CORE training framework ties reinforcement learning to explicit math concepts
The arXiv paper introduces Concept-Oriented REinforcement, a method that gives language models fine-grained conceptual supervision during math training instead of rewarding only final answers.
In plain words
A research team proposes a reinforcement learning method that trains language models on short quizzes tied to specific math concepts, not just final answers. The goal is to close the gap between models that can restate a definition and models that can actually apply it when solving problems.
What was announced
The authors argue that popular RLVR pipelines reinforce final answers but give little fine-grained conceptual signal, so models improve at pattern reuse rather than genuine application. As a sanity probe, they report that LLMs can restate definitions yet fail concept-linked quizzes, which they use to quantify the conceptual reasoning gap. CORE starts from a textbook resource linking verifiable exercises to concise concept descriptions, then synthesizes concept-aligned quizzes, injects brief concept snippets during rollouts to elicit concept-primed trajectories, and reinforces conceptual reasoning through trajectory replacement after group failures, a forward-KL constraint aligning unguided with concept-primed policies, or standard GRPO on the quizzes.
According to the paper, across several models CORE delivers consistent gains over vanilla and SFT baselines on both in-domain concept-exercise suites and diverse out-of-domain math benchmarks. The framework unifies direct training on concept-aligned quizzes and concept-injected rollouts under outcome regularization and is described as algorithm- and verifier-agnostic. The claims come from the paper's abstract alone, so independent replication and full experimental detail are not covered by this source.
What it does not show
- Evidence is limited to the paper's abstract; full experimental setup, model sizes and benchmark numbers are not available in the supplied passages.
- The reported gains are the authors' own claims and have not been independently confirmed.
Sources
Source-support review completed 19 Nov 2026. This story was drafted with AI assistance and checked against retained source passages. Automated checking is not independent verification.
Coverage: arXiv: abstract.
Updates and corrections
No corrections so far.
First published 19 Nov 2026
New to the subject? AI from the Ground Up starts from the beginning.