LLMs and Alignment. Research talks: Eric Malmi

​LLMs and Alignment is launching a new series of recurring events—“Research Talks”—featuring invited speakers. At our first event, Professor Eric Malmi will present his research:
Test-time scaling for mastering board games with LLMs.

​Test-time scaling is increasingly recognised as an effective paradigm for enhancing the capabilities of large language models (LLMs). This talk describes an ICML 2025 paper from Professor Eric Malmi’s team that introduces and contrasts two test-time search strategies for board games (Chess, Chess960, Connect Four, and Hex). In the external-search approach, an LLM steers Monte Carlo Tree Search (MCTS) rollouts and evaluations without relying on an external game engine. In the internal-search approach, the model is trained to generate, in context, a linearised search tree and then produce a final move selection. Together, these methods demonstrate strong test-time scaling, culminating in state-of-the-art chess performance among LLM-based systems.