How VPO Trains LLMs to Win at Test-Time Search with Ryan Bahlous-Boldi from MIT
Deep Learning with Yacine · 124:43 · Watch on YouTube
VPO (Vector Policy Optimization), a method from Ryan at MIT's Improbable AI Lab, changes RL post-training so a language model learns to produce several different good answers instead of one. Standard GRPO sums the rew...