Shanbo Cheng

Shanbo Cheng

程善伯

Senior Principal Engineer

Seed Model, ByteDance

Singapore

About me

I’m a Senior Principal Engineer on the Seed Model team at ByteDance in Singapore. My research focuses on multilingual language models and multimodal understanding and generation.

I led the development of Seed LiveInterpret 2.0, which interprets live speech between English and Chinese in the speaker’s own voice. My team also released Seed-X, a family of open-weight text translation models covering 28 languages.

2026

  1. ACL

    Improving Long-Context Translation via Self-Supervised Dual Learning

    LongDu trains on monolingual documents. It favors translations that retain enough information to reconstruct the original document.

    BibTeX

    LongDu back-translates a candidate document translation to evaluate reconstruction.
  2. EMNLP

    GRRM: Group Relative Reward Modeling for Machine Translation

  3. ICLR

    DuPO: Enabling Reliable Self-Verification via Dual Preference Optimization

2025

  1. arXiv preprint

    Seed LiveInterpret 2.0: End-to-end Simultaneous Speech-to-speech Translation with Your Voice

    A duplex model interprets between English and Chinese while preserving each speaker’s voice, with an average delay of about three seconds in our evaluation.

    Project

    Overlapping English speech and Chinese interpretation with a short delay.
  2. arXiv preprint

    Seed-X: Building Strong Multilingual Translation LLM with 7B Parameters

    A family of open-weight 7B models for translation across 28 languages, including instruction-tuned and reasoning variants.

    Code & models

    Average COMET scores for Seed-X Instruct, DuPO and PPO are 89.05, 89.96 and 90.81, respectively (Table 4 of the report).

2024

  1. arXiv preprint

    Towards Achieving Human Parity on End-to-end Simultaneous Speech Translation via LLM Agent

    CLASI learns when to listen and when to translate. It retrieves domain terminology to interpret specialized speech.

    ProjectBenchmark

    Streaming speech, memory and retrieved terms provide context to the CLASI translation agent.

2023

  1. Findings of ACL

    BigVideo: A Large-scale Video Subtitle Translation Dataset for Multimodal Machine Translation

2022

  1. ICLR

    switch-GLAT: Multilingual Parallel Machine Translation Via Code-Switch Decoder

2016

  1. NAACL

    PRIMT: A Pick-Revise Framework for Interactive Machine Translation

Academic Service

Area Chair

IJCNLP-AACL 2023 · NLP Applications track

Conference reviewing

ACL
2019, 2020, 2021, 2023 (primary reviewer)
EMNLP
2021, 2022, 2025
NAACL
2021, 2022
EACL
2023
NeurIPS
2024
ICLR
2025
ICASSP
2023

Program committee member

AAAI
2023
CCMT
2025
NLPCC
2022

ACL Rolling Review

Reviewer for the January and December 2022 cycles and the May 2025 cycle. Reviewer roster