Latest stories
D-JEPA Aligns Latent World Models With Real Execution Outcomes
A new latent world model, D-JEPA, learns decision-relevant relations among candidate futures from executed outcomes, improving act...
IndicBankBench Tests Whether Banking Assistants Are Reliably Safe, Not Just Plausible
A new 799-case benchmark for Indian retail banking evaluates language model assistants across four stages — safety, tool use, resp...
AgentWorld Benchmark Exposes Multi-Agent LLM Collaboration Gaps
Hugging Face researchers introduce AgentWorld, a 100-task benchmark for long-horizon multi-agent collaboration in an MMORPG sandbo...
Diffusion Model Refines Satellite 3D Maps, Cutting Elevation Error by Up to 43%
Researchers adapt Stable Diffusion 3 to refine noisy photogrammetric Digital Surface Models using Pléiades imagery, reducing RMSE...
Tactile-JEPA Brings Topology-Aware Self-Supervision to Robot Electronic Skins
A new self-supervised pre-training method for distributed tactile sensors predicts masked sensing elements using the sensor connec...
LastOPD Fixes Latent Distillation Collapse in Small Language Models
Hugging Face paper LastOPD identifies a collapse failure in latent on-policy distillation and proposes applying latent supervision...
TimeEvo Lets Time Series Agents Build Their Own Tool Libraries From Failure
A new Hugging Face paper argues that pre-supplied tool libraries misalign with agent needs and can silently degrade answers, propo...
RayOrch Brings Lineage-Aware Dataflows to Foundation-Model Data Prep
A new Hugging Face paper introduces RayOrch, a programming model and distributed engine that preserves parent-child relationships...
Hugging Face Paper Proposes Disaggregated Quantization to Split LLM Prefill and Decode
A new research paper introduces disaggregated quantization (DQ), which assigns separate precision formats and weights to the prefi...
Basis Completes 50-Tab Tax Workbook Twice as Fast with GPT-6 Astra
OpenAI reports that GPT-6 Astra finished a 50-tab tax workbook twice as fast as GPT-5.6 Sol, with improved intent understanding th...
OpenAI Calls on Builders to Share Their Codex Stories for 'Codex Originals'
OpenAI is gathering real stories from builders, tinkerers, researchers, and creators who use Codex, inviting them to submit their...
CARD Brings Cluster-Level Personalization to LLM Text Generation
A new hierarchical framework clusters users by shared writing style, learns group-specific LoRA adapters, and applies personalizat...