TLDR: A new framework called DURIT improves Small Language Models (SLMs) by separating problem understanding from reasoning. It maps complex questions into a simpler, standardized format, allowing SLMs to focus on solving. This iterative training method significantly boosts SLM performance and robustness on math and logic tasks, making them more efficient and capable.
Small Language Models (SLMs), despite their compact size and efficiency, have long faced a significant hurdle in complex reasoning tasks. Unlike their larger counterparts, Large Language Models (LLMs), SLMs struggle with the dual challenge of first understanding the nuances of natural language problems and then performing multi-step reasoning. This often leads to a “noisy problem space” that makes it difficult for these smaller models to learn and optimize their reasoning abilities.
A recent research paper titled “Decoupling Understanding from Reasoning via Problem Space Mapping for Small-scale Model Reasoning” introduces a groundbreaking framework designed to tackle this very issue. Authored by Li Wang, Changhao Zhang, Zengqi Xiu, Kai Lu, Xin Yu, Kui Zhang, Wenjun Wu, and others from institutions like Beihang University and University College London, the paper proposes a novel approach that separates the act of understanding a problem from the act of reasoning to solve it.
The core idea is to transform complex natural language problems into a simpler, standardized “canonical problem space.” Imagine taking a convoluted, wordy math problem and rephrasing it into a clear, concise, and consistent format. This simplification allows SLMs to focus their limited capacity primarily on the reasoning process, rather than getting bogged down by linguistic variations or irrelevant details.
The framework introduces an innovative algorithm called DURIT (Decoupled Understanding from Reasoning via Iterative Training). DURIT operates in three alternating steps:
1. Problem Mapper Training
First, a dedicated “problem mapper” (itself an LLM) is trained using reinforcement learning. This mapper learns to take original natural language questions and convert them into the standardized problem space. It uses clever techniques like implicit templates and clustering to ensure that similar problems are mapped to similar, simplified forms. The training process even includes a “cheating penalty” to prevent the mapper from directly solving the problem, ensuring it only focuses on the transformation.
2. Self-Distillation Training
Once the mapper is proficient, its ability to transform questions is “internalized” into the SLM. This means the SLM learns to perform this simplification on its own, without needing the external mapper during actual use. This step uses a combination of supervised fine-tuning and knowledge distillation, essentially teaching the SLM to understand and process complex questions as if they were already simplified.
Also Read:
- Rethinking LLM Reasoning: A Data-Smart Approach to Distillation
- Streamlining Large Reasoning Models: A New Approach to Shorter, Smarter Outputs
3. Reinforcement Learning Training
Finally, the SLM is further optimized using reinforcement learning directly on the original, complex problems. Because it has internalized the understanding component, it can now reason more effectively and robustly. The beauty of DURIT is its iterative nature: the mapper and the reasoner (SLM) are co-trained in a loop, continuously refining each other and progressively enhancing the SLM’s reasoning capabilities.
Experiments conducted on various mathematical and logical reasoning tasks, including GSM8K, MAWPS, MATH500, and LogiQA, demonstrated significant improvements. DURIT consistently outperformed existing methods, even with limited training data. For instance, it showed substantial gains on both in-domain and out-of-domain tasks for models like Qwen2.5-0.5B-Instruct and Llama3.2-1B-Instruct. Beyond just accuracy, DURIT also notably improved the robustness of SLMs’ reasoning, meaning they were less susceptible to minor changes in problem phrasing.
The research highlights that this problem space mapping leads to more compact and efficient representations of problems within the model, allowing for faster learning and better generalization. This approach offers a promising new direction for making powerful reasoning capabilities accessible to smaller, more efficient language models, opening doors for their deployment in resource-constrained environments. You can read the full paper here: Decoupling Understanding from Reasoning via Problem Space Mapping for Small-scale Model Reasoning.
