AI Science Tracker
EN

RLVR training expands hard-problem coverage but can harm easier-task performance

University of Notre Dame; The Pennsylvania State University; Virginia Tech; VinUniversity; University of Stuttgart · 2 Nov 2026 · computer science

Discovery

RLVR training expands hard-problem coverage but can harm easier-task performance

Source check: Source checked. Not independent scientific confirmation.

Finding: Controlled RLVR experiments found that training on harder problems could expand observed solution coverage, while prolonged training also reduced performance on easier problems. The result concerns the tested models, training schedules and problem sets; it does not establish a general increase in reasoning capability.

Technology: Reinforcement learning with verifiable rewards (RLVR)

Publication: Preprint · Test context: direct test. Catalog categories, not an evidence grade.

AI contribution
RLVR training changes model parameters from verifiable task rewards; the study measures resulting problem-solving performance.
Test and result
Controlled model-training experiments compared solution coverage across hard and easier problems and observed both increased hard-problem coverage and degradation on easier problems under prolonged training.
Limitation / still unknown
The finding is bounded to the tested models, training schedules and problem sets; broad reasoning gains are not established.
Evidence sources and annotation context

Evidence context (descriptions follow the cited sources; test type is a catalog classification):

Evidence sources: https://arxiv.org/abs/2611.04061v1 · https://arxiv.org/html/2611.04061v1

Source checked · editorial details

The source page was checked against this entry. Independent confirmation is not assessed by this label.

Source type: paper. Paper: linked separately.

Review date: 22 Nov 2026. Review scope: Primary result-bearing material for the stated method, result and evaluation, with bibliographic version information..

No follow-up recorded; this is not a claim that no follow-up exists. Follow-up review scope and date are not recorded. No source-supported relationships recorded.

Editorial assessment: Impact 3/5 · Validation 3/5 · Novelty 3/5. Impact and novelty are editorial judgments; validation follows the annotated test type where one exists. Not an evidence grade. Scoring criteria.

Open in the full catalog for filters, charts and related discoveries. Browse computer science discoveries.