ICLR 2026

CORE: Concept-Oriented Reinforcement for Bridging the Definition–Application Gap in Mathematical Reasoning

Turning textbook concepts into training signals for stronger mathematical reasoning

Zijun Gao1    Zhikun Xu2    Xiao Ye2    Ben Zhou2
1University of Illinois Urbana–Champaign    2Arizona State University

Abstract

Large language models (LLMs) have demonstrated impressive mathematical problem-solving capabilities. However, a significant gap exists between an LLM's ability to recite mathematical definitions and its capacity to effectively apply these concepts in novel problem-solving scenarios, a phenomenon we term the definition–application gap. To systematically investigate and address this gap, we leverage a classical mathematical textbook with clear associations between concepts and exercises, enabling both qualitative gap analysis and targeted training.


We introduce CORE (Concept-Oriented REinforcement), an RL-based framework with three training recipes: CORE-Base trains directly on concept-aligned quizzes; CORE-CR (Concept Replacement) dynamically replaces failed trajectories with concept-guided ones; and CORE-KL uses forward KL divergence to distill concept-primed reasoning into the base policy. Experiments across 11 math benchmarks show consistent improvements over strong baselines, with gains of up to +9.6% on TheoremQA and +9.3% on the in-domain Textbook test set.


Framework

CORE Framework Overview

Overview of the CORE framework. For a given query, the policy model generates multiple candidate solutions. If any is correct, CORE-Base proceeds with standard policy update. When all solutions fail, CORE activates concept-guided correction: the Concept Recall module retrieves relevant domain knowledge, and Concept Injection re-prompts the model. CORE-CR replaces failed trajectories with concept-primed ones, while CORE-KL distills the reasoning via forward KL loss.


Three Training Recipes

CORE provides three complementary approaches to inject conceptual understanding into RL training:

Concept Quizzes

CORE-Base

Standard GRPO trained directly on 1,110 concept-aligned quizzes generated from the textbook. The data-level recipe: the model learns to implicitly connect concepts with problem-solving through rich concept-grounded training signals.

Trajectory Replacement

CORE-CR

When all N responses fail, retrieves the relevant concept, re-prompts the model, and replaces failed trajectories with concept-primed ones. Adds a bonus reward r=0.4 to guide learning.

KL Alignment

CORE-KL

Instead of replacing trajectories, adds a forward KL divergence loss to align the base policy with concept-primed reasoning. Uses asymmetric weights: λ=0.03 for correct, λ=0.005 for incorrect.


Results

CORE achieves consistent improvements across in-domain and out-of-domain math benchmarks.

+9.3
CORE-KL
Textbook (in-domain)
+9.6
CORE-KL
TheoremQA
+5.8
CORE-CR
OlympiadBench
+3.4
CORE-CR
TabMWP
Results improvement chart

Qwen2-Math-7B (Table 2)

MethodTextbookGSM8KASDivMAWPSTabMWPMATHMMLU-STEMGaokaoCounterMathTheoremQAOlympiadBench
Vanilla46.489.895.196.890.269.172.955.313.234.628.7
SFT45.086.694.196.685.659.472.446.516.744.219.7
CORE-Base50.790.895.497.292.671.172.959.513.540.433.9
CORE-CR52.191.195.797.393.671.472.658.415.542.334.5
CORE-KL55.790.795.597.590.670.573.159.515.844.232.9

All results under SC@21 (self-consistency with 21 samples, T=0.7). Bold = best in column.


Data

We curate a corpus from Advanced Algebra (3rd Edition, Yao & Xie, 2015), a canonical Chinese textbook manually translated to English.

236
Concept DefinitionsCore mathematical concepts from 10 chapters
703
Illustrative ExamplesWorked examples with step-by-step solutions
140
Textbook ExercisesMultiple-choice test set (in-domain benchmark)
1,110
Concept QuizzesTraining set — generated by Qwen2.5-72B, validated by GPT-4o

Citation

If you find this work useful, please cite our paper:


@inproceedings{gao2026core, title = {{CORE}: Concept-Oriented Reinforcement for Bridging the Definition--Application Gap in Mathematical Reasoning}, author = {Gao, Zijun and Xu, Zhikun and Ye, Xiao and Zhou, Ben}, booktitle = {International Conference on Learning Representations}, year = {2026}, url = {https://arxiv.org/abs/2512.18857} }