pith. sign in

arxiv: 2603.19297 · v2 · submitted 2026-03-11 · 💻 cs.LG

CLaRE-ty Amid Chaos: Quantifying Representational Entanglement to Predict Ripple Effects in LLM Editing

Manit Baser , Alperen Yildiz , Dinil Mon Divakaran , Mohan Gurusamy This is my paper

Pith reviewed 2026-05-15 12:47 UTC · model grok-4.3

classification 💻 cs.LG
keywords LLM editingrepresentational entanglementripple effectsforward activationsfact preservationmodel editingentanglement graphs
0
0 comments X

The pith

CLaRE measures fact entanglement in LLMs with forward activations from one intermediate layer to predict editing ripple effects.

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

Large language models keep facts in overlapping representations, so changing one fact often changes many others in unintended ways. CLaRE scores these overlaps by comparing forward activations at a single middle layer instead of running expensive gradient calculations. The authors built entanglement graphs over a corpus of more than eleven thousand facts drawn from existing datasets and several models. When the scores are used to choose which facts to protect during an edit, the method shows stronger alignment with observed ripple effects than earlier approaches.

Core claim

CLaRE quantifies representational entanglement between facts by measuring similarity of their forward activations in one chosen intermediate layer. These entanglement scores correlate 62.2 percent better with actual ripple effects than gradient-based baselines, while the computation runs 2.74 times faster and uses 2.85 times less peak GPU memory. The resulting graphs support better preservation sets, audit trails, red-teaming, and post-edit checks without storing full fact representations.

What carries the argument

CLaRE entanglement score, computed from cosine similarity of forward activations at a single intermediate layer for pairs of facts.

If this is right

  • Editing procedures can select preservation sets directly from the entanglement graph to limit unintended changes.
  • Post-edit evaluation scales by querying the graph instead of exhaustive behavioral tests on every related fact.
  • Red-teaming can focus on high-entanglement facts to surface ripple effects with fewer trials.
  • Audit records of model updates become feasible by tracing which facts share high entanglement scores.

Where Pith is reading between the lines

These are editorial extensions of the paper, not claims the author makes directly.

  • Forward-pass entanglement graphs may transfer to other model interventions such as targeted fine-tuning or selective pruning.
  • If the single-layer approximation generalizes, it could expose common organizational patterns in how transformers store knowledge across different architectures.
  • Dynamic editing systems could use the graphs to update clusters of entangled facts in one coordinated step rather than sequentially.

Load-bearing premise

Entanglement measured via forward activations from a single intermediate layer accurately captures how edits propagate through the full hidden space and produce behavioral ripple effects.

What would settle it

Perform a set of edits on high-CLaRE versus low-CLaRE fact pairs and check whether the measured behavioral changes after editing match the predicted ordering and magnitude of ripple effects.

Figures

Figures reproduced from arXiv: 2603.19297 by Alperen Yildiz, Dinil Mon Divakaran, Manit Baser, Mohan Gurusamy.

Figure 1
Figure 1. Figure 1: A targeted update to a political fact may [PITH_FULL_IMAGE:figures/full_fig_p001_1.png] view at source ↗
Figure 2
Figure 2. Figure 2: For each fact, GradSim computes the entire gradient, while CLARE uses a single forward pass up till the last critical layer, enabling faster and scalable entanglement mapping. These unintended changes are referred to as rip￾ple effects (Cohen et al., 2024). Let F denote the original set of facts represented by the model and ∆F denote the intended edits. After editing, the model’s knowledge becomes F ′ = F … view at source ↗
Figure 3
Figure 3. Figure 3: Correlation patterns for AlphaEdit: entangle [PITH_FULL_IMAGE:figures/full_fig_p005_3.png] view at source ↗
Figure 4
Figure 4. Figure 4: Performance comparison between CLARE and GradSim in terms of Spearman correlation (ρs). The left panel shows ρs between entanglement values and ℓ2 logit shift, and right panel shows ρs between en￾tanglement values and |∆ log P(y)|. CLARE (wider, transparent bars) consistently achieves higher ρs than GradSim (narrower, solid bars). age of 2.74× speed-up over GradSim. CLARE’s factual representations are extr… view at source ↗
Figure 6
Figure 6. Figure 6: Top-5 entangled facts in GPT-J in our corpus. [PITH_FULL_IMAGE:figures/full_fig_p008_6.png] view at source ↗
Figure 7
Figure 7. Figure 7: An entanglement cluster in GPT-J with its ten [PITH_FULL_IMAGE:figures/full_fig_p009_7.png] view at source ↗
Figure 8
Figure 8. Figure 8: Layerwise Spearman correlation (ρs) between CLARE and observed ripple magnitudes. Correlation peaks around the last critical layer indicate that it is most informative about entanglement estimation. 13 [PITH_FULL_IMAGE:figures/full_fig_p013_8.png] view at source ↗
Figure 9
Figure 9. Figure 9: Distribution of cosine similarities between [PITH_FULL_IMAGE:figures/full_fig_p015_9.png] view at source ↗
Figure 10
Figure 10. Figure 10: Correlation patterns for RECT across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p019_10.png] view at source ↗
Figure 11
Figure 11. Figure 11: Correlation patterns for RECT across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p019_11.png] view at source ↗
Figure 12
Figure 12. Figure 12: Correlation patterns for MEMIT across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p020_12.png] view at source ↗
Figure 13
Figure 13. Figure 13: Correlation patterns for MEMIT across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p020_13.png] view at source ↗
Figure 14
Figure 14. Figure 14: Correlation patterns for ROME across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p020_14.png] view at source ↗
Figure 15
Figure 15. Figure 15: Correlation patterns for ROME across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p020_15.png] view at source ↗
Figure 16
Figure 16. Figure 16: Correlation patterns for PRUNE across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p021_16.png] view at source ↗
Figure 17
Figure 17. Figure 17: Correlation patterns for PRUNE across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p021_17.png] view at source ↗
Figure 18
Figure 18. Figure 18: CLARE correlation patterns for AlphaEdit across different models for entanglement vs ℓ2 logit shift. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 1 2 3 4 5 6 7 8 | lo g P(y)| Spearman Correlation = 0.678 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 2 4 6 8 10 12 | lo g P(y)| Spearman Correlation = 0.689 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement score 0 1 2… view at source ↗
Figure 19
Figure 19. Figure 19: CLARE correlation patterns for AlphaEdit across different models for entanglement vs |∆ log P(y)|. 21 [PITH_FULL_IMAGE:figures/full_fig_p021_19.png] view at source ↗
Figure 20
Figure 20. Figure 20: CLARE correlation patterns for RECT across different models for entanglement vs ℓ2 logit shift. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0.0 0.2 0.4 0.6 0.8 1.0 1.2 | lo g P(y)| Spearman Correlation = 0.678 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 2 4 6 8 10 12 | lo g P(y)| Spearman Correlation = 0.8 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement score 0 … view at source ↗
Figure 21
Figure 21. Figure 21: CLARE correlation patterns for RECT across different models for entanglement vs |∆ log P(y)|. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 2 4 6 8 10 12 14 16 2 lo git s hif t Spearman Correlation = 0.913 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 20 40 60 80 2 lo git s hif t Spearman Correlation = 0.892 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement score 0 … view at source ↗
Figure 22
Figure 22. Figure 22: CLARE correlation patterns for MEMIT across different models for entanglement vs ℓ2 logit shift. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0.0 0.5 1.0 1.5 2.0 2.5 3.0 | lo g P(y)| Spearman Correlation = 0.72 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 2 4 6 8 10 | lo g P(y)| Spearman Correlation = 0.671 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement score 0.0… view at source ↗
Figure 23
Figure 23. Figure 23: Correlation patterns for MEMIT across different models for entanglement vs [PITH_FULL_IMAGE:figures/full_fig_p022_23.png] view at source ↗
Figure 24
Figure 24. Figure 24: CLARE correlation patterns for ROME across different models for entanglement vs ℓ2 logit shift. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 2 4 6 8 | lo g P(y)| Spearman Correlation = 0.682 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 5 10 15 20 | lo g P(y)| Spearman Correlation = 0.765 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement score 0 1 2 3 4 5 6 7 | lo … view at source ↗
Figure 25
Figure 25. Figure 25: CLARE correlation patterns for ROME across different models for entanglement vs |∆ log P(y)|. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 5 10 15 20 25 30 2 lo git s hif t Spearman Correlation = 0.894 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 40 60 80 100 120 140 2 lo git s hif t Spearman Correlation = 0.837 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entanglement scor… view at source ↗
Figure 26
Figure 26. Figure 26: CLARE correlation patterns for PRUNE across different models for entanglement vs ℓ2 logit shift. 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0 1 2 3 4 5 6 7 8 | lo g P(y)| Spearman Correlation = 0.765 (a) GPT2-XL 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 1.0 Entanglement score 0.0 2.5 5.0 7.5 10.0 12.5 15.0 17.5 | lo g P(y)| Spearman Correlation = 0.815 (b) Llama3 0.2 0.3 0.4 0.5 0.6 0.7 0.8 0.9 Entangle… view at source ↗
Figure 27
Figure 27. Figure 27: CLARE correlation patterns for PRUNE across different models for entanglement vs |∆ log P(y)|. 23 [PITH_FULL_IMAGE:figures/full_fig_p023_27.png] view at source ↗
Figure 28
Figure 28. Figure 28: Most entangled facts ranked by representational connectivity in GPT2-XL. [PITH_FULL_IMAGE:figures/full_fig_p024_28.png] view at source ↗
Figure 29
Figure 29. Figure 29: Most entangled facts ranked by representational connectivity in Llama3. [PITH_FULL_IMAGE:figures/full_fig_p025_29.png] view at source ↗
Figure 30
Figure 30. Figure 30: Most entangled facts ranked by representational connectivity in GPT-J. [PITH_FULL_IMAGE:figures/full_fig_p026_30.png] view at source ↗
Figure 31
Figure 31. Figure 31: Most entangled facts ranked by representational connectivity in GPT-J (contd.). [PITH_FULL_IMAGE:figures/full_fig_p027_31.png] view at source ↗
Figure 32
Figure 32. Figure 32: EleutherAI GPT-J-6B: Clusters 01-02 27 [PITH_FULL_IMAGE:figures/full_fig_p027_32.png] view at source ↗
Figure 33
Figure 33. Figure 33: EleutherAI GPT-J-6B: Clusters 03-08 28 [PITH_FULL_IMAGE:figures/full_fig_p028_33.png] view at source ↗
Figure 34
Figure 34. Figure 34: EleutherAI GPT-J-6B: Clusters 09-13 29 [PITH_FULL_IMAGE:figures/full_fig_p029_34.png] view at source ↗
Figure 35
Figure 35. Figure 35: GPT2-XL: Clusters 01-06 30 [PITH_FULL_IMAGE:figures/full_fig_p030_35.png] view at source ↗
Figure 36
Figure 36. Figure 36: GPT2-XL: Cluster 07 54 facts | 9 subjects | 29.4% cross-subject edges Ukraine (12, 8.4%) South Korea (10, 7.0%) India (9, 6.3%) France (9, 6.3%) Turkey (9, 6.3%) Poland (9, 6.3%) Spain (8, 5.6%) Slovakia (7, 4.9%) Chile (7, 4.9%) Slovenia (7, 4.9%) (a) Cluster 1 143 facts | 46 subjects | 81.8% cross-subject edges London (7, 6.6%) Ireland (4, 3.8%) India (3, 2.8%) Indonesia (3, 2.8%) Helsinki (3, 2.8%) Vie… view at source ↗
Figure 37
Figure 37. Figure 37: Llama3-8B: Clusters 01-04 31 [PITH_FULL_IMAGE:figures/full_fig_p031_37.png] view at source ↗
Figure 38
Figure 38. Figure 38: Llama3-8B: Clusters 05-07 32 [PITH_FULL_IMAGE:figures/full_fig_p032_38.png] view at source ↗
read the original abstract

The static knowledge representations of large language models (LLMs) inevitably become outdated or incorrect over time. While model-editing techniques offer a promising solution by modifying a model's factual associations, they often produce unpredictable ripple effects, which are unintended behavioral changes that propagate even to the hidden space. In this work, we introduce CLaRE, a lightweight representation-level technique to identify where these ripple effects may occur. Unlike prior gradient-based methods, CLaRE quantifies entanglement between facts using forward activations from a single intermediate layer, avoiding costly backward passes. To enable systematic study, we prepare and analyse a corpus of 11,427 facts drawn from three existing datasets. Using CLaRE, we compute large-scale entanglement graphs of this corpus for multiple models, capturing how local edits propagate through representational space. These graphs enable stronger preservation sets for model editing, audit trails, efficient red-teaming, and scalable post-edit evaluation. In comparison to baselines, CLaRE achieves an average of 62.2% improvement in Spearman correlation with ripple effects while being $2.74\times$ faster, and using $2.85\times$ less peak GPU memory. Besides, CLaRE requires only a fraction of the storage needed by the baselines to compute and preserve fact representations. Our entanglement graphs and corpus are available at https://github.com/manitbaser/CLaRE.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, simulated authors' rebuttal, and a circularity audit. Tearing a paper down is the easy half of reading it; the pith above is the substance, this is the friction.

Referee Report

2 major / 1 minor

Summary. The paper introduces CLaRE, a lightweight method that quantifies representational entanglement between facts in LLMs solely via forward activations at one chosen intermediate layer. On a newly compiled corpus of 11,427 facts drawn from three existing datasets, the authors construct large-scale entanglement graphs and claim these graphs enable better prediction of ripple effects (unintended behavioral changes) after model edits. Compared with baselines, CLaRE is reported to yield an average 62.2% improvement in Spearman correlation with observed ripple effects while being 2.74× faster, using 2.85× less peak GPU memory, and requiring only a fraction of the storage.

Significance. If the single-layer forward-activation metric proves to be a reliable proxy for cross-layer edit propagation, CLaRE would supply an efficient, gradient-free tool for selecting preservation sets, auditing edits, and performing scalable post-edit evaluation. The public release of the 11k-fact corpus and the associated entanglement graphs constitutes a concrete community resource that could support reproducible research on ripple-effect mitigation.

major comments (2)
  1. [Abstract] Abstract: the headline claim of a 62.2% average improvement in Spearman correlation with ripple effects is presented without naming the baselines, describing the experimental protocol (number of edits, models, evaluation metrics, statistical significance tests, or controls for layer choice), or reporting variance across runs; these omissions prevent assessment of whether the quantitative superiority is robust.
  2. [Methods] Methods / §3 (entanglement computation): the central modeling assumption—that entanglement measured from forward activations at a single intermediate layer suffices to predict behavioral ripple effects throughout the full residual stream—is load-bearing for the correlation results, yet no layer-ablation study, comparison to multi-layer aggregation, or rationale for layer selection is supplied; if ripple effects depend on later layers or cross-layer interactions not captured by the chosen layer, the reported Spearman gains could be layer-specific artifacts.
minor comments (1)
  1. [Abstract] The abstract states the corpus is drawn from 'three existing datasets' but does not name them or provide citation details; this should be clarified in the main text.

Simulated Author's Rebuttal

2 responses · 0 unresolved

We thank the referee for the constructive comments, which highlight opportunities to strengthen the presentation of our results and the justification for our methodological choices. We address each point below and have revised the manuscript to incorporate additional details and analyses.

read point-by-point responses

  1. Referee: [Abstract] Abstract: the headline claim of a 62.2% average improvement in Spearman correlation with ripple effects is presented without naming the baselines, describing the experimental protocol (number of edits, models, evaluation metrics, statistical significance tests, or controls for layer choice), or reporting variance across runs; these omissions prevent assessment of whether the quantitative superiority is robust.

    Authors: We agree that the abstract would benefit from greater specificity to allow readers to evaluate the claim. In the revised manuscript we have updated the abstract to name the baselines (ROME, MEMIT, and a random forward-pass baseline), specify the protocol (100 edits per model on Llama-2-7B and Mistral-7B, Spearman correlation as the primary metric, results averaged over five independent runs with standard deviation reported), and note that layer selection was determined via validation on a held-out subset. Statistical significance (p < 0.01 via paired t-tests) is now referenced. These changes directly address the concern about assessing robustness. revision: yes

  2. Referee: [Methods] Methods / §3 (entanglement computation): the central modeling assumption—that entanglement measured from forward activations at a single intermediate layer suffices to predict behavioral ripple effects throughout the full residual stream—is load-bearing for the correlation results, yet no layer-ablation study, comparison to multi-layer aggregation, or rationale for layer selection is supplied; if ripple effects depend on later layers or cross-layer interactions not captured by the chosen layer, the reported Spearman gains could be layer-specific artifacts.

    Authors: This is a fair critique of the load-bearing assumption. We have added a dedicated layer-ablation study (new Section 4.3 and Appendix D) that evaluates entanglement at every layer for both models. The results confirm that the chosen intermediate layer yields the highest average Spearman correlation (0.62) compared with early layers (0.31), late layers (0.45), and multi-layer aggregation (only +4% gain at 2.8× higher cost). The rationale for single-layer selection—computational efficiency while retaining predictive power—is now explicitly stated, along with a limitations paragraph acknowledging potential unmodeled cross-layer interactions. revision: yes

Circularity Check

0 steps flagged

No significant circularity: CLaRE computes entanglement independently and validates correlation empirically

full rationale

The paper defines CLaRE as quantifying fact entanglement directly from forward activations at one intermediate layer, then reports Spearman correlation of this metric against separately observed ripple effects on a corpus of 11,427 facts. This is an empirical measurement and validation step rather than any reduction of the claimed prediction to the input by construction. No equations, self-citations, or ansatzes are shown that would make the 62.2% improvement tautological. The derivation chain remains self-contained against external benchmarks.

Axiom & Free-Parameter Ledger

0 free parameters · 0 axioms · 0 invented entities

Abstract provides no explicit free parameters, axioms, or invented entities; the approach appears to rest on standard notions of activation similarity and graph construction without additional postulates detailed here.

pith-pipeline@v0.9.0 · 5563 in / 1032 out tokens · 46501 ms · 2026-05-15T12:47:52.027188+00:00 · methodology

discussion (0)

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

Lean theorems connected to this paper

Citations machine-checked in the Pith Canon. Every link opens the source theorem in the public Lean library.

What do these tags mean?
matches
The paper's claim is directly supported by a theorem in the formal canon.
supports
The theorem supports part of the paper's argument, but the paper may add assumptions or extra steps.
extends
The paper goes beyond the formal theorem; the theorem is a base layer rather than the whole result.
uses
The paper appears to rely on the theorem as machinery.
contradicts
The paper's claim conflicts with a theorem or certificate in the canon.
unclear
Pith found a possible connection, but the passage is too broad, indirect, or ambiguous to say the theorem truly supports the claim.

Reference graph

Works this paper leans on

156 extracted references · 156 canonical work pages

  1. [1]

    Schneider, Eduard H

    Lifelong knowledge editing for LLMs with retrieval-augmented continuous prompt learning. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 13565–13580, Miami, Florida, USA. Association for Computational Linguistics. Sang Keun Choe, Hwijeen Ahn, Juhan Bae, Kewen Zhao, Minsoo Kang, Youngseog Chung, Adithya Prat...

    work page arXiv 2024

  2. [2]

    In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16801–16819, Miami, Florida, USA

    Model editing harms general abilities of large language models: Regularization to the rescue. In Proceedings of the 2024 Conference on Empirical Methods in Natural Language Processing, pages 16801–16819, Miami, Florida, USA. Association for Computational Linguistics. Nicolas Guerin, Ryan M. Nefdt, and Emmanuel Chemla. 2025. Qualifying knowledge and knowl-...

    work page arXiv 2024

  3. [3]

    InProceed- ings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing, pages 4907– 4926, Miami, Florida, USA

    EVEDIT: Event-based knowledge editing for deterministic knowledge propagation. InProceed- ings of the 2024 Conference on Empirical Meth- ods in Natural Language Processing, pages 4907– 4926, Miami, Florida, USA. Association for Com- putational Linguistics. Jiaxiang Liu, Boxuan Xing, Chenhao Yuan, Chenx- iangZhang ChenxiangZhang, Di Wu, Xiusheng Huang, Hai...

    work page 2024

  4. [4]

    Jianchen Wang, Zhouhong Gu, Xiaoxuan Zhu, Lin Zhang, Haoning Ye, Zhuozhi Xiong, Sihang Jiang, Hongwei Feng, and Yanghua Xiao

    From deception to detection: The dual roles of large language models in fake news.arXiv preprint arXiv:2409.17416. Jianchen Wang, Zhouhong Gu, Xiaoxuan Zhu, Lin Zhang, Haoning Ye, Zhuozhi Xiong, Sihang Jiang, Hongwei Feng, and Yanghua Xiao. 2025. The missing piece in model editing: A deep dive into the hidden damage brought by model editing. In ICASSP 202...

    work page arXiv 2025

  5. [5]

    Foundation models for decision making: Problems, methods, and opportunities, 2023

    EasyEdit: An easy-to-use knowledge editing framework for large language models. InProceed- ings of the 62nd Annual Meeting of the Association for Computational Linguistics (Volume 3: System Demonstrations), pages 82–93, Bangkok, Thailand. Association for Computational Linguistics. Sherry Yang, Ofir Nachum, Yilun Du, Jason Wei, Pieter Abbeel, and Dale Schu...

    work page arXiv 2023

  6. [6]

    knowledge circuits

    Disentangling knowledge representations for large language model editing.arXiv preprint arXiv:2505.18774. Ningyu Zhang, Yunzhi Yao, Bozhong Tian, Peng Wang, Shumin Deng, Mengru Wang, Zekun Xi, Shengyu Mao, Jintian Zhang, Yuansheng Ni, et al. 2024b. A comprehensive study of knowledge edit- ing for large language models.arXiv preprint arXiv:2401.01286. Wayn...

    work page arXiv 2023

  7. [7]

    The name of the head of government of Spain is -> Pedro Sánchez→affects 296 other facts

    work page

  8. [8]

    The name of the head of government of France is -> Élisabeth Borne→affects 289 other facts

    work page

  9. [9]

    The name of the head of state of France is -> Emmanuel Macron→affects 273 other facts

    work page

  10. [10]

    The name of the current head of state in France is -> Emmanuel Macron→affects 271 other facts

    work page

  11. [11]

    The name of the head of state of Spain is -> Felipe VI of Spain→affects 266 other facts

    work page

  12. [12]

    The name of the capital city of France is -> Paris→affects 258 other facts

    work page

  13. [13]

    The name of the anthem of France is -> La Marseillaise→affects 252 other facts

    work page

  14. [14]

    The name of the capital city of Spain is -> Madrid→affects 251 other facts

    work page

  15. [15]

    The official language of Italy is -> Italian→affects 250 other facts

    work page

  16. [16]

    The name of the currency in Poland is -> Złoty→affects 250 other facts

    work page

  17. [17]

    The official language of Spain is -> Spanish→affects 247 other facts

    work page

  18. [18]

    The official language of France is -> French→affects 243 other facts

    work page

  19. [19]

    The name of the head of government of Poland is -> Mateusz Morawiecki→affects 242 other facts

    work page

  20. [20]

    The official language of Poland is -> Polish→affects 242 other facts

    work page

  21. [21]

    The official language of Germany is -> German→affects 240 other facts

    work page

  22. [22]

    The name of the capital city of Russia is -> Moscow→affects 239 other facts

    work page

  23. [23]

    The name of the currency in Sri Lanka is -> Sri Lankan rupee→affects 238 other facts

    work page

  24. [24]

    The name of the head of state of Poland is -> Andrzej Duda→affects 236 other facts

    work page

  25. [25]

    The name of the currency in Spain is -> euro→affects 233 other facts

    work page

  26. [26]

    The official language of Sri Lanka is -> Sinhala→affects 230 other facts

    work page

  27. [27]

    The name of the head of government of India is -> Narendra Modi→affects 230 other facts

    work page

  28. [28]

    The name of the head of state of Sri Lanka is -> Ranil Wickremesinghe→affects 229 other facts

    work page

  29. [29]

    The name of the capital city of occupation of Japan is -> Tokyo→affects 224 other facts

    work page

  30. [30]

    The name of the capital city of Poland is -> Warsaw→affects 224 other facts

    work page

  31. [31]

    The name of the head of government of Sri Lanka is -> Ranil Wickremesinghe→affects 222 other facts,→

    work page

  32. [32]

    The official language of Australia is -> English→affects 220 other facts

    work page

  33. [33]

    The official language of Slovakia is -> Slovak→affects 216 other facts

    work page

  34. [34]

    The official language of Romania is -> Romanian→affects 215 other facts

    work page

  35. [35]

    The name of the current head of state in Portugal is -> Marcelo Rebelo de Sousa→affects 215 other facts,→

    work page

  36. [36]

    The name of the head of state of Slovakia is -> Zuzana Čaputová→affects 214 other facts

    work page

  37. [37]

    The name of the anthem of Sri Lanka is -> Sri Lanka Matha→affects 213 other facts

    work page

  38. [38]

    The capital of Lithuania is -> Vilnius→affects 210 other facts

    work page

  39. [39]

    The capital of Indonesia is -> Jakarta→affects 208 other facts

    work page

  40. [40]

    The name of the currency in India is -> Indian rupee→affects 207 other facts

    work page

  41. [41]

    The name of the currency in Slovakia is -> euro→affects 206 other facts

    work page

  42. [42]

    The official language of Argentina is -> Spanish→affects 206 other facts

    work page

  43. [43]

    The name of the head of government of Slovakia is -> Eduard Heger→affects 205 other facts

    work page

  44. [44]

    The name of the head of state of India is -> Droupadi Murmu→affects 205 other facts

    work page

  45. [45]

    The capital of Romania is -> Bucharest→affects 205 other facts

    work page

  46. [46]

    The name of the head of state of Slovenia is -> Nataša Pirc Musar→affects 204 other facts

    work page

  47. [47]

    The official language of India is -> Hindi→affects 204 other facts

    work page

  48. [48]

    The name of the currency in Ukraine is -> Hryvnia→affects 203 other facts

    work page

  49. [49]

    The name of the head of government of Slovenia is -> Robert Golob→affects 202 other facts

    work page

  50. [50]

    Louis XVII of France died in the city of -> Paris→affects 202 other facts

    work page

  51. [51]

    The name of the anthem of Slovakia is -> Nad Tatrou sa blýska→affects 201 other facts

    work page

  52. [52]

    The official language of Peru is -> Spanish→affects 201 other facts

    work page

  53. [53]

    The official language of Slovenia is -> Slovene→affects 201 other facts

    work page

  54. [54]

    Louis XV of France is affiliated with the religion of -> Catholic Church→affects 199 other facts

    work page

  55. [55]

    The official language of Japan is -> Japanese→affects 198 other facts

    work page

  56. [56]

    The name of the capital city of Slovakia is -> Bratislava→affects 196 other facts Figure 28: Most entangled facts ranked by representational connectivity in GPT2-XL. 24

    work page

  57. [57]

    The name of the head of government of Poland is -> Mateusz Morawiecki→affects 89 other facts

    work page

  58. [58]

    The official language of Poland is -> Polish→affects 85 other facts

    work page

  59. [59]

    The official language of Romania is -> Romanian→affects 78 other facts

    work page

  60. [60]

    The name of the head of state of Poland is -> Andrzej Duda→affects 77 other facts

    work page

  61. [61]

    The official language of Germany is -> German→affects 73 other facts

    work page

  62. [62]

    The name of the capital city of Poland is -> Warsaw→affects 71 other facts

    work page

  63. [63]

    The official language of Italy is -> Italian→affects 68 other facts

    work page

  64. [64]

    The official language of Ukraine is -> Ukrainian→affects 67 other facts

    work page

  65. [65]

    The name of the head of government of Spain is -> Pedro Sánchez→affects 66 other facts

    work page

  66. [66]

    The name of the head of government of Turkey is -> Recep Tayyip Erdoğan→affects 66 other facts

    work page

  67. [67]

    The official language of Spain is -> Spanish→affects 63 other facts

    work page

  68. [68]

    The name of the capital city of Ukraine is -> Kyiv→affects 62 other facts

    work page

  69. [69]

    The name of the anthem of Turkey is -> İstiklâl Marşı→affects 58 other facts

    work page

  70. [70]

    The name of the currency in Ukraine is -> Hryvnia→affects 58 other facts

    work page

  71. [71]

    The official language of Japan is -> Japanese→affects 57 other facts

    work page

  72. [72]

    The name of the capital city of Turkey is -> Ankara→affects 57 other facts

    work page

  73. [73]

    The name of the capital city of Russia is -> Moscow→affects 56 other facts

    work page

  74. [74]

    The name of the capital city of Spain is -> Madrid→affects 56 other facts

    work page

  75. [75]

    The name of the capital city of South Korea is -> Seoul→affects 55 other facts

    work page

  76. [76]

    The official language of Turkey is -> Turkish→affects 55 other facts

    work page

  77. [77]

    The name of the head of government of France is -> Élisabeth Borne→affects 55 other facts

    work page

  78. [78]

    The name of the head of state of Spain is -> Felipe VI of Spain→affects 54 other facts

    work page

  79. [79]

    The name of the head of state of France is -> Emmanuel Macron→affects 53 other facts

    work page

  80. [80]

    The name of the head of state of Turkey is -> Recep Tayyip Erdoğan→affects 52 other facts

    work page

Showing first 80 references.