This digest covers post-training and fine-tuning developments from 2026-11-06 to 2026-11-13, across arXiv, the Hugging Face blog, Together AI, Interconnects, and Latent Space.
🔥 Highlights
Human or Synthetic? What a Fixed Budget Buys for Post-Training — synthetic preference data wins hard at low budgets.
The Surprising Effectiveness of Deleting Weights in LLM Post-Training — zeroing 0.05% of weights matches full fine-tuning.
Rethinking "RL Generalizes, SFT Memorizes": The Role of SFT Data — data mix, not algorithm, drives generalization.
SafetyPulse: Catching Safety Drift During Fine-Tuning by Reading the Model's Refusal Circuit — a live monitor for alignment decay.
A Theory of Policy Optimization in Post-Training: Coverage Collapse and Reference Resets — a theoretical fix for stalled RLHF/GRPO runs.
arXiv (cs.CL / cs.LG)
PORT: Preference Optimization via Robust Token-Level Reweighting — 2026-11-09.…