Do Localization Methods Actually Localize Memorized Data in LLMs? A Tale of Two Benchmarks (2024.naacl-long)
Copied to clipboard
| Challenge: | Existing studies on the ability of localization methods to pinpoint LLM components for memorized data are lacking. |
| Approach: | They propose to use a subset of LLM weights to evaluate localization methods . they propose to measure how much dropping out identified neurons deletes a memorized sequence. |
| Outcome: | The proposed methods show promising localization ability, despite differences in their evaluations. |
Similar Papers
Generalisation First, Memorisation Second? Memorisation Localisation for Natural Language Classification Tasks (2024.findings-acl)
Copied to clipboard
| Challenge: | Memorisation in neural models is concerned due to overfitting and privacy concerns . a dominant hypothesis based on image classification is that lower layers learn generalisable features and deeper layers specialise and memorise. |
| Approach: | They apply 4 techniques to localise and edit models' memories. |
| Outcome: | The proposed method shows that memorisation is a gradual process rather than a localised one. |
Learnable Privacy Neurons Localization in Language Models (2024.acl-short)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) memorize and disclose private information, especially Personally Identifiable Information (PII) concerns regarding privacy and security within human society remain poorly understood. |
| Approach: | They propose to use learnable binary weight masks to localize PII-sensitive neurons within LLMs by deactivating localized privacy neurons. |
| Outcome: | The proposed method localizes PII-sensitive neurons across all layers and shows the property of PI I specificity. |
From Remembering to Metacognition: Do Existing Benchmarks Accurately Evaluate LLMs? (2025.findings-emnlp)
Copied to clipboard
Geng Zhang, Yizhou Ying, Sihang Jiang, Jiaqing Liang, Guanglei Yue, Yifei Fu, Hailin Hu, Yanghua Xiao
| Challenge: | Existing benchmark datasets focus on low-level cognitive tasks while providing limited coverage of higher-level reasoning skills. |
| Approach: | They analyze the cognitive depth of popular LLM benchmarks using Bloom’s Taxonomy to evaluate both the cognitive and knowledge dimensions. |
| Outcome: | The results show that incorporating higher-level cognitive instructions into the current instruction fine-tuning process improves model performance. |
Neuron-Level Differentiation of Memorization and Generalization in Large Language Models (2025.emnlp-main)
Copied to clipboard
Ko-Wei Huang, Yi-Fu Fu, Ching-Yu Tsai, Yu-Chieh Tu, Tzu-ling Cheng, Cheng-Yu Lin, Yi-Ting Yang, Heng-Yi Liu, Keng-Te Liao, Da-Cheng Juan, Shou-De Lin
| Challenge: | Existing models exhibit memorization and generalization behaviors in ways that are not easily interpretable or controllable. |
| Approach: | They propose to use a GPT-2 and LLaMA-3.2 model to identify distinct neuron subsets responsible for each behavior to steer the model toward memorization or generalization. |
| Outcome: | The proposed models show that inference-time interventions on these neurons can steer the model’s behavior toward memorization or generalization. |
Can LLM Graph Reasoning Generalize beyond Pattern Memorization? (2024.findings-emnlp)
Copied to clipboard
| Challenge: | Existing studies seek to enhance the graph reasoning capabilities of Large Language Models (LLMs) by specialized instruction tuning. |
| Approach: | They propose to evaluate LLM graph reasoning generalization using in-distribution settings . they propose to use three strategies to improve LLM generalization . |
| Outcome: | The proposed benchmark evaluates LLM graph reasoning generalization with in-distribution settings only . it shows that LLMs struggle to generalize across reasoning and real-world patterns . |
Generalization or Memorization? Multi-Agent vs. Baseline LLMs and AutoML Models for Tabular Classification (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) are increasingly used for structured tabular data. |
| Approach: | They evaluate a representative modular Multi-Agent LLM framework against state-of-the-art AutoML systems and established baselines. |
| Outcome: | The proposed model outperforms AutoML on pre-cutoff and post-cut off datasets. |
Memorisation Cartography: Mapping out the Memorisation-Generalisation Continuum in Neural Machine Translation (2023.emnlp-main)
Copied to clipboard
| Challenge: | Using the counterfactual memorisation metric, we find that when training neural networks, models will memorise some inputs but not others. |
| Approach: | They use the counterfactual memorisation metric to build a resource that places 5M NMT datapoints on a memorisations-generalisation map and describe how the datapoint’s surface-level characteristics and a models’ per-datum training signals are predictive of memorising in NMT. |
| Outcome: | The proposed model places 5M NMT datapoints on a memorisation-generalisation map and shows how their surface-level characteristics and models’ per-datum training signals are predictive of memorising in NMT. |
On Generalization across Measurement Systems: LLMs Entail More Test-Time Compute for Underrepresented Cultures (2025.acl-long)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) should be able to provide accurate information irrespective of the measurement system at hand . |
| Approach: | They use newly compiled datasets to test if this is true for seven open-source LLMs. |
| Outcome: | The proposed model can provide accurate information regardless of the measurement system at hand. |
Memorization vs. Generalization : Quantifying Data Leakage in NLP Performance Evaluation (2021.eacl-main)
Copied to clipboard
| Challenge: | Public datasets are often used to evaluate the efficacy and generalizability of state-of-the-art methods for many tasks in natural language processing (NLP). |
| Approach: | They identify leakage of training data into test data on several publicly available datasets used to evaluate NLP tasks, including named entity recognition and relation extraction. |
| Outcome: | The proposed model can memorize and generalize data on several publicly available datasets and is compared against previously unseen data. |
Fine-Grained Detection of Context-Grounded Hallucinations Using LLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing representations of hallucinations limit the types of errors that can be expressed, so we propose a new representation based on free-form textual descriptions, capturing the full range of possible errors. |
| Approach: | They propose a benchmark for localizing hallucinations using LLMs with a human annotation of over 1,000 examples and a protocol to verify its quality in a humans evaluation. |
| Outcome: | The proposed representation captures the full range of possible errors, and the best model achieves an F1 score of 0.67. |