R-HORIZON: Uncovering Long-Range Reasoning Bottlenecks and Breakthrough Paths for Large Reasoning Models

The LongCat team from Meituan's work R-HORIZON, accepted at ICLR 2026, constructs a long-range reasoning benchmark using a problem combination method, reveals the performance degradation of current large models in multi-step dependent reasoning, and provides effective training improvement solutions, which is of great significance for AI reasoning ability evaluation and model optimization.

Current mainstream reasoning benchmarks (such as MATH, AIME) focus on independent single-step reasoning tasks, with samples isolated from each other, which cannot simulate complex real-world scenarios with multi-step correlations (e.g., pre-steps of scientific experiments, interactions between software development modules). This leads to the inability to evaluate the real long-range dependent reasoning ability of models, forming a blind spot in performance assessment.

R-HORIZON proposes the Query Combination method to construct long-range reasoning tasks, with a three-step process: 1. Filter problems containing valid integers (to ensure variable replacement feasibility); 2. Identify key variables (as connectors between problems); 3. Concatenate problems to form chain dependencies (the answer of the previous step serves as the parameter for the next step, enforcing long-range logical consistency).

Evaluation of over 20 advanced models shows that all models experience a sharp performance drop in long-range reasoning. Taking DeepSeek-R1 as an example: the pass rate for single AIME25 problems is 87.3%, while it drops to only 24.6% for 5 concatenated problems; larger models have stronger resilience, but the degradation in code generation tasks is steeper, and models have the problem of unbalanced allocation of thinking resources.

The team trained models using R-HORIZON combined data with GRPO reinforcement learning. The results show: Training with 2-problem combinations improved AIME24 (n=2) by 17.4 points and single problems by 7.5 points (positive transfer); training with n=4 combinations increased the pass rate of MATH500 (n=8) from 8.4% to 50.6%, proving the effectiveness of the training method.

This research implies: 1. Need to construct a more comprehensive long-range reasoning evaluation framework; 2. Reveal a new dimension of Scaling Law—the length of reasoning chains; 3. Provide theoretical and data foundations for Agent systems (multi-step planning and execution).

The team has open-sourced: the paper (arXiv:2510.08189), benchmark datasets (Hugging Face includes subsets like Math500), combined training data, and trained models to help researchers reproduce and improve.

R-HORIZON reveals the capability boundary of current large models in long-range reasoning, but also proves that significant improvement can be achieved through targeted training. We look forward to the community using open-source resources to jointly push AI reasoning ability to new heights.