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
Physics of Language Models: Part 3.1, Knowledge Storage and Extraction
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.
Forward citations
Cited by 22 Pith papers
-
DataComp-VLM: Improved Open Datasets for Vision-Language Models
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).
-
ContinuousBench: Can Differentially Private Synthetic Text Improve Capabilities?
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.
-
Can a Language Model Learn Facts Continually in Its Weights?
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...
-
Self-Study Reconsidered: The Hidden Fragility of Learning from Self-Generated QA
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%.
-
Vibe Calibration: Autonomous Bring-up of a 112-Qubit Superconducting Quantum Processor by a Skill-Orchestrating Language Agent
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.
-
Enough is as good as a feast: A Comprehensive Analysis of How Reinforcement Learning Mitigates Task Conflicts in LLMs
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.
-
Can a Language Model Learn Facts Continually in Its Weights?
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.
-
Pretraining Curricula Enable Selective Fine-tuning
Imbalanced pretraining curricula disentangle task circuits in transformers, improving in-context learning and the selectivity of refusal fine-tuning relative to balanced training.
-
DataComp-VLM: Improved Open Datasets for Vision-Language Models
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.
-
Boosting Self-Consistency with Ranking
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...
-
A Systematic Study of Behavioral Cloning for Scientific Data Annotation
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...
-
Synthetic Pre-Pre-Training Improves Language Model Robustness to Noisy Pre-Training Data
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.
-
Cram Less to Fit More: Training Data Pruning Improves Memorization of Facts
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...
-
TRINITY: An Evolved LLM Coordinator
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.
-
How Training Data Shapes the Use of Parametric and In-Context Knowledge in Language Models
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.
-
Provable Knowledge Acquisition and Extraction in One-Layer Transformers
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 ...
-
Understanding and Mitigating Bias Inheritance in LLM-based Data Augmentation on Downstream Tasks
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.
-
Data Quality over Capacity: Internalizing Documents into LoRA Adapters for Closed-Book QA
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%.
-
PASTA: A Paraphrasing And Self-Training Approach for Knowledge Updating in LLMs
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...
-
Do Activation Verbalization Methods Convey Privileged Information?
Activation verbalization methods for LLMs largely reflect the verbalizer model's parametric knowledge rather than privileged information from the target model's activations.
-
Analyzing the Correlation Between Hallucinations and Knowledge Conflicts in Large Language Models
Probing experiments indicate that hallucination patterns in LLMs are not fully reducible to knowledge conflict representations.
-
Beyond Inference-Only Deployment: Comparing Weight-Based Consolidation Against Cascading Compaction
Nightly LoRA consolidation retains 80.4% of knowledge vs 36.8% for cascading compaction across 10 conversations with 1146 test questions.
discussion (0)
Sign in with ORCID, Apple, or X to comment. Anyone can read and Pith papers without signing in.