On the Role of Reasoning Patterns in the Generalization Discrepancy of Long Chain-of-Thought Supervised Fine-Tuning

Dimensions overall score 7.0

• Lower training loss in SFT does not necessarily translate to better generalization performance in reasoning models.result 90%
  Evidencepartial

  Despite their comparable performance, we uncover a striking paradox: lower training loss does not translate to better generalization.

  Implicationpartial

  Directly stated as a discovered paradox with supporting evidence

  Verificationpartial

  partial

• Supervised Fine-Tuning (SFT) on long Chain-of-Thought (CoT) trajectories from DeepSeek-R1-0528 achieves lower training loss but exhibits worse generalization performance compared to SFT on trajectories from gpt-oss-120b.result 95%
  Evidencepartial

  SFT on \texttt{DeepSeek-R1-0528} data achieves remarkably lower training loss, yet exhibits significantly worse generalization performance on reasoning benchmarks compared to those trained on \texttt{gpt-oss-120b}.

  Implicationpartial

  Explicitly stated in the abstract with clear comparative results

  Verificationpartial

  partial

• The generalization discrepancy stems from differences in reasoning patterns between DeepSeek-R1-0528 (divergent, branch-heavy exploration) and gpt-oss-120b (convergent, deductive trajectories).technical 90%
  Evidencepartial

  Our analysis reveals a difference in reasoning patterns. \texttt{gpt-oss-120b} exhibits highly convergent and deductive trajectories, whereas \texttt{DeepSeek-R1-0528} favors a divergent and branch-heavy exploration pattern.

  Implicationpartial

  Directly stated in the abstract as the key finding from multi-faceted analysis

  Verificationpartial

  partial

• Models trained with DeepSeek-R1-0528 data inherit inefficient exploration behaviors that trap them in redundant exploratory branches, hindering correct solutions.technical 85%
  Evidencepartial

  Consequently, models trained with \texttt{DeepSeek-R1} data inherit inefficient exploration behaviors, often getting trapped in redundant exploratory branches that hinder them from reaching correct solutions.

  Implicationpartial

  Directly stated in the abstract as a consequence of the reasoning pattern difference

  Verificationpartial

  partial

• Filtering out frequently branching trajectories from DeepSeek-R1-0528 data improves generalization performance of SFT.method 90%
  Evidencepartial

  Building upon this insight, we propose a simple yet effective remedy of filtering out frequently branching trajectories to improve the generalization of SFT.

  Implicationpartial

  Explicitly stated as a proposed remedy with specific performance improvements

  Verificationpartial

  partial

• Training on selected DeepSeek-R1-0528 subsets improves reasoning performance by up to 5.1% on AIME25, 5.5% on BeyondAIME, and on average 3.6% on five benchmarks.result 95%
  Evidencepartial

  Experiments show that training on selected \texttt{DeepSeek-R1-0528} subsets surprisingly improves reasoning performance by up to 5.1% on AIME25, 5.5% on BeyondAIME, and on average 3.6% on five benchmarks.

  Implicationpartial

  Explicitly stated with specific numeric results in the abstract

  Verificationpartial

  partial

• The influence of CoT trajectory sources on model generalization performance was previously an open question in building large reasoning models.limitation 80%
  Evidencepartial

  However, how CoT trajectories from different sources influence the generalization performance of models remains an open question.

  Implicationpartial

  Directly stated in the abstract as the research motivation

  Verificationpartial

  partial

• The study used identical problem sets to compare CoT trajectories from DeepSeek-R1-0528 and gpt-oss-120b, controlling for content variation.method 85%
  Evidencepartial

  with their problem sets controlled to be identical

  Implicationpartial

  Explicitly stated in the abstract as part of the experimental design

  Verificationpartial

  partial

Competitive landscape

This research identifies a critical flaw in how large language models learn reasoning from diverse data sources and proposes a filtering method to significantly improve generalization performance on complex reasoning tasks.