FADE separates ability and difficulty for RLVR scheduling
Original title: FADE: Factored Ability--Difficulty Filtering for Online RLVR Scheduling
Bilinear filter updates policy competence instantly while stabilizing item intercepts.
Group-based RLVR suffers from degenerate rollouts when pass rates entangle fast-evolving policy skills with slow-changing prompt difficulty. FADE uses a bilinear state-space model with whitened skill loadings and a fast-drifting ability vector updated via Pólya-Gamma augmentation and Woodbury steps. This yields closed-form utility for non-degenerate groups and consistently improves accuracy and calibration across benchmarks and schedulers.
- Single counters scramble ordering at short or long memory
- Closed-form posterior updates transfer competence to shared skills
- Less than 0.6 ms per step overhead
From the original article
1 min read at arxiv.orgGroup-based reinforcement learning with verifiable rewards (RLVR) learns only from prompts whose sampled completions disagree: all-pass and all-fail groups produce zero relative advantage and burn generation budget without moving the policy.
By Tomas Duncun · arxiv.org
A short excerpt; the article and its images belong to the publisher. Summary, key facts and assessment above are AI-written from the article and can be wrong.