Pith. sign in

REVIEW 14 cited by

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

Not yet reviewed by Pith; the record is open.

This paper has not been read by Pith yet. Machine review is queued; the pith claim, tier, and objections will appear here once it completes.

SPECIMEN: schema-true, not a live event

T0 review · schema-true

One-sentence machine reading of the paper's core claim.

pith:XXXXXXXX · record.json · timestamp

arxiv 2503.21676 v2 pith:N7BATLN5 submitted 2025-03-27 cs.CL cs.LG

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

classification cs.CL cs.LG
keywords knowledgedynamicslanguagemodelsdataacquisitiondistributionfactual
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
Share X Bluesky LinkedIn Reddit HN
read the original abstract

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.

discussion (0)

Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.

Forward citations

Cited by 14 Pith papers

Reviewed papers in the Pith corpus that reference this work. Sorted by Pith novelty score.

  1. Bridging Compute- and Data-Optimal Pretraining

    cs.LG 2026-07 conditional novelty 7.0

    Pretraining loss obeys a single law in which repeated or paraphrased tokens count as η(N, data-per-parameter, expansion-ratio) fresh tokens, with total effective data saturating as derived tokens grow.

  2. Can a Language Model Learn Facts Continually in Its Weights?

    cs.CL 2026-07 conditional novelty 7.0

    Facts written into LLM weights are stored but question-keyed; broad data creates usable knowledge while later writes destroy access rather than storage, leaving context more reliable than weights for composition and r...

  3. Structure Before Collapse: Transient semantic geometry in next-token prediction

    cs.LG 2026-06 unverdicted novelty 7.0

    Semantic geometry emerges transiently early in next-token prediction training before collapsing to Neural Collapse symmetry in synthetic settings with latent semantic factors.

  4. Natural Ungrokking: Asymmetric Control of Which Rules Survive Pretraining

    cs.LG 2026-06 unverdicted novelty 7.0

    During pretraining, language models exhibit natural ungrokking where learned rules are forgotten based on their support frequency in the corpus, with asymmetric editability of rule survival.

  5. Can a Language Model Learn Facts Continually in Its Weights?

    cs.CL 2026-07 unverdicted novelty 6.0

    Weight writes create question-keyed knowledge that later sequential writes redirect, so context—not weights—is the reliable channel for facts that must survive or be composed.

  6. Pretraining Curricula Enable Selective Fine-tuning

    cs.LG 2026-07 conditional novelty 6.0

    Imbalanced pretraining curricula disentangle task circuits in transformers, improving in-context learning and the selectivity of refusal fine-tuning relative to balanced training.

  7. Reliability Scales Inversely: Hallucinations Snowball Faster in Bigger Language Models

    cs.LG 2026-06 conditional novelty 6.0

    Hallucination in LLMs is driven by an oracle-invisible “decoding risk” term that grows with scale and causally compounds errors within a response.

  8. Why Fine-Tuning Encourages Hallucinations and How to Fix It

    cs.CL 2026-04 unverdicted novelty 6.0

    Supervised fine-tuning increases LLM hallucinations via interference among overlapping semantic representations; self-distillation mitigates this by regularizing output-distribution drift while freezing parameters pre...

  9. Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts

    cs.CL 2026-04 conditional novelty 6.0

    Loss-based pruning of training data to limit facts and flatten their frequency distribution enables a 110M-parameter GPT-2 model to memorize 1.3 times more entity facts than standard training, matching a 1.3B-paramete...

  10. Deep sequence models tend to memorize geometrically; it is unclear why

    cs.LG 2025-10 unverdicted novelty 6.0

    Deep sequence models develop geometric memory in embeddings that encodes novel global relationships, transforming l-fold composition tasks into 1-step navigation via a natural spectral bias connected to Node2Vec.

  11. How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models

    cs.CL 2025-09 unverdicted novelty 6.0

    Balanced parametric and in-context knowledge use in LLMs is an emergent property requiring intra-document repetition, moderate inconsistency, and skewed distributions in training data.

  12. The Future of Facts: Tracing the Factual Generation-Verification Gap

    cs.CL 2026-05 unverdicted novelty 5.0

    Empirical tracing across model families shows verification precedes and outlasts generation for facts, with updates producing simultaneous verification of old and new answers.

  13. Fine-Tuning Without Forgetting via Loss-Adaptive Learning Rates

    cs.LG 2026-05 unverdicted novelty 5.0

    FINCH is a loss-adaptive learning-rate schedule that reduces forgetting by 93% on average during LLM fine-tuning while matching standard task performance across several benchmarks.

  14. Do Activation Verbalization Methods Convey Privileged Information?

    cs.CL 2025-09 unverdicted novelty 5.0

    Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.