Chia-chun Liang

PhD Student, Data Science Degree Program

National Taiwan University

About

I am a PhD student in the Data Science Degree Program at National Taiwan University, where I go by Anna. My research centers on large language models, in particular on understanding when they should be trusted and how reliably they can express what they do not know.

My earlier work approached machine learning from the systems side, on making deep networks cheaper to run through tensor decomposition, and on building speech recognition systems for languages with little training data. Those projects left me with a lasting interest in the gap between a model's reported confidence and its actual competence, the question I now work on directly.

If you are also a Christian speaking Chinese, welcome to see my Christian blog.

Recently I just started my journey on how to make some contribution (not trouble) in some open-source projects. (See my github)

Research Interests

Large Language Models

Evaluating and improving how large language models reason, generalize, and behave under distribution shift, with an emphasis on what their outputs actually reveal about their internal state.

Uncertainty Estimation

Calibration, confidence, and selective prediction: measuring when a model knows that it does not know, and turning that signal into systems that can abstain or defer rather than fail silently.

Projects

  • How LLM reads your Python?

    PyCon TW 2026 poster. From the Transformer Circuits perspective to hands-on tracing of an open-source model.

    An attribution-graph study of how gemma-2-2b completes a Python function. Cut off after return, the model reliably predicts the right local variable, and control prompts show how: each variable name has its own pathway to its own logit, binding sets how much drive that name accumulates, and scaling one pathway down flips the answer. Theory (the 2020 to 2025 timeline, attention as a kernel, the word "circuit") and practice (the write-up, the interactive graph, and where the rules break).

    Mechanistic Interpretability Circuit Tracing Code Models
  • Daily stock direction forecasts

    A forecasting page that retrains itself every trading day. Twelve Taiwan and fourteen US blue chips, ten years of daily bars, one XGBoost classifier per stock and horizon, walk-forward accuracy printed next to every probability, and a running track record of how the published forecasts actually turned out. Runs on GitHub Actions and commits its own data back to this repository.

    XGBoost Time Series GitHub Actions
  • Mock trading

    Simple trading rules (momentum, reversal, low volatility, a trend filter) run on one basket of Taiwan large caps since 2006 with brokerage and transaction tax, against the market and an equal-weight benchmark. Pick any window of years and compare growth, drawdown and calendar-year returns; nothing places an order. Reruns itself weekly on GitHub Actions from FinMind data.

    Backtesting Taiwan Equities GitHub Actions
  • Macro dashboard

    A dashboard, in Traditional Chinese with an English toggle, of world equity indices, index futures, oil, gold, currencies, Treasury yields, the Fed's target range and U.S. inflation and jobs since 2020, with every central-bank decision linked to its press release and a growing archive of public headlines. Two companion pages track institutional flows in Taiwan, China and Hong Kong, Japan and the U.S. at each market's real publication cadence, and hot IPOs from offer price to today. Public data only, refreshed three times a day on GitHub Actions.

    Macro Public Data GitHub Actions

Tools

  • Headset check

    Three tests to run before a call, all in the browser with nothing uploaded: a microphone level meter with record-and-listen-back, a hard left / hard right channel test, and a clarity report that measures speech level, noise floor, clipping and the microphone's effective bandwidth, the number that tells a Bluetooth headset stuck in telephone-quality hands-free mode from one that is fine.

    Web Audio API Signal Processing No Dependencies

Publications

  1. A Whisper-Based System with Multi-Faceted Data Augmentation for Low-Resource Language

    P.-C. Chen, Y.-C. Chen, C.-C. Liang, C.-Y. Lin, P.-J. Tsai, W.-Y. Ma
    Proceedings of the 37th Conference on Computational Linguistics and Speech Processing (ROCLING), 2025

    Speech Recognition Low-Resource Languages Data Augmentation
  2. Automatic Selection of Tensor Decomposition for Compressing Convolutional Neural Networks: A Case Study on VGG-type Networks

    C.-C. Liang, C.-R. Lee
    IEEE International Parallel and Distributed Processing Symposium Workshops (IPDPSW), 2021, pp. 770-778

    Model Compression Tensor Decomposition

    DOI

Teaching & Outreach

  • june-concierge-Sci-Eng

    An AI-generated lecture series in science and engineering. The current run works through Advanced Calculus (Loomis & Sternberg) section by section, from logic and quantifiers to vector spaces and linear transformations, with every episode released in both Mandarin and English.

    The videos come out of a generation pipeline originally built for the inaugural 教學怪物 (Teaching Monster) competition, where our team, PhD.ICU, placed fifth. The channel is where that pipeline has kept running since, as much an experiment in what current models can carry with little supervision as it is a set of lecture notes.

    AI-Generated Video Mathematics Bilingual