Talks: Hive Mind

Synopsis: Counting is one of the simplest forms of visual reasoning: to count, a system must identify individual objects, keep them distinct, and aggregate them into a representation of number. Yet even state-of-the-art vision-language models often fail at this seemingly elementary task.

Visual counting provides a window into the internal computations underlying multimodal intelligence. Large vision-language models exhibit a striking behavioral pattern reminiscent of human numerical cognition: highly accurate judgments for small numerosities give way to increasingly noisy estimates as the number of objects grows. Using mechanistic interpretability tools, specialized attention mechanisms that support object localization, individuation, and numerical aggregation can be identified. Together, these mechanisms form a structured “counting circuit” that is shared with broader visual reasoning tasks.

What accounts for these limitations in counting?  Visual patchification can fragment objects and impair object individuation, while attention-based aggregation progressively compresses distinctions between neighboring numerosities.  Lightweight architectural interventions can substantially improve counting and, surprisingly, transfer to broader spatial reasoning.Together, these findings raise broader questions about how artificial visual systems construct objects, represent quantity, and perform compositional reasoning

Together, these findings raise broader questions about how artificial visual systems construct objects, represent quantity, and perform compositional reasoning

• What can the successes and failures of vision-language models in counting tell us about how they represent objects and number?
• How do the computational mechanisms underlying counting relate to broader visual reasoning?
• What can comparisons with human numerical cognition reveal about how artificial systems represent quantity?
• What do these findings suggest about the relationship between model architecture and visual reasoning?

Synopsis: Public discussions of artificial intelligence often swing between enthusiasm and alarm: AI is described as a transformative technology that may expand human capabilities, or as a threat to work, creativity, education, truth, and even humanity itself. But evaluating many of these claims requires us to think carefully about concepts that cognitive scientists have studied for decades: intelligence, learning, reasoning, representation, understanding, memory, and agency.

.