Tag: on-policy
All the articles with the tag "on-policy".
-
Notes on: OPD+
The actual bug OPD+ finds — stop-gradient drops a real gradient term for every f-divergence except reverse KL, which is why MiniLLM-style training "just happened" to work — plus the fix and the Table 2 numbers.
-
Notes on: Rethinking On-Policy Distillation of LLMs
The actual mechanism behind when on-policy distillation works — token overlap ratio rising toward 91%, a "stronger teacher can fail while a weaker one succeeds" result, and a reverse-distillation experiment that undercuts the whole premise.
-
Notes on: MiniLLM
The reverse-KL policy-gradient derivation behind MiniLLM — per-token rewards, teacher-mixed sampling, length normalization — plus the actual GPT-2/OPT/LLaMA numbers.
-
Notes on: On-Policy Distillation of Language Models (GKD)
The actual GKD objective — a λ-weighted mix of fixed-dataset and student-sampled sequences, a tunable JSD(β) divergence family, and where each choice wins across T5-family experiments.