Introduction
The recent success of Large Reasoning Models (LRMs) like OpenAI’s o1 and DeepSeek-R1 has demonstrated the power of "slow thinking"—long chain-of-thought reasoning—for solving complex mathematical, coding, and logical problems. However, the massive scale of these models (e.g., DeepSeek-R1 at 671B parameters) creates significant barriers for research and deployment. This survey reviews the emerging landscape of Small Reasoning Models (SRMs), typically under 10 billion parameters, which achieve comparable reasoning capabilities through efficient training and inference techniques.
Taxonomy of Small Reasoning Models
SRMs are categorized by their architectural focus, training pipelines, and inference strategies. The following figure provides a structured overview of the current SRM landscape, from training methodologies to domain-specific applications.
Training Methodologies and Datasets
To overcome parameter constraints, SRMs rely on sophisticated training pipelines:
- Knowledge Distillation: This is currently the most effective strategy. SRMs use LRMs as teachers to generate high-quality, long-form reasoning trajectories. Datasets like
OpenThoughts-114kandOpenR1-Mathprovide the necessary synthetic data for this process. - Supervised Fine-Tuning (SFT): Models are trained on Chain-of-Thought (CoT) data to force the explicit generation of intermediate steps.
- Reinforcement Learning (RL): Beyond standard RLHF, SRMs benefit from Outcome Reward Supervision (checking the final answer) using efficient algorithms like Group Relative Policy Optimization (GRPO), and Process Reward Models (PRMs) that evaluate intermediate reasoning steps.
Inference-Time Scaling
A core insight of the survey is that "inference compute" can often compensate for model size. By allocating more time for a model to "think" during test time, a 7B model can outperform much larger counterparts. Key techniques include:
- Advanced Prompting: Moving from simple CoT to Tree of Thoughts (ToT) or Self-Consistency (CoT-SC).
- Search and Verification: Using Monte Carlo Tree Search (MCTS) or step-wise verifiers to iteratively explore and refine solution paths.
- Agent-Based Reasoning: Employing multi-agent collaboration where different SRM agents debate or check each other's work.
Empirical Performance and Benchmark Parity
Aggregated community findings show that distilled SRMs are routinely matching or beating older, much larger models on reasoning benchmarks.
As illustrated above, the DeepSeek-R1-Distill-Qwen-7B model demonstrates exceptional performance, particularly in mathematics (AIME2024 and MATH-500), where it significantly outperforms frontier models like GPT-4o. It also remains highly competitive in coding (LiveCodeBench) and general reasoning (GPQA Diamond).
Domain-Specific Applications
SRMs are particularly effective when adapted to specialized domains. In healthcare, specialized models like Hippo-Mistral-7B have achieved 59.9% on MedQA, outperforming the 70B Meditron model. Similar successes are noted in finance (e.g., Bloomberg-related corpora) and law (e.g., ChatLaw), where narrow domain expertise is combined with distilled reasoning logic.
Limitations and Future Directions
While SRMs democratize high-end AI research, they face challenges such as limited capacity for "exploration" beyond their training data and the potential to inherit biases from their LRM teachers. Future research is focused on developing adaptive RL strategies that balance exploration within limited parameter budgets, low-resource learning, and optimizing multi-agent communication to reduce token overhead.
Summary updated