DPPO
Notes on DPPO - a trust-region rethink for LLM RL that replaces sampled-token ratio clipping with direct divergence constraints such as TV or KL, using binary and top-k approximations to keep it practical.
Notes on DPPO - a trust-region rethink for LLM RL that replaces sampled-token ratio clipping with direct divergence constraints such as TV or KL, using binary and top-k approximations to keep it practical.