TLDR: A computational framework models social learning in AI by integrating linguistic guidance and direct experience. Using large language models as probabilistic speaker models, agents interpret advice and refine their understanding of game rules. Experiments across 10 video games show that this approach enables human-level learning efficiency, accelerates learning with linguistic input, facilitates knowledge accumulation across generations, and supports effective knowledge transfer between humans and AI.
Humans possess a remarkable ability to learn from both direct experience and the advice of others. Imagine learning to forage for mushrooms: you gain insights from each outing, but an experienced forager’s warning about “red ones with white spots” or guidance to “look for chanterelles near oak trees” can dramatically accelerate your learning and prevent fatal mistakes. This integration of linguistic guidance with hands-on experience is fundamental to human intelligence, enabling safer, more efficient learning and the accumulation of knowledge across generations. However, a general computational model explaining how AI systems can achieve this dual learning capability has been largely missing.
Current AI models often fall short in this regard. Reinforcement learning (RL) agents can master complex tasks but require millions of trial-and-error steps. While large language models (LLMs) excel at processing linguistic input, they struggle with interactive planning and embodied learning. Other Bayesian models of social cognition typically operate in simpler, non-interactive tasks.
A new computational framework addresses this challenge by treating linguistic guidance and direct experience as complementary sources of evidence. This framework allows AI agents to infer structured, executable world models, much like humans form mental models of their environment. The researchers introduce three key innovations to make this possible:
LMs as Speaker Models
The framework leverages powerful language models to estimate the probability that a human, holding specific beliefs about the world, would produce particular linguistic advice. This allows AI agents not only to generate advice for others but also to interpret human language as valuable evidence during their learning process.
Inference from Linguistic Evidence
The speaker model is used to evaluate how plausible received advice is under different potential world models. This means linguistic input directly shapes the AI’s Bayesian inference, working alongside experiential learning to refine its understanding of the environment.
Also Read:
- Agentic Reinforcement Learning: Empowering LLMs as Autonomous Decision-Makers
- Unlocking Deeper Understanding: How Multi-Agent LLMs Are Revolutionizing Causal AI
LM-Accelerated Inference
Language models are also employed to transform advice into targeted suggestions, guiding the AI’s learning updates towards the most promising areas of its hypothesis space, leading to faster convergence and more efficient learning.
To validate this framework, researchers conducted behavioral experiments and computational simulations across 10 different video games. These games, built using the Video Game Description Language (VGDL), provided rich, causal environments where players had to discover rules and objectives through exploration, sometimes aided by advice.
The findings were compelling. First, the model achieved human-level learning efficiency when relying solely on experience, and its efficiency improved by approximately 40% with linguistic guidance, mirroring how advice accelerates human exploration. Second, an analysis of message content revealed that detailed descriptions, especially those containing information about win conditions, were particularly effective. Successful messages significantly shaped learners’ exploration patterns, helping them avoid costly mistakes and discover key game mechanics faster.
Third, the study demonstrated generational learning through iterated experiments. Agents with limited individual experience were able to pass on advice across generations, leading to a slow accumulation of knowledge and performance improvements over time, much like cultural transmission in human communities. Finally, the framework enabled successful knowledge transfer between humans and models, suggesting that structured, language-compatible representations could pave the way for human-machine collaborative learning.
Interestingly, while models learned better from model-generated advice than human-generated advice, human learners showed only a modest advantage from model advice. This asymmetry highlights differences in communication styles: human messages often included metacognitive strategies, analogies, or emotional content, which human learners readily use but the model found harder to interpret. This points to both the potential and challenges of language-mediated human-machine collaboration.
This research opens exciting avenues for future work, including extending the framework to interpret richer forms of linguistic guidance (like high-level strategies), enabling agents to adaptively trust information sources, and scaling the approach to more open-ended game spaces. The ability of the model to both benefit from human guidance and provide effective, pedagogical advice back to humans marks a significant step towards bidirectional learning systems, fostering enhanced collaboration and collective intelligence in both artificial and human communities. You can read the full research paper here.
