Reading Depth Does Not Transfer: The Distillation Ceiling in Computer Go

Abstract Knowledge distillation has been spectacularly successful in large language models, where small students efficiently approximate large teachers because both operate at the same task abstraction: next-token prediction. We report a structurally different outcome when distillation is applied to game AI for the game of Go. Using the pure-distillation pipeline introduced in Kojima (2026, in preparation; hereafter Paper I), we trained a 7M-parameter student on three successive rounds of KataGo b18c192 labels at 4000 visits per position; the training corpus grew from 45 thousand positions (the Paper I endpoint) to 120 thousand and then to 200 thousand. Across this 4.4× data scaling, the student’s top-1 oracle agreement on the fixed Move Quality evaluation set monotonically degraded : 0.170 → 0.120 → 0.040 (top-3: 0.270 → 0.160 → 0.080). Value MAE was non-monotonic (0.303 → 0.193 → 0.290) and showed no net improvement across the 8× increase from generation 2 to generation 3. We interpret this as evidence for a structural difference between LLM and game-AI distillation. The teacher’s strength in Go is roughly half neural-network evaluation and half MCTS search depth; the search-depth component does not transfer through single-position labels , regardless of how many such labels are provided. A complementary five-intervention study (Section 5) tested value-head capacity, β_z target blending, score-head re-activation, and ownership-weight tuning; four interventions failed and one auxiliary intervention (ownership-weight tuning) produced measurable improvement, consistent with the same ceiling.