This is the repo for ACL 2026 paper "Hierarchical Policy Optimization for Simultaneous Translation of Unbounded Speech". It only contains the RL post-training part. For SFT, please refer to InfiniSST repo.

We provide slurm script to run it on 3 8xH100 nodes.

bash docker_sbatch_3node.sh YAML_NAME 

You can find the following yaml files under examples/configs

For En-Zh

• grpo_infinisst_4b_laal0.5_as_tgtq-5.0 (multiplier=1)
• grpo_infinisst_4b_laal0.5_as_tgtq-5.0_m{2...6} (multiplier=2...6)

For En-De

• grpo_infinisst_4b_laal0.5_as_tgtq-5.0_de (multiplier=1)
• grpo_infinisst_4b_laal0.5_as_tgtq-5.0_de_m{2...6} (multiplier=2...6)

For En-Ja

• grpo_infinisst_4b_laal0.5_as_tgtq-5.0_ja (multiplier=1)
• grpo_infinisst_4b_laal0.5_as_tgtq-5.0_ja_m{2...6} (multiplier=2...6)

Running this requires the SFT model checkpoint, the speech data pre-encoded into features with the speech encoder, and a docker container.

We provide the example SFT checkpoint for en-zh [here], and the example pre-encoded data [manifest] [encoded feature] and the docker image [here].