AllLLMRoboticsStartupsResearchEthicsBusinessHardwareAutonomousMedicineCreative

All news

10671
APEX Compresses Diffusion Latents 128x and Beats a 256-Token Model on ImageNet-512
Research·arXiv cs.CV

APEX Compresses Diffusion Latents 128x and Beats a 256-Token Model on ImageNet-512

Oct 12, 2026·2 min
SWIM saves compute: audio-video generation gains up to 30.7% over Best-of-N
Research

SWIM saves compute: audio-video generation gains up to 30.7% over Best-of-N

Oct 12, 2026·2 min
SAM 2.1 fine-tuned on facades and interiors: mIoU rises from 49.9% to 69.6%
LLM

SAM 2.1 fine-tuned on facades and interiors: mIoU rises from 49.9% to 69.6%

Oct 12, 2026·2 min
SAIL Keeps Its Target in Frame Using Audio and Text, and Beats Rivals on Two Benchmarks
Research

SAIL Keeps Its Target in Frame Using Audio and Text, and Beats Rivals on Two Benchmarks

Oct 12, 2026·2 min
Wearable Stereo Camera Estimates Food Volume With 9.1% Error, No Reference Object Needed
Hardware·arXiv cs.CV

Wearable Stereo Camera Estimates Food Volume With 9.1% Error, No Reference Object Needed

A pipeline of FoundationStereo, SAM 2 and TSDF builds a 3D model of a meal from a single stereo clip from a wearable. Mean volume error across five dishes: 9.1%.

Oct 12, 2026·2 min
Researchers Compare CNN, ViT and Mamba Attention Against Human Gaze
Research·arXiv cs.CV

Researchers Compare CNN, ViT and Mamba Attention Against Human Gaze

Three architectures, three domains, three metrics: eye tracking shows CNNs look broadly, ViTs lock onto micro-features, and Mamba is strong in medicine but diverges…

Oct 12, 2026·3 min
Stroop Test Shows VLMs Read the Text in an Image Before They See the Color
Research·arXiv cs.CV

Stroop Test Shows VLMs Read the Text in an Image Before They See the Color

Researchers ran contrastive and generative VLMs through the classic Stroop test. When text and color conflict, the models more often answer with the written word.

Oct 12, 2026·2 min
ForcingBooster Halves Denoising Steps in Autoregressive Video Generation
Research·arXiv cs.CV

ForcingBooster Halves Denoising Steps in Autoregressive Video Generation

ForcingBooster assigns denoising steps per patch, so simple regions stop early. The result: roughly 50% fewer steps at quality comparable to a four-step baseline.

Oct 12, 2026·2 min
Top themes
01MemReg Outperforms Pairwise Methods and SLAM in LiDAR Point Cloud RegistrationOct 12, 2026·2 min02Video-Language Models Confuse Car Maneuvers: LoRA and Probes Fix the ErrorOct 12, 2026·2 min03MIRAGE anchors video-based 3D reconstruction to LiDAR and beats COLMAP on three datasetsOct 12, 2026·2 min04OMNI: One Compact Encoder Absorbs the Knowledge of 10 Models for PathologyOct 12, 2026·2 min05TerraAgent checks whether the AI actually looks at the image: hallucinations drop from 17.1% to 9.8%Oct 12, 2026·2 min
AVA improves multimodal ICL accuracy by 1.7 points without any training
Research·arXiv cs.CV

AVA improves multimodal ICL accuracy by 1.7 points without any training

Oct 12, 2026·2 min
Coral-KV Compresses Video Model KV-Cache by 37–48% and Lifts PSNR by Up to 4.95 dB
Research·arXiv cs.CV

Coral-KV Compresses Video Model KV-Cache by 37–48% and Lifts PSNR by Up to 4.95 dB

Oct 12, 2026·2 min
Black Lake: Chinese Factories Skip Digitization and Go Straight to AI Agents
Business·SCMP Tech

Black Lake: Chinese Factories Skip Digitization and Go Straight to AI Agents

Oct 12, 2026·2 min
T. Rowe Price: 55% of Hong Kong Residents Expect AI to Hit Their Retirement, but Only One in Five Feels Ready
Business·SCMP Tech

T. Rowe Price: 55% of Hong Kong Residents Expect AI to Hit Their Retirement, but Only One in Five Feels Ready

Oct 12, 2026·2 min
Genetic Optimizer Cuts Aerosol Sensor Placement Time in the UK by 60%
Research·arXiv cs.LG

Genetic Optimizer Cuts Aerosol Sensor Placement Time in the UK by 60%

Researchers built a neural process into DeepSensor and cut the time to compute aerosol monitoring station placement in the UK from 25 to 10 minutes.

Oct 12, 2026·2 min
Kronecker Product Attention Compresses the KV-Cache Up to 10x and Matches MLA on Quality
Research·arXiv cs.LG

Kronecker Product Attention Compresses the KV-Cache Up to 10x and Matches MLA on Quality

An arXiv preprint proposes KPA, which stores keys and values as sums of Kronecker products. It claims up to 10x KV-cache compression with quality on par with MHA…

Oct 12, 2026·2 min
TimesFM 3.0 Cuts University Building Energy Forecast Error by 43%
Research·arXiv cs.LG

TimesFM 3.0 Cuts University Building Energy Forecast Error by 43%

Pretrained TimesFM 3.0, with no local fine-tuning, cut day-ahead MAE for a university building by 43% versus a naive baseline. XGBoost stays close behind.

Oct 12, 2026·2 min
Wasserstein BaryBooster Turns Diffusion Realism into Event Forecast Accuracy
Research·arXiv cs.LG

Wasserstein BaryBooster Turns Diffusion Realism into Event Forecast Accuracy

Oct 12, 2026·2 min
DiffusionART speeds up RL fine-tuning of video models 12x, with no EMA tuning
LLM·arXiv cs.LG

DiffusionART speeds up RL fine-tuning of video models 12x, with no EMA tuning

Oct 12, 2026·2 min
A Physical Transformer Built Entirely From Transistors Learns In-Context Learning in Simulation
Research·arXiv cs.LG

A Physical Transformer Built Entirely From Transistors Learns In-Context Learning in Simulation

Oct 12, 2026·2 min
Theory Shows Masked Diffusion on Transformers Learns from Õ(LK/ε) Samples
Research·arXiv cs.LG

Theory Shows Masked Diffusion on Transformers Learns from Õ(LK/ε) Samples

Oct 12, 2026·2 min
SAZO speeds up LLM fine-tuning without backprop: up to 6× fewer queries to reach the same loss
Research·arXiv cs.LG

SAZO speeds up LLM fine-tuning without backprop: up to 6× fewer queries to reach the same loss

Oct 12, 2026·2 min
Holomorphic Neural Networks With Singular Functions Beat All Methods on Laplace Benchmarks
Research

Holomorphic Neural Networks With Singular Functions Beat All Methods on Laplace Benchmarks

Oct 12, 2026·2 min
Cox Regression Learns From Truncated Data Orders of Magnitude Faster Than EM
Research

Cox Regression Learns From Truncated Data Orders of Magnitude Faster Than EM

Oct 12, 2026·2 min
CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%
Research

CacheScope speeds up diffusion models by up to 3.57x and cuts LPIPS by 70%

Oct 12, 2026·2 min
AI Error in Estimating Hidden Solar Generation Rises from 0.104 to 0.185 kW
Research·arXiv cs.LG

AI Error in Estimating Hidden Solar Generation Rises from 0.104 to 0.185 kW

An arXiv study tested deep models on Ausgrid and Pecan Street data: daytime MAE is 0.185 kW for new homes vs 0.104 kW for known customers. Fine-tuning on five homes…

Oct 12, 2026·2 min
CircuitTree assembles 39 HuBERT phoneme circuits into a syllable transcriber with 64.8% accuracy
Research·arXiv cs.LG

CircuitTree assembles 39 HuBERT phoneme circuits into a syllable transcriber with 64.8% accuracy

A new algorithm merges small HuBERT subnetworks into a single model. It recognizes open syllables with 64.8% accuracy, and 99.5% of the circuit components are…

Oct 12, 2026·2 min