KAI-CHOU YANG · 楊凱州
From data to capability.
I build data and training systems for language models and agents, from large-scale data curation to agent post-training.
I’m a machine learning engineer working on how language models learn and act. At Taiwan AI Labs, I lead the NLP team working on FedGPT, across data engineering, post-training, and agent systems. Lately, I’ve been interested in training agents in synthetic environments, where their actions can be checked against verifiable outcomes.
I came to the field through competitions and open source, and I still think of the work as a craft—something to practise with care. Outside work, I make small things and publish what comes out: models, datasets, and benchmarks, mostly for Chinese.
I also write, mostly in Chinese. I’m interested in technology where it meets everyday life, and when I try to explain something difficult, I look for the story around it and for examples beyond the field itself. If something I’ve written is wrong, unclear, or could simply be better, I’d be glad to hear about it.
Deep research · Personal project
Research you can follow, claim by claim.
A deep-research system where specialist agents gather evidence and freeze it before any writing begins, so every sentence can cite only what was read. A verification gate checks each citation, figure and quote, and every report ships with its receipts.
Evidence-first research · Multi-agent orchestration
Language models · Open source
Better Traditional Chinese through better data.
Retrieval-augmented data synthesis and post-training for Traditional Chinese: 2B and 9B models trained on 1M synthesized conversations and 150K preference pairs.
2nd placeGoogle / Kaggle2025
Chinese writing · Open source
Writing as plan, draft and revision.
A 24B Chinese writing model trained on its own process: planning, drafting, review and revision form a data-synthesis loop, with rejection sampling selecting the examples worth training on. Released alongside Zhiyin, a Chinese writing benchmark of 280 prompts across 18 tasks.
Data synthesis · Model evaluation
Code, models and learning data, from early NLP tools to post-training and evaluation.
Code & implementations
Retrieval-based Chinese conversation, with a PTT-derived dialogue corpus.
Solution code from four competition entries in text matching and classification.
Hands-on Chinese word embeddings and sequence-to-sequence learning.
Models & training data
Learning data built around execution feedback and chosen–rejected pairs.
Chinese science QA and mathematical reasoning, with answer-level checks.
Multi-turn tool-use trajectories for training agents: ToolScale rollouts from two frontier models, and a larger set filtered against ground-truth tool calls.
Google · KaggleAdapting Gemma to a language or cultural context.
Create a Gemma 2 model variant for a specific language or cultural context, with a clear, reproducible account of the tuning approach.
Kyara focuses on Traditional Chinese: retrieval-grounded instruction synthesis, translation and rewriting, followed by supervised fine-tuning and preference optimization. The work includes released models, training recipes, and instruction and preference datasets.
Google AI · KaggleResolving ambiguous pronouns across genders.
Given a passage, an ambiguous pronoun and two candidate entities, predict A, B or neither. The challenge evaluates pronoun resolution across genders using multiclass log loss.
ByteDance · WSDM CupClassifying the relationship between two headlines.
Given a known fake-news headline A and a new headline B, classify their relation: agreed, disagreed or unrelated. The task compares headline pairs against a known claim, rather than judging an isolated article as true or false.
Alibaba · CIKMAdapting question-pair matching from English to Spanish.
Determine whether two questions have the same meaning in Spanish, using English as the source language for adaptation. The challenge focuses on short-text semantic matching under limited target-language supervision.
Journal of Medical Internet Research
Background Vaccination is an important intervention to prevent the incidence and spread of serious diseases. Many factors including information obtained from the internet influence individuals' decisions to vaccinate. Misinformation is a critical issue and can be hard to detect, although it can change people's minds, opinions, and decisions. The impact of misinformation on public health and vaccination hesitancy is well documented, but little research has been conducted on the relationship between the size of the population reached by misinformation and the vaccination decisions made by that population. A number of fact-checking services are available on the web, including the Islander news analysis system, a free web service that provides individuals with real-time judgment on web news. In this study, we used such services to estimate the amount of fake news available and used Google Trends levels to model the spread of fake news. We quantified this relationship using official public data on COVID-19 vaccination in Taiwan.
Objective In this study, we aimed to quantify the impact of the magnitude of the propagation of fake news on vaccination decisions.
Methods We collected public data about COVID-19 infections and vaccination from Taiwan's official website and estimated the popularity of searches using Google Trends. We indirectly collected news from 26 digital media sources, using the news database of the Islander system. This system crawls the internet in real time, analyzes the news, and stores it. The incitement and suspicion scores of the Islander system were used to objectively judge news, and a fake news percentage variable was produced. We used multivariable linear regression, chi-square tests, and the Johnson-Neyman procedure to analyze this relationship, using weekly data.
Results A total of 791,183 news items were obtained over 43 weeks in 2021. There was a significant increase in the proportion of fake news in 11 of the 26 media sources during the public vaccination stage. The regression model revealed a positive adjusted coefficient (β=0.98, P=.002) of vaccine availability on the following week's vaccination doses, and a negative adjusted coefficient (β=-3.21, P=.04) of the interaction term on the fake news percentage with the Google Trends level. The Johnson-Neiman plot of the adjusted effect for the interaction term showed that the Google Trends level had a significant negative adjustment effect on vaccination doses for the following week when the proportion of fake news exceeded 39.3%.
Conclusions There was a significant relationship between the amount of fake news to which the population was exposed and the number of vaccination doses administered. Reducing the amount of fake news and increasing public immunity to misinformation will be critical to maintain public health in the internet age.
arXiv · Preprint
With thousands of news articles from hundreds of sources distributed and shared every day, news consumption and information acquisition have been increasingly difficult for readers. Additionally, the content of news articles is becoming catchy or even inciting to attract readership, harming the accuracy of news reporting. We present Islander, an online news analyzing system. The system allows users to browse trending topics with articles from multiple sources and perspectives. We define several metrics as proxies for news quality, and develop algorithms for automatic estimation. The quality estimation results are delivered through a web interface to newsreaders for easy access to news and information. The website is publicly available at https://islander.cc/
In this paper, we propose Self Inference Neural Network (SINN), a simple yet efficient sentence encoder which leverages knowledge from recurrent and convolutional neural networks. SINN gathers semantic evidence in an interaction space which is subsequently fused by a shared vector gate to determine the most relevant mixture of contextual information. We evaluate the proposed method on four benchmarks among three NLP tasks. Experimental results demonstrate that our model sets a new state-of-the-art on MultiNLI, Scitail and is competitive on the remaining two datasets over all sentence encoding methods. The encoding and inference process in our model is highly interpretable. Through visualizations of the fusion component, we open the black box of our network and explore the applicability of the base encoding methods case by case.
GeBNLP Workshop at ACL 2019
In this paper, we describe our entry in the gendered pronoun resolution competition which achieved fourth place without data augmentation. Our method is an ensemble system of BERTs which resolves co-reference in an interaction space. We report four insights from our work: BERT's representations involve significant redundancy; modeling interaction effects similar to natural language inference models is useful for this task; there is an optimal BERT layer to extract representations for pronoun resolution; and the difference between the attention weights from the pronoun to the candidate entities was highly correlated with the correct label, with interesting implications for future work.
WSDM Cup 2019 · arXiv
This report describes the entry by the Intelligent Knowledge Management (IKM) Lab in the WSDM 2019 Fake News Classification challenge. We treat the task as natural language inference (NLI). We individually train a number of the strongest NLI models as well as BERT. We ensemble these results and retrain with noisy labels in two stages. We analyze transitivity relations in the train and test sets and determine a set of test cases that can be reliably classified on this basis. The remainder of test cases are classified by our ensemble. Our entry achieves test set accuracy of 88.063% for 3rd place in the competition.
Data Science Collective · Medium
In this article, we summarize our approach that achieved second place in the Kaggle Competition: Google — Unlock Global Communication with Gemma, which is a LLM tuning competition featured a $150,000 prize pool.
I’ll break down the data recipe behind our solution: what worked, what didn’t, and how you can reproduce the same ideas for your own LLM.
TDS Archive · Medium
This is my solution summary for CIKM AnalytiCUP 2018, a competition about short-text semantic similarity and knowledge transfer.
The goal of this challenge is to build a cross-lingual short-text matching model. The source language is English and the target language is Spanish. Participants could train their models by applying advanced techniques to classify whether question pairs are the same meaning or not.
More implementation notes and derivations live at 雷德麥的藏書閣.
Taiwan AI Labs
Mar 2023 — Present
Lead LLM and agent development behind FedGPT, from trillion-token data pipelines and knowledge synthesis to agent post-training and an in-house multi-agent harness.
Data curation, knowledge synthesis, supervised fine-tuning, rejection sampling, reinforcement learning, and stateful agent systems.
Sep 2019 — Mar 2023
Built retrieval, question-answering and language-understanding systems, with an emphasis on weak supervision, domain adaptation and reusable training pipelines.
Multilingual dense retrieval, passage reranking, RAG, and Kubeflow training pipelines.
National Cheng Kung University
Computer Science and Information Engineering
2017 — 2019
GPA 4.30 / 4.30 Ranked 1st
Honorary member of the Phi-Tau-Phi Scholastic Honor Society.
Speaker for organizations and communities including the Industrial Technology Research Institute (ITRI), Delta Electronics, and PyTorch Tainan.
2013 — 2017
GPA 3.94 / 4.30 Ranked 4th
Academic Excellence Awards, 2015 & 2016.
Honorable mention at the graduation exhibition.
Research assistant on a question-answering project funded by the Ministry of Science and Technology.
Qwen Dev Ambassador
Contributing to Qwen’s agentic tool-calling ecosystem.