Mind Lab pushes RL past 2M tokens without more GPUs

Mind Lab says its LongStraw method trains reinforcement learning models past 2 million tokens of context without adding GPUs, per the July 17 paper.

Long-context RL work has mostly been gated by cluster access, not algorithmic ideas. If the fixed GPU budget holds, smaller teams get a shot at training runs that used to need bigger clusters.

The paper landed on Hugging Face's daily papers list with 174 upvotes, per the July 17 posting. "Fixed GPU budget" is the phrase carrying the claim.

No bigger model, no new architecture, just the same compute stretched over a longer context window during RL training.

Wall-clock cost isn't in the paper, and neither is whether the technique holds past 2 million tokens.

What's worth watching isn't the 2 million token number, it's whether anyone outside Mind Lab reproduces the result on the same GPU count. A paper that claims to solve the compute wall should be checkable by teams that don't have a frontier lab's cluster. Nobody has yet.