
EAGC Fixes the Hidden Gradient War Holding Back Open-World AI Discovery
A new plug-and-play module fixes a subtle optimization conflict in Generalized Category Discovery, delivering up to 9.6 percent accuracy gains.
20 stories

A new plug-and-play module fixes a subtle optimization conflict in Generalized Category Discovery, delivering up to 9.6 percent accuracy gains.

VideoLingo is an open-source pipeline that replaces an entire subtitle team — transcription, translation, and voice dubbing — with one automated workflow

AgentAlphaAGI open-sources Idea2Paper, a modular agent framework that turns a raw research idea into a structured scientific manuscript using a pre-built knowledge graph

dots.ocr / dots.mocr is a 3B open-source VLM that parses documents, tables, formulas, and 126 languages in a single model pass, beating Gemini 2.5 Pro on multilingual benchmarks.

AgentScope 2.0 ships a full-stack agent platform with sandboxed execution, triple-layer permissions, and built-in multi-tenant deployment — all open-source from Alibaba's Tongyi Lab.

Alibaba's Qwen team ships an 80B mixture-of-experts coder that activates just 3B parameters yet matches Claude Sonnet on agentic coding tasks.

A hierarchical memory system inspired by OS memory management gives LLM agents persistent, personalized recall across long conversations.

A new detection transformer family plugs DINOv3 features into eight model sizes, with the S variant becoming the first sub-10M model above 50 AP on COCO.

Strix ships autonomous AI agents that run your code, exploit real vulnerabilities, and hand back working proof-of-concept exploits instead of static-analysis noise.

MiniMind-3 lets you train a working 64M-parameter LLM from scratch in 2 hours on a single GPU, with full RL, tool calling, and agentic training included

The HKUDS graph-enhanced RAG framework accepted at EMNLP 2025 crosses 30K GitHub stars, with multimodal support and a growing ecosystem of extensions now in active development.

Rakuten open-sourced a 671B-parameter Mixture-of-Experts model tuned for Japanese, then got caught reusing DeepSeek-V3 weights without proper attribution.

Datalab's new 5B parameter OCR model hits state of the art on olmOCR, adds 90+ languages, and doubles throughput over Chandra 1.

ChatDev 2.0 (DevAll) turns multi-agent AI orchestration into a zero-code visual experience, backed by NeurIPS-accepted research from Tsinghua University's OpenBMB lab.

Meta and CMU released a single transformer that handles 12+ metric 3D reconstruction tasks from any mix of images, poses, depth, and calibration inputs.

A four-agent runtime uses ACT-R cognitive twins and Bayesian safety gates to keep LLM copilots from overriding nuclear operators.

Six frontier language models independently audited a published gesture-recognition paper and all converged on the same verdict: the 99% accuracy came from data leakage.

Browser-Use ships its own open-source browser agent model and a public benchmark, staking a position at the center of the web automation stack

OpenWolf is open-source middleware that gives Claude Code persistent project memory and token tracking, cutting usage by up to 80% through six invisible lifecycle hooks.

Newton, a free GPU-accelerated physics engine from NVIDIA, Google DeepMind, and Disney Research, is closing the sim-to-real gap in robotics with 475x speedups and production deployments at Samsung and Skild AI.