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

• URL: http://arxiv.org/abs/2604.01702v2
• Date: Sat, 04 Apr 2026 13:28:43 GMT
• Title: On the Role of Reasoning Patterns in the Generalization Discrepancy of Long Chain-of-Thought Supervised Fine-Tuning
• Abstract summary: Supervised Fine-Tuning (SFT) on long Chain-of-Thought (CoT) trajectories has become a pivotal phase in building large reasoning models.<n>We conduct a comparative study using two sources of verified CoT trajectories generated by two competing models.<n>We find that SFT on textttDeepSeek-R1-0528 data achieves remarkably lower training loss, yet exhibits significantly worse generalization performance.
• Score: 63.41902113656453
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Supervised Fine-Tuning (SFT) on long Chain-of-Thought (CoT) trajectories has become a pivotal phase in building large reasoning models. However, how CoT trajectories from different sources influence the generalization performance of models remains an open question. In this paper, we conduct a comparative study using two sources of verified CoT trajectories generated by two competing models, \texttt{DeepSeek-R1-0528} and \texttt{gpt-oss-120b}, with their problem sets controlled to be identical. Despite their comparable performance, we uncover a striking paradox: lower training loss does not translate to better generalization. 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}. To understand this paradox, we perform a multi-faceted analysis probing token-level SFT loss and step-level reasoning behaviors. 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. 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. Building upon this insight, we propose a simple yet effective remedy of filtering out frequently branching trajectories to improve the generalization of SFT. 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.

Related papers

• Learning from Contrasts: Synthesizing Reasoning Paths from Diverse Search Trajectories [26.441136942925056]
  Contrastive Reasoning Path Synthesis (CRPS) is a framework that transforms supervision extraction into a synthesis procedure.<n>CRPS uses a structured reflective process to analyze the differences between high- and low-quality search trajectories, extracting explicit information about strategic pivots and local failure modes.<n>We show empirically that models fine-tuned on just 60K CRPS-synthesized examples match or exceed the performance of baselines trained on 590K examples derived from standard rejection sampling.
  arXiv  Detail & Related papers  (2026-04-13T12:05:34Z)
• Bridging SFT and RL: Dynamic Policy Optimization for Robust Reasoning [10.4404175363648]
  Post-training paradigms for Large Language Models (LLMs) face a fundamental dilemma.<n>SFT provides stability (low variance) but suffers from high fitting bias, while RL enables exploration (low bias) but grapples with high gradient variance.<n>Existing unified optimization strategies often employ naive loss weighting, overlooking the statistical conflict between these distinct gradient signals.<n>We propose textbfDYPO (Dynamic Policy Optimization), a unified framework designed to structurally mitigate this conflict.
  arXiv  Detail & Related papers  (2026-04-10T03:42:16Z)
• When Actions Teach You to Think: Reasoning-Action Synergy via Reinforcement Learning in Conversational Agents [2.689316553293938]
  Supervised fine-tuning (SFT) has emerged as one of the most effective ways to improve the performance of large language models (LLMs) in downstream tasks.<n>We propose a pipeline in which LLMs generate reasoning steps that guide both the invocation of tools and the final answer generation for conversational agents.
  arXiv  Detail & Related papers  (2025-12-12T04:44:40Z)
• Efficient Thought Space Exploration through Strategic Intervention [54.35208611253168]
  We propose a novel Hint-Practice Reasoning (HPR) framework that operationalizes this insight through two synergistic components.<n>The framework's core innovation lies in Distributional Inconsistency Reduction (DIR), which dynamically identifies intervention points.<n> Experiments across arithmetic and commonsense reasoning benchmarks demonstrate HPR's state-of-the-art efficiency-accuracy tradeoffs.
  arXiv  Detail & Related papers  (2025-11-13T07:26:01Z)
• From Harm to Help: Turning Reasoning In-Context Demos into Assets for Reasoning LMs [58.02809208460186]
  We revisit this paradox using high-quality traces from DeepSeek-R1 as demonstrations.<n>We find that adding more exemplars consistently degrades accuracy, even when demonstrations are optimal.<n>We introduce Insight-to-solve (I2S), a sequential test-time procedure that turns demonstrations into explicit, reusable insights.
  arXiv  Detail & Related papers  (2025-09-27T08:59:31Z)
• When Inverse Data Outperforms: Exploring the Pitfalls of Mixed Data in Multi-Stage Fine-Tuning [19.873408132011132]
  We construct a high-quality reverse reasoning dataset derived by inverting 1,000 forward examples from s1k.<n>We find that naively mixing forward and reverse data during supervised fine-tuning weakens the directional distinction.<n>These findings suggest that mixed reasoning data introduce conflicting supervision signals, underscoring the need for robust and direction-aware alignment strategies.
  arXiv  Detail & Related papers  (2025-09-16T13:36:36Z)
• Fractured Chain-of-Thought Reasoning [61.647243580650446]
  We introduce Fractured Sampling, a unified inference-time strategy that interpolates between full CoT and solution-only sampling.<n>We show that Fractured Sampling consistently achieves superior accuracy-cost trade-offs, yielding steep log-linear scaling gains in Pass@k versus token budget.
  arXiv  Detail & Related papers  (2025-05-19T11:30:41Z)
• Long-Short Chain-of-Thought Mixture Supervised Fine-Tuning Eliciting Efficient Reasoning in Large Language Models [23.34070841541423]
  We propose Long-Short Chain-of-Thought Mixture Supervised Fine-Tuning (LS-Mixture SFT)<n>Our experiments demonstrate that models trained using LS-Mixture SFT, compared to those trained with direct SFT, achieved an average accuracy improvement of 2.3%.<n>This work offers an approach to endow non-reasoning models with reasoning capabilities through supervised fine-tuning.
  arXiv  Detail & Related papers  (2025-05-06T12:18:11Z)
• SEAL: Steerable Reasoning Calibration of Large Language Models for Free [58.931194824519935]
  Large Language Models (LLMs) have demonstrated compelling capabilities for complex reasoning tasks via the extended chain-of-thought (CoT) reasoning mechanism.<n>Recent studies reveal substantial redundancy in the CoT reasoning traces, which negatively impacts model performance.<n>We introduce SEAL, a training-free approach that seamlessly calibrates the CoT process, improving accuracy while demonstrating significant efficiency gains.
  arXiv  Detail & Related papers  (2025-04-07T02:42:07Z)
• When More is Less: Understanding Chain-of-Thought Length in LLMs [51.631483479081645]
  Large Language Models (LLMs) employ Chain-of-Thought (CoT) reasoning to deconstruct complex problems.<n>This paper argues that longer CoTs are often presumed superior, arguing that longer is not always better.
  arXiv  Detail & Related papers  (2025-02-11T05:28:59Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.