Identifies a 'dual-capability bottleneck' where low-rated training data is essential for state tracking while high-rated data is needed for decision quality.
April 1, 2026
Original Paper
Tracking vs. Deciding: The Dual-Capability Bottleneck in Searchless Chess Transformers
arXiv · 2603.29761
The Takeaway
It provides a non-obvious dataset composition strategy for searchless transformers, demonstrating how to reach expert-level performance by balancing diversity for tracking and weight for quality rather than just scaling size.
From the abstract
A human-like chess engine should mimic the style, errors, and consistency of a strong human player rather than maximize playing strength. We show that training from move sequences alone forces a model to learn two capabilities: state tracking, which reconstructs the board from move history, and decision quality, which selects good moves from that reconstructed state. These impose contradictory data requirements: low-rated games provide the diversity needed for tracking, while high-rated games pr