DAPO trains a step-level value critic and regresses the policy log-ratio to the critic-derived advantage, improving LLM math and code reasoning over the base model on several benchmarks.
On the Linear Convergence of Policy Gradient under Hadamard Parameterization
1 Pith paper cite this work. Polarity classification is still indexing.
abstract
The convergence of deterministic policy gradient under the Hadamard parameterization is studied in the tabular setting and the linear convergence of the algorithm is established. To this end, we first show that the error decreases at an $O(\frac{1}{k})$ rate for all the iterations. Based on this result, we further show that the algorithm has a faster local linear convergence rate after $k_0$ iterations, where $k_0$ is a constant that only depends on the MDP problem and the initialization. To show the local linear convergence of the algorithm, we have indeed established the contraction of the sub-optimal probability $b_s^k$ (i.e., the probability of the output policy $\pi^k$ on non-optimal actions) when $k\ge k_0$.
citation-role summary
citation-polarity summary
fields
cs.AI 1years
2024 1verdicts
CONDITIONAL 1roles
background 1polarities
unclear 1representative citing papers
citing papers explorer
-
Improving Multi-Step Reasoning Abilities of Large Language Models with Direct Advantage Policy Optimization
DAPO trains a step-level value critic and regresses the policy log-ratio to the critic-derived advantage, improving LLM math and code reasoning over the base model on several benchmarks.