Choosing which problems an AI model should practice could make reinforcement learning more productive. In a November 2, 2026 arXiv preprint, researchers affiliated with the universities of Michigan and Edinburgh report that FADE improves average mathematical reasoning accuracy by tracking changing model ability separately from each prompt’s difficulty. Peer review is not verified, and the supplied evidence contains no external replication.
The method targets reinforcement learning with verifiable rewards, where an answer checker scores generated responses. In group-relative training, a prompt whose responses all pass or all fail supplies no within-group relative reward signal. Schedulers therefore seek prompts likely to produce a mixture. Tracking each prompt independently becomes difficult when the model improves between infrequent visits: short memory can disrupt difficulty rankings, while long memory leaves success estimates stale.
FADE uses frozen text embeddings reduced to 32 whitened dimensions as shared skill coordinates. A rapidly changing ability vector captures model progress, while slowly updated offsets represent individual prompt difficulty. Pólya–Gamma-based filtering lets outcomes from one batch update predictions for other prompts with related coordinates. An uncertainty-aware utility then helps select prompts or allocate generated responses. The Gaussian update is exact conditional on fixed auxiliary precisions and its assumptions; success-probability integration and scheduling utility remain approximations.
The authors trained Qwen2.5-Math-1.5B and Llama-3.2-1B-Instruct on 150,000 NuminaMath problems. They replaced estimators within four scheduling configurations—DAPO-Filter, PCL, GRESO and GVM—while retaining each host’s scheduling rule. Evaluation covered MATH-500, AMC 2023, Minerva Math, text-only OlympiadBench mathematics and combined AIME 2024–2025. Average@8 measures mean exact-match success across eight draws per problem at temperature 0.7, rather than the chance of obtaining at least one correct answer. The reported overall score weights the five benchmarks equally.
FADE raised that average in all eight scheduler–model combinations, with gains over native estimators of 1.8–2.9 percentage points for Qwen and 1.3–1.8 points for Llama. Under GVM, Qwen rose from 42.7% to 45.5%, and Llama from 12.3% to 14.1%. FADE did not win every individual benchmark comparison.
In a separate replay of a GRPO training trajectory, FADE’s prediction error against eight-response reference estimates was 0.118 for Qwen and 0.102 for Llama, compared with CDAS’s 0.153 and 0.138. Spearman ranking correlations increased from 0.69 to 0.82 and from 0.67 to 0.79, respectively. Reported estimation overhead was 0.06 hours, versus 42.8 and 45.1 hours for on-policy probing; probing nevertheless delivered better prediction error and ranking correlation.
The evidence leaves important uncertainties. Independent training-run counts, between-run variability and confidence intervals are not provided in the supplied results. Small competition datasets make individual gains uncertain. Claimed budget matching also needs clarification: 450 steps × 256 prompts × eight responses equals 921,600 responses, not the stated 1.44 million. Tracking used noisy reference rates on 2,048 held-out training prompts, without described overlap checks or tuning separation. Frozen embeddings may poorly represent some skills, and correlated rewards require extending the model. These results support FADE within the tested configurations; they do not establish universal superiority or prove that better calibration caused the accuracy gains.
Source: https://arxiv.org/abs/2611.03625
Cover illustration generated with AI; conceptual, not research evidence.