|
Cultural evolution of efficient semantic systems in humans and AI | title | Cultural evolution of efficient semantic systems in humans and AI |
|---|
| start_date | 2025/10/15 |
|---|
| schedule | 16h30-17h30 |
|---|
| online | no |
|---|
| location_info | En ligne |
|---|
| summary | Human languages efficiently compress meanings into words, but how did our semantic systems evolve to be that way? Are AI systems capable of evolving efficient semantic systems and representing meaning as we do? In this talk, I address these open questions from cognitive, cultural, and computational perspectives. First, I show that individual human learners favor efficiently compressed semantic representations. This inductive learning bias, when amplified via cultural transmission,drives the evolution of near-optimally efficient semantic systems. Second, I consider large language models (LLMs) and show that while
they vary widely in their semantic alignment with humans, theynevertheless exhibit a similar tendency toward efficient compression: when simulating cultural evolution with LLMs, they iteratively restructure initially random semantic systems towards greater efficiency. Finally, I show that introducing an explicit pressure for efficient compression, grounded in the information bottleneck principle, enables multi-agent reinforcement learning systems to evolve efficient, human-like semantic systems without any human supervision. Taken together, these results demonstrate how humans and AI can evolveefficient systems through social interaction and cultural transmission, and more broadly, they suggest that efficient compression may be a fundamental principle of intelligenc |
|---|
| responsibles | Bernard |
|---|
Workflow history| from state (1) | to state | comment | date |
|---|
| submitted | published | | 2025/10/13 13:04 UTC |
| |
|