Achieving Exact Data Deletion from Language Model Memory

Vishwajith Ramesh· July 31, 2026 View original

Key takeaways

  • Exact data deletion from LLMs is achievable, depending on memory representation.
  • Algebraic decrement works for addressable records in certain memory architectures.
  • Checkpoint-based rewind-and-replay enables exact deletion for entangled writes.
  • This capability is crucial for privacy, compliance, and model security.

Who benefits

HealthcareBFSILegalTechGovernment

Summary

New research demonstrates exact deletion from language model memory, distinguishing between addressable influence and entangled writes. It shows that records can be precisely removed from models like Gemma 3 and Kimi Linear hybrids through algebraic decrement or checkpointed rewind-and-replay.

The ability to precisely delete specific data from a language model's persistent memory is a critical challenge, especially concerning privacy and compliance. This research explores the mechanisms required for exact deletion, differentiating between data influences that can be algebraically subtracted and those that are deeply entangled within the model's recurrent state through subsequent writes. The study first examines Gemma 3, where global-attention layers were replaced with support-vector memory. For this architecture, algebraic decrement and refitting retained keys resulted in next-token outputs that were nearly identical to a never-ingested baseline after deleting support tokens. While utility costs increased with model size, the method proved effective. A masked-refit proxy also showed indistinguishable performance from the baseline under various attacks. Secondly, the research investigated a 48B Kimi Linear hybrid model. It found that additive writes allowed for fixed or corrected decrements, whereas the delta rule made a significant portion of a record's contribution suffix-dependent. Crucially, a checkpointed rewind-and-replay method achieved bit-for-bit exact deletion of real clinical records, matching never-ingested logits and recurrent states. This work establishes that exact deletion is fundamentally tied to the memory representation: addressable records can be subtracted, while entangled writes necessitate a replay mechanism.

Why it matters

Exact data deletion is paramount for data privacy, regulatory compliance (e.g., GDPR 'right to be forgotten'), and model security. This research provides a pathway for organizations to manage sensitive information within large language models more effectively.

How to implement this in your domain

  1. 1Evaluate current LLM architectures for their capacity to support exact data deletion based on memory representation.
  2. 2Develop internal protocols for data deletion requests, incorporating 'subtract' or 'replay' strategies as appropriate for model types.
  3. 3Implement checkpointing and version control for LLM training to enable rewind-and-replay deletion methods.
  4. 4Collaborate with model developers to integrate memory representations that facilitate addressable record removal.

Original post by Vishwajith Ramesh

"arXiv:2607.27539v1 Announce Type: new Abstract: Exact deletion from persistent language-model memory depends on how that memory represents a record. Addressable influence can be removed by algebraic decrement; influence transformed by later writes inside shared recurrent state re…"

View on X

Originally posted by Vishwajith Ramesh on X · view source

Want to go deeper?

Turn these trends into skills with Learnijoy's hands-on AI & tech courses.

Explore courses

More in AI Research

AI ResearchAI Engineering & DevTools

New Framework Improves Partial Multi-View Clustering Performance.

DAS-PMVC is a novel framework for partial multi-view clustering that addresses view asymmetry and irrelevant samples by leveraging dual alignment and structure enhancement. It uses anchor graph structure alignment, structure-enhanced feature learning, and a dual alignment strategy to achieve superior clustering performance on various datasets.

Shubin Ma, Liang Zhao, Chuanye He, Zhenjiao Liu, Liang Zou, Lin Yuanbo Wu, Yu ShaoJul 31, 2026
AI Engineering & DevToolsAI Research

Dual Teachers Improve Adversarial Robustness and Accuracy.

This work extends Information Bottleneck Distillation (IBD) by introducing a "clean teacher" alongside a robust teacher to improve the robustness/accuracy tradeoff against adversarial attacks. The proposed method transfers features from both teachers to a student model, achieving better clean accuracy while maintaining adversarial robustness, outperforming original IBD and competing with state-of-the-art approaches.

Vincent Ryusuke Takahashi, Yoshinari Takeishi, Jun'ichi Takeuchi, Kave SalamatianJul 31, 2026
AI Engineering & DevToolsAI Research

Dynamic Batch Sizes Improve Large Language Model Training Efficiency.

This paper proposes a new approach to deep learning dynamics, deriving joint scaling laws for loss based on both learning rate and batch size schedules. It introduces an optimal dynamic batch size schedule that consistently outperforms static batch size baselines, highlighting its importance for large language model training.

Jiaxiang Li, Zhiqi Bu, Shiyun XuJul 31, 2026