Maia-2: A Unified Model for Human-AI Alignment in Chess

TL;DR

Recognizing the complex, non-linear nature of human learning, this work introduces a skill-aware attention mechanism to dynamically integrate players' strengths with encoded chess positions, enabling the model to be sensitive to evolving player skill.

Abstract

There are an increasing number of domains in which artificial intelligence (AI) systems both surpass human ability and accurately model human behavior. This introduces the possibility of algorithmically-informed teaching in these domains through more relatable AI partners and deeper insights into human decision-making. Critical to achieving this goal, however, is coherently modeling human behavior at various skill levels. Chess is an ideal model system for conducting research into this kind of human-AI alignment, with its rich history as a pivotal testbed for AI research, mature superhuman AI systems like AlphaZero, and precise measurements of skill via chess rating systems. Previous work in modeling human decision-making in chess uses completely independent models to capture human style at different skill levels, meaning they lack coherence in their ability to adapt to the full spectrum of human improvement and are ultimately limited in their effectiveness as AI partners and teaching tools. In this work, we propose a unified modeling approach for human-AI alignment in chess that coherently captures human style across different skill levels and directly captures how people improve. Recognizing the complex, non-linear nature of human learning, we introduce a skill-aware attention mechanism to dynamically integrate players' strengths with encoded chess positions, enabling our model to be sensitive to evolving player skill. Our experimental results demonstrate that this unified framework significantly enhances the alignment between AI and human players across a diverse range of expertise levels, paving the way for deeper insights into human decision-making and AI-guided teaching tools.

Figures (10)

• Figure 1: Overview of the Maia-2 model architecture.
• Figure 2: Move prediction accuracy across diverse skill levels. Colors represent performance, with warmer tones indicating higher accuracy.
• Figure 3: (A). (Top) Joint probability assigned to human moves played by Maia-2 ($x$) and Maia 1900 ($y$), split by move quality. Blunders (left) reduce the expected win-rate by $\geq$ 10%, Errors (middle) by 5--10%, and Optimal (right) by $\leq 0\%$. (Bottom) Log odds ratio of $p(x,y)$ and $p(y, x)$ from top. (B). Move prediction agreement as (left) active player and (right) opponent player skill are varied. All cells are evaluated on the same set of positions but with altered skill level configurations.
• Figure 4: Maia-2's chess concept recognition as a function of skill level, as measured by linear activation probes right before (blue) and after (orange) skill-aware attention. (a) Stockfish overall board evaluation for middle-game positions. (b) Stockfish evaluation of middle-game bonuses and penalties to pieces for white. (c) Does the active player own two bishops? (d) Can the active player capture the opponent's queen?
• Figure 5: Maia-2 and Maia-1 solving a Mate-in-1 chess puzzle of rating 1500. Green arrows represent correct move predictions, while red arrows indicate incorrect predictions. The darkness of the green color correlates with the model's confidence, with darker arrows denoting a higher probability of making the correct move.