Geometry Conflict: Explaining and Controlling Forgetting in LLM Continual Post-Training

Benchmark Model Rank Results
arithmetic-reasoning-on-gsm8kQwen3-14B with GCWM#15Accuracy: 95.8
code-generation-on-mbppQwen3-14B with GCWM#42Accuracy: 76.7
code-generation-on-humanevalQwen3-14B with GCWM#43Pass@1: 86.6
mathematical-reasoning-on-math-500Qwen3-14B with GCWM#98Accuracy: 78.2
multiple-choice-question-answering-mcqa-on-gpqa-diamondQwen3-14B with GCWM#126Accuracy: 39.9