What it is
A 4B encoder-decoder chess LM that plugs a strong silent chess encoder into an instruction-tuned LM, then iteratively distills search-based explanations to improve play and commentary. Instead of teaching a plain LM chess from text alone, it cross-attends to a strong silent chess encoder, then improves explanations via Bellman-style search distillation from child-position analyses.
Key results
- 2697 Elo after 7 iterations, up from 1782 (+915 Elo)
- Beats GPT-5.6-Sol at 2071 Elo and Gemini-3.1-Pro at 2201
- Higher no-mistake rate on tactical puzzles than all baselines
Shows that coupling an LM to expert latent states and distilling search-improved explanations can raise both playing strength and explanation quality.