Breaking01 / AI
FlashREINFORCE trains agents with one run per input
FlashREINFORCE researchers said their method reduced repeated runs per input while maintaining training stability.

Researchers announced FlashREINFORCE on 17 Nov, an agent training method that uses one run per input.
According to the paper, the method is designed to train without a separate model that estimates the value of actions. It removes the process of running the same input repeatedly to collect results for comparison.
In researcher-published benchmark figures, FlashREINFORCE achieved a mean accuracy of 38.0 on Qwen2.5-Math-1.5B. The paper lists 256k runs for FlashREINFORCE and 512k for GRPO, the comparison method.
The researchers said the method controlled changes in the agent's action-selection rules during training. It also controlled negative training updates amplified by long execution histories.
The researcher-published benchmark figures put FlashREINFORCE 1.7 percentage points ahead of GRPO in mean accuracy. The paper explains that methods which group repeated runs of the same input limit the range of inputs used for training under a fixed run budget.
The researchers did not give a release schedule for the training code in the abstract.