Bits and Memories: Measuring Verbatim Extraction Across LLM Quantization

arXiv:2607.25451 · cs.LG, cs.CR · Submitted 2026-07-28 · Read on arXiv

Akshay Sasi

cs.LG, cs.CR

Submitted: 2026-07-28

Comments: 8 pages, 3 figures. Code: https://github.com/AkshaySasi/bits-and-memories. Data and results: https://huggingface.co/datasets/AkshaySasi/bits-and-memories

Code: https://github.com/AkshaySasi/bits-and-memories

License: http://arxiv.org/licenses/nonexclusive-distrib/1.0/

Terminology

• Quantization — 57× in this paper · explained in 12 episodes
• Perplexity — 12× in this paper · explained in 3 episodes
• Accuracy — 4× in this paper · explained in 1 episode
• In International Conference on Learning Representations — 3× in this paper
• In International Conference on Machine Learning — 2× in this paper
• Pruning — 2× in this paper · explained in 4 episodes
• Scaling Laws — 2× in this paper · explained in 4 episodes
• Transformer — 2× in this paper · explained in 2 episodes
• Decode — 1× in this paper · explained in 2 episodes
• Generalization — 1× in this paper · explained in 13 episodes
• Interpretability — 1× in this paper · explained in 15 episodes
• Redundancy — 1× in this paper · explained in 2 episodes

Sources

• Bits for Privacy: Evaluating Post-Training Quantization via Membership Inference
• Membership Inference Risks in Quantized Models: A Theoretical and Empirical Study
• Pythia: A Suite for Analyzing Large Language Models Across Training and Scaling
• Emergent and Predictable Memorization in Large Language Models
• Extracting Training Data from Large Language Models
• Quantifying Memorization Across Neural Language Models
• LLM.int8(): 8-bit Matrix Multiplication for Transformers at Scale
• QLoRA: Efficient Finetuning of Quantized LLMs
• The case for 4-bit precision: k-bit Inference Scaling Laws
• GPTQ: Accurate Post-Training Quantization for Generative Pre-trained Transformers
• The Pile: An 800GB Dataset of Diverse Text for Language Modeling
• Measuring memorization in language models via probabilistic extraction
• How Quantization Impacts Privacy Risk on LLMs for Code?
• CompLeak: Deep Learning Model Compression Exacerbates Privacy Leakage
• AWQ: Activation-aware Weight Quantization for LLM Compression and Acceleration
• Pointer Sentinel Mixture Models
• Scalable Extraction of Training Data from (Production) Language Models
• Membership Inference Attacks against Machine Learning Models
• Identifying and Mitigating Privacy Risks Stemming from Language Models: A Survey
• SmoothQuant: Accurate and Efficient Post-Training Quantization for Large Language Models

Related papers

• Polynomial-Augmented Neural Networks (PANNs) with Weak Orthogonality Constraints for Enhanced Function and PDE Approximation
• AIRL-S: Unifying Reinforcement Learning and Search-Based Test-Time Scaling via Adversarial Inverse Reinforcement Learning
• Transformers as Bayesian In-Context Experimenters: Smoothness-Adaptive Efficient ATE Estimation
• Convergence issues in Relational Concept Analysis based on AOC-posets
• Beliefs Beyond Posteriors: Local-Consistency Optimisation for Bayesian Neural Networks
• Understanding Diffusion Models via Ratio-Based Function Approximation with SignReLU Networks