Skip to content

Qwen3-4B Results Reproducibility #41

@rishiskhare

Description

@rishiskhare

Hi LatentMAS authors,

I'm currently unable to replicate the values presented in the full arXiv for Qwen3-4B. I've tried gsm8k and medqa with varying parameters, but in all my runs, the baseline somehow outperforms LatentMAS in accuracy. Could you share the specific hyperparameters you used for Qwen3-4B on an example benchmark for reproducibility (i.e. number of latent steps, latent realignment enabled/disabled, maximum token generation)? If you could share the full config for one specific benchmark to reproduce your numbers, it would greatly help - right now, I'm running medqa on the single-agent baseline which was reported to have 47% accuracy on your paper, though achieves ~66% locally for me, which is even higher than my LatentMAS run on 40 steps.

Sharing the full commands you ran to reproduce Qwen3-4B would be awesome.

Edit: I noticed that this is related to #29

Metadata

Metadata

Assignees

No one assigned

    Labels

    No labels
    No labels

    Type

    No type

    Projects

    No projects

    Milestone

    No milestone

    Relationships

    None yet

    Development

    No branches or pull requests

    Issue actions