sciwork 2026

To see our schedule with full functionality, like timezone conversion and personal scheduling, please enable JavaScript and go here.
09:00
09:00
30min
Registration

Registration

402
09:30
09:30
10min
Opening

Opening

402
09:40
09:40
30min
On the Tenth Anniversary of AlphaGo: An Introduction to Reinforcement Learning and Planning
Sheng-Wen Cheng

Ten years after AlphaGo demonstrated the remarkable potential of combining learning and search, reinforcement learning has become an important framework for building agents that learn to make sequential decisions.

In this talk, we will give an introduction to the core ideas behind reinforcement learning and planning. We will start from rewards and the basic formulation of sequential decision-making, then introduce value functions, Q-value functions, and policies. From there, we will discuss policy evaluation, policy iteration, and policy-based methods, building an intuitive picture of how an agent can learn to improve its behavior through interaction and experience.

Finally, we will move from learning to planning and introduce Monte Carlo Tree Search (MCTS), one of the key components behind AlphaGo. By connecting these concepts, the talk aims to provide a conceptual roadmap for understanding how reinforcement learning and search algorithm can work together to solve complex decision-making problems.

403
09:40
30min
打開 AI 的黑盒子:機制可解釋性(Mechanistic Interpretability)的最新進展
Junli Xiao

這場演講將帶您穿透 LLM 黑箱,探索機制可解釋性(MI)的最新進展,它本質上是一項結合多個研究領域的挑戰。我們將援引各個假說來剖析模型內部的概念運作,討論如何近似手術般精準去觀察、操控模型,並理性展望延伸問題。

402
10:20
10:20
30min
AI Agent x Open Source Compiler
Kito Cheng

三年前,聽到有人想用 AI 幫 GCC 加上新的指令支援時,我的第一反應是:X! 笑死!

三年後,AI Agent 完全融入我的日常 compiler 開發流程:修 bug、加測試、研究效能問題、分析 IR 與 assembly。

這場演講會從一位 Open Source Compiler Developer 的角度,分享在 GCC、LLVM 等大型開源專案中使用 AI Agent 的經驗。

402
10:20
30min
PyVista on WebAssembly: Enabling Serverless 3D Visualization
Tetsuo Koyama

PyVista is a VTK-based Python library for 3D data visualization. In this talk, we introduce the technology stack, implementation details, and performance characteristics of pyvista-wasm, which enables PyVista to run entirely on WebAssembly (WASM). By combining Pyodide and VTK-WASM, we can deliver full-featured 3D scientific visualization directly in the browser—without requiring a backend server. We will share hands-on technical insights from development alongside live browser-based demos.


Background and Motivation

Historically, scientific visualization has relied on local desktop environments or heavy server-side rendering pipelines. With the rise of WebAssembly, heavyweight scientific libraries like VTK are following NumPy and SciPy into browser environments, opening up entirely new architectural possibilities.

Talk Outline (30 mins)

  1. Overview of PyVista and WASM (5 mins) * What is PyVista? * Why WASM? What changes for scientific workflows?

  2. Technology Stack & Implementation Details (12 mins) * Emscripten builds of VTK * Running PyVista on Pyodide * Integrating with WebGL rendering * Challenges and solutions in local development and CI pipelines

  3. In-Browser Live Demo (8 mins) * Running inside JupyterLite * Interactive 3D mesh manipulation

  4. Performance & Limitations (5 mins) * Performance comparison: WASM vs. native * Current WASM constraints and practical workarounds * Future roadmap

403
11:00
11:00
30min
Beyond Calling BLAS: Implementing and Optimizing Matrix Multiplication
Chun-Shih Chang

Matrix multiplication may look like three nested loops or a single BLAS call, but implementing it efficiently in an array library requires more than a fast computational kernel. Shapes, strides, broadcasting, and operand reuse can change which execution strategy works best—even when the arithmetic workload remains the same.

This talk explores the implementation and optimization of matrix multiplication in Solvcon’s SimpleArray. We follow the path from direct computation to layout-aware execution, covering BLAS-compatible layouts, signed-stride operands, and batched matrix multiplication. Particular attention is given to small-matrix execution, reuse of broadcast operands, and the trade-offs between complete packing and streamed packing with reusable scratch storage. We also discuss how separating operation semantics, execution selection, and computation makes these optimizations easier to integrate and maintain.

To evaluate the implementation, we briefly demonstrate a Qt Benchmark Inspector and explain how identical operands, numerical comparisons, and balanced timing help assess performance improvements. The central lesson is that efficient matrix multiplication depends not only on kernel speed, but also on how data layout, reuse, and execution strategy fit together.

403
11:00
30min
從自然語言到空間音訊:用 Codex 打造可驗證的瀏覽器聲學原型
Kevin

當 AI Agent 能夠協助產生大量程式碼,開發者如何確認一個跨越幾何、聲學與即時互動的應用,真的符合原先的設計?

本演講以 OpenAI Build Week 期間開發的 Echo Canvas 為案例,分享如何與 Codex 協作,將自然語言描述與可編輯的空間場景,轉換成能在瀏覽器中操作與試聽的空間音訊原型。使用者可以調整牆面、開口、音源與聆聽位置,觀察聲音路徑,並比較遮蔽、早期反射與殘響帶來的變化。

演講將說明兩個互相配合的設計:以規格、資料契約與測試約束 Codex 的開發工作;在應用內,讓 GPT-5.6 負責編譯候選場景及解釋結果,並透過結構與領域驗證,把模型輸出交給確定性的幾何與聲學計算,再由 Web Audio 完成瀏覽器 HRTF 渲染。

透過架構拆解與 Demo,我們將討論如何界定 AI 的責任、檢查生成結果,以及區分自動測試與人耳驗收所能證明的事情。

Echo Canvas 定位為用於空間音訊原型設計與預視的互動式聲學近似;演講也會交代近似方法與使用限制。聽眾將帶走一套把 AI 協作、可驗證計算與互動展示連接起來的工程實作思路。

402
11:30
11:30
80min
Lunch

Lunch

403
11:30
80min
Lunch

Lunch

402
12:50
12:50
50min
Project Sharing

Project Sharing

403
13:50
13:50
30min
Werilog: Toward a Bidirectional Coding Architecture
Sivan

Werilog is a browser-based environment for Hardware Description Languages that removes the "local toolchain tax" of hardware design — no installs, no simulator setup, just coding in the browser.

At first glance, Werilog looks like another EDA tool: write Verilog, see it simulate. But the real question we're exploring is the opposite direction — can you draw a diagram and get working codes back out? Most tools only go one way: code in, diagram out, read-only. We're building toward a loop where the diagram is not just a visualization but an editable, valid representation of the code itself — and where structured, regular designs (like datapaths and control logic) can be constructed visually and compiled down to correct HDL.

This talk covers why we think bidirectionality matters for how people learn and design hardware, what we've built so far (live HDL-to-diagram visualization, in-browser simulation, an early diagram-to-HDL path for structural designs), the architectural choices and constraints that shaped it, and where the project is headed — including where AI assistance fits into closing the loop.

This talk covers why we think bidirectionality matters for how people learn and design hardware, what we've built so far (live HDL-to-diagram visualization, in-browser simulation, an early diagram-to-HDL path for structural designs), and the architectural choices and constraints that shaped it.

403
13:50
160min
體驗火箭導控 Balloon Popping Challenge
William Mou

小時候是不是也曾幻想過寫程式控制火箭升空?現在,你的 code 真的能主宰火箭的飛行軌跡!隨著臺灣入軌火箭計畫的推進,我們正徵招「導航、制導與控制(GNC)」的跨領域軟體人才,並分享國家太空中心(TASA)與陽明交通大學前瞻火箭研究中心(ARRC)共同推動的「火箭導控軟體競賽(Balloon Popping Challenge)」。

這是個結合 Gymnasium 訓練環境與 ActiveRocketPy 模擬系統(基於開源專案 RocketPy) 的火箭導控競賽。活動中會手把手帶領與會者嘗試控制火箭的推力向量控制(TVC)系統,並進一步探討如何應對陣風、感測器雜訊等複雜的真實情境。

「Be a User -> Be a Contributor」是這場工作坊的核心精神!我們非常鼓勵大家參與開源貢獻:無論你想發揮 Computer Science 背景幫忙打造飛行軌跡的資料視覺化工具,還是對物理學有熱忱,想發 PR 改善火箭 Python 模組的模擬特性,這裡都有你大展身手的舞台。

本工作坊無需事先報名,敬請攜帶筆電,一起成為太空開源專案的貢獻者!

402
14:30
14:30
30min
Shall We Test? BBGO 可測試交易策略開發
Dboy Liao

我們會在這場 talk 裡會以 bbgo 為基礎,希望以深入淺出的方式,介紹以下主題:
- 回測引擎的元件與開發
- 可測試交易策略開發
但不會包含:
- 交易策略相關指標介紹與優化

403
15:00
15:00
60min
Tea Break

Tea Break

403
16:00
16:00
30min
從 MicroGPT 到真實模型:拆解 LLM 原理,再用 LoRA 微調打造 Python 文件繁中翻譯 LLM
Timmy

大型語言模型對大多數開發者來說,常常是一個「輸入文字、吐出答案」的黑盒子。我們每天呼叫 API、串接 Chatbot,卻很少真正拆開來看:這些模型內部到底在算什麼?當我們說「微調一個模型」,實際上又是在調整什麼、可能會踩到哪些坑?這場分享想從這兩個角度,帶大家從「懂原理」走到「真的做得出來」。

這次分享主要會分成兩個段落:

第一段會用一個極簡版的 GPT 實作(MicroGPT),搭配一份刻意設計過的小型測試資料,現場展示一個語言模型從「什麼都不懂」到「學會特定規律」的完整過程。測試資料來自 python-docs-zh-tw 的術語列表,並刻意把原本的命名機器改成翻譯機器(例如把「access」對應到「存取」),讓資料規模小到可以在一般筆電上、幾分鐘內完成訓練,同時保留真實的語言對應關係。透過這個極簡模型,我們可以清楚看到:文字是怎麼被切成 token、token 怎麼變成向量、注意力機制怎麼決定「這個詞該參考哪些上下文」。因為模型夠小、資料夠少,整個過程是完全透明、可以逐行程式碼解釋的。

第二段會實際用開源預訓練模型,微調出一個「Python 技術文件英翻中」的專用翻譯模型。訓練資料完全來自 python/python-docs-zh-tw 這個開源翻譯專案:透過解析專案裡的 .po 翻譯檔案,把每一組已經由社群人工翻譯、審核過的英文原文與繁體中文譯文,轉換成訓練用的格式。

在模型選擇上,會分享實際嘗試過的幾個 Hugging Face 開源模型並使用 LoRA(透過 Unsloth 加速)這種參數高效微調技術,只調整模型裡一小部分參數,就能讓模型學會特定領域的翻譯能力,而不需要重新訓練整個模型;以及在模型訓練時實際踩過的各種坑。

403
16:50
16:50
10min
Closing

Closing

402