Relations, Negations, and Numbers: Looking for Logic in Generative Text-to-Image Models

TL;DR

Three forms of logical operators are examined: relations, negations, and discrete numbers, which are asked human respondents to evaluate images generated by a state-of-the-art image-generating AI prompted with these `logical probes', and find that none reliably produce human agreement scores greater than 50\%.

Abstract

Despite remarkable progress in multi-modal AI research, there is a salient domain in which modern AI continues to lag considerably behind even human children: the reliable deployment of logical operators. Here, we examine three forms of logical operators: relations, negations, and discrete numbers. We asked human respondents (N=178 in total) to evaluate images generated by a state-of-the-art image-generating AI (DALL-E 3) prompted with these `logical probes', and find that none reliably produce human agreement scores greater than 50\%. The negation probes and numbers (beyond 3) fail most frequently. In a 4th experiment, we assess a `grounded diffusion' pipeline that leverages targeted prompt engineering and structured intermediate representations for greater compositional control, but find its performance is judged even worse than that of DALL-E 3 across prompts. To provide further clarity on potential sources of success and failure in these text-to-image systems, we supplement our 4 core experiments with multiple auxiliary analyses and schematic diagrams, directly quantifying, for example, the relationship between the N-gram frequency of relational prompts and the average match to generated images; the success rates for 3 different prompt modification strategies in the rendering of negation prompts; and the scalar variability / ratio dependence (`approximate numeracy') of prompts involving integers. We conclude by discussing the limitations inherent to `grounded' multimodal learning systems whose grounding relies heavily on vector-based semantics (e.g. DALL-E 3), or under-specified syntactical constraints (e.g. `grounded diffusion'), and propose minimal modifications (inspired by development, based in imagery) that could help to bridge the lingering compositional gap between scale and structure. All data and code is available at https://github.com/ColinConwell/T2I-Probology

Figures (11)

• Figure 1: Example layout of a typical trial in all behavioral experiments: Participants were presented with 10 grids of images, each grid paired with a target sentence, and asked to select images that matched the sentence. This particular trial/grid is an example from Experiment 1 (Relations), and is the result of randomly prompting DALL·E 3 with 'a potato under a spoon'.
• Figure 2: Results for Experiment 1 (Relations): participant agreement that images matched a prompt. Each dot is an individual prompt. Boxes show overall means and deviation within a relation. The top of the image shows density distributions for each relation.
• Figure 3: Examples of 'Modified' and 'Unmodified' stimuli used in the Negation Experiment. Unmodified prompts involve the injunction to not change the text. Modified prompts are modified by handing them to ChatGPT for rendering without instructions to leave them as is.
• Figure 4: Results for Experiment 2 (Negation): Participant agreement that images matched a prompt. Each dot is an individual prompt. Boxes show overall means and deviation for each entity. The top of the image shows density distributions for each prompt sub-type.
• Figure 5: Illustrative examples of number progressions. Each row shows an entity class, and each column shows a progression of requested entities from 1 to 6. Images are sampled at random from the 18 images available for each Entity x Number combination.