SkillOpt: Microsoft Research's Optimizer That Trains Agent Skills, Not Model Weights
Microsoft Research has introduced SkillOpt, an optimizer that treats natural-language agent skills as trainable parameters instead of fine-tuning model weights. It achieves best or tied-best performance in 52 out of 52 settings across 6 benchmarks and 7 models — including GPT-5.5 with Codex and Claude Code.
Read more →
Poetiq's Recursive Self-Improvement Tops LiveCodeBench Pro: Flash Model Beats Gemini Deep Think
Poetiq's Meta-System has set a new state-of-the-art on LiveCodeBench Pro (LCB Pro) by automatically constructing and optimizing a coding harness through recursive self-improvement. The system improved Gemini 3.1 Pro by 12.3%, pushed GPT-5.5 to 93.9%, and surpassed Google's own Gemini Deep Think — all without fine-tuning or privileged model access.
Read more →
Khala 1.0: The Open-Source Music Generation Model From Beijing's Central Conservatory
The Central Conservatory of Music in Beijing has released Khala 1.0, a fully open-source music generation model that scales acoustic token language models toward high-fidelity output. The release includes the research paper, model weights, code, and a live demo — a complete open-source music generation stack from one of the world's leading music institutions.
Read more →
AstraFlow: Infini-AI-Lab's Open-Source Dataflow RL System for Multi-Agentic LLM Training
Infini-AI-Lab has released AstraFlow, an open-source dataflow-oriented reinforcement learning system built specifically for training multi-agentic and multi-policy LLMs. It achieves 2.7x faster multi-policy collaborative RL training, reduces remote rollout sync from 28 GB to 1.5 GB, and supports elastic deployment across heterogeneous GPUs with zero-code configuration.
Read more →
Prime Intellect Introduces Renderers: 3x Throughput for Agentic RL Training
Prime Intellect's Renderers fix the token-message mismatch between RL trainers and agent environments, unlocking more than 3x throughput improvement on popular open models without changing the underlying architecture.
Read more →
TRIBE v2: Meta's Predictive Foundation Model That Predicts Brain Responses to Sight, Sound, and Language
Meta FAIR's TRIBE v2 is a trimodal foundation model trained on fMRI data that predicts high-resolution human brain activity in response to video, audio, and language — enabling zero-shot predictions for new subjects and tasks.
Read more →