Abstract
Large foundation models face challenges in acquiring transferable, structured thinking abilities, especially when supervised with rigid templates or crowd-annotated instruction datasets. Unlike prior approaches, we focus on a thinking-centric data synthesis paradigm that enables models to evolve through self-generated, cognitively guided data. We propose MindGYM, a structured and scalable framework for question synthesis, composed of: (1) Cognitive Thinking Process Injection, which infuses high-level reasoning objectives to shape the model's synthesis behavior; (2) Seed Single-Hop Question Synthesis, generating atomic questions from diverse semantic types to encourage broader thinking; and (3) Challenging Multi-Hop QA Synthesis, composing more complex multi-hop questions based on QA seeds for deeper reasoning. Detailed analysis shows that synthetic data generated by our method achieves 16.7% higher average quality and 67.91% lower quality variance compared to baseline sources, highlighting that both high-quality and self-contained data are essential for effective, thinking-oriented fine-tuning. MindGYM improves performance on six reasoning benchmarks, achieving gains of up to 16% on MathVision using only 400 data samples, and generalizable improvements across different model sizes and architectures.
Introduction
While Large Foundation Models (LFMs) and Vision-Language Models (VLMs) are increasingly capable, they often encounter a "reasoning ceiling." Traditional methods to elevate reasoning capabilities—such as manual data annotation—are prohibitively expensive and slow to scale. Conversely, existing automated synthetic data pipelines frequently produce repetitive or shallow questions that fail to challenge a model's internal cognitive boundaries. MindGYM addresses this gap by shifting the focus from data quantity to "thinking-centric" data quality, targeting the model's internal reasoning bottlenecks through self-challenging mechanisms.
Methodology: The MindGYM Framework
MindGYM employs a four-stage pipeline designed to inject structured thinking into the data generation process, enabling models to act as their own "personal trainers."
- Cognitive Thinking Injection: Utilizing a "Stream of Consciousness" approach, the model is prompted to externalize internal reasoning steps, ensuring generated data is logically grounded.
- Seed Single-Hop Synthesis: The model generates "atomic" questions based on background contexts across diverse categories including arithmetic, logic, and causality to ensure thinking breadth.
- Challenging Multi-Hop QA Synthesis: These atomic units are composed into complex, multi-step questions using operators such as Bridging (linking distant facts), Comparison, and Temporal reasoning to increase thinking depth.
- Structured Extraction: The final output is formatted into clean Question-Answer-Rationale triplets for training. To maintain diversity, Reject Sampling via semantic vectorization is used to prune redundant questions.
Experimental Results
MindGYM was evaluated across six reasoning-intensive benchmarks, including GSM8K, MATH, GPQA, and multimodal benchmarks like MathVision. The framework was tested against state-of-the-art reasoning datasets such as LIMO and Open-O1.
Key Findings
- Extreme Data Efficiency: Using only 400 self-generated Chinese samples (MindGYM-Text), the framework boosted Qwen2.5-VL-7B's performance on MathVision-Mini by 16%, outperforming baselines that utilized ten times more data.
- Quality and Consistency: MindGYM data achieved 16.7% higher average quality and a 67.91% reduction in quality variance compared to standard baseline sources. This suggests that reducing quality variance is as critical as improving mean quality for stable fine-tuning.
- Cross-Modal Thinking Transfer: Training on text-only synthetic data significantly improved performance on vision-heavy benchmarks. This indicates that "thinking" is a modality-agnostic capability; enhancing the model's internal logic engine directly benefits its visual reasoning.
- Generalizability: Improvements were consistent across model sizes (7B to 38B) and architectures (Qwen and InternVL), demonstrating the robustness of the self-challenging approach.
Conclusion and Future Work
MindGYM demonstrates that self-challenging mechanisms can effectively refine large model capabilities with minimal human intervention. By focusing on high-quality, low-variance synthetic data, models can overcome traditional reasoning plateaus. Future work will explore fully generative pipelines—where models synthesize both visual context and reasoning tasks from scratch—and the integration of Reinforcement Learning (RL) to provide automated correctness verification for self-generated challenges.
Summary updated