FADE
FADE is a bilinear state-space filtering method for online RLVR training scheduling. By decoupling model capability from prompt difficulty, it predicts non-degenerate sampling groups more accurately and improves training efficiency.
Timeline (last 90 days)
FADE (Factored Ability–Difficulty Estimation) is an online RLVR scheduling method that separates model capability drift from intrinsic prompt difficulty via bilinear state-space filtering.
FADE leverages Pólya–Gamma augmentation and the Woodbury matrix identity to achieve closed-form updates of the mini-batch capability posterior, with per-step overhead below 0.6 milliseconds.
The paper evaluates FADE on five reasoning benchmarks, two base models, and four scheduling paradigms, with all experiments conducted under matched generation budgets, and FADE consistently improves downstream accuracy and calibration.
FADE improves performance under different scheduling paradigms, indicating that its approach of decoupling capability and difficulty is general.
The plug-and-play solution provided by FADE requires no modification to existing RLVR pipelines and yields more stable training signals solely by improving the scheduler.
In the future, FADE may be extended to more complex reward structures or multimodal tasks.
All Facts (6)
FADE (Factored Ability–Difficulty Estimation) is an online RLVR scheduling method that separates model capability drift from intrinsic prompt difficulty via bilinear state-space filtering.
50%UnverifiedFADE leverages Pólya–Gamma augmentation and the Woodbury matrix identity to achieve closed-form updates of the mini-batch capability posterior, with per-step overhead below 0.6 milliseconds.
50%UnverifiedThe paper evaluates FADE on five reasoning benchmarks, two base models, and four scheduling paradigms, with all experiments conducted under matched generation budgets, and FADE consistently improves downstream accuracy and calibration.
50%UnverifiedFADE improves performance under different scheduling paradigms, indicating that its approach of decoupling capability and difficulty is general.
50%UnverifiedThe plug-and-play solution provided by FADE requires no modification to existing RLVR pipelines and yields more stable training signals solely by improving the scheduler.
50%UnverifiedIn the future, FADE may be extended to more complex reward structures or multimodal tasks.
50%