mlsanityData to 14 Oct 2026

Context-Free Synthetic Data Mitigates Forgetting

arXiv 2505.13811 ↗First posted 20 May 2025cs.LGcited by 5

On the map: Synthetic data and self-improvement for LLMs → Model collapse from synthetic data

Abstract (arXiv)

Fine-tuning a language model often results in a degradation of its existing performance on other tasks, due to a shift in the model parameters; this phenomenon is often referred to as (catastrophic) forgetting. We are interested in mitigating this, in settings where we only have access to the model weights but no access to its training data/recipe. A natural approach is to penalize the KL divergence between the original model and the new one. Our main realization is that a simple process - which we term context-free generation - allows for an approximate unbiased estimation of this KL divergence. We show that augmenting a fine-tuning dataset with context-free generations mitigates forgetting, in two settings: (a) preserving the zero-shot performance of pretrained-only models, and (b) preserving the reasoning performance of thinking models. We show that contextual synthetic data, and even a portion of the pretraining data, are less effective. We also investigate the effect of choices like generation temperature, data ratios etc. We present our results for OLMo-1B for pretrained-only setting and R1-Distill-Llama-8B for the reasoning setting.

Key facts. Cited by 5 papers in the map (data through 14 Oct 2026). Placed in Model collapse from synthetic data, one of 13 directions in Synthetic data and self-improvement for LLMs. Builds on 29 of its 36 references that are in the map. First posted on arXiv 20 May 2025.

Builds on

29 of 36 references.

References · 29

  1. 2021LoRA: Low-Rank Adaptation of Large Language Modelscited by 14,380
  2. 2024DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Modelscited by 9,379
  3. 2023Efficient Memory Management for Large Language Model Serving with PagedAttentioncited by 5,944
  4. 2016Overcoming catastrophic forgetting in neural networkscited by 4,787
  5. 2020What Disease does this Patient Have? A Large-scale Open Domain Question Answering Dataset from Medical Examscited by 1,356
  6. 2022Editing Models with Task Arithmeticcited by 1,120
  7. 2018Experience Replay for Continual Learningcited by 1,095
  8. 2023A Comprehensive Survey of Continual Learning: Theory, Method and Applicationcited by 872
  9. 2021Robust fine-tuning of zero-shot modelscited by 797
  10. 2023MetaMath: Bootstrap Your Own Mathematical Questions for Large Language Modelscited by 735
  11. 2024OLMo: Accelerating the Science of Language Modelscited by 613
  12. 2024Dolma: an Open Corpus of Three Trillion Tokens for Language Model Pretraining Researchcited by 483
17 more
  1. 2019Dreaming to Distill: Data-free Knowledge Transfer via DeepInversioncited by 423
  2. 2024LoRA Learns Less and Forgets Lesscited by 328
  3. 2019Episodic Memory in Lifelong Language Learningcited by 188
  4. 2023Mitigating the Alignment Tax of RLHFcited by 183
  5. 2017FearNet: Brain-Inspired Model for Incremental Learningcited by 177
  6. 2021Always Be Dreaming: A New Approach for Data-Free Class-Incremental Learningcited by 131
  7. 2024Mitigating Catastrophic Forgetting in Large Language Models with Self-Synthesized Rehearsalcited by 118
  8. 2025MedReason: Eliciting Factual Medical Reasoning Steps in LLMs via Knowledge Graphscited by 110
  9. 2025NaturalReasoning: Reasoning in the Wild with 2.8M Challenging Questionscited by 84
  10. 2023Maintaining Plasticity in Continual Learning via Regenerative Regularizationcited by 65
  11. 2024Towards Lifelong Learning of Large Language Models: A Surveycited by 64
  12. 2018Incremental Classifier Learning with Generative Adversarial Networkscited by 32
  13. 2024Lottery Ticket Adaptation: Mitigating Destructive Interference in LLMscited by 31
  14. 2021How do Quadratic Regularizers Prevent Catastrophic Forgetting: The Role of Interpolationcited by 19
  15. 2025Upweighting Easy Samples in Fine-Tuning Mitigates Forgettingcited by 16
  16. 2024MoFO: Momentum-Filtered Optimizer for Mitigating Forgetting in LLM Fine-Tuningcited by 14
  17. 2025Forget About It: Mitigating Forgetting Without Replay or Auxiliary Objectivescited by 14

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2025: 1; 2026: 4.

5 papers.

  1. 2025SFT Doesn't Always Hurt General Capabilities: Revisiting Domain-Specific Fine-Tuning in LLMscited by 24
  2. 2026(How) Learning Rates Regulate Catastrophic Overtrainingcited by 6
  3. 2026Decouple Searching from Training: Scaling Data Mixing via Model Merging for Large Language Model Pre-trainingcited by 2
  4. 2026What LLMs Think When You Don't Tell Them What to Think About?cited by 1
  5. 2026Fine-Tuning Without Forgetting via Loss-Adaptive Learning Ratescited by 1

Similar, not cited

Close in wording, no citation either way

  1. 2026Forgetting in Language Models: Capacity, Optimization, and Self-Generated Replay
  2. 2026Talking to Yourself: Defying Forgetting in Large Language Models
  3. 2026Learning from Synthetic Data without Model Collapse in Iterative Instruction Tuning
  4. 2025Forget to Know, Remember to Use: Context-Aware Unlearning for Large Language Models
  5. 2024Preserving Knowledge in Large Language Model with Model-Agnostic Self-Decompression
  6. 2025Retaining by Doing: The Role of On-Policy Data in Mitigating Forgetting
  7. 2024Scaling Laws for Forgetting When Fine-Tuning Large Language Models
  8. 2025ReLearn: Unlearning via Learning for Large Language Models