LLM Post-Training Theory Conversation

Verbal MLE round covering LLM and post-training fundamentals, from regularization to tokenization and preference-optimization methods. No coding or whiteboarding is required in the reported format. Requirements Format: verbal Q&A, roughly 60 minutes for the MLE onsite theory round. Topics include: Dropout. L1 and L2 regularization. Tokenization, especially vocabulary size versus sequence length trade-offs. Supervised fine-tuning: training objective and data preparation. RL / preference optimization: DPO, GRPO, PPO, method differences, and comparison. Adjacent MLE screens can ask broader post-training fundamentals. Notes Prepare both definitions and trade-offs. Interviewers can move from basic to advanced quickly. For tokenization, connect vocabulary size to embedding table size, sequence length, compression, rare-token behavior, and downstream latency. For SFT, be ready to describe prompt / response formatting, loss masking, quality filtering, and why data preparation matters as much as optimizer choice. For DPO / GRPO / PPO, focus on what signal each method consumes, whether it needs a reward model, and the stability / implementation trade-offs. Preparation Build a one-page comparison table for DPO, GRPO, and PPO: objective, data required, reward model requirement, stability, and implementation complexity. Rehearse a concise explanation of tokenization trade-offs with one numeric example involving vocab size and sequence length. Review dropout and L1/L2 regularization at the level of equations and practical behavior during training.