Yiran Rex Ma*, Yuxiao Ye*, and Huiyuan Xie(Co-first authors (*equal contribution). Hybrid, reference-free evaluation of stylistic fidelity in Chinese legal text, combining linguistic features with experience-guided LLM-as-a-judge scoring; experiments on 200 legal documents.)
LREC 2026 (ACL Anthology)
Yuxiao Ye, Yiming Zhang, Yiran Rex Ma, Huiyuan Xie, Huining Zhu, and Zhiyuan Liu(Dialogue modelled as a signalling game over communicative intents and strategies; human-expert evaluations in simulated courtrooms and debates show dialogue quality gains of 6.3%–20.1% across clarity, conciseness, argumentation, and tactics.)
ACL 2026 Findings
Yiran Rex Ma(Position paper charting the computational turn in Chinese paleography, from automating isolated visual tasks toward integrated, human-centric digital research ecosystems.)
arXiv preprint
Yiran Rex Ma(Benchmark for Chinese-English pun translation over text and images, with the Constant-Variable Optimization (CVO) translation strategy and the Overlap (Ovl) metric.)
ACL 2025 SRW
Yiran Rex Ma(Philosopher and Symposium frameworks simulate self- and group-reflection for multi-persona agents; identifies an "overthinking threshold" beyond which deliberation yields diminishing returns.)
ACL 2025 SRW
Yiran Rex Ma(Five LLM scientist personas investigate, self-review, and peer-review their way through table QA and fact verification, outperforming vanilla LLMs and rivaling supervised models without task-specific training.)
International Joint Conference on Neural Networks (IJCNN)
Yiran Rex Ma(LLM-annotated comparable corpora show English news favors core-information-first narration with postposed adverbials, while Chinese news favors background-first presentation with preposed ones.)
arXiv preprint
2026
Hybrid, reference-free evaluation framework for the stylistic quality of Chinese legal text, combining linguistic features with experience-guided LLM-as-a-judge scoring learned from contrastive document pairs.
Present
Company co-founded with his partner to scale continual learning and personal models — foundation models that swiftly adapt to diverse domains and align with individual human preferences beyond verifiable rewards.
