Can 3D/4D Spatial Representation Beat Test-Time Compute? Zero-CoT Three.js Scene Graphs Outperform 345,000-Token Reasoning on 2D Grids and Eliminate Spatial Drift

Recent advances in Large Language Models (LLMs) heavily rely on scaling Test-Time Compute (Chain-of-Thought, CoT) to compensate for architectural weaknesses in spatial reasoning. In our companion study (Scene-Code-as-Topology, DOI: 10.5281/zenodo.23691649), we showed under symmetric inference settings that compiling a 2D Gomoku grid into executable Three.js 3D/4D procedural scene graph code overcomes 1D token-stride and diagonal blindness (23W–3D–4L across 30 games). However, a critical question remains: Can a superior spatial representation with zero test-time reasoning beat a flawed 2D representation backed by massive test-time Chain-of-Thought compute? In this study, we introduce an Asymmetric Compute-vs-Representation Handicap Benchmark (Group 4, 10 games, 205 plies on DeepSeek-V4.1-Flash-1017), pitting Agent A (2D ASCII Matrix + Full Thinking ENABLED, consuming 345,052 reasoning tokens at 28.08s/move) against Agent B (Three.js 3D/4D Scene Code + Thinking DISABLED, consuming 0 reasoning tokens at 5.52s/move). Despite a 0 vs. 345,052 reasoning-token handicap, Zero-CoT Agent B defeats Full-CoT Agent A by 6 Wins to 4 Losses (60.0% win rate, including a 6-game winning streak from Game 4 to Game 9 and 3/5 wins as second-mover White), while running 5.09× faster and reducing occupied-cell hallucinations by 94.1% (3 vs. 51 retries). Furthermore, we uncover and formalize two novel spatial-reasoning phenomena: (1) Disconnected Spatial Drift, where 2D ASCII agents hallucinate tactical relevance in empty board deserts and drop stones at Chebyshev distance ≥ 2 from all existing stones (equivalent to forfeiting a turn in Gomoku), which occurs 18 times in Full-CoT 2D vs. 0 times in Zero-CoT 3D (18:0 in Group 4; 31:1 across 20 DeepSeek games, a 96.8% reduction); and (2) the Input-Length vs. Thinking-Time Inversion Law, showing that a 5.1×–7.4× longer 3D code prompt slashes end-to-end latency by 32%–80% by trading serial autoregressive CoT decoding for parallel GPU prefill. Across all 40 games (924 moves in Groups 1–4), Scene-Code-as-Topology achieves 29W–3D–8L (78.4% decisive win rate, two-tailed exact binomial p = 0.00077 < 0.001). All theoretical explanations are presented as cautious preliminary conjectures, accompanied by an open call for compute sponsorship and academic collaboration ([email protected]).