A flow-wide method improves low-temperature generation

When a flow language model is pushed toward more likely answers by lowering temperature, does its generation process still match the intermediate states it encounters?

Standard tempering changes the implied states but keeps the original sampling motions, leaving a mismatch between what the model was trained on and how it generates.

BCT aligns temperature-aware training and sampling with the same intermediate-state bridge, then tests whether this improves low-temperature generation.

The researchers compared BCT-L with standard tempering after both models continued from the same released S-arch checkpoint for 2,000 updates.

On GSM8K, where generated programs are checked against reference answers, BCT-L reached 27.0% mean accuracy at temperature 0.1 across three generation seeds, versus 17.5% for the continued standard-tempering model.

A control found that continuing the original training recipe alone changed standard-tempering accuracy by no more than one percentage point at tested temperatures on either S-FLM backbone.

Together, these results support the view that temperature-aware training contributes to BCT-L’s advantage in this setting, though the samplers also differ.

The GSM8K comparison changes both the continuation training objective and the inference sampler, so it does not isolate training as the cause of the gain. The evidence covers specific models, tasks and two bridge families; it does not establish results for other settings.