Understanding or Memorizing? A Case Study of German Definite Articles in Language Models (2026.acl-long)
Copied to clipboard
| Challenge: | grammatical agreement is unclear whether language models rely on memorization or generalization . morphologically rich languages such as German have syncretic forms that are syncretically arranged . |
| Approach: | They use a GRADIEND-based interpretability method to learn parameter update directions for gender-case specific article transitions. |
| Outcome: | Using GRADIEND, we find that updates learned for gender-case specific article transitions affect unrelated gender- case settings . |
Similar Papers
GerEO: A Large-Scale Resource on the Syntactic Distribution of German Experiencer-Object Verbs (2022.lrec-1)
Copied to clipboard
| Challenge: | Psych verbs and their properties in multiple languages have ignited discussions among linguists for several decades . Psych-verbs are often considered syntactically deviant, although this has occasionally been called into question . |
| Approach: | They propose to use a large-scale database of more than 10,000 examples for 64 verbs from a newspaper corpus annotated for several syntactic and semantic features relevant for their analysis. |
| Outcome: | The proposed database contains 10,000 examples for 64 verbs from a newspaper corpus and includes syntactic construction, semantic stimulus type, and form of possible stimulus preposition. |
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. |
Do Grammatical Error Correction Models Realize Grammatical Generalization? (2021.findings-acl)
Copied to clipboard
| Challenge: | Existing models for grammatical error correction use pseudo data, but they are inconvenient for realworld deployment due to large amounts of training data. |
| Approach: | They propose a method to evaluate whether GEC models can generalize to unseen errors by using synthetic and real GEC datasets with controlled vocabularies. |
| Outcome: | The proposed model fails to realize grammatical generalization even in simple settings with limited vocabulary and syntax, suggesting it lacks the generalization ability required to correct errors from provided training examples. |
Memorisation versus Generalisation in Pre-trained Language Models (2022.acl-long)
Copied to clipboard
| Challenge: | State-of-the-art pre-trained language models have been shown to memorise facts and perform well with limited amounts of training data. |
| Approach: | They propose to extend pre-trained language models to generalise and memorise facts in noisy and low-resource scenarios. |
| Outcome: | The proposed extension improves performance in low-resource named entity recognition tasks. |
Circuit Stability Characterizes Language Model Generalization (2025.acl-long)
Copied to clipboard
| Challenge: | Rapid development of state-of-the-art models induce benchmark saturation, while creating more challenging datasets is labor-intensive. |
| Approach: | They propose to introduce circuit stability as a new way to assess model performance. |
| Outcome: | The proposed methods characterize and predict different aspects of generalization. |
How Do Language Models Acquire Character-Level Information? (2026.eacl-long)
Copied to clipboard
| Challenge: | Language models (LMs) implicitly encode character-level information, despite not being explicitly provided during training. |
| Approach: | They analyze how language models acquire character-level knowledge by comparing them to standard settings. |
| Outcome: | The results show that LMs do not treat words as opaque tokens, but instead treat them as tokens. |
How Conservative are Language Models? Adapting to the Introduction of Gender-Neutral Pronouns (2022.naacl-main)
Copied to clipboard
| Challenge: | a recent study shows that gender-neutral pronouns are not associated with processing difficulties . linguistic scholars have observed how technology has altered the course of language evolution . |
| Approach: | They show that gender-neutral pronouns in Danish, English and Swedish are not associated with processing difficulties. |
| Outcome: | a new study shows that gender-neutral pronouns are not associated with human processing difficulties . the findings suggest that such conservativity in language models may limit widespread adoption . |
Do LLMs learn a true syntactic universal? (2024.emnlp-main)
Copied to clipboard
| Challenge: | linguistics literature has debated whether large multilingual language models learn language universals . Typological generalizations are a key battleground in such debates - e.g. van der Hulst, 2023, chapter 7). |
| Approach: | They consider a candidate universal for language universals, the Final-over-Final Condition . they suggest that modern language models may need additional sources of bias to become truly human-like . |
| Outcome: | The proposed model only seems to recognize the Final-over-Final Condition in German, Russian, Hungarian and Serbian . |
Are Transformers a Modern Version of ELIZA? Observations on French Object Verb Agreement (2021.emnlp-main)
Copied to clipboard
| Challenge: | Recent studies have shown that unsupervised sentence representations of neural networks encode syntactic information by observing that neural language models are able to predict the agreement between a verb and its subject. |
| Approach: | They propose to take an alternative look at these results by studying whether neural networks are able to build an abstract sentence representation rather than capture surface statistical regularities. |
| Outcome: | The proposed model can achieve high accuracy on the long-range French object-verb agreement, indicating a possible flaw in the model's syntactic ability. |
How Well Do Text Embedding Models Understand Syntax? (2023.findings-emnlp)
Copied to clipboard
| Challenge: | Existing text embedding models have not addressed syntactic understanding challenges, highlighting ineffectiveness and enhancing generalization ability. |
| Approach: | They propose to examine the ability of text embedding models to generalize across syntactic contexts. |
| Outcome: | The proposed models exhibit high similarity socres at this simple task. |