Group Pattern Selection Optimization: Let LRMs Pick the Right Pattern for Reasoning

Hanbin Wang, Jingwei Song, Jinpeng Li, Fei Mi, Lifeng ShangarXiv 2026 · arXiv · OpenReview

TL;DR

Large language models (LLMs) often get stuck using a single reasoning strategy, even when it's not the best for a given problem. This paper introduces Group Pattern Selection Optimization (GPSO), a new training method that teaches LLMs to dynamically choose the most effective reasoning pattern for each specific task, leading to significant performance improvements.

What Problem Does This Paper Address?

Large reasoning models (LRMs) have shown impressive capabilities in complex tasks like math and science. However, current training methods often lead these models to favor a few "dominant" reasoning patterns (like direct step-by-step solutions or reflection-and-verification) and stick to them. The authors' analysis reveals that this reliance on a single, default pattern can be suboptimal, as different problems benefit from different reasoning strategies. For instance, some problems might require exploring multiple solutions to find the best insight, while others are best solved with a direct, verified approach. The core problem is that LLMs don't automatically know which reasoning pattern to use for which problem, leading to missed performance opportunities.

Methods

The paper proposes Group Pattern Selection Optimization (GPSO), a novel training framework built upon reinforcement learning with verifiable rewards (RLVR). Here's how it works:

  • Multi-Pattern Rollouts: For each problem, GPSO prompts the model to generate solutions using multiple predefined reasoning patterns (e.g., "Direct Solution," "Reflection and Verification," "Explore Multiple Solutions").
  • Verifier-Guided Selection: A "verifier" (which checks the correctness of the answer) provides feedback on the success of each generated solution. GPSO then identifies the optimal pattern for that specific problem based on this feedback.
  • Selective Policy Update: The model's policy is updated only using the rollouts from the identified optimal pattern. This focuses learning on the most effective strategies.
  • Attention Masking: To prevent the model from simply memorizing the explicit pattern prompts (e.g., "Explore Multiple Solutions"), GPSO uses attention masking. This ensures that the pattern "scaffolding" doesn't leak into the model's learned reasoning process, forcing it to internalize the mapping from problem to pattern.

Essentially, GPSO teaches the LLM to become a meta-reasoner, learning to select the best tool (reasoning pattern) from its toolbox for each job.

Datasets

The paper uses the following datasets:

  • Training: DAPO-Math-17K (a collection of competition-level math problems).
  • Evaluation: AIME2024, AIME2025, MATH-500, and GPQA datasets.

No new benchmarks were introduced; the focus was on evaluating existing models on established reasoning tasks.

Key Insights and Ideas

  • Reasoning Patterns Aren't One-Size-Fits-All: The most striking finding is that a model's default reasoning pattern is often not the best one for a given problem. Performance can vary significantly depending on the chosen strategy.
  • LLMs Can Learn to Choose: GPSO demonstrates that LLMs can be trained to dynamically select the most appropriate reasoning strategy on a per-instance basis, rather than relying on a single default.
  • Internalizing vs. Memorizing: The attention masking technique is crucial. It ensures that models learn when to use a pattern, not just how to generate the pattern's specific text, leading to more robust reasoning.
  • Model-Agnostic Improvement: GPSO provides consistent performance gains across various model architectures, suggesting it's a generalizable enhancement for LLM reasoning.

Key Results

  • Consistent Performance Gains: GPSO consistently improved the average performance of several LLMs. For example, Nemotron-Research-Reasoning-Qwen-1.5B saw a +2.6% improvement, and DeepSeek-R1-Distill-Qwen-7B improved by +3.1%.
  • Best Overall Performance: The Qwen3-8B (Thinking) model, when trained with GPSO, achieved the best overall average score of 75.3.
  • Stronger on Challenging Tasks: Improvements were particularly notable on challenging reasoning tasks like AIME2024 and AIME2025, with average gains of 4.0 and 2.7 points, respectively.
  • Generalization: GPSO showed strong generalization, improving performance even on domains it wasn't explicitly trained on (e.g., a 2.1-point average improvement on GPQA).
  • Ablation Studies: Removing any key component of GPSO (multi-pattern rollout, optimal pattern selection, or attention masking) led to noticeable performance degradation, highlighting the importance of each element.

Limitations and Future Work

The authors acknowledge two main limitations:

  • Computational Cost: The multi-pattern rollout process for each problem increases computational cost compared to standard single-path RL.
  • Predefined Patterns: The set of candidate reasoning patterns is predefined. Future work could explore mechanisms for automatically discovering or evolving new patterns.

Summary updated

Abstract

Large reasoning models (LRMs) exhibit diverse high-level reasoning patterns (e.g., direct solution, reflection-and-verification, and exploring multiple solutions), yet prevailing training recipes implicitly bias models toward a limited set of dominant patterns. Through a systematic analysis, we identify substantial accuracy variance across these patterns on mathematics and science benchmarks, revealing that a model's default reasoning pattern is often sub-optimal for a given problem. To address this, we introduce Group Pattern Selection Optimization (GPSO), a reinforcement learning framework that extends GRPO by incorporating multi-pattern rollouts, verifier-guided optimal pattern selection per problem, and attention masking during optimization to prevent the leakage of explicit pattern suffixes into the learned policy. By exploring a portfolio of diverse reasoning strategies and optimizing the policy on the most effective ones, GPSO enables the model to internalize the mapping from problem characteristics to optimal reasoning patterns. Extensive experiments demonstrate that GPSO delivers consistent and substantial performance gains across various model backbones and benchmarks, effectively mitigating pattern sub-optimality and fostering more robust, adaptable reasoning. All data and codes are available at https://github.com/wanghanbinpanda/GPSO.
BibTeX
@misc{wang2026group,
  author = {Hanbin Wang and Jingwei Song and Jinpeng Li and Fei Mi and Lifeng Shang},
  title = {{Group Pattern Selection Optimization: Let LRMs Pick the Right Pattern for Reasoning}},
  year = {2026},
  eprint = {2601.07238},
  archivePrefix = {arXiv},
  url = {https://arxiv.org/abs/2601.07238},
}
Figure