Humans vs AI: How we adapt language under extreme vocabulary constraints
When limited to just 250 words, humans talk like greedy algorithms—but experts revise like LLMs.
Deep Dive
Researchers found that humans communicating under severe vocabulary limits—down to just 250 words—tend to behave like greedy lexical planners, but more skilled individuals backtrack and revise, a non-greedy behavior. Both humans and models lean on semantically light words when constrained, offering insights into psycholinguistics, second-language communication, and language impairments.
Key Points
- Humans under a 250-word limit behave like greedy algorithms, making short-term optimal word choices.
- More skilled speakers show non-greedy backtracking, similar to globally optimal LLM sampling using Sequential Monte Carlo.
- Both humans and models rely on semantically light words (e.g., 'get', 'thing') to maintain communication under high constraint.
Why It Matters
Insights into constrained language production can improve assistive communication tools and second-language learning strategies.