Quotient Recurrent Reasoners hit 97.9% on Sudoku-Extreme with 78% fewer loops
burny_tech · x · 2026-10-16
TK Rusch shares his student Sumin Park's new work: Quotient Recurrent Reasoners (QRR). Looped reasoning models like TRM and FPRM waste many loops on oscillatory dynamics that don't contribute to reasoning. QRR uses gauge equivariance to define convergence in a quotient space instead of state-space, letting symmetry-generated rotations persist while task-relevant dynamics converge faster — oscillations belong to the same equivalence class. Result: 97.9% on Sudoku-Extreme with 78% fewer loops than FPRM, with strong results across reasoning benchmarks and a striking visualization.
More from Research
- Logit lens on a Beat Saber finetune reveals it reuses language grammar circuits — ctjlewis · 2026-10-16
- Handwired compiles plain JavaScript into an exact, fully differentiable neural net — ctjlewis · 2026-10-16
- 128×128 record stands: how intermediate-work experiments foreshadowed reasoning models — ctjlewis · 2026-10-16
- TeamFusion: preference-based AI proxies debate to preserve minority views in team decisions — burkov · 2026-10-16
- MIT's C-GMT gives humanoid robots whole-body compliance without sacrificing motion tracking — MIT_CSAIL · 2026-10-16
- SUNY Albany to offer Topological Deep Learning course, from Perceptrons to WL test — burny_tech · 2026-10-16