Latest

Qwen2.5-Math-7B posted a reported +20.10% accuracy gain on AIME2024, and self-supervised confidence training is what earned it. The method is called RLSC, short for Reinforcement Learning via Self-Confidence. And it uses the model's own confidence in its answers as the reward signal, which the paper says

AgentWorld grades multi-agent LLM teamwork inside a persistent 2D RPG. And that framing fixes a real gap in how agents get evaluated. AgentWorld is an open-source research platform built on the Kaetram game engine, designed to study multi-agent coordination in a persistent 2D multiplayer environment (project page). Agents connect through

Google DeepMind's Dream-RSI claims 162x fewer discovery-agent calls, and the trick is not a smarter model. It's cheaper practice. A VentureBeat report credits the framework with cutting discovery-agent calls by up to 162 times versus a baseline called SimpleTES, not by making the underlying model better.

56% to 95% mean success. Hand-engineered planner: 47%. That's the gap coding agents just opened on generalized task-and-motion planning, and it isn't subtle. The source is "Coding Agents for Generalized Task and Motion Planning Problems" (arXiv:2609.30233), written by Matteo Merler, Bowen Li,

TypeSafe launched Jev on September 15, 2026. And developers answered by shipping nearly 500 open-source projects around it within days. pg-jev and duckdb-jev showed up on GitHub almost immediately, LangChain published its own integration package. And by September 21 an official Browser Use pairing arrived with a documented 7.1-second

AI supply chain attacks on JavaScript packages stopped being theoretical on February 28, 2026. That day, a commit co-authored by Anthropic's Claude Opus added the malicious package @solana-launchpad/sdk to an open-source autonomous crypto trading project, planting a credential-stealing payload through a transitive dependency, per the Cloud Security