On the Additive Compositionality of Task Vectors in Vision–Language Models (2026.eacl-short)
Copied to clipboard
| Challenge: | In-context learning (ICL) in large language models (LLMs) has been shown to operate through task vectors, but its extension to vision-language models (VLMs) remains underexplored. |
| Approach: | They construct visual reasoning tasks with clearly defined subtasks and extract task vectors from few-shot demonstrations. |
| Outcome: | The proposed model can be extended to vision-language models (VLMs) by adding the vectors of its constituent subtasks. |
Similar Papers
Label Words as Local Task Vectors in In-Context Learning (2026.findings-acl)
Copied to clipboard
| Challenge: | Large Language Models (LLMs) have demonstrated remarkable abilities, one of the most important being in-context learning (ICL). |
| Approach: | They hypothesized that the network creates a task vector in specific positions during ICL, which can be computed by averaging across the dataset. |
| Outcome: | The proposed model can achieve zero-shot performance with dummy inputs comparable to few-shot learning by patching the global task vector. |
Learning Task Representations from In-Context Learning (2025.findings-acl)
Copied to clipboard
| Challenge: | Existing methods for generalizing tasks to modalities beyond text fail to generalize effectively to linguistic tasks. |
| Approach: | They propose a method for encoding task information in ICL prompts as a function of attention heads within the transformer architecture. |
| Outcome: | The proposed method extracts task-specific information from in-context demonstrations and excels in both text and regression tasks. |
In-Context Learning Creates Task Vectors (2023.findings-emnlp)
Copied to clipboard
| Challenge: | In-context learning (ICL) is a powerful new learning paradigm for Large Language Models (LLMs). |
| Approach: | They propose to use a model with a prompt and a query to learn a mapping based on two examples to produce the output. |
| Outcome: | The proposed model can learn functions from a simple structure based on a training set and a single task vector calculated from the training set. |
One Task Vector is not Enough: A Large-Scale Study for In-Context Learning (2026.acl-srw)
Copied to clipboard
| Challenge: | Existing studies limit comprehensive analysis of large language models based on task vectors . recent work points to "task vectors" as mechanism for encoding task rules . |
| Approach: | They propose a novel task vector with 30 input-output pairs for in-context learning . they use a few prompt-based examples to adapt to new tasks without weight updates . |
| Outcome: | Experiments with Llama-3-8B on QAF show task vector performance peaks at intermediate layer . complex tasks rely on multiple, subtask-specific vectors rather than a single vector . |
In-Context Compositional Generalization for Large Vision-Language Models (2024.emnlp-main)
Copied to clipboard
| Challenge: | Recent work shows that in-context learning for large language models exhibits compositional generalization capacity. |
| Approach: | They propose a method to exhibit in-context compositional generalization in large vision-language models by combining visual and linguistic modalities. |
| Outcome: | The proposed method reduces redundancy and complexity in in-context learning with LVLMs. |
Exploring the Compositional Deficiency of Large Language Models in Mathematical Reasoning Through Trap Problems (2024.emnlp-main)
Copied to clipboard
| Challenge: | Current LLMs lack systematic compositionality, and therefore cannot serve as reliable cognitive models. |
| Approach: | They propose to introduce logical traps into the original problems of MATH and GSM8K to investigate the compositionality of large language models in mathematical reasoning. |
| Outcome: | The proposed model can generate infinite combinations from finite learned components. |
Quantifying Compositionality of Classic and State-of-the-Art Embeddings (2025.findings-emnlp)
Copied to clipboard
| Challenge: | Static word embeddings make strong claims about compositionality, but the SOTA generative models go too far in the other direction. |
| Approach: | a new study evaluates the compositionality of word embeddings by canonical correlation analysis . strong compositional signals are observed in later training stages across data modalities . |
| Outcome: | a new evaluation of compositional models shows that they exploit access meanings when justified . strong compositional signals are observed in later training stages and in deeper layers of the transformer-based model before a decline at the top layer. |
Skills-in-Context: Unlocking Compositionality in Large Language Models (2024.findings-emnlp)
Copied to clipboard
| Challenge: | eliciting compositional generalization capabilities in large language models is challenging for advanced LLMs because they lack foundational skills and compositional examples in the same prompt context. |
| Approach: | They propose to use compositional generalization capabilities in large language models to elicit compositional skills in a prompt context. |
| Outcome: | The proposed structure enables LLMs to tackle more challenging problems with as few as two exemplars and unlocks their latent potential. |
Long Story Short: Disentangling Compositionality and Long-Caption Understanding in Contrastive VLMs (2026.findings-acl)
Copied to clipboard
| Challenge: | Existing benchmarks for vision-language models treat compositionality and long-caption understanding in isolation. |
| Approach: | They analyze when compositional reasoning and long-caption understanding transfer across tasks and when this relationship fails. |
| Outcome: | The proposed model can generalize on poorly grounded captions and with strong visual grounding, while architectural choices can limit compositional learning. |
How do Transformer Embeddings Represent Compositions? A Functional Analysis (2025.findings-acl)
Copied to clipboard
| Challenge: | Despite the popularity of transformer-based models, little is known about how they represent compound words and whether they are compositional. |
| Approach: | They evaluate compositionality in mistral, OpenAI Large, and Google embedding models and compare them with BERT. |
| Outcome: | The proposed models perform best in addition, multiplication, dilation, regression, and the classic vector addition model performs almost as well as any other model. |