What if AI agents could not only collaborate but also learn from their own failures and evolve? Researchers from Xi'an Jiaotong University, Lenovo, and the University of Sydney present a new survey. They introduce the LIFE progression: build agent capabilities, integrate them
@jiqizhixin
-
RL Scaling Laws for LLMs: From Pretraining to RL
By
–
Awesome blogpost! RL Scaling Laws for LLMs How scaling laws have evolved from pretraining to reinforcement learning… https://
cameronrwolfe.substack.com/p/rl-scaling-l
aws
… -
Presentation of the paper ‘Agents’ Last Exam’ on arXiv
By
–
—
Agents' Last Exam Paper: https://
arxiv.org/abs/2606.05405
— -
Agents’ Last Exam: Benchmarking AI on real-world professional tasks for economy
By
–
Are AI benchmarks really measuring what matters for the economy? Enter Agents’ Last Exam (ALE) — a benchmark that tests AI agents on long, real-world professional tasks, not just puzzles. It covers 1,000+ tasks across 55 fields mapped to U.S. job classifications. The
-
Fast KV Compaction reduces LLM memory by 50x without loss
By
–
What if you could reduce the memory of a language model by 50x in seconds without losing performance? Researchers from MIT present Fast KV Compaction via Attention Matching. They build compact key-value caches in the latent space that preserve
-
Gamma-World: NVIDIA’s generative multi-agent world model
By
–
What if your video game world could manage multiple agents — players, robots or AI — without any effort? NVIDIA, @Tsinghua_Uni, and Vector Institute present Gamma-World, a generative multi-agent world model. They invented two tricks: a 'simplex' encoding
-
Apple develops faster image codec using neural architecture search
By
–
What if your photos compressed 3x better while decoding faster than most ML codecs run on a V100 GPU? Apple researchers present a new learned image codec built for both perceptual quality and on-device speed. They used neural architecture search over millions of backbones to
-
Pion Optimizer Rotates Weight Matrices for Stable LLM Training
By
–
Can a new optimizer keep your LLM training stable without sacrificing performance? Researchers at Chinese University of Hong Kong, Max Planck Institute, and Westlake University introduce Pion. Instead of adding adjustments like Adam or Muon, Pion rotates each weight matrix
-
Syll: Open-source harness unifying APIs, CLI, GUI for AI agent learning
By
–
What if your AI agent could learn your workflow by watching you click, type, and run commands across apps, terminals, and websites? Tsinghua University and GigaAI present Syll — an open-source, self-hosted harness that unifies APIs, CLI, and GUI in one runtime. Users teach
-
WaDi: One-step image generation with better quality than multi-step diffusion
By
–
What if you could generate an image in one step with better quality than multi-step diffusion models? Researchers from Nankai University, Nanjing University, and NKIARI present WaDi. They discovered that during distillation, weight direction changes matter far more than