mlsanityData to 7 Oct 2026

How do language models learn facts? Dynamics, curricula and hallucinations

arXiv 2503.21676 ↗First posted 27 Mar 2025cs.CL cs.LGcited by 33

Not on the map. Why?

Abstract (arXiv)

Large language models accumulate vast knowledge during pre-training, yet the dynamics governing this acquisition remain poorly understood. This work investigates the learning dynamics of language models on a synthetic factual recall task, uncovering three key findings: First, language models learn in three phases, exhibiting a performance plateau before acquiring precise factual knowledge. Mechanistically, this plateau coincides with the formation of attention-based circuits that support recall. Second, the training data distribution significantly impacts learning dynamics, as imbalanced distributions lead to shorter plateaus. Finally, hallucinations emerge simultaneously with knowledge, and integrating new knowledge into the model through fine-tuning is challenging, as it quickly corrupts its existing parametric memories. Our results emphasize the importance of data distribution in knowledge acquisition and suggest novel data scheduling strategies to accelerate neural network training.

Key facts. Cited by 33 papers in the map (data through 7 Oct 2026). Builds on 35 of its 52 references that are in the map. First posted on arXiv 27 Mar 2025.

Builds on

35 of 52 references.

References · 35

  1. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,204
  2. 2020Language Models are Few-Shot Learnerscited by 32,199
  3. 2016Overcoming catastrophic forgetting in neural networkscited by 4,718
  4. 2022Training Compute-Optimal Large Language Modelscited by 2,927
  5. 2021TruthfulQA: Measuring How Models Mimic Human Falsehoodscited by 2,919
  6. 2022Locating and Editing Factual Associations in GPTcited by 2,609
  7. 2022Emergent Abilities of Large Language Modelscited by 2,483
  8. 2019Language Models as Knowledge Bases?cited by 1,922
  9. 2020Transformer Feed-Forward Layers Are Key-Value Memoriescited by 1,307
  10. 2022In-context Learning and Induction Headscited by 930
  11. 2023Progress measures for grokking via mechanistic interpretabilitycited by 822
  12. 2013Exact solutions to the nonlinear dynamics of learning in deep linear neural networkscited by 748
23 more
  1. 2022What Can Transformers Learn In-Context? A Case Study of Simple Function Classescited by 700
  2. 2023Siren's Song in the AI Ocean: A Survey on Hallucination in Large Language Modelscited by 695
  3. 2022Grokking: Generalization Beyond Overfitting on Small Algorithmic Datasetscited by 599
  4. 2023Are Emergent Abilities of Large Language Models a Mirage?cited by 515
  5. 2023Dissecting Recall of Factual Associations in Auto-Regressive Language Modelscited by 512
  6. 2023Scalable Extraction of Training Data from (Production) Language Modelscited by 415
  7. 2022Data Distributional Properties Drive Emergent In-Context Learning in Transformerscited by 303
  8. 2023Physics of Language Models: Part 3.1, Knowledge Storage and Extractioncited by 264
  9. 2024Griffin: Mixing Gated Linear Recurrences with Local Attention for Efficient Language Modelscited by 232
  10. 2024Does Fine-Tuning LLMs on New Knowledge Encourage Hallucinations?cited by 227
  11. 2023Zoology: Measuring and Improving Recall in Efficient Language Modelscited by 188
  12. 2021Open-Ended Learning Leads to Generally Capable Agentscited by 177
  13. 2023Fine-Tuning or Retrieval? Comparing Knowledge Injection in LLMscited by 175
  14. 2024Physics of Language Models: Part 3.3, Knowledge Capacity Scaling Lawscited by 137
  15. 2023The mechanistic basis of data dependence and abrupt learning in an in-context classification taskcited by 115
  16. 2024How Do Large Language Models Acquire Factual Knowledge During Pretraining?cited by 86
  17. 2023The Transient Nature of Emergent In-Context Learning in Transformerscited by 72
  18. 2013Active Learning of Inverse Models with Intrinsically Motivated Goal Exploration in Robotscited by 56
  19. 2024Competition Dynamics Shape Algorithmic Phases of In-Context Learningcited by 52
  20. 2024Understanding Factual Recall in Transformers via Associative Memoriescited by 45
  21. 2024Understanding Transformers via N-gram Statisticscited by 23
  22. 2025How Do LLMs Acquire New Knowledge? A Knowledge Circuits Perspective on Continual Pre-Trainingcited by 22
  23. 2024Emergent properties with repeated examplescited by 11

Built on by

Cited by year (2026 to 7 Oct)

Citing papers per year: 2025: 17; 2026: 16.

33 papers.

Sort:

By area:

  1. 2025On the generalization of language models from in-context learning and finetuning: a controlled studycited by 39
  2. 2025Tracing the Representation Geometry of Language Models from Pretraining to Post-trainingcited by 33
  3. 2025Pixels Versus Priors: Controlling Knowledge Priors in Vision-Language Models through Visual Counterfactscited by 30
  4. 2025The emergence of sparse attention: impact of data distribution and benefits of repetitioncited by 25
  5. 2025Robust Hallucination Detection in LLMs via Adaptive Token Selectioncited by 19
  6. 2025Data Mixing Can Induce Phase Transitions in Knowledge Acquisitioncited by 15
  7. 2025Do Activation Verbalization Methods Convey Privileged Information?cited by 8
  8. 2025Comprehension Without Competence: Architectural Limits of LLMs in Symbolic Computation and Reasoningcited by 7
  9. 2025RuleReasoner: Reinforced Rule-based Reasoning via Domain-aware Dynamic Samplingcited by 6
  10. 2026Why Fine-Tuning Encourages Hallucinations and How to Fix Itcited by 6
  11. 2025Rote Learning Considered Useful: Generalizing over Memorized Data in LLMscited by 5
  12. 2026Training AI Scientists to Replicate Researchcited by 4
21 more
  1. 2025Pierce the Mists, Greet the Sky: Decipher Knowledge Overshadowing via Knowledge Circuit Analysiscited by 3
  2. 2025How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Modelscited by 3
  3. 2025Facts in Stats: Impacts of Pretraining Diversity on Language Model Generalizationcited by 3
  4. 2026Universal Redundancies in Time Series Foundation Modelscited by 3
  5. 2026Understanding Contextual Recall in Transformers: How Finetuning Enables In-Context Reasoning over Pretraining Knowledgecited by 2
  6. 2026Cram Less to Fit More: Training Data Pruning Improves Memorization of Factscited by 2
  7. 2026Structure Before Collapse: Transient semantic geometry in next-token predictioncited by 2
  8. 2025Probability Consistency in Large Language Models: Theoretical Foundations Meet Empirical Discrepanciescited by 1
  9. 2026How Do Large Language Models Learn Concepts During Continual Pre-Training?cited by 1
  10. 2026Hierarchical Latent Structures in Data Generation Process Unify Mechanistic Phenomena across Scalecited by 1
  11. 2026Fine-Tuning Without Forgetting via Loss-Adaptive Learning Ratescited by 1
  12. 2026The Future of Facts: Tracing the Factual Generation-Verification Gapcited by 1
  13. 2026Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretrainingcited by 1
  14. 2025FictionalQA: A Dataset for Studying Memorization and Knowledge Acquisitioncited by 0
  15. 2025Tracing Multilingual Knowledge Acquisition Dynamics in Domain Adaptation: A Case Study of English-Japanese Biomedical Adaptationcited by 0
  16. 2025Rewriting History: A Recipe for Interventional Analyses to Study Data Effects on Model Behaviorcited by 0
  17. 2026Gradient Flow Polarizes Softmax Outputs towards Low-Entropy Solutionscited by 0
  18. 2026Understanding Data Temporality Impact on Large Language Models Pre-trainingcited by 0
  19. 2026Pretraining Curricula Enable Selective Fine-tuningcited by 0
  20. 2026Can a Language Model Learn Facts Continually in Its Weights?cited by 0
  21. 2026Reliability Scales Inversely: Hallucinations Snowball Faster in Bigger Language Modelscited by 0