Zhenghua Bao

About

I work in the research department at FlashLabs, where I co-lead Chroma, a real-time end-to-end spoken dialogue model that reasons over what it hears. My research asks how AI systems detect and recover from errors in the evidence they use.

In cascaded spoken question answering, speech becomes text, text guides retrieval, and retrieved evidence supports reasoning. An error at one stage can be amplified by later components, even when those components score better on clean input. My EMNLP 2026 paper measures that amplification in multi-hop RAG, and IndexRAG asks what belongs in the index rather than the query path.

Earlier work explored multi-agent reinforcement learning with graph neural networks, specifically generalization across unseen network topologies, and test-time adaptation for EEG foundation models.

Spoken dialogue. End-to-end models that hold a real-time conversation and reason over speech.
Retrieval-augmented generation. Architectures that pre-compute cross-document inference at index time.
Robustness. How errors introduced upstream by speech recognition are amplified rather than absorbed as a question travels through multi-hop retrieval.

Education

2023–2025 MSc Computer Science, focus on artificial intelligence^†
Technische Universität Darmstadt

2022–2024 MSc Internet and Web-based Systems^†
Technische Universität Darmstadt

† One thesis, credited to both MSc programmes.

2018–2022 BSc Computer Science
Technische Universität Darmstadt

Zhenghua Bao at his master's graduation at TU Darmstadt, in front of a slide naming his thesis — Master’s graduation at TU Darmstadt, 6 June 2025. The thesis on the slide is *Integration of Multi-Agent Reinforcement Learning in Dynamic Networks*, supervised by Dr. Jannis Weil at the Multimedia Communications Lab of Prof. Dr.-Ing. Dr. h.c. Ralf Steinmetz.

Zhenghua Bao in cap and gown against a TU Darmstadt backdrop — The same ceremony, at the photo wall in cap and gown.

News

2026.09 IndexRAG was accepted to AACL-IJCNLP 2026 Findings.

2026.08 Better Retrieval, Worse Robustness was accepted to EMNLP 2026 Main as sole author.

2026.08 Released the preprint, code, and dataset for the spoken multi-hop RAG study.

2026.06 OrcaRouter-Lite reached 650+ GitHub stars.

2026.05 The OrcaRouter technical report is on arXiv.

2026.03 Chroma 1.0 reached 540+ GitHub stars and 15k+ HuggingFace downloads.

2026.03 IndexRAG preprint released on arXiv.

2026.01 FlashLabs Chroma 1.0 open-sourced and released as a preprint.

Earlier

2025.11 Joined FlashLabs as an AI Research Engineer.

2025.09 NeuroTTT preprint released on arXiv.

2025.07 Completed MSc in Computer Science (AI) at TU Darmstadt.

2024.10 Completed MSc in Internet and Web-based Systems at TU Darmstadt.

2024.05 Towards Generalizability of Multi-Agent Reinforcement Learning in Graphs was accepted to AAMAS 2024 as an oral presentation.

2024.02 AAMAS preprint released on arXiv.

Publications

Better Retrieval, Worse Robustness: How Multi-hop RAG Amplifies Upstream ASR Errors

Zhenghua Bao

EMNLP 2026 MainPoster

Budapest, Hungary

arXiv code dataset

BibTeX

@misc{bao2026better,
      title={Better Retrieval, Worse Robustness: How Multi-hop RAG Amplifies Upstream ASR Errors},
      author={Zhenghua Bao},
      year={2026},
      eprint={2608.22872},
      archivePrefix={arXiv},
      primaryClass={cs.CL},
      url={https://arxiv.org/abs/2608.22872},
}

IndexRAG: Index-Time Reasoning for Multi-Hop Retrieval-Augmented Generation

Zhenghua Bao, Yi Shi

Findings of AACL-IJCNLP 2026Poster

Hengqin, China

arXiv code

BibTeX

@inproceedings{bao2026indexrag,
  title     = {IndexRAG: Index-Time Reasoning for Multi-Hop Retrieval-Augmented Generation},
  author    = {Bao, Zhenghua and Shi, Yi},
  booktitle = {Findings of the Association for Computational Linguistics: AACL-IJCNLP 2026},
  year      = {2026},
  note      = {To appear},
  url       = {https://arxiv.org/abs/2603.16415}
}

OrcaRouter: A Production-Oriented LLM Router with Hybrid Offline-Online Learning

Zhenghua Bao, Fengya Tian, Chris Zhang, Zhenjun Chen, Xile Ma, Yi Shi

Technical report, 2026

1.5k+ stars

arXiv code

BibTeX

@misc{bao2026orcarouter,
      title={OrcaRouter: A Production-Oriented LLM Router with Hybrid Offline-Online Learning},
      author={Zhenghua Bao and Fengya Tian and Chris Zhang and Zhenjun Chen and Xile Ma and Yi Shi},
      year={2026},
      eprint={2605.30736},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2605.30736},
}

FlashLabs Chroma 1.0: A Real-Time End-to-End Spoken Dialogue Model with Personalized Voice Cloning

Tanyu Chen*, Tairan Chen*, Kai Shen*, Zhenghua Bao*, Zhihui Zhang, Man Yuan, Yi Shi

Preprint, 2026

550+ stars, 17k+ downloads, 380+ likes

arXiv code model

BibTeX

@misc{chen2026chroma,
      title={FlashLabs Chroma 1.0: A Real-Time End-to-End Spoken Dialogue Model with Personalized Voice Cloning},
      author={Tanyu Chen and Tairan Chen and Kai Shen and Zhenghua Bao and Zhihui Zhang and Man Yuan and Yi Shi},
      year={2026},
      eprint={2601.11141},
      archivePrefix={arXiv},
      primaryClass={cs.SD},
      url={https://arxiv.org/abs/2601.11141},
}

NeuroTTT: Bridging Pretraining-Downstream Task Misalignment in EEG Foundation Models via Test-Time Training

Suli Wang, Yangshen Deng, Zhenghua Bao, Xinyu Zhan, Yiqun Duan

Preprint, 2025

arXiv code

BibTeX

@misc{wang2025neurottt,
      title={NeuroTTT: Bridging Pretraining-Downstream Task Misalignment in EEG Foundation Models via Test-Time Training},
      author={Suli Wang and Yangshen Deng and Zhenghua Bao and Xinyu Zhan and Yiqun Duan},
      year={2025},
      eprint={2509.26301},
      archivePrefix={arXiv},
      primaryClass={cs.LG},
      url={https://arxiv.org/abs/2509.26301},
}

Towards Generalizability of Multi-Agent Reinforcement Learning in Graphs with Recurrent Message Passing

Jannis Weil, Zhenghua Bao, Osama Abboud, Tobias Meuser

AAMAS 2024Oral

Auckland, New Zealand

arXiv pdf code

BibTeX

@inproceedings{weil2024towards,
  title={Towards Generalizability of Multi-Agent Reinforcement Learning in Graphs with Recurrent Message Passing},
  author={Weil, Jannis and Bao, Zhenghua and Abboud, Osama and Meuser, Tobias},
  booktitle={Proceedings of the 23rd International Conference on Autonomous Agents and Multiagent Systems (AAMAS 2024)},
  pages={1919--1927},
  year={2024},
  publisher={IFAAMAS},
  url={https://www.ifaamas.org/Proceedings/aamas2024/pdfs/p1919.pdf}
}

* denotes equal contribution.