TLDR: R-HORIZON is a new method and benchmark that evaluates and improves Large Reasoning Models’ (LRMs) ability to solve complex, multi-step problems by composing simple queries into interdependent tasks. It reveals that current LRMs suffer significant performance degradation on these long-horizon tasks due to limited effective reasoning length, narrow reflection scope, and poor thinking budget allocation. However, training LRMs with R-HORIZON-generated data significantly enhances their performance on both multi-step and single-step reasoning tasks, promoting more efficient and deeper reasoning.
Large Reasoning Models (LRMs) like OpenAI’s o1 and DeepSeek-R1 have shown impressive capabilities, especially with techniques like Chain-of-Thought (CoT) that allow them to “think” through problems. However, a new research paper introduces a critical question: how well do these models truly perform on complex, multi-step reasoning tasks that mimic real-world scenarios?
The paper, titled “R-Horizon: How Far Can Your Large Reasoning Model Really Go in Breadth and Depth?”, by researchers Yi Lu, Jianing Wang, Linsen Guo, Wei He, Hongyin Tang, Tao Gui, Xuanjing Huang, Xuezhi Cao, Wei Wang, and Xunliang Cai from Fudan University and Meituan, highlights a significant gap in current AI evaluation. Existing benchmarks often focus on isolated, single-step problems, which don’t fully test an AI’s ability to handle a sequence of interdependent challenges over a longer period.
Introducing R-HORIZON: A New Approach to Long-Horizon Reasoning
To address this, the researchers propose R-HORIZON, a novel method designed to stimulate and evaluate long-horizon reasoning in LRMs. R-HORIZON works by composing simple, single-step tasks into complex, multi-step problems with explicit dependencies. Imagine a math problem where the answer to the first part is crucial for solving the second, and so on. This creates a “long reasoning horizon” that forces the model to maintain context and accuracy across multiple steps.
The R-HORIZON framework is used to build a new benchmark dataset spanning various domains, including mathematics, code generation, and agentic tasks. This benchmark allows for a more comprehensive assessment of how LRMs perform when faced with problems that require sustained, sequential reasoning.
Key Findings: Performance Degradation and Model Limitations
The evaluation of 25 mainstream LRMs on the R-HORIZON benchmark revealed some striking insights. Even the most advanced models experienced significant performance drops as the number of composed queries (the “reasoning horizon”) increased. For example, DeepSeek-R1’s accuracy on AIME25 tasks plummeted from 87.3% for a single problem to 24.6% for five interdependent problems. Smaller models showed even more severe degradation.
The analysis pinpointed several limitations in current LRMs:
- Limited Effective Reasoning Length: Models struggle to maintain performance beyond a certain “thinking budget” or length of reasoning.
- Constrained Reflection Scope: LRMs often reflect only on the immediate problem, failing to identify or correct errors from earlier steps in a multi-step sequence.
- Overthinking Phenomenon: Models tend to allocate excessive computational resources to early problems, leaving insufficient “thinking budget” for subsequent, equally important steps.
Also Read:
- Enhancing LLMs’ Multi-Step Reasoning with Composed Data and Curriculum Learning
- Empowering Language Models: How TAPO Integrates Reasoning and Adaptive Tool Use
Enhancing Reasoning with R-HORIZON Data in Reinforcement Learning
Beyond evaluation, R-HORIZON also serves as a powerful tool for improving LRMs. The researchers used R-HORIZON to construct long-horizon reasoning data for reinforcement learning with verified rewards (RLVR). This means training models with these complex, multi-step problems, rather than just isolated ones.
The results were highly encouraging. Training models with R-HORIZON data not only substantially improved their performance on multi-horizon reasoning tasks but also boosted their accuracy on standard, single-horizon reasoning tasks. Models trained with this approach demonstrated more efficient reasoning, better allocation of their “thinking budget” across problems, and an increased ability to engage in longer-range reflection.
In essence, R-HORIZON offers a scalable, controllable, and cost-effective way to both evaluate and enhance the long-horizon reasoning capabilities of Large Reasoning Models. This research paves the way for future AI systems that can tackle real-world problems requiring sustained, complex thought processes. You can read the full paper here.
