SynLogic-Mix-3-32B: Multi-Domain Reasoning Model

SynLogic-Mix-3-32B, developed by MiniMaxAI, is an advanced 32 billion parameter model based on Qwen2.5-32B-Base. It stands out due to its unique Zero-RL (reinforcement learning from scratch) training methodology, applied to a diverse dataset encompassing logical reasoning, mathematics, and coding tasks. This approach enables the model to achieve enhanced generalization and superior cross-domain transfer compared to single-domain training.

Key Capabilities & Features

• Multi-Domain Training: Jointly trained on 35k mathematical, 9k coding, and 17k SynLogic logical reasoning samples.
• Zero-RL Training: Utilizes Group Relative Policy Optimization (GRPO) from a base model, without instruction tuning.
• Enhanced Generalization: Demonstrates improved performance across various reasoning domains.

Performance Highlights

SynLogic-Mix-3-32B shows strong performance on challenging benchmarks:

• BBEH: Achieves 28.6, matching or surpassing DeepSeek-R1-Distill-Qwen-32B.
• KOR-Bench: Scores 65.0, comparable to leading models.
• GPQA-Diamond: Outperforms DeepSeek-R1-Zero-Qwen-32B by +2.5 points, scoring 57.5.
• Ablation studies confirm that the inclusion of SynLogic logical reasoning data significantly boosts performance on logical reasoning (e.g., +10.1 points on BBEH) and out-of-domain reasoning tasks.

Ideal Use Cases

This model is particularly well-suited for applications requiring robust performance in:

• Complex Logical Reasoning: Solving intricate logical puzzles and problems.
• Mathematical Problem Solving: Handling diverse mathematical queries and computations.
• Code Generation & Understanding: Assisting with coding tasks and understanding programming logic.