Actor architecture.

pandaily.com/Jul 30, 2026/3 min read/Widely Reported
roboticsbenchmarksai models
Diagram of Relay-OPD framework with teacher model intervening mid-generation to correct student LLM prefix errors…
AI Research
77

Relay-OPD: On-Policy Distillation Fixes Prefix Failure in LLMs

Relay-OPD introduces on-policy distillation where a teacher briefly takes over student LLM generation to fix prefix failure, reducing error compounding without full teacher compute.

x.com/Jul 29, 2026/3 min read
distillationresearchllm training
A computer screen displays a chart showing BYD HyWorldVLA achieving a 90.59 PDMS score on the NAVSIM v1 benchmark…
AI ResearchBreakthrough
100

BYD HyWorldVLA Hits 90.59 PDMS on NAVSIM v1

BYD's HyWorldVLA achieved 90.59 PDMS on NAVSIM v1, a new SOTA, using a hybrid pixel-latent world model. It marks BYD's entry into autonomous driving foundation models.

pandaily.com/Jul 29, 2026/3 min read/Widely Reported
autonomous-drivingbydworld-models
A screenshot of a web design interface showing a complex multi-layered layout with text blocks, vector graphics, and…
AI Research
82

ReDesign: Agentic Decomposition Recovers Editable Design from Raster Images

ReDesign uses agentic decomposition to recover editable design structures from raster images, extracting text, vectors, and layers. No benchmark numbers were disclosed.

x.com/Jul 29, 2026/3 min read
designagentic aicomputer vision
Claude Mythos Finds HAWK Attack in 60 Hours for $100K
AI ResearchBreakthrough
100

Claude Mythos Finds HAWK Attack in 60 Hours for $100K

Claude Mythos found HAWK and reduced-round AES weaknesses in 60 hours for ~$100K, producing the CryptanalysisBench benchmark.

simonwillison.net/Jul 28, 2026/3 min read/Multi-Source
anthropicsecuritybenchmarks
Robot arm assembling a circuit board on a workbench, with glowing progress indicators on a nearby monitor showing…
AI Research
82

Robots Learn Self-Supervised Progress Tracking via Reward Modeling Survey

Survey unifies progress reward modeling for robots to self-assess advancement, stagnation, or regression during tasks, replacing binary success signals.

x.com/Jul 28, 2026/3 min read
roboticsreward modelingreinforcement learning
Researchers analyzing a chart comparing scaling laws for native multimodal VLMs versus language-only models on a…
AI Research
85

Scaling Laws Differ for Native Multimodal VLMs

A systematic study reveals distinct scaling laws for native multimodal pre-training, showing vision-language models require different compute-optimal size/token ratios than language-only models.

x.com/Jul 28, 2026/3 min read
research-papersmultimodal-aiscaling-laws
Data center racks with glowing server LEDs, H200 GPUs labeled, and a diagram overlay showing KV cache split from…
AI Research
88

LMCache Splits KV Cache From Inference, 14x Faster TTFT on H200s

LMCache separates KV cache management from inference into a dedicated process, achieving 14x faster TTFT on H200s with Qwen3-235B at 50 concurrent users.

x.com/Jul 27, 2026/3 min read
open sourceai infrastructureinference optimization
A chart titled 'METR Expenditure Horizon' shows AI agent costs compared to human labor for NanoGPT speedrun tasks…
AI ResearchBreakthrough
90

METR's 'Expenditure Horizon': AI Agents Break Even at $3,300

METR's expenditure horizon metric shows AI agents break even at $0–$3,300 on NanoGPT, vs $2,500 per 1% speedup for humans. GPT-5 and Opus-4.1 pro lead, but blind spots remain.

the-decoder.com/Jul 27, 2026/3 min read/Multi-Source
cost analysisbenchmarkingresearch
A diagram showing the ZhiJing social intelligence system with Zing model outperforming GPT-5.5 on SoMBench social…
AI ResearchBreakthrough
88

CAS ZhiJing Beats GPT-5.5 on Social Cognition with FLARE Training

CAS ICT releases ZhiJing social intelligence system, Zing model beats GPT-5.5 on social cognition via FLARE training.

pandaily.com/Jul 27, 2026/3 min read
benchmarksai modelschina ai
NVIDIA researchers at a workstation display a diagram of the Molt RL framework scaling to trillion-parameter MoE…
AI Research
91

NVIDIA's Molt: 9.2K-Line RL Framework Scales to 1T-Parameter MoE Models

NVIDIA released Molt, a 9.2K-line PyTorch RL framework scaling to 1T-parameter MoE models via vLLM, targeting agentic tasks with fully-async rollout.

x.com/Jul 27, 2026/3 min read
open-sourcellm-trainingnvidia
Alibaba engineer points at dashboard showing RecGPT-V3 metrics on Taobao, with GMV up 3.97% and costs down 52.4%
AI Research
85

Alibaba's RecGPT-V3 Boosts GMV 3.97%, Cuts Serving Cost 52.4% on Taobao

Alibaba's RecGPT-V3, a stateful hybrid-modal recommender with continual memory, boosts GMV by 3.97% and cuts serving costs by 52.4% on Taobao.

x.com/Jul 26, 2026/3 min read
recommender systemsproduction aialibaba
Two Conformer encoder architecture diagrams labeled 220M and 600M parameters, with word error rate reduction…
AI Research
92

GigaAM Multilingual: 600M Conformer cuts WER 32-86% vs Whisper on 70+ languages

GigaAM Multilingual open-sources two Conformer encoders (220M/600M params) pre-trained on 2M hours across 70+ languages, achieving 32–86% relative WER reduction vs Whisper-large-v3 on Central Asian languages.

x.com/Jul 26, 2026/3 min read
open-sourcespeech-recognitionmultilingual-ai
Claude interface showing a browser agent executing tasks while a graph tracks prompt injection attempts falling to…
AI ResearchBreakthrough
100

Opus 5 Hits 0% Prompt Injection Rate in Browser Agents

Anthropic's Opus 5 with Auto Mode achieved 0% prompt injection success across 129 tests, challenging OpenAI's view that the problem is unsolvable.

the-decoder.com/Jul 25, 2026/3 min read/Widely Reported
anthropicai securityprompt injection
Bar chart comparing AI model scores on DeepSWE benchmark, GPT-5.6 Sol at 72.7% and Opus 5 at 68.8%, with a 3.9-point…
AI Research
100

GPT-5.6 Sol Leads DeepSWE at 72.7%, Beating Opus 5's 68.8%

GPT-5.6 Sol scores 72.7% on DeepSWE, beating Opus 5's 68.8%. The undocumented benchmark tests autonomous SWE agents.

x.com/Jul 25, 2026/3 min read/Widely Reported
gpt-5ai benchmarkscoding agents
Alibaba's RynnBrain 1.1 model card on Hugging Face with robot arm handling objects, showcasing embodied AI for…
AI Research
100

Alibaba Releases RynnBrain 1.1 Embodied AI Models at 2B-122B Scales

Alibaba released RynnBrain 1.1 on Hugging Face with 2B, 9B, and 122B-A10B MoE models for robot manipulation, but disclosed no benchmarks.

x.com/Jul 25, 2026/3 min read/Widely Reported
roboticsopen source modelsembodied ai
Two side-by-side images show a grid of colored shapes; left is a question prompt, right displays the pixel-based…
AI Research
85

Benchmark lets image models answer in pixels, not text

New 'Show, Don't Tell' benchmark tests spatial cognition via pixel-level outputs. GPT Image 2 solves 37% of cases missed by GPT-5.4, highlighting a gap in text-based spatial reasoning.

x.com/Jul 24, 2026/3 min read
spatial reasoningaibenchmarks