MindGYM: What Matters in Question Synthesis for Thinking-Centric Fine-Tuning?

TL;DR

MindGYM tackles the challenge of instilling transferable, structured thinking in large foundation models by proposing a thinking-centric data synthesis framework. It integrates Cognitive Thinking Injection, Seed Single-Hop Question Synthesis, and Challenging Multi-Hop QA Synthesis to produce self-generated training data with interpretable thinking traces. Empirical results across six reasoning benchmarks show significant performance gains, high data efficiency with as few as 400 synthetic samples, and robust cross-modal benefits even when training is text-only. The approach reduces human annotation needs, scales across model sizes, and demonstrates the viability of self-challenging data evolution for reasoning-centric fine-tuning. The authors also release code and data to foster ongoing data-centric research in self-evolving foundation models.

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. MindGYM underscores the viability of self-challenging mechanisms in refining large model capabilities while minimizing human intervention and resource demands. Code and data are released to promote data-centric research into self-evolving foundation models driven by their internal reasoning capabilities.

Figures (2)

• Figure 1: The proposed MindGYM framework incorporates a cognitively guided data synthesis pipeline with four stages: ① Context Generation, ② Single-Hop Question Synthesis, ③ Multi-Hop Composition with Thinking Trace, and ④ Structured Extraction. Starting from a meta-topic and guided by a shared cognitive objective, the model iteratively builds background context, atomic reasoning steps, composite questions with interpretable thinking traces, and final structured QA samples for downstream use. Color-coded regions and arrows illustrate the hierarchical progression from simple reasoning to advanced problem-solving and emphasizing conceptual depth.
• Figure 2: An end-to-end example of MindGYM. For multimodal data, we generate five seed questions first, and make the model self-challenging itself via synthesizing multi-hop questions and multi-hop answers while preserving its internal thinking process.