All news

Matrix conference talk on redesigning the chat timeline data model for the next protocol generation.

After shipping Video.js v10, the Mux team looks at which new web APIs on the iPhone Duo matter for web video.

Analyzes what memory is actually worth to language agents.

OpenAI will train models on cached and simulated copies of the web instead of the live internet, citing risks from capable models.

Optimal learning of discrete distributions under person-level differential privacy.

Toolkit to audit differentially private graph releases.

TraceForge distills with the student own reasoning traces.

Finds token collapse in graph soft prompting for LLMs.

Position on recursive self-improvement in physical AI.

Generalist dexterous bimanual tracking from human demos.

Cut-topology FPGA logic architecture with better area efficiency.

Federated DP optimization via low-rank preconditioning.

Step-sequence reasoning to recognize mistakes in procedural videos.

Harness-evolving recursive distillation for continual learning.

Traces membership imprints in on-policy self-distillation.

Pre-history memory for humanoid loco-manipulation.

Advantage-based rule for when on-policy distillation should follow the teacher.

Explainable artifact detection in AI-generated videos.

Compares search space complexity for end-to-end ASR.

DIAL distills controllable instruction-adherence levels into LLMs.

Moonshot launches Kimi K3.1 Pro, its largest 3T-class open-weight model, plus a Highspeed variant.

IBM shows ensemble forecasting across four Granite time-series models behind one interface.

Linux kernel module unlocks full FP and BF16 instruction issue rates on Nvidia RTX 3000 cards.

Pose-aware risk mitigation for end-effector actions in vision-language-action robot models.

Agentic system for 3D virtual film cinematography.

Survey and audit of recommender research where LLM agents edit recommender configs or code and retrain them.

A user found Samsung TVs run an INT8 TensorFlow Lite model on 960x540 screenshots every few seconds to detect what game is being played.

Foundation4D unified perception and reconstruction for autonomous vehicles.

Change-aware multi-view commit message generation.

vLLM supports MiniMax-M3.1-Flash on day 0, including its NVFP4 KV cache, NVFP4 MoE weights and a draft model for speculative decoding.