MindLoom: Composing Thought Modes for Frontier-Level Reasoning Data Synthesis

Although LLMs have made substantial progress in reasoning, systematically producing frontier-level reasoning data remains difficult. Existing synthesis methods often have limited visibility into the structural factors that govern problem difficulty, which can result in narrow diversity and unstable difficulty control. In this work, we view the difficulty of a reasoning problem as arising from the accumulation of atomic knowledge-reasoning transformations, which we term thought modes. Building on this perspective, we propose MindLoom, a framework for synthesizing frontier-level reasoning data through compositional thought mode engineering. Given a collection of hard problems with verified solutions, MindLoom first decomposes those solutions into thought mode chains that reveal each problem's construction logic. It then trains a retrieval model that matches problem states to compatible thought modes, providing guidance on which reasoning challenges to introduce during synthesis. New problems are composed by iteratively applying retrieved thought modes to seed questions, with distribution-aligned sampling to encourage diverse reasoning coverage. Finally, a rollout-based judging stage labels generated questions by difficulty and supplies judged-correct responses for supervised fine-tuning. We evaluate MindLoom on nine benchmarks covering five STEM disciplines and four mathematical reasoning tasks across multiple model families and sizes. Models fine-tuned on MindLoom-generated data achieves favorable performances over base models, distillation, and external-data baselines across the reported benchmarks. Ablation studies indicate the contribution of each component, and further analysis suggests that MindLoom covers a broad range of reasoning patterns while maintaining useful difficulty control. We have open-sourced our implementation at https://github.com/EachSheep/MindLoom.

📋 한 줄 요약

**[Reasoning Data Synthesis / Thought Modes]** MindLoom이 hard problem 솔루션을 thought mode chain으로 분해·retrieval로 매칭·반복 합성, 9 STEM 벤치·4 수학 task에서 base·distillation·external-data baseline 능가.

🎯 핵심 기여도

LLM의 reasoning 진전에도 frontier 수준 reasoning 데이터의 systematic 생성은 어려움을 지적 — 기존 합성 방법이 문제 난이도를 결정하는 구조 요인에 대한 가시성이 제한되어 좁은 다양성·불안정 난이도 통제 야기.
Reasoning 문제 난이도를 atomic knowledge-reasoning transformation의 누적으로 보는 관점 제안 — 이를 thought modes로 명명.
MindLoom 프레임워크 제안 — compositional thought mode engineering을 통한 frontier-level reasoning 데이터 합성.
검증된 솔루션의 hard problem 모음을 thought mode chain으로 분해해 각 문제 구성 logic 노출, problem state→compatible thought mode 매칭 retrieval model 학습, seed question에 retrieved thought mode 반복 적용으로 새 문제 합성·distribution-aligned sampling으로 다양 reasoning coverage, rollout-based judging으로 난이도 라벨·정답 응답 제공.

💡 핵심 아이디어

Reasoning 데이터 합성의 난이도와 다양성은 문제를 thought mode chain — atomic knowledge-reasoning transformation의 누적 — 으로 분해해 구성 logic을 명시한 뒤 retrieval로 합성하는 compositional engineering으로 systematically 통제 가능하다.

🔬 기술적 접근법

**방법론**: MindLoom — thought mode 분해 + retrieval + 반복 합성 + rollout judging.
**핵심 기법**: (1) Hard problem 솔루션을 thought mode chain으로 분해, (2) Problem state→compatible thought mode 매칭 retrieval model 학습, (3) Seed question에 retrieved thought mode 반복 적용으로 새 문제 합성, (4) Distribution-aligned sampling으로 다양 reasoning pattern coverage, (5) Rollout-based judging으로 난이도 라벨·정답 응답 SFT 데이터 제공.

📊 주요 결과

9 벤치마크(5 STEM 도메인 + 4 수학 task), 다양 모델 family·size에서 평가.
MindLoom 데이터로 fine-tune한 모델이 base·distillation·external-data baseline 능가.
Ablation으로 각 component 기여 확인.
분석에서 광범위 reasoning pattern coverage와 유용한 난이도 통제 시연.
Open-source: github.com/EachSheep/MindLoom.

💭 의의 및 한계

**의의**: Frontier reasoning 데이터의 systematic·통제 가능 합성 패러다임 제공, thought mode 개념의 일반화 가능성, open-source로 재현성. **한계**: Seed로 verified hard problem 필요(획득 비용), retrieval model 품질에 의존, rollout judging의 false-positive 가능성, STEM·수학 외 일반화는 후속.

🚀 실용적 활용

LLM reasoning SFT 데이터 합성.
난이도 통제 가능 학습 corpus 생성.
Open-source 합성 파이프라인 활용.