Clauday.

A new paper called SkillRise (arXiv 2607.26784) goes after one of the more honest weaknesses of today's agents: they relearn everything from scratch. Solve a task, throw away what …

NVIDIA's NeMo team dropped Molt today and it shot to the top of Hugging Face papers with 647 upvotes. It's a PyTorch-native training framework for agentic reinforcement learning, a…

The default recipe for a good search agent right now is: take a small model, distill trajectories out of a frontier model, ship it. HKUST's DeepSearch-World, arXiv 2607.07820, does…

Here's an embarrassing asymmetry in the agent stack: inference systems happily serve million-token contexts, but RL post-training — the thing that actually makes agents better at l…

Long-horizon agents trained with RL have a dumb problem: you only find out at the very end whether the whole episode worked, so the model gets almost no signal about which of its h…

Prime Intellect announced a 130 million dollar Series A at a 1 billion dollar valuation on July 8, led by Radical Ventures with Nvidia Ventures, Intel Capital, Dell Technologies Ca…

Everyone wants agents that run for hundreds of steps. Almost nobody can train them without the reinforcement learning going unstable. A new paper from the Zhipu/GLM team — Single-R…

Cognition dropped SWE-1.7 today, the coding model that runs inside Devin. The headline claim is not that it beats everyone, it doesn't. On FrontierCode 1.1 it hits 42.3% against GP…

A team out of Alibaba's Taobao group put out a paper with a blunt title: The Mirage of Optimizing Training Policies. The claim is that in RL for LLMs, we've been optimizing the wro…

Most benchmarks score the final answer. EvoPolicyGym scores the climb. It hands an agent a chunk of Python policy code, 16 environments, and a fixed 128-episode interaction budget,…