← Back to Research
arxiv
LLM February 22, 2026 Impact: 9/10

ReasonFlux: Hierarchical LLM Reasoning via Scaling Thought Templates

By Ling Yang, Zhaochen Yu, Bin Cui, Mengdi Wang
We present that hierarchical LLM reasoning via scaling thought templates can effectively optimize the reasoning search space and outperform the mathematical reasoning capabilities of powerful LLMs like OpenAI o1-preview and DeepSeek V3. We train our ReasonFlux-32B model with only 8 GPUs and introduces three innovations: (i) a structured and generic thought template library, containing around 500 high-level thought templates capable of generalizing to similar or relevant reasoning problems; (ii) performing hierarchical reinforcement learning on a sequence of thought templates instead of long CoTs, optimizing a base LLM to plan out an optimal template trajectory for gradually handling complex problems; (iii) a brand new inference scaling system that enables hierarchical LLM reasoning by adaptively scaling thought templates at inference time. ReasonFlux-32B achieves 91.2% on MATH benchmark and solves 56.7% of AIME problems, surpassing o1-preview and DeepSeek-V3 by 27% and 45% respectively.
reasoningthought templateshierarchical RLinference scalingmathematical reasoning
Analysis

Summary

ReasonFlux introduces a fundamentally different approach to LLM reasoning compared to the dominant chain-of-thought and long-form reasoning paradigms. Instead of training models to produce extended reasoning traces, the team built a library of approximately 500 high-level "thought templates" — reusable, generalizable reasoning patterns that can be composed hierarchically to solve complex problems. The model learns to select and sequence these templates rather than generating raw reasoning text, which dramatically constrains the search space and makes the reasoning process more structured and interpretable.

Why It Matters

The efficiency story here is remarkable: ReasonFlux-32B was trained with just 8 GPUs, yet it outperforms models backed by far greater compute resources. On the MATH benchmark it hits 91.2% accuracy, beating o1-preview by 6.7 percentage points. On AIME (the USA Math Olympiad benchmark — one of the hardest reasoning tests for LLMs), it solves 56.7% of problems compared to o1-preview's ~30% and DeepSeek-V3's ~12%. These are not marginal improvements; they represent a qualitative leap in mathematical reasoning capability achieved through architectural cleverness rather than brute-force scaling.

Implications

The hierarchical RL approach — optimizing over sequences of thought templates rather than raw token sequences — suggests a promising direction for making LLM reasoning more sample-efficient and verifiable. When a model's reasoning trace is a sequence of named, interpretable templates, it becomes much easier to audit, debug, and trust. This has significant implications for high-stakes domains like scientific research, legal analysis, and financial modeling where reasoning transparency matters as much as accuracy.

Open Questions

The key questions going forward are about generalizability: how well do thought templates transfer beyond mathematics into more open-ended reasoning domains like coding, planning, or scientific hypothesis generation? And who curates the template library — is this a human-in-the-loop process, or can templates be automatically discovered? If template discovery can be automated, ReasonFlux's approach could become a general-purpose reasoning architecture rather than a specialized math solver. Either way, this paper should be required reading for anyone building reasoning systems in 2025.

Key Findings
1
ReasonFlux-32B achieves 91.2% on MATH benchmark, surpassing o1-preview by 6.7%
2
Solves 56.7% of AIME problems — beats o1-preview by 27% and DeepSeek-V3 by 45%
3
Trained with only 8 GPUs using hierarchical RL over thought template sequences, not raw token sequences
4
500-template library enables generalizable reasoning patterns that transfer across similar problem types
Impact Score
9/10
Expert Commentary
ReasonFlux is one of the most interesting efficiency-vs-performance results of early 2025 — the idea of reasoning over structured templates rather than raw token space is elegantly simple and the empirical gains are hard to argue with.
D
Dong Tran
AI Researcher. Deep-dive analysis of frontier AI research papers.