mlsanityData to 14 Oct 2026

Rethinking Generalization in Reasoning SFT: A Conditional Analysis on Optimization, Data, and Model Capability

arXiv 2604.06628 ↗First posted 8 Apr 2026cs.AIcited by 14

On the map: LLM reasoning → Test-time compute and efficient thinking

Abstract (arXiv)

A prevailing narrative in LLM post-training holds that supervised finetuning (SFT) memorizes while reinforcement learning (RL) generalizes. We revisit this claim for reasoning SFT with long chain-of-thought (CoT) supervision and find that cross-domain generalization is not absent but conditional, jointly shaped by optimization dynamics, training data, and base-model capability. Some reported failures are under-optimization artifacts: cross-domain performance first degrades before recovering and improving with extended training (a dip-and-recovery pattern), so shorttraining checkpoints can underestimate generalization. Data quality and structure both matter: low-quality solutions broadly hurt generalization,while verified long-CoT traces yield consistent cross-domain gains. Model capability is essential: stronger models internalize transferable procedural patterns (e.g., backtracking) even from a toy arithmetic game, while weaker ones imitate surface verbosity. This generalization is asymmetric, however: reasoning improves while safety degrades, reframing the question from whether reasoning SFT generalizes to under what conditions and at what cost.

Key facts. Cited by 14 papers in the map (data through 14 Oct 2026). Placed in Test-time compute and efficient thinking, one of 17 directions in LLM reasoning. Builds on 24 of its 56 references that are in the map. First posted on arXiv 8 Apr 2026.

Builds on

24 of 56 references.

Closest ancestors · 3

Same direction, newest first

  1. 2026On the Role of Reasoning Patterns in the Generalization Discrepancy of Long Chain-of-Thought Supervised Fine-Tuning2 Apr
  2. 2026Data Repetition Beats Data Scaling in Long-CoT Supervised Fine-Tuning11 Feb
  3. 2025Wait, Wait, Wait... Why Do Reasoning Models Loop?15 Dec

Other references · 21

  1. 2025Qwen3 Technical Reportcited by 8,820
  2. 2024Qwen2.5 Technical Reportcited by 4,942
  3. 2024MMLU-Pro: A More Robust and Challenging Multi-Task Language Understanding Benchmarkcited by 1,848
  4. 2025Gemma 3 Technical Reportcited by 1,824
  5. 2023Instruction-Following Evaluation for Large Language Modelscited by 1,228
  6. 2024Length-Controlled AlpacaEval: A Simple Way to Debias Automatic Evaluatorscited by 897
  7. 2023HaluEval: A Large-Scale Hallucination Evaluation Benchmark for Large Language Modelscited by 341
  8. 2025Learning to Reason under Off-Policy Guidancecited by 281
  9. 2024On Memorization of Large Language Models in Logical Reasoningcited by 117
  10. 2025Does Math Reasoning Improve General LLM Capabilities? Understanding Transferability of LLM Reasoningcited by 106
  11. 2025Retaining by Doing: The Role of On-Policy Data in Mitigating Forgettingcited by 94
  12. 2025Reinforcement Fine-Tuning Naturally Mitigates Forgetting in Continual Post-Trainingcited by 47
9 more
  1. 2025Supervised Fine Tuning on Curated Data is Reinforcement Learning (and can be improved)cited by 38
  2. 2025Towards a Unified View of Large Language Model Post-Trainingcited by 36
  3. 2025Beyond Log Likelihood: Probability-Based Objectives for Supervised Fine-Tuning across the Model Capability Continuumcited by 18
  4. 2025RL Is Neither a Panacea Nor a Mirage: Understanding Supervised vs. Reinforcement Learning Fine-Tuning for LLMscited by 17
  5. 2025Self-Jailbreaking: Language Models Can Reason Themselves Out of Safety Alignment After Benign Reasoning Trainingcited by 12
  6. 2025Shape of Thought: When Distribution Matters More than Correctness in Reasoning Taskscited by 11
  7. 2026Towards On-Policy SFT: Distribution Discriminant Theory and its Applications in LLM Trainingcited by 11
  8. 2025Debunk the Myth of SFT Generalizationcited by 8
  9. 2025When Models Outthink Their Safety: Unveiling and Mitigating Self-Jailbreak in Large Reasoning Modelscited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2026: 14.

14 papers.

Sort:

By area:

  1. 2026Purified OPSD: On-Policy Self-Distillation Without Losing How to Thinkcited by 31
  2. 2026InfoSFT: Learn More and Forget Less with Information-Aware Token Weightingcited by 4
  3. 2026Zone of Proximal Policy Optimization: Teacher in Prompts, Not Gradientscited by 3
  4. 2026Zipping the Thought: When and How Compressed Reasoning Data Works in LLM Post-Trainingcited by 2
  5. 2026Capsule Lens: Locating and Tracking Concept Geometry in Model Representationscited by 1
  6. 2026Unlocking Fine-Grained Translation Quality Estimation in LRMs through Mutually Boosting Implicit and Explicit Reasoningcited by 0
  7. 2026Everything in Moderation: Per-Domain Coverage Optima and Alignment-Resistant Domain Gaps in Multi-Domain Mid-Trainingcited by 0
  8. 2026Reach or Solve? Attributing Agentic RL Gains with Checkpoint Handoffscited by 0
  9. 2026Towards Efficient Reasoning: Learning Causal Shortcuts for Diffusion Language Modelscited by 0
  10. 2026Generalization Dynamics of LM Pre-trainingcited by 0
  11. 2026Learning to Steer, Steering to See: Unveiling the Geometry of RLVR in Large Language Models via Trainable Vectorscited by 0
  12. 2026On-Policy Parameter Update Direction Underlies Generalization in LLM Post-Trainingcited by 0
2 more
  1. 2026Facts Don't Speak Louder than Words: The Behavioral Essence of Long-CoT Distillationcited by 0
  2. 2026Geometric Calibration of Supervised Fine-Tuningcited by 0

Similar, not cited

Close in wording, no citation either way

  1. 2025How and Why LLMs Generalize: A Fine-Grained Analysis of LLM Reasoning from Cognitive Behaviors to Low-Level Patterns
  2. 2026Memorize Theorems, Not Instances: Probing SFT Generalization through Mathematical Reasoning
  3. 2026Generalization in LLM Problem Solving: The Case of the Shortest Path
  4. 2025SFT Memorizes, RL Generalizes: A Comparative Study of Foundation Model Post-training
  5. 2026When Can LLMs Learn to Reason with Weak Supervision?
  6. 2026DeReason: A Difficulty-Aware Curriculum Improves Decoupled SFT-then-RL Training for General Reasoning
  7. 2025Does Learning Mathematical Problem-Solving Generalize to Broader Reasoning?
  8. 2025Revisiting Reinforcement Learning for LLM Reasoning from A Cross-Domain Perspective