Year: 2025
Type: dissertation
Abstract: This thesis focuses on the fine-tuning of Large Language Models (LLMs) for cooperative play in the card game Hanabi, a benchmark for imperfect-information and multi-agent coordination. To produce our dataset, we use an open-source Hanabi environment developed in C++ with diverse bot strategies, ranging from simple heuristics to adaptive and meta-reasoning agents. We introduce an integer-based encoding system that compresses game transcripts into compact token sequences suitable for LLM training. Using this representation, we fine-tune Transformer-based LLMs to predict partner moves, evaluating their capacity to model implicit communication, infer hidden information and adapt to different partner strategies. Our experimental pipeline measures accuracy across different partner and move types. Results demonstrate that LLMs capture recurring conventions and style-dependent behaviors, achieving competitive prediction performance compared to heuristic baselines. These findings advance our understanding of how LLMs can be adapted for sequential decision-making under uncertainty, with implications for cooperative AI in imperfect-information domains.
Source: Aaltodoc (Aalto University)
Author: Garcia Fierro, Diego
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: green
OpenAlex ID: W7161932820