Greedy or not, here I come - Language production under vocabulary constraints in humans and resource-rational models

Sep 29, 2026

Communicating using only a limited vocabulary is a demanding cognitive challenge that forces speakers to plan for intelligibility while navigating a constrained lexicon. This study tests how humans answer many questions when allowed as few as 250 of the most frequent English words. Comparisons to greedy and globally optimal sampling use Sequential Monte Carlo inference with large language models. Results show human behavior resembles greedy sampling overall, with skilled participants more likely to backtrack and revise, yielding an anon-greedy pattern under high constraint.

LanguageEnglish

CategoryResearch & Report

FormatPDF

Pages8

Reading Time20 min

AccessFree to download

Abstract

Introduction

Greedy or not, here I come: Language production under vocabulary constraints in humans and resource-rational models Thomas Hikaru Clark 1 , Sihan Chen 1 , Laura Nicolae2 1Department of Brain and Cognitive Sciences, MIT 2Department of Economics, Harvard University {thclark, [sihanc](sihanc}@mit.edu)[}](sihanc}@mit.edu)[@mit.edu](sihanc}@mit.edu) , [lauranicolae@g.harvard.edu](lauranicolae@g.harvard.edu) arXiv :2605 . 15365v1 [ cs .CL] 14 May 2026 Abstract Communicating using only a limited vocabulary is a common but challenging cognitive phenomenon, requiring an ideal communicator to plan carefully to optimize for intelligibility while circumventing a constrained lexicon. In this work, we investigate how humans respond to a broad array of questions under variable vocabulary limitations, consisting of only 250 highly frequent words at the most restrictive. We provide theoretically motivated comparisons to greedy and globally optimal sampling algorithms using Sequential Monte Carlo inference with large language models. Humans generally resemble greedy sampling more than globally optimal sampling, though more skilled humans are more likely to backtrack and revise – anon-greedy behavior. An observed human pattern of leaning on semantically light words in high-constraint settings falls out of both greedy and globally optimal sampling. We discuss the results and their broader implications for resource-rational cognition, psycholinguistics, L2 communication, and language impairments. Keywords: Psycholinguistics; Large Language Models; Explanations; Planning; Resource-Rationality Introduction Humans use language to communicate (Fedorenko et al., 2024) . Using insights from information theory (Shannon, 1948), researchers model language as communication code between a speaker and a listener: the speaker converts their meaning to utterances, and the listener receives the utterances andrecovers the meaning (see Gibson et al., 2019, for a review) . The production process is incremental: instead of starting with a full utterance in mind, a person typically plans what to say as they speak or write (e.g. Bock & Levelt, 1994; Ferreira & Swets, 2002; Pickering & Garrod, 2013) . At the computational level, Futrell (2023) modeled language production asan action planning problem in which a speaker’s incremental production is influenced both by a pressure to say contextually predictable things and a pressure to say things that further the communicative goal. In this work, we specifically investigate how constraints on the available set of vocabulary words affect incremental language production. Communication under a vocabulary constraint is a common cognitive phenomenon. Many societies contain linguistic minorities with limited knowledge of the majority language, which affects their ability to access basic services or participate in the economy (Bleakley & Chin, 2004; Grogger et al., 2020) . Compared to native speakers, learners of a language (L2 learners) tend to use fewer and more frequent words (Laufer, 1991) . When L2 learners do not know a specific word, they may employ various strategies to get a meaning across, such as describing key properties of the concept (Poulisse, 2011) or using “all-purpose” or semantically light words as substitutes (Drnyei & Scott, 1997) . This phenomenon suggests that there may be diminishing marginal returns to learning new vocabulary words above a certain point, since a subset of the vocabulary is typically sufficient for communication. Another common case of vocabulary-constrained communication is when an expert explains a complex phenomenon to a layperson. Past work has considered what makes for a good explanation (e.g. Brewer et al., 2000; Cruz & Lombrozo, 2025; Chandra et al., 2024; McCarthy & Keil, 2023) . For example, Sulik et al. (2023) suggests that a good explanation contains functional (i.e., what is something for?) or mechanistic (i.e., how does something work?) information. Besi

FAQs

What cognitive problem does the study focus on?

The study investigates how people communicate and produce language when constrained to a limited vocabulary, requiring careful planning to remain intelligible.

How does the research compare human performance to computational algorithms?

It compares humans to greedy sampling and globally optimal sampling, implemented using Sequential Monte Carlo inference with large language models.

What is the main finding about human behavior under vocabulary constraints?

Humans generally resemble greedy sampling more than globally optimal sampling, though higher-skilled participants more often backtrack and revise, showing an anon-greedy behavior.

Sep 29, 2026

Upload by Riley West