Skip to content

m2po.md

Latest commit

 

History

History
73 lines (47 loc) · 2.82 KB

m2po.md

File metadata and controls

73 lines (47 loc) · 2.82 KB
Raw

Second-Moment Trust Policy Optimization (M2PO)

Last updated: Oct 23, 2025

Author: Jingyuan Ma

m2po figure

Second-Moment Trust Policy Optimization (M2PO) (Zheng et al., 2025) is an RL method that achieves stable off-policy training even when data is stale by at least 256 model updates and matches on-policy performance by constraining the second moment of importance weights to suppress only extreme outliers while preserving informative updates.

The first step of M2PO is to compute the second moment: $$ \hat{M_2}=\frac{1}{N}\sum_{i=1}^NM_{2,i}=\frac{1}{N}\sum_{i=1}^N(\log{r_i})^2=\frac{1}{N}\sum_{i=1}^N\left(\log\frac{\pi_\theta (a_i|s_i)}{\pi_{behav}(a_i|s_i)}\right)^2 $$

The second step is to compute the second-moment mask:

m2po masking

The final step is to optimize the objective:

$$ J_{\text{M2PO}}(\theta) = \frac{1}{\sum_{i=1}^G|o_i|}\sum_{i=1}^G\sum_{t=1}^{|o_i|}M_{i,t}\frac{\pi_\theta(o_i|q)}{\pi_{\theta_{old}}(o_i|q)}A_{i,t}. $$

Where $M$ is computed in the second step and

$$ A_{i,t}=\frac{r_i-mean({R_i}_{i=1}^G)}{std({R_i}_{i=1}^G)}. $$

For more details:

Core Parameters

  • actor.m2_threshold: The threshold for the mean of the second moment, used in computing the M2PO mask as $\tau_{M_2}$

Example Usage

We recommend changing the parameters in the configuration file (examples/math/gsm8k_m2po.yaml).

Backend CMD
local python3 examples/math/gsm8k_rl.py --config examples/math/gsm8k_m2po.yaml scheduler.type=local --<other_args_to_overwrite>
ray python3 examples/math/gsm8k_rl.py --config examples/math/gsm8k_m2po.yaml scheduler.type=ray --<other_args_to_overwrite>
slurm python3 examples/math/gsm8k_rl.py --config examples/math/gsm8k_m2po.yaml scheduler.type=slurm --<other_args_to_overwrite>

Test Result

m2po test figure

In this test, trial names follow these conventions:

  • stale: the value of max_head_offpolicyness
  • dx+dy: x is the number of rollout workers and y is the number of training workers
  • rollout: the value of max_concurrent_rollout

The GRPO setting is stale 256 d2+d1 rollout 96.

The key findings across the trials are as follows:

  • The grad_norm of GRPO is higher than M2PO, which may cause training instability.
  • The evaluation reward of M2PO is higher than GRPO.