Small Models, Big Wisdom: How Qwen3-1.7B Breaks Through the 'Reasoning Gap' in Vietnamese Mathematical Reasoning

A groundbreaking study focuses on the potential and challenges of small language models (SLMs) in non-English reasoning tasks, using Qwen3-1.7B as the research object. By constructing the Vietnamese elementary math dataset Vi-S1K and the evaluation benchmark Vi-Elementary-Bench, it was found that supervised fine-tuning (SFT) can unlock the hidden reasoning capabilities of the model, while complex agent frameworks (such as ReAct) instead become a cognitive burden, providing a new path for edge AI to achieve complex reasoning.

Reasoning Dilemma of Edge AI

The vision of ubiquitous AI requires models to run on edge devices, but small language models (SLMs) face a "reasoning gap" and struggle to maintain a coherent chain of thought. Non-English environments (such as Vietnamese's unique grammar and tones) add further complexity.

Comparison Between Large and Small Models

Large models (like GPT-4) have strong reasoning abilities but rely on the cloud, with high costs and data security concerns; 1.7B-scale small models can run on ordinary devices, and if they have reasoning capabilities, they can promote AI democratization.

Underestimated Challenges of Non-English Languages

Existing research is English-centric, and the impact of grammar and cultural differences in non-English languages on reasoning far exceeds translation issues.

Vi-S1K Dataset

Contains 1000 carefully curated Vietnamese elementary math problems, each with detailed solution steps and explanations; localized via the Gemini 2.5 Flash-Lite pipeline to ensure terms comply with Vietnamese textbook standards, problems are culturally relevant, and solution steps align with local teaching traditions.

Vi-Elementary-Bench Benchmark

Two-dimensional evaluation: computational accuracy (whether the correct answer is obtained) and explanation quality (whether the problem-solving思路 can be clearly explained), reflecting the math education goal of "knowing not only the result but also the reason".

Hidden Reasoning Capabilities

The Qwen3-1.7B base model achieves a computational accuracy of 4.05/5, with a "format gap"—it has correct knowledge but cannot output it in the format expected by humans.

Unlocking Effect of SFT

Supervised fine-tuning improves explanation quality by 77%, proving that SFT is a reasoning unlocker. High-quality small-scale datasets (like Vi-S1K) are more effective than large-scale low-quality data, and domain-specific fine-tuning yields significant benefits.

Cognitive Tax of Complex Frameworks

Agent frameworks like ReAct reduce small model performance due to attention distraction, format overhead, and error accumulation; the pure Chain of Thought (CoT) + self-consistency strategy performs best.

Hierarchical Strategy for Edge Deployment

1. Supervised fine-tuning (essential, unlocks reasoning capabilities); 2. Simplified test-time scaling (CoT + self-consistency, controllable overhead); 3. Avoid complex agent frameworks (suitable for 7B+ models).

Implications for AI Democratization

• Language diversity: The Vietnamese experience can be extended to other underserved languages;
• Small model strategy: Well-fine-tuned small models are more effective in resource-constrained scenarios;
• Data engineering: High-quality domain-specific datasets are key.

Big Future of Small Models

Small models are expected to allow non-English users to enjoy AI services without relying on the cloud, which is a key path to AI democratization.

Research Limitations

• Evaluation only covers the field of Vietnamese elementary math;
• Only uses the single architecture of Qwen3-1.7B.

Future Directions

• Expand to more non-English languages and subject areas;
• Explore the impact of model compression and quantization techniques on reasoning capabilities;
• Study whether multilingual joint training improves monolingual reasoning performance.