Papers by Sondre Wold
More room for language: Investigating the effect of retrieval on language models (2024.naacl-short)
Copied to clipboard
| Challenge: | Retrieval-augmented language models are a promising alternative to standard pretraining, but little attention has been put into understanding what this type of training scheme does to the underlying language model when analyzed as a standalone -separated from the overall retrieval pipeline. |
| Approach: | They propose an ‘ideal retrieval’ methodology to study these models in a fully controllable setting and propose a retrieval augmentation methodology to examine their effects. |
| Outcome: | The proposed model saves substantially less world knowledge in their weights, but is worse at comprehending global context. |
Circuit Compositions: Exploring Modular Structures in Transformer-Based Language Models (2025.acl-long)
Copied to clipboard
| Challenge: | Recent advances in mechanistic interpretability have made progress in identifying circuits, the minimal computational subgraphs responsible for a model’s behavior on specific tasks. |
| Approach: | They propose to analyze circuits for highly compositional subtasks within a transformer-based language model to determine their modularity and how they relate to each other. |
| Outcome: | The proposed approach shows that the circuits identified exhibit notable node overlap and cross-task faithfulness. |
Measuring Idiomaticity in Text Embedding Models with epsilon-compositionality (2026.eacl-long)
Copied to clipboard
| Challenge: | Existing studies on compositionality of text embedding models have limited understanding of the principle . idioms have traditionally been seen as non-compositional . |
| Approach: | They propose to use formal definitions to define compositionality in text embedding models . they find that most models differentiate between idiomatic and non-idiomatic phrases . |
| Outcome: | The proposed model is able to differentiate between idiomatic and non-idiomatic phrases, the authors show . |
Systematic Generalization in Language Models Scales with Information Entropy (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for assessing compositional behavior are unclear on how to measure the difficulty of a systematic generalization problem. |
| Approach: | They propose a framework for measuring entropy in a sequence-to-sequence task and a method for measuring it. |
| Outcome: | The proposed framework scales with the entropy of the distribution of component parts in the training data. |
Estimating Lexical Complexity from Document-Level Distributions (2024.lrec-main)
Copied to clipboard
| Challenge: | Existing methods for complexity estimation are limited to entire documents . health assessment tools are too short for existing methods to apply . |
| Approach: | They propose a two-step approach for estimating lexical complexity that does not rely on pre-annotated data. |
| Outcome: | The proposed method is tested on the Norwegian language and compares with other assessment tools. |
Compositional Generalization with Grounded Language Models (2024.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for combining language models with knowledge graphs struggle with generalization to sequences of unseen lengths and novel combinations of seen base components. |
| Approach: | They propose a procedure for generating natural language questions paired with knowledge graphs that targets different aspects of compositionality and avoids grounding models in information already encoded in their weights. |
| Outcome: | The proposed method fails to generalize to unseen lengths and to novel combinations of seen base components. |