Chihsheng Jin

Chihsheng Jin, William Jurayj, William Walden

We developed a general pipeline for text steganography that is robust to extreme semantic preserving attacks with the help of large language models. We implemented the pipeline with three NLG tasks, namely long-form question answering, story generation, and literature review. The generated stegotexts are generally discrete to common texts generated for the respective NLG tasks.

William Walden, Pavlo Kuchmiichuk, Alexander Martin, Chihsheng Jin, Angela Cao, Claire Sun, Curisia Allen, Aaron Steven White

We constructed a cross-document summarization dataset on top of the FAMuS dataset. Results from extensive experiments show that cross-document summarization is a non-trivial task for language models, and smaller models can even outperform larger models when fine-tuned with the dataset.

Presented at XLLM@ACL2025 and PEER2025  ·  Slides

Senior thesis

I used a character-naming paradigm to test whether the phonological mapping status of cognates in Mandarin and Shanghai Dialect would affect the speech production process. Besides the common cognate facilitation effect, there are two novel findings: the phonological mapping complexity between languages interferes with speech production, and whether phonemes are the most frequent match for the latent language also affects production.

Note: This experiment was conducted during a COVID lockdown with limited sample size, so I decided against publishing.

A minimalist PDF reader extension designed for people who spend way too much time reading academic papers in the browser, packed with superpowers. It intentionally keeps only the most essential features, but aims to make the actual experience of reading papers smoother, faster, and more immersive.

Live Demo (Desktop Only)

I conducted extensive experiments on GPT-4o and Claude Sonnet 3.5's capability of extracting event roles from the FAMuS dataset using various prompt strategies. Using CEAF metrics, results suggest large language models cannot extract roles accurately, indicating they lack human-level event understanding. However, with post-processing, extractions can still be useful for some applications.

Read the paper

A Python module for parsing multiple context-free grammars. MCFG is a class of grammars strictly more expressive than context-free grammar, useful for cross-serial dependencies, agreement within relative clauses, and more.

Undisclosed.

Worked on Robust Semantic Steganography with LLMs and Multi-modal RAG benchmarks.

Worked on a series of projects related to structured representations of events, event extraction and event retrieval.

Project management intern. Used coding skills to process files and improve workflow efficiency.

Chief editor and designer for a bimonthly in-house newspaper dedicated to news on language regulations and language research at Fudan University.