18-Month Test of 93 Models Suggests Distillation, Quantization and Narrow RL Hurt General Ability

Over 18 months, Kevin Bryan of the University of Toronto evaluated 93 models on his HardGeoBench image geolocation benchmark, finding that distillation, quantization and narrow RL training may harm general capabilities, with one DeepSeek version ranking last.

2026-10-19 ~ 2026-10-19 · 2 related posts

• HardGeoBench: 93 models tested over 18 months, distillation plus RL hurts general ability — Afinetheorem · 2026-10-19
• 93-Model Benchmark Over 1.5 Years: Distillation + Quantization + RL Hurts General AI — Afinetheorem · 2026-10-19