Pith. sign in

REVIEW 22 cited by

Physics of Language Models: Part 3.1, Knowledge Storage and Extraction

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 2309.14316 v3 pith:LNW4HX7D submitted 2023-09-25 cs.CL cs.AIcs.LG

Physics of Language Models: Part 3.1, Knowledge Storage and Extraction

classification cs.CL cs.AIcs.LG
keywords knowledgepretrainingdatamodelstrainingaugmentationcorrelationduring
verification ladder T0 review T1 audit T2 compute T3 formal T4 reserved
0 comments
read the original abstract

Large language models (LLMs) can store a vast amount of world knowledge, often extractable via question-answering (e.g., "What is Abraham Lincoln's birthday?"). However, do they answer such questions based on exposure to similar questions during training (i.e., cheating), or by genuinely learning to extract knowledge from sources like Wikipedia? In this paper, we investigate this issue using a controlled biography dataset. We find a strong correlation between the model's ability to extract knowledge and various diversity measures of the training data. $\textbf{Essentially}$, for knowledge to be reliably extracted, it must be sufficiently augmented (e.g., through paraphrasing, sentence shuffling, translations) $\textit{during pretraining}$. Without such augmentation, knowledge may be memorized but not extractable, leading to 0% accuracy, regardless of subsequent instruction fine-tuning. To understand why this occurs, we employ (nearly) linear probing to demonstrate a strong connection between the observed correlation and how the model internally encodes knowledge -- whether it is linearly encoded in the hidden embeddings of entity names or distributed across other token embeddings in the training text. This paper provides $\textbf{several key recommendations for LLM pretraining in the industry}$: (1) rewrite the pretraining data -- using small, auxiliary models -- to provide knowledge augmentation, and (2) incorporate more instruction-finetuning data into the pretraining stage before it becomes too late.

discussion (0)

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

Forward citations

Cited by 22 Pith papers

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

  1. DataComp-VLM: Improved Open Datasets for Vision-Language Models

    cs.CV 2026-06 conditional novelty 8.0

    DataComp-VLM benchmark shows instruction-heavy data mixing outperforms filtering for VLM training, with DCVLM-Baseline achieving 63.6% on 33 tasks for 8B models (+5.4pp over FineVision).

  2. ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?

    cs.LG 2026-06 unverdicted novelty 8.0

    ContinuousBench shows non-private synthetic text transfers corpus-specific capabilities while state-of-the-art DP methods fail to do so even at ε=100.

  3. 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...

  4. Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA

    cs.AI 2026-06 unverdicted novelty 7.0

    Self-generated QA supervision for language models is fragile due to non-uniform question selection and instruction compliance during answering, with mitigations that reduce compliance from 88% to 13%.

  5. Vibe Calibration: Autonomous Bring-up of a 112-Qubit Superconducting Quantum Processor by a Skill-Orchestrating Language Agent

    quant-ph 2026-06 unverdicted novelty 7.0

    Vibe Calibration uses LLM agents to orchestrate reusable decision-tree Skills distilled from expert knowledge, autonomously calibrating 108/112 qubits in 4.7 hours with 4-5x speedup and transferable workflows.

  6. Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs

    cs.CL 2026-07 conditional novelty 6.0

    RL-trained LLMs keep most of their skills after weight merging, while SFT-trained LLMs drop about 19% on average, because RL keeps parameter updates smaller and more task-compatible.

  7. 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.

  8. 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.

  9. DataComp-VLM: Improved Open Datasets for Vision-Language Models

    cs.CV 2026-06 unverdicted novelty 6.0

    DataComp-VLM benchmark shows instruction-heavy data mixtures outperform caption-heavy ones for VLM training, with DCVLM-Baseline reaching 63.6% on 33 tasks using 200B tokens, +5.4pp over FineVision.

  10. Boosting Self-Consistency with Ranking

    cs.CL 2026-06 unverdicted novelty 6.0

    RISC reformulates self-consistency answer selection as a ranking task solved by a lightweight LambdaRank model with five hand-designed features, yielding better accuracy-efficiency trade-offs than majority voting on Q...

  11. A Systematic Study of Behavioral Cloning for Scientific Data Annotation

    cs.HC 2026-05 unverdicted novelty 6.0

    Introduces 9 synthetic annotation tasks and benchmarks for behavioral cloning, finding hierarchical skill learning, scaling benefits, effective multi-task pretraining, and shared internal representations of task phase...

  12. Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data

    cs.CL 2026-05 unverdicted novelty 6.0

    Synthetic pre-pre-training on structured data improves LLM robustness to noisy pre-training, matching baseline loss with up to 49% fewer natural tokens for a 1B model.

  13. 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...

  14. TRINITY: An Evolved LLM Coordinator

    cs.LG 2025-12 unverdicted novelty 6.0

    A compact 0.6B-parameter coordinator with a 10K-parameter head uses evolutionary strategy to dynamically delegate roles to LLMs, achieving SOTA results such as 86.2% on LiveCodeBench.

  15. 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.

  16. Provable Knowledge Acquisition and Extraction in One-Layer Transformers

    cs.LG 2025-07 unverdicted novelty 6.0

    In a stylized one-layer transformer, pre-training encodes factual knowledge via relation-specific feature directions and attention patterns; fine-tuning extracts it through a relation-covering mechanism that succeeds ...

  17. Understanding and Mitigating Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks

    cs.LG 2025-02 unverdicted novelty 6.0

    Empirical study across 10 tasks showing bias inheritance from LLM-augmented data harms related downstream performance, with three misalignment factors and three mitigation strategies identified.

  18. Data Quality over Capacity: Internalizing Documents into LoRA Adapters for Closed-Book QA

    cs.CL 2026-07 conditional novelty 5.0

    For closed-book QA via LoRA adapters, once capacity is sufficient, training-data quality—especially answer brevity—outperforms rank, learning rate, and architecture changes, lifting 15-doc accuracy from 57.7% to 85.7%.

  19. PASTA: A Paraphrasing And Self-Training Approach for Knowledge Updating in LLMs

    cs.CL 2026-06 unverdicted novelty 5.0

    PASTA combines data augmentation and a self-learning DPO process to integrate new factual knowledge from news articles into LLMs, raising accuracy from 0.02 to 0.82 on post-cutoff questions while preserving general ca...

  20. 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.

  21. Analyzing the Correlation Between Hallucinations and Knowledge Conflicts in Large Language Models

    cs.CL 2026-06 unverdicted novelty 4.0

    Probing experiments indicate that hallucination patterns in LLMs are not fully reducible to knowledge conflict representations.

  22. Beyond Inference-Only Deployment: Comparing Weight-Based Consolidation Against Cascading Compaction

    cs.AI 2026-05 unverdicted novelty 4.0

    Nightly LoRA consolidation retains 80.4% of knowledge vs 36.8% for cascading compaction across 10 conversations with 1146 test questions.