Topic

#Reinforcement Learning

12 articles on Reinforcement Learning — news, releases, guides and analysis from the SourceFeed engine.

Fine-Tuning Can't Teach What Pretraining Never Saw

A 5B model raised strictly on grade-school text shrugs off every trick meant to push it further.

A model raised on nothing past fifth grade shows RL, scaling, and prompting only amplify what pretraining put there.

Neon and Castform's benchmark-free demo still marks a real shift in who needs frontier models for RAG.

A tiny GRPO run on two workstation GPUs outscored five frontier models at catalog review — asterisks included.

The self-scored benchmark deserves skepticism, but the playbook behind it is real, cheap, and already running at Shopify scale.

Effort levels won because reasoning depth is trained, not dialed — and your cost model needs to catch up.

A from-scratch Super Mario world model clears short-horizon tests, then stalls on distant goals.

Open weights, real Runpod jobs, and held-out transfer make meta-RL a forkable template, not theater.

Cognition's new model matches frontier performance at a fraction of the cost by solving core RL scaling bottlenecks.

By optimizing both the reasoning loop and the code output, these MIT-licensed models bring native agentic capabilities to local hardware.

Qwen-AgentWorld replaces heavy external simulators with a native language world model to train and evaluate AI agents.

VibeThinker-3B matches flagship math and code reasoning, then collapses on general knowledge. That gap is the real lesson.