Chen
/

If AI is a nuclear blast, I choose to witness it firsthand.

If it is a bubble, I’ll be on the front line, laughing at the absurdity.

曹铸臣

Zhuchen Cao.

Language-model research. Clinical AI. Systems engineering.

Head of AI · Jinxin Technology

Centered natural-color portrait of Zhuchen Cao in a black T-shirt, without sunglasses, on a white background
EMNLP 2026Oral · Award nominationCo-author, main-conference paperCodeRSAFirst author · 10 / 12Best mean accuracy among compared rerankers · 10/12 settingsMPI InformaticsResearch with Vera DembergPragmatic reasoning & language modelsJinxin TechnologyHead of AIClinical speech, AI monitoring, document systems

A conversation with Chen

Interview my double.

Zhuchen Cao wearing round sunglasses

Ask about my work. Ask something awkward.

Enter to ask · Shift + Enter for a new line0 / 1000

An AI double. Factual answers link to sources.

Questions go to an AI provider; this site does not save the conversation.

01

Research

2026Co-authorEMNLP · ORALAward nomination

Summarization is Not Dead Yet

Dongqi Liu, Chenxi Whitehouse, Zheng Zhao, Zhuchen Cao, Jian Li, Yabiao Wang

Do fluent summaries actually preserve the information that matters? A multi-track evaluation examines factuality, information coverage and linguistic diversity.

Method

Human assessment, bias-mitigated LLM judges, external fact verification and corpus-level linguistic analysis.

Finding

Across five datasets and five LLMs, human reference summaries retained advantages in informativeness and faithfulness.

My contribution

Led the implementation and refinement of the evaluation algorithms, and contributed to project design.

5 datasets × 5 modelsPaper
2025 / 26First authorPreprint

CodeRSA

Pragmatic Reasoning improves LLM Code Generation

Zhuchen Cao, Sven Apel, Adish Singla, Vera Demberg

An ambiguous instruction can fit several plausible programs. CodeRSA ranks candidates by reasoning about what the user meant.

Method

Generate alternative instructions from candidate code. Compare candidates through local pragmatic contests, then combine that signal with global support.

Evaluation

HumanEval+, MBPP+ and BigCodeBench across four open-weight models. Highest mean accuracy among the compared rerankers in 10 of 12 model–benchmark settings.

3 benchmarks × 4 modelsResearch notes
02

Engineering

JINXIN TECHNOLOGY

I lead the AI work across these four projects at Jinxin Technology.

01

Clinical Notes

Hospital pilotProject overview

AI Lead

A clinical record of the consultation, the patient's concerns and what needs to happen next.

  • Clinical documentation

    Turn consultations into structured draft notes for clinician review.

  • Patient context

    Keep patient concerns and follow-up needs alongside the medical facts.

  • Traceable evidence

    Retain the original audio so clinicians can check a draft against the conversation.

Hospital-record integration is in progress. Communication analysis runs in batches; named-nurse handoff is the next step.

02

GEO Observatory

In operationProject overview

AI Lead

An operational system for examining how AI platforms describe hospitals—and checking the evidence behind their answers.

  • Across platforms

    Collect answers and citations from multiple AI platforms for comparison.

  • Back to the source

    Connect each answer to its sources, making missing or inconsistent information visible.

  • Review & follow-up

    Bring answers, evidence and follow-up into one review workflow for hospital teams.

Collection and review are deployed. Changes in model answers are measured through observation.

03

IVF Digital Passport

Working prototypeProject overview

AI Lead

An IVF-specific record that organizes reports across treatment cycles into a coherent, verifiable history.

  • Across cycles

    Bring scattered reports into a treatment history, with each finding linked to its source.

  • Uncertainty stays visible

    Keep conflicting reports visible and corrections traceable.

  • Records that travel

    Create source-linked summaries patients can take to another hospital.

Designed for free patient access. Clinical and data-governance validation is the next step.

04

IVF-Copilot

Research prototype

AI Lead

Decision support that follows an IVF stimulation cycle, revising its assessment as new evidence arrives.

  • Prediction & reasoning

    Project follicle and hormone trajectories under stated treatment assumptions; weigh the evidence and alternative explanations.

  • Longitudinal updates

    Each visit updates the assessment with observed results and actual medication, replacing outdated predictions.

  • Clinical review

    Present predicted trajectories with their reasoning for clinician review.

Implemented prototype with end-to-end workflow validation. Prospective clinical evaluation is pending; treatment decisions remain with clinicians.

03

Experience

JUN 2026 — PRESENT

Jinxin Technology

Head of AI

Leading the AI work across Clinical Notes, GEO Observatory, IVF Digital Passport and IVF-Copilot.

RESEARCH FROM NOV 2023

Max Planck Institute for Informatics

Research with Prof. Vera Demberg’s group · Saarland University

Rational Speech Act models, pragmatic reasoning, LLM evaluation and code generation.

Institute profile
EDUCATION

BSc (Hons), Applied Mathematics — Liverpool · MSc, Data Analytics (Distinction) — Queen’s University Belfast