規則 Rules

第一屆 AI 教學挑戰賽：教學怪物挑戰賽 The 1st AI Teaching Challenge: Teaching Monster Challenge

1. 競賽宗旨 1. Competition Mission

「尋找最能根據學習需求，生成適性化影音教材的 AI Agent。」 "To find the AI Agent best capable of generating adaptive audiovisual educational materials based on learning needs."

隨著生成式人工智慧技術的快速演進，現有模型已展現出根據文字指令生成高品質影片的卓越能力。然而，單純的影像生成並不等同於具備有效的教學成效。教育影片的製作門檻遠高於一般影像創作，其核心價值不僅在於內容的正確性，更取決於系統是否具備「學科教學知識」。 With the rapid evolution of generative AI technology, existing models have demonstrated remarkable abilities to generate high-quality videos based on text instructions. However, simple image generation does not equate to effective teaching. The barrier for creating educational videos is much higher than general content creation, as its core value lies not only in accuracy but in the system's "Pedagogical Content Knowledge."

本競賽旨在挑戰人工智慧在教育領域的應用極限，核心任務在於驗證人工智慧系統是否能掌握人類教師的教學智慧，具體包含以下三大面向： This competition aims to challenge the limits of AI applications in education. The core task is to verify whether AI systems can master the pedagogical wisdom of human teachers, specifically including the following three dimensions:

  1. 知識結構搭建：運用「鷹架理論 」 將複雜概念進行邏輯拆解與結構化，建立新舊知識間的連結。 Knowledge Scaffolding: Applying "Scaffolding Theory" to logically deconstruct and structure complex concepts, building connections between old and new knowledge.
  2. 適性化教學：針對學習者的先備知識與背景差異，動態調整講解策略與語言深度，實現真正的「適性學習 」。 Adaptive Teaching: Dynamically adjusting explanation strategies and language depth according to the learner's prior knowledge and background differences to achieve true "Adaptive Learning."
  3. 認知投入維持：透過有效的教學敘事與多模態呈現，引導並維持學習者的專注力與「認知投入」。 Cognitive Engagement Maintenance: Guiding and maintaining the learner's focus and "Cognitive Engagement" through effective pedagogical narrative and multi-modal presentation.

我們期許透過此挑戰賽，尋求能深刻理解人類學習需求、並將其轉化為高品質教學內容的前瞻性人工智慧系統。 Through this challenge, we hope to seek forward-looking AI systems that can deeply understand human learning needs and transform them into high-quality educational content.

2. 競賽背景與範疇 2. Background and Scope

  • 核心任務：參賽團隊須開發一套全自動化之「人工智慧教學系統」，透過 API 接收學習需求（文字描述），並在完全無須人工介入的情況下，全自動生成對應的教學影片。 Core Task: Participating teams must develop a fully automated "AI Teaching System" that receives learning needs (text descriptions) via API and generates corresponding educational videos automatically without any human intervention.
  • 學科領域：中等教育階段之物理、生物、電腦科學、數學
    為確保競賽內容具備跨國界的通用性與客觀評測標準，本競賽將以美國「大學預修課程 (AP)」之知識作為參考基準。     Subject Domains: Secondary education level Physics, Biology, Computer Science, Mathematics.
    To ensure that the competition content has cross-border universality and objective evaluation standards, this competition will use the knowledge of the US "Advanced Placement (AP)" courses as a reference.
  • 語言要求：以英文 為主。 Language Requirement: Primarily in English.
  • 呈現形式： Presentation Format:
    • 時長上限：30 分鐘（由人工智慧系統視內容深度自行判斷最佳教學長度）。 Maximum Video Duration: 30 minutes (best length determined by the AI system based on content depth).
    • 風格：不拘，參賽者可自由發揮最佳教學形式，包含但不限於： Style: Flexible. Participants can choose the best pedagogical format, including but not limited to:
      • 動態簡報配合旁白：以結構化的圖文投影片搭配語音解說。 Dynamic Slides with Voiceover: Structured slides with voice explanation.
      • 虛擬講師導讀：運用數位虛擬化身模擬真人講師之授課互動。 Virtual Instructor: Using digital avatars to simulate interaction with a human instructor.
      • 視覺化動態演示：針對抽象數學公式或演算法流程，生成動態圖形或動畫進行直觀解構。 Visual Dynamic Demonstration: Generating dynamic graphics or animations for intuitive deconstruction of abstract formulas or algorithms.

3. 參賽技術規格 3. Technical Specifications

3.1 提交形式與系統架構 3.1 Submission Format and System Architecture

參賽者須採用「雙階段提交」模式： Participants must follow a "Two-Stage Submission" model:

  1. 即時推論介面（競賽期間）：參賽團隊需架設一組符合大會規範之應用程式介面（API）服務，供評測系統即時呼叫。 Real-time Inference Interface (During Competition): Participating teams must set up an API service following competition standards for real-time evaluation.
  2. 程式碼封裝與環境（得獎隊伍）：獲獎候選團隊須於賽後提交完整的系統程式碼映像檔（如 Docker 映像檔），以供主辦方在隔離環境中進行全自動化流程的複驗。 Code Packaging (Winning Teams): Candidate winning teams must submit a complete system image (e.g., Docker) after the competition for verification in an isolated environment.

3.2 系統介面 3.2 System Interface

所有資料傳輸均採用標準化的資料交換格式（JSON），並透過超文本傳輸協定（HTTP/HTTPS）進行溝通。 All data transmissions use standardized data exchange formats (JSON) and communicate via Hypertext Transfer Protocol (HTTP/HTTPS).

請注意參賽者的伺服器接收到生成請求後，請確保內容生成完成可供下載後再進行回應，否則平台接收到回應會馬上進行檔案下載導致評估為生成失敗。 Please note that after receiving the generation request, participants should ensure that the content is generated and available for download before responding. Otherwise, the platform will immediately download the file upon receiving the response, resulting in the generation failing the evaluation.

  • 輸入 (Input)：主辦方傳送 JSON 格式的「學習需求」，包含： Input: The organizer sends a JSON format "Learning Requirement," containing:
    • request_id （字串）：唯一識別碼，用於追蹤該次生成任務。 (String): Unique identifier used to track the generation task.
    • course_requirement （字串）：學習內容的具體指示與範圍。會包含具體的學習目標，以及期望涵蓋的關鍵知識或核心概念。 (String): Specific instructions and scope of the learning content. This will include specific learning objectives, as well as the key knowledge or core concepts expected to be covered.
    • student_persona （字串）: 學生背景自述（例如：「我是高中一年級學生，沒學過微積分，但懂基礎矩陣運算」）。 (String): Student background description (e.g., "I am a first-year high school student, haven't learned calculus, but understand basic matrix operations.")
  • 輸出 (Output)： Output:
    • video_url （字串）：生成影片的可下載直連網址。 (String): Downloadable direct link to the generated video.
    • subtitle_url （字串，optional）：同步字幕檔（SRT 或 VTT 格式），用於輔助內容正確性分析。 (String, optional): Synchronized subtitle file (SRT or VTT format) used to assist in content accuracy analysis.
    • supplementary_url （字串清單或字串，optional，也接受空清單）：輔助教材（講義、投影片）的下載連結。 (String list or string, optional, empty lists also accepted): Download links for auxiliary teaching materials (handouts, slides).

3.3 輸出媒體規格限制 3.3 Output Media Specifications

為確保評測一致性，生成之檔案須符合以下標準： To ensure evaluation consistency, generated files must meet the following standards:

  • 視訊格式：MP4 封裝。 Video Format: MP4 container.
  • 解析度：至少為高畫質（1280x720 像素）。 Resolution: At least High Definition (1280x720 pixels).
  • 音訊格式：取樣率不低於 16 kHz，需確保語音清晰可辨。 Audio Format: Sampling rate not lower than 16 kHz, ensuring clear and intelligible voice.
  • 檔案存取：下載連結必須在回傳後 48 小時內保持有效。 File Access: Download links must remain valid for 48 hours after being returned.
  • 檔案大小限制：影片檔案上限為 3GB，每部影片的輔助教材與字幕檔案上限為合計 100MB，且每部影片輔助教材數量不超過 5 個檔案。 File Size Limits: Max 3GB per video file. Combined limit of 100MB and 5 files for supplementary materials and subtitles per video.
  • 影片長度上限：30 分鐘 Maximum Video Duration: 30 minutes.

3.4 系統限制 3.4 System Constraints

  1. 運算時限：針對單一生成請求，系統必須在 30 分鐘內完成所有運算並回傳結果。若超過時限，系統將自動判定該題為失敗。 Computation Timeout: For a single generation request, the system must complete all computations and return the result within 30 minutes. If the time limit is exceeded, the system will automatically judge the item as a failure.
  2. 全自動化規範：從接收題目到產出影片的完整流程，必須由演算法自主完成。禁止任何形式的人工介入（如人工撰寫腳本、手動剪輯或配音）。主辦方將透過賽後的程式碼複驗來嚴格執行此規定。 Full Automation Policy: The entire process from receiving the topic to producing the video must be completed autonomously by the algorithm. Any form of human intervention (such as manually writing scripts, manual editing, or voice-over) is prohibited. The organizer will strictly enforce this rule through post-competition code verification.
  3. 外部工具與聯網權限： External Tools and Internet Permissions:
    • 允許系統進行即時網路搜尋以獲取最新資訊。 Allows the system to perform real-time web searches to obtain the latest information.
    • 允許呼叫第三方大型語言模型介面或使用程式碼直譯器。 Allows calling third-party Large Language Model interfaces or using code interpreters.
    • 透明度要求：若使用外部付費服務或模型，必須在系統文件中詳細揭露，並確保該服務在複驗階段亦可被存取。 Transparency Requirement: If external paid services or models are used, they must be detailed in the system documentation, ensuring the service remains accessible during the verification phase.
  4. 由於教學內容強調正確性，系統若使用了網路上的第三方素材（圖片、圖表、數據），必須遵守以下規範： Since teaching content emphasizes accuracy, if the system uses third-party materials from the internet (images, charts, data), it must comply with the following standards:
    1. 來源標示：在影片畫面中或輔助教材中，清楚標示資料來源。 Source Attribution: Clearly mark data sources within the video frame or auxiliary materials.
    2. 幻覺檢核：嚴禁捏造不存在的論文、案例或數據。 Hallucination Check: Strictly forbid fabricating non-existent papers, cases, or data.
    3. 版權合規：優先使用公開授權（CC授權）或由人工智慧原創生成的素材。若發生版權爭議，將由參賽團隊自行承擔法律責任。 Copyright Compliance: Prioritize the use of public licenses (CC licenses) or materials originally generated by AI. In case of copyright disputes, the participating team will bear legal responsibility.

4. 評分標準 4. Judging Criteria

此標準將適用於「熱身賽 (AI 評分)」與「正式賽 (人類評分)」。 These standards apply to both the "Warm-up (AI Scoring)" and the "Official Competition (Human Scoring)."

(1) 內容正確性與實證基礎 (1) Content Accuracy and Empirical Foundation

此為教學之核心底線。 This is the core baseline of teaching.

  • 零幻覺原則：嚴格檢視所有技術術語、數學推導過程、歷史數據與程式碼範例。任何似是而非、誤導性的虛構資訊將導致該維度直接判定為失敗。 Zero Hallucination Principle: Strictly review all technical terms, mathematical derivations, historical data, and code examples. Any plausible but misleading fictional information will result in a direct failure for this dimension.
  • 知識深度與驗證：評估內容是否觸及核心原理，而非僅流於表面名詞堆砌。引用的學術案例或論文必須真實存在且與主題高度相關。 Knowledge Depth and Verification: Evaluate whether the content touches on core principles rather than just superficial jargon. Cited academic cases or papers must exist and be highly relevant to the topic.
  • 引用規範性：檢視外部資料的引用是否準確、合乎學術倫理，且無抄襲嫌疑。不得違反著作權相關之法律。 Citation Standard: Review whether citations of external data are accurate, ethically sound, and free from plagiarism suspicions. Must not violate copyright laws.

(2) 教學邏輯與架構流暢度 (2) Pedagogical Logic and Structural Fluency

評估知識傳遞的結構化程度。 Evaluating the degree of structural organization in knowledge delivery.

  • 教學鷹架搭建：檢視影片是否遵循「由淺入深」的原則。是否從學生已知的概念出發，逐步搭建通往新知識的橋樑，而非突兀地跳躍至進階概念。 Pedagogical Scaffolding: Checking whether the video follows the principle of "from simple to deep." Does it start from concepts already known to the student and gradually build bridges to new knowledge, rather than jumping abruptly to advanced concepts?
  • 敘事連貫性：段落間的轉折是否自然流暢？知識點之間是否環環相扣，形成完整的邏輯鏈條，而非破碎的資訊拼湊。 Narrative Cohesion: Are transitions between segments natural and smooth? Are knowledge points interconnected to form a complete logical chain, rather than a fragmented patchwork of information?

(3) 學習者需求適配度 (3) Learner Needs Adaptation

評估系統對「學生畫像」的理解與回應能力。 Evaluating the system's ability to understand and respond to the "Student Persona."

  • 近側發展區識別：人工智慧是否精準識別學生的知識邊界？（例如：面對僅懂代數的高中生，系統應避免直接使用未經解釋的偏微分方程，而改用直觀的幾何圖解。） Zone of Proximal Development Identification: Does the AI precisely identify the student's knowledge boundaries? (e.g., facing a high schooler who only knows algebra, the system should avoid unexplained partial differential equations and instead use intuitive geometric diagrams.)
  • 語言與譬喻調整：解說的語言風格、用詞深淺及舉例是否符合該目標受眾的生活經驗與認知水準。 Language and Analogy Adjustment: Whether the explanatory language style, vocabulary depth, and examples match the target audience's life experience and cognitive level.

(4) 認知投入與多模態呈現 (4) Cognitive Engagement and Multi-modal Presentation

評估教材的吸引力與傳達效率。 Evaluating the appeal and delivery efficiency of the educational materials.

  • 敘事魅力：是否能將生硬的知識融入生動的案例或比喻中，引發學習動機，讓觀眾產生「想繼續看下去」的內在驅動力。 Narrative Charm: Can stiff knowledge be integrated into vivid cases or analogies to trigger learning motivation, giving the audience an intrinsic drive to "keep watching"?
  • 認知專注力維持：影片是否運用有效的教學策略（如：懸念設計、蘇格拉底式提問、階段性小結）來維持學習者的高強度注意力，避免枯燥的單向輸出。 Cognitive Focus Maintenance: Does the video employ effective teaching strategies (e.g., suspense design, Socratic questioning, periodic summaries) to maintain the learner's high-intensity attention, avoiding dull one-way output.
  • 多模態一致性：視覺畫面（圖表、動畫、游標軌跡）與聽覺解說是否同步且互相增強？畫面不應僅是裝飾，而應能具體輔助抽象概念的理解（例如：講到矩陣旋轉時，畫面應同步呈現旋轉動畫）。 Multi-modal Consistency: Are visual elements (charts, animations, cursor tracks) and auditory explanations synchronized and mutually reinforcing? Visuals should not be mere decoration but should specifically aid in understanding abstract concepts (e.g., when discussing matrix rotation, the screen should simultaneously show rotation animation).

5. 賽制流程 5. Competition Process

第一階段：熱身賽 Phase 1: Warm-up

本階段旨在協助參賽團隊進行系統壓力測試、熟悉 API 串接流程，並透過 AI 評審的回饋優化系統。熱身賽將分為兩輪進行： This phase aims to assist participating teams in system stress testing, familiarizing themselves with API integration processes, and optimizing systems through feedback from AI reviewers. The warm-up phase will be divided into two rounds:

  • 第一輪：3 月 1 日開始 Round 1: Starting March 1st

    • 題目來源：由主辦方透過語言模型（LLM）自動生成的學習需求。 Topic Source: Learning requirements automatically generated by the organizer through Large Language Models (LLMs).
    • 目的：提供大量樣本供參賽團隊進行大規模壓力測試與 API 穩定性驗證。 Purpose: To provide a large number of samples for participating teams to conduct large-scale stress testing and API stability verification.

  • 第二輪：4 月 14 日開始 (共 32 題) Round 2: Starting in April 14th (There are 32 questions in total.)

    • 題目來源：由評審委員親自設計的模擬學習需求。 Topic Source: Simulated learning requirements personally designed by the judging committee.
    • 目的：提供更具挑戰性、貼近人類教師需求的題目，作為初賽前的「模擬考」，讓團隊調整 Agent 對於複雜指令的理解能力。 Purpose: To provide more challenging topics that align closely with human teacher needs, serving as a "mock exam" before the preliminary round, allowing teams to adjust their Agents' ability to understand complex instructions.

  • 形式： Format:

    • 主辦方提供學習需求。 The organizer provides learning requirements.
    • 參賽 API 需針對每一需求產生影片。 The participating API must generate a video for each requirement.

  • 評審機制：AI 學生 Evaluation Mechanism: AI Student

    • 由主辦方特製的 LLM (AI 學生) 擔任評審。 A specially developed LLM (AI Student) by the organizer will act as the judge.
    • AI 學生將嚴格根據上述四大維度，提供詳細的文字回饋與分數。 The AI Student will provide detailed textual feedback and score based strictly on the four aforementioned dimensions.
    • 此階段分數僅供團隊參考優化模型，不計入最終排名。 Scores in this phase are only for teams to refer to for model optimization and will not count towards the final ranking.
    • AI 學生使用上限由主辦方訂定，並保留按系統負荷調整的權限。 The maximum usage limit for AI Student is set by the organizer, who reserves the right to adjust it according to system load.

第二階段：初賽 (第一部分 16 題（5/1）、第二部分 16 題（5/12）) Phase 2: Preliminary Round (Part 1: 16 questions (5/1) Part 2: 16 questions (5/12))

  • 時間：5 月 1 日 - 5 月 15 日 Time: May 1st - May 15th
  • 目的：篩選出具備實戰能力的優秀 Agent。 Purpose: To screen for outstanding Agents with practical capabilities.

  • 形式： Format:

    • 主辦方釋出評審委員設計的學習需求。 The organizer releases learning requirements designed by the judging committee.
    • 參賽 API 於時限內回傳影片。 The participating API returns videos within the time limit.

  • 評審機制：兩階段篩選 Evaluation Mechanism: Two-stage screening

    （一）AI 學生自動評測（初篩） (I) AI-powered student automatic assessment (initial screening)

    • 所有參賽隊伍生成的教學影片，先由 AI 學生（LLM）依據四大評分標準（內容正確性、教學邏輯、學習者適配、認知投入）進行自動化評分。 All the instructional videos generated by the participating teams will be first automatically scored by the AI ​​student (LLM) based on four scoring criteria (content accuracy, instructional logic, learner suitability, and cognitive engagement).
    • 依據 AI 學生評分排名，選出最多 10 名 隊伍進入真人評審階段。 Based on the AI ​​student rating rankings, a maximum of 10 teams will be selected to enter the human review stage.

    （二）真人評分競技場 (II) Real-person rating arena

    • 評審團：透過 Prolific 平台招募的真人評測者，以盲測方式進行評比。 Judging Panel: The evaluation was conducted by real reviewers recruited through the Prolific platform using a blind testing method.

    • 評分模式： Scoring Mode:

      • 評測介面將左右並列播放針對「同一題目」由「兩個不同系統」生成的影片（影片 A vs. 影片 B）。 The evaluation interface will play videos generated by "two different systems" for the "same topic" side-by-side (Video A vs. Video B).
      • 評審判斷哪一個影片教學效果更好。 Judges determine which video has better teaching effectiveness.
      • 評審將採用與 AI 學生一致的四大評分標準進行判斷。 Judges will make judgments using the same four scoring criteria as the AI Student.
      • 系統會計算出每個參賽系統的潛在實力分數（Elo 等級分）。 The system will calculate the potential strength score (Elo rating) for each participating system.
    • 晉級：入選隊伍的影片交由透過 Prolific 平台招募的真人評測者，以 Arena 盲測方式進行 pairwise 比較。系統依據 Elo 積分排名，選出前三名隊伍進入決賽。 Advancement: The selected teams' videos were compared pairwise by real reviewers recruited through the Prolific platform using a blind Arena test. The system then ranked the top three teams based on their Elo ratings to advance to the finals.
    • 每個隊伍在初賽期間只能提交一個模型產出的影片進行評比。初賽結束（5/15 23:59:59 AoE）後不可再提交新影片。排名與獎項皆以隊伍為單位。 Each team may submit only one video of their model for evaluation during the preliminary round. No new videos may be submitted after the preliminary round ends (May 15, 23:59:59 AoE). Rankings and awards are based on teams.

第三階段：決賽 (共 16 題) Phase 3: Finals (16 questions)

  • 時間：6 月 12 日 - 6 月 13 日 Time: June 12th - June 13th
  • 目的：由頂尖專家驗證 AI 的教學極限。 Purpose: Top experts verify the teaching limits of AI.
  • 評審團：由國高中教師、校長、大學教授組成。 Judging Panel: Composed of junior and senior high school teachers, principals, and university professors.

  • 形式： Format:

    • 專家出題：每位評審提出數個具挑戰性的學習需求，共 16 題。 Expert Topic Design: Each judge proposes several challenging learning requirements, there are 16 questions in total.
    • 限時挑戰：所有題目集合後傳送給晉級的三套系統，系統需在限定時間內（依據 30 分鐘/題的原則）回傳所有影片。 Time-limited Challenge: All topics are collected and sent to the three advanced systems, which must return all videos within the specified time (based on the principle of 30 minutes/topic).
    • 每題回應時限 30 分鐘，逾時該題判定未完成；測試次數無上限；允許重傳（以最後一次提交為準）。 There is a 30-minute time limit for responding to a question; if the time limit is exceeded, the question will be considered incomplete. There is no limit to the number of tests; resubmission is allowed (based on the last submission).

  • 評選流程： Selection Process:

    • 評審團觀看所有生成影片。 The judging panel watches all generated videos.
    • 進行綜合討論與獨立排名。 Comprehensive discussion and independent rankings are conducted.
    • 決勝：每位評審各自對一部分影片進行排名（1-3 名），以所有排名的平均值決定最終名次。平均排名最低者為冠軍。 Winning: Each judge ranks a selection of films (1st-3rd place), and the final ranking is determined by the average of all rankings. The film with the lowest average ranking is the champion.
  • 結尾：舉辦線上工作坊進行頒獎與技術分享。 Conclusion: An online workshop will be held for awards and technical sharing.

6. 獎項與榮譽 6. Awards and Honors

所有晉級決賽並完成展示的團隊，將獲得由主辦方頒發之官方證書。 All teams that advance to the finals and complete their presentation will receive official certificates issued by the organizer.

  • 金獎：總排名前 1 名之團隊，獲頒金獎證書。 Gold Medal: The team ranked 1st overall will be awarded a Gold Medal certificate.
  • 銀獎：總排名第 2-3 名之團隊，獲頒銀獎證書。 Silver Medal: Teams ranked 2nd-3rd overall will be awarded Silver Medal certificates.

7. 比賽時程表 (時區：Any where on Earth, AoE) 7. Competition Schedule (Timezone: Anywhere on Earth, AoE)

以競賽首頁之時程表為準 Please refer to the timeline on the competition's homepage.

8. 主辦單位與聯繫方式 8. Organizer and Contact Information

  • 主辦單位：台灣大學人工智慧頂尖研究中心 (National Taiwan University AI Center of Research Excellence, NTU AI-CoRE) Organizer: National Taiwan University AI Center of Research Excellence (NTU AI-CoRE)
  • Email Group：[email protected] Email Group: [email protected]
返回首頁 Return Home