FADE: Factored Ability--Difficulty Filtering for Online RLVR Scheduling
The authors report better accuracy and calibration at the same rollout budget, with less than 0.6 milliseconds of added work per step.
In group-based RLVR, prompts whose sampled answers all pass or all fail produce no relative learning signal, wasting generation budget. FADE separates the policy’s changing abilities from each prompt’s more stable difficulty to estimate which prompts will yield informative groups. An ability update also adjusts estimates for unseen prompts that share the same skills. The reported gains span five reasoning benchmarks, two base models and four scheduling approaches.
ArXiv · AI/CL/LG's note
In group-based RLVR, prompts whose sampled answers all pass or all fail produce no relative learning signal, wasting generation budget. FADE separates the policy’s changing abilities from each prompt’s more stable difficulty to estimate which prompts will yield informative groups. An ability update also adjusts estimates for unseen prompts that share the same skills. The reported gains span five reasoning benchmarks, two base models and four scheduling approaches.
ArXiv · AI/CL/LG's note
score 4