Language and Experience: A Computational Model of Social Learning in Complex Tasks

TL;DR

This work presents a Bayesian framework that unifies linguistic guidance and direct experience to learn executable world models in complex tasks. By treating language as evidence via a probabilistic speaker model and using LM-enabled guidance to shape inference and proposals, the approach achieves sample-efficient learning and robust knowledge transfer across humans and models. Across 10 VGDL-based video games, linguistic guidance accelerates exploration, reduces dangerous interactions, and enables generational and bidirectional transfer of knowledge, illustrating a path toward effective human–AI collaborative learning. The findings highlight the potential of language-augmented, theory-based agents to teach, adapt, and co-evolve within social learning networks, with implications for scalable, collaborative intelligence.

Abstract

The ability to combine linguistic guidance from others with direct experience is central to human development, enabling safe and rapid learning in new environments. How do people integrate these two sources of knowledge, and how might AI systems? We present a computational framework that models social learning as joint probabilistic inference over structured, executable world models given sensorimotor and linguistic data. We make this possible by turning a pretrained language model into a probabilistic model of how humans share advice conditioned on their beliefs, allowing our agents both to generate advice for others and to interpret linguistic input as evidence during Bayesian inference. Using behavioral experiments and simulations across 10 video games, we show how linguistic guidance can shape exploration and accelerate learning by reducing risky interactions and speeding up key discoveries in both humans and models. We further explore how knowledge can accumulate across generations through iterated learning experiments and demonstrate successful knowledge transfer between humans and models -- revealing how structured, language-compatible representations might enable human-machine collaborative learning.

Figures (12)

• Figure 1: Overview of the model and experiments a) Experimental design: players (participants or models) are given N=10 lives to learn to play a new video game, either from experience only (Player 1) or from experience and advice written by a previous player (Player 2). b) Example learning trajectory: The model maintains beliefs about possible game rules and objectives (programs at bottom), and constantly refines them based on an initial linguistic guidance and new incoming experience.
• Figure 2: Game example (beesAndBirds). Players must discover rules and objectives on their own, sometimes helped with advice from others, to solve 4 game levels (right: levels 1 and 2/4).
• Figure 3: Learning from experience. Median across games (N=10, IQ range).
• Figure 4: Learning from experience and human advice. In most games, both humans (blue) and our model (orange) learn significantly faster with human advice (plain lines) than without (dashed lines), see median in Figure \ref{['fig:median_plot']}. The pure LLM baseline does not learn significantly better with advice, solving the first level in only 3/10 games (N=20, median $\pm$ interquartile range).
• Figure 5: Bidirectional human--model learning. a) Median performance across games for humans and models learning from experience alone (blue), or experience combined with human-generated (orange) or model-generated (green) linguistic guidance (N=10, shaded: interquartitle range). b) Performance improvements over individual learning baselines for each game (error bar: sem).