Papers
Topics
Authors
Recent
Gemini 2.5 Flash
Gemini 2.5 Flash
119 tokens/sec
GPT-4o
56 tokens/sec
Gemini 2.5 Pro Pro
43 tokens/sec
o3 Pro
6 tokens/sec
GPT-4.1 Pro
47 tokens/sec
DeepSeek R1 via Azure Pro
28 tokens/sec
2000 character limit reached

PMET: Precise Model Editing in a Transformer (2308.08742v6)

Published 17 Aug 2023 in cs.CL, cs.AI, and cs.LG

Abstract: Model editing techniques modify a minor proportion of knowledge in LLMs at a relatively low cost, which have demonstrated notable success. Existing methods assume Transformer Layer (TL) hidden states are values of key-value memories of the Feed-Forward Network (FFN). They usually optimize the TL hidden states to memorize target knowledge and use it to update the weights of the FFN in LLMs. However, the information flow of TL hidden states comes from three parts: Multi-Head Self-Attention (MHSA), FFN, and residual connections. Existing methods neglect the fact that the TL hidden states contains information not specifically required for FFN. Consequently, the performance of model editing decreases. To achieve more precise model editing, we analyze hidden states of MHSA and FFN, finding that MHSA encodes certain general knowledge extraction patterns. This implies that MHSA weights do not require updating when new knowledge is introduced. Based on above findings, we introduce PMET, which simultaneously optimizes Transformer Component (TC, namely MHSA and FFN) hidden states, while only using the optimized TC hidden states of FFN to precisely update FFN weights. Our experiments demonstrate that PMET exhibits state-of-the-art performance on both the COUNTERFACT and zsRE datasets. Our ablation experiments substantiate the effectiveness of our enhancements, further reinforcing the finding that the MHSA encodes certain general knowledge extraction patterns and indicating its storage of a small amount of factual knowledge. Our code is available at https://github.com/xpq-tech/PMET.

Definition Search Book Streamline Icon: https://streamlinehq.com
References (31)
  1. GPT-NeoX-20B: An Open-Source Autoregressive Language Model. In Proceedings of BigScience Episode #5 – Workshop on Challenges & Perspectives in Creating Large Language Models, 95–136. virtual+Dublin: Association for Computational Linguistics.
  2. Knowledgeable or Educated Guess? Revisiting Language Models as Knowledge Bases. In Proceedings of the 59th Annual Meeting of the Association for Computational Linguistics and the 11th International Joint Conference on Natural Language Processing (Volume 1: Long Papers), 1860–1874. Online: Association for Computational Linguistics.
  3. Evaluating the Ripple Effects of Knowledge Editing in Language Models. arXiv preprint arXiv:2307.12976.
  4. Editing Factual Knowledge in Language Models. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 6491–6506. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics.
  5. Dissecting Recall of Factual Associations in Auto-Regressive Language Models. arXiv:2304.14767.
  6. Transformer Feed-Forward Layers Build Predictions by Promoting Concepts in the Vocabulary Space. In Proceedings of the 2022 Conference on Empirical Methods in Natural Language Processing, 30–45. Abu Dhabi, United Arab Emirates: Association for Computational Linguistics.
  7. Transformer Feed-Forward Layers Are Key-Value Memories. In Proceedings of the 2021 Conference on Empirical Methods in Natural Language Processing, 5484–5495. Online and Punta Cana, Dominican Republic: Association for Computational Linguistics.
  8. Self-Attention Attribution: Interpreting Information Interactions Inside Transformer. Proceedings of the AAAI Conference on Artificial Intelligence, 35(14): 12963–12971.
  9. How much does attention actually attend? Questioning the Importance of Attention in Pretrained Transformers. arXiv:2211.03495.
  10. Language Models as Knowledge Bases: On Entity Representations, Storage Capacity, and Paraphrased Queries. In Proceedings of the 16th Conference of the European Chapter of the Association for Computational Linguistics: Main Volume, 1772–1791. Online: Association for Computational Linguistics.
  11. Inspecting and Editing Knowledge Representations in Language Models. arXiv:2304.00740.
  12. Survey of Hallucination in Natural Language Generation. ACM Computing Surveys, 55(12): 1–38.
  13. Feed-Forward Blocks Control Contextualization in Masked Language Models. arXiv:2302.00456.
  14. Revealing the Dark Secrets of BERT. In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 4365–4374. Hong Kong, China: Association for Computational Linguistics.
  15. Zero-Shot Relation Extraction via Reading Comprehension. In Proceedings of the 21st Conference on Computational Natural Language Learning (CoNLL 2017), 333–342. Vancouver, Canada: Association for Computational Linguistics.
  16. Knowledge Graph Contrastive Learning Based on Relation-Symmetrical Structure. IEEE Transactions on Knowledge and Data Engineering, 1–12.
  17. Learn from relational correlations and periodic events for temporal knowledge graph reasoning. In Proceedings of the 46th International ACM SIGIR Conference on Research and Development in Information Retrieval, 1559–1568.
  18. Knowledge Consistency between Neural Networks and Beyond. arXiv:1908.01581.
  19. Locating and Editing Factual Associations in GPT. In Koyejo, S.; Mohamed, S.; Agarwal, A.; Belgrave, D.; Cho, K.; and Oh, A., eds., Advances in Neural Information Processing Systems, volume 35, 17359–17372. Curran Associates, Inc.
  20. Mass-Editing Memory in a Transformer. arXiv:2210.07229.
  21. Fast Model Editing at Scale. In International Conference on Learning Representations.
  22. Memory-Based Model Editing at Scale. arXiv:2206.06520.
  23. Murphy, A. H. 1996. The Finley affair: A signal event in the history of forecast verification. Weather and forecasting, 11(1): 3–20.
  24. Language Models as Knowledge Bases? In Proceedings of the 2019 Conference on Empirical Methods in Natural Language Processing and the 9th International Joint Conference on Natural Language Processing (EMNLP-IJCNLP), 2463–2473. Hong Kong, China: Association for Computational Linguistics.
  25. Editable Neural Networks. arXiv:2004.00345.
  26. GPT-J-6B: A 6 Billion Parameter Autoregressive Language Model. https://github.com/kingoflolz/mesh-transformer-jax.
  27. Interpretability in the wild: a circuit for indirect object identification in gpt-2 small. arXiv:2211.00593.
  28. Editing Large Language Models: Problems, Methods, and Opportunities. arXiv:2305.13172.
  29. A Survey of Large Language Models. arXiv:2303.18223.
  30. Can We Edit Factual Knowledge by In-Context Learning? arXiv:2305.12740.
  31. Modifying Memories in Transformer Models. arXiv:2012.00363.
User Edit Pencil Streamline Icon: https://streamlinehq.com
Authors (6)
  1. Xiaopeng Li (166 papers)
  2. Shasha Li (57 papers)
  3. Shezheng Song (12 papers)
  4. Jing Yang (320 papers)
  5. Jun Ma (347 papers)
  6. Jie Yu (98 papers)
Citations (95)
View on Semantic Scholar

Summary

Analysis of PMET: Precise Model Editing in a Transformer

The paper "PMET: Precise Model Editing in a Transformer" introduces a novel model editing technique specifically developed for refining LLMs without extensive retraining. This paper makes significant strides in enhancing the precision of model editing by addressing inadequacies in current approaches like ROME and MEMIT, thereby optimizing the hidden states of Transformer components such as Multi-Head Self-Attention (MHSA) and Feed-Forward Networks (FFN).

The authors critique existing methods that primarily optimize Transformer Layer (TL) hidden states, largely overlooking that those states incorporate a broader information flow beyond FFN requirements. Consequently, this oversight undermines the potential accuracy when updating FFN weights, often leading to broader performance deficits. Contrarily, PMET posits a dual optimization paradigm, enhancing Transformer Component (TC) hidden states, and leverages optimized FFN hidden states for weight updates. Notably, their analyses highlight that MHSA functions more as a knowledge extractor, embedding general knowledge extraction heuristics, reducing the necessity for weight modification within it.

These insights underpin PMET's foundational methodology, where the model is subject to optimization solely of the FFN hidden states rather than a blanket modification approach. Testing on datasets like counterfact and zsRE, the efficacy was evident as PMET surpassed state-of-the-art performance standards, particularly demonstrating superiority in knowledge retention and update precision. The methodology recorded a 3.3% improvement in reliability over conventional models in the counterfact dataset, emphasizing its advanced accuracy levels.

Furthermore, ablation experiments reinforced the robustness of PMET’s optimizations, particularly with insights into MHSA’s storage of limited factual knowledge, underscoring that FFN should remain the focal point for updates. While the paper doesn’t attempt to sensationalize outcomes as revolutionary, it underscores the nuanced and specific improvements made possible through PMET’s approach, hinting at broader implications for future AI developments, especially within LLMs.

This paper is positioned at the intersection of precision modifications in neural networks and the practical aspects of maintaining fidelity in LLM updates. It raises prospects for refining model updates in a way that harmonizes performance fidelity with practical utility, providing valuable perspectives for academic and industry researchers aiming to enhance LLM capabilities while minimizing resource expenditures. Future discourse may explore extending the PMET methodology across more sophisticated frameworks, potentially paving the way for nuanced understanding and deployment of LLMs in highly dynamic or evolving knowledge domains.

File Document Download Save Streamline Icon: https://streamlinehq.com PDF File Document Download Save Streamline Icon: https://streamlinehq.com Markdown
Github Logo Streamline Icon: https://streamlinehq.com

GitHub

  1. GitHub - xpq-tech/PMET: This is a repository for "PMET: Precise Model Editing in a Transformer" (52 stars)