Adapted DINOv2 generalises better than task-specific medical segmentation nets
Only a small part of DINOv2 was adapted; in-distribution U-Nets stayed sharper and caught small lesions more often.
Stories per day · 30 days
4 RSS
Relatedforecasting4time-series4continual-learning3pretraining3robotics3agent-evaluation2anomaly-detection2open-weights2reinforcement-learning2supervised-fine-tuning2work-agents2agent-harness1agentic-robotics1agents1ai-startups1aws1azure1benchmarks1
67 stories, newest first · one card per story
Stories tagged #foundation-models
Bonbon shows zero-shot molecular skills without structural training
A 1.6B model trained only on protein-ligand sequences reports binding-site, mechanism, and bond predictions plus a 61.5% wet-lab hit rate.
TimeSeek pretrains one model for multi-task time series analytics
After large-scale pretraining, authors report state-of-the-art forecasting, anomaly detection, and classification in few-shot and zero-shot settings.
Paper measures sampling-frequency invariance of four time-series models
MOMENT, Chronos-2, Moirai-MoE and TimesFM are scored with new FSI, FDI and CDI metrics.
LEASH cuts harm rates in agentic robotics safety benchmark
RoboRisk runs 92 simulated and 13 physical tasks; six foundation models generally showed high harm rates.
NVIDIA Announces Investments in Korean AI Startups
NVIDIA will invest in Korean startups Point2, Motif Technologies, RLWRLD and Upstage following the MSIT-NVIDIA forum.
Channels4World releases 100,000 outdoor RF scenes for wireless models
The 29.14-TB corpus spans 3.5 and 7 GHz with a 90K/10K scene-level train-test split.
Study confirms 103 of 233 AI-discovered cloud threats on live clouds
Three foundation models reached 75.2% precision, but multi-model agreement did not improve confirmation rates.
PhysioWave-T cuts biosignal token length with wavelet scale folding
A 30M EEG encoder leads 8 of 9 metrics versus BENDR, BIOT, LaBraM and EEGPT, and RALF lifts seizure sensitivity from 61.4% to 74.4%.
Drex Open Decision Models Achieve State of the Art
Drex family of decision models including 8B to 35B parameters sets new benchmarks in accuracy and speed over autoregressive LLMs.
TailSteer Improves Multi-turn Image Editing Robustness
TailSteer reduces cumulative drift in multi-turn editing while preserving base model performance.
Geometric Morphology Learning integrates geometry into foundation models
New framework enables AI models to learn geometric and topological structures for reasoning and action.
Paper unifies foundation-model pretraining, SFT and RL under one algorithm
A single likelihood trainer mixes static data and RL generations and is reported to improve the accuracy-compute tradeoff.
Paper unifies pretraining, SFT, and RL under one likelihood algorithm
The authors recast RL stages as maximum-likelihood fits to model-generated targets and mix them in one trainer.
Timestral models lead GIFT-Eval, TIME and fev-bench
Four probabilistic forecasters from 5M to 209M parameters ship with Hugging Face weights and GitHub inference code.
Liquid AI releases open-weight multimodal LFM3 models around 2B
LFM3-1.7B-Base was pre-trained on 22T tokens, with a 256K context after mid-training.
Paper frames work agents as model-plus-harness systems for verifiable tasks
It defines work persistence and openness, five work forms, and model, harness, and joint optimization.
Paper frames work agents as systems that deliver verifiable outcomes
It defines work persistence and openness, five work forms, and five capability families for reliable execution.
Flora open robotics data engine reports SOTA caption and reward scores
Authors say Flora labels 200,000 hours a day and is about 125 times cheaper than GPT-6 Astra.
F2G-Pose estimates category-level RGB-D object pose in one pass
It reports state-of-the-art accuracy on SOPE and ROPE at 28 FPS for pose-only inference, with code and weights released.
Paper proposes memory-centric continual learning for foundation-model agents
MCCL judges memory by later task gains, with a taxonomy, two-loop lifecycle, eval protocol, and scaling model.
TS2Vec-FM freezes a 1.6M-parameter time-series encoder
Kept frozen, it scores 83.3% on UCR-128 and 75.3% on UEA-27, within 0.3 points of MantisV2.
CERN and Mistral plan a foundation model for collider physics
The planned Large Physics Model would train on billions of simulated LHC collisions and is not released.
ConSDL improves sparse dictionary consistency over SAE baselines
On 16 synthetic settings recovery was near perfect; consistency rose 17.6% for dictionaries and 35.9% for coefficients.