Vague quantification
and object representation
in humans vs. Large Vision Language Models
Vague quantification in human language
Vague quantifiers display a huge amount of variability. Despite this, psycholinguists and theoretical linguists agree that human quantification is affected/regulated by:
- Typicality
(Tiel and Geurts 2014; Ramotowska et al. 2024, a.o.)
- Ordered Scales: <some, many, most, all>
(Horn 1972; Katsos et al. 2016; Pagliarini et al. 2018; Gotzner and Romoli 2022, a.o.)
- Cognitive biases: clustering, size, etc.
(Krueger 1972; Pauw 2013; Bertamini et al. 2018, a.o.)
Quantification in LVLM: Testoni, Sprott, and Pezzelle (2024)
Study comparing the language used by humans and Large Vision Language Models (LVLMs) when describing images.
Conclusion: models deviated most from humans in the quantifier task.
Methodology
- Languages: English, Greek, Russian, Spanish, Italian and Catalan
- Data collection per language:
| `Participants’ | Method | |
|---|---|---|
| Humans | 40 | PCIbex |
| GPT-4o | 40 | openAI API |
| o4-mini | 40 | openAI API |
Internal respresentations in LVLM - embeddings
Cosine similarity
\(\cos(\vec{q},\vec{p}) = \dfrac{\vec{q}\cdot \vec{p}}{||\vec{q}|| \cdot||\vec{p}||}\)
Gives a number between -1 and 1.
The closer to 1, the higher the similarity.
Results: internal respresentations - embeddings
Model: text-embedding-3-large \(\rightarrow\) only accessible embedding model in OpenAI’s API
Some models face problems with the internal representation of quantification.
Internal representations in LVLM - reasoning
Reasoning models are trained via Reinforcement Learning (on verifiable problems) to create long chains of thought before providing a final answer (OpenAI 2025).
Hypothesis: given the inter-speaker variability in quantifier usage, these techniques may not be enough to replicate typicality values and ranges of usage.
Recap
Results: approximate number system
Catalan: quantifiers’ variation
