Research map / Large language models / Agents, code and math

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

NeurIPS 2026 · Large language models › Agents, code and math · approach: Inductive and abstract reasoning evaluation · 0 citations (OpenAlex)

Read the paperCheck your idea against 108,967 papers

AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.

Proposes MindLoom, which reverse-engineers verified hard-problem solutions into thought-mode chains, retrieves and composes them with distribution-aligned sampling, and SFTs Qwen3/Qwen3.5 models on 9,230 judged-correct synthesized problems.

Problem

Systematically producing frontier-level reasoning training data with controllable difficulty and diverse reasoning coverage is hard; existing synthesis methods lack visibility into structural difficulty factors.

Approach

Model problem difficulty as accumulation of atomic knowledge-reasoning transformations ("thought modes") extracted by reverse-engineering solutions into chains; train an embedding retriever to match problem states to compatible modes and compose new problems with a similarity-plus-scarcity (cluster-coverage) selection score.

Key results

• On Qwen3-4B, MindLoom raises pass@3 on AIME 2025 from 33.33 (base) to 53.33 and HMMT-Feb from 26.67 to 43.33.
• On Qwen3.5-9B, MindLoom reaches 73.96 pass@1 on CS-Bench vs 69.60 for base and 73.13 for OpenThought.
• Ablation on Qwen3-4B: removing rollout-based filtering drops MATH-500 pass@1 from 85.20 to 67.80 and AIME 2025 pass@1 from 40.00 to 16.67.

Surprising findings

• Removing the scarcity reward raises MATH-500 pass@1 for Qwen3-4B (92.40 vs 85.20 for full MindLoom), contrary to the claimed overall benefit.
• Qwen3-4B MindLoom does not beat OpenThought on HMMT-Feb pass@1 (33.33 vs 40.00).

Limitations

• Diversity of synthesized problems is bounded by the thought-mode bank derived from the reference corpus.
• No experiments on larger models or combined with reinforcement learning.
• Heavy reliance on DeepSeek V3.2 as the chat model for extraction, synthesis and judging.

Experimental setup

Datasets: 58,526 source problems from 16 STEM/expert reasoning datasets SuperGPQA MMLU-Pro Hendrycks-MATH ARC-AGI-2 OlympiadBench

Benchmarks: CS-Bench ChemBench HLE MedQA SciBench MATH-500 HMMT February 2025 HMMT November 2025 AIME 2025

Metrics: pass@1 pass@3

Baselines: DS-V3.2 Distill MegaScience OpenThought w/o Scarcity w/o Filter Random Retriever w/o Reverse Eng.

Builds on: ANCE Qwen3-Embedding FAISS ms-swift

Closest papers by meaning

• MindGYM: What Matters in Question Synthesis for Thinking-Centric Fine-Tuning? (NeurIPS 2025 · similarity 0.91)
• The Illusion of Thinking: Understanding the Strengths and Limitations of Reasoning Models via the Lens of Problem Complexity (NeurIPS 2025 · similarity 0.90)
• RATT: A Thought Structure for Coherent and Correct LLM Reasoning (AAAI 2025 · similarity 0.89)
• Scalable Chain of Thoughts via Elastic Reasoning (ICLR 2026 · similarity 0.88)
• Decompose, Analyze and Rethink: Solving Intricate Problems with Human-like Reasoning Cycle (NeurIPS 2024 · similarity 0.88)
• Reasoning Scaffolding: Distilling the Flow of Thought from LLMs (ICLR 2026 · similarity 0.88)