夜雨聆风学习资料网

ARTICLE · 1387385

Potato IOAI2026真题解读 Task3:只知道哪个词更接近答案,怎样完成三十轮猜词.

Potato IOAI2026真题解读 Task3:只知道哪个词更接近答案,怎样完成三十轮猜词.

魔都IOAI.NOAI · 国际真题逐段解读 03

PotatoIOAI2026真题解读 Task3只知道哪个词更接近答案,怎样完成三十轮猜词

IOAI是International Olympiad in Artificial Intelligence,即国际人工智能奥林匹克,面向中学生开展人工智能学术挑战。2026年比赛在哈萨克斯坦阿斯塔纳举行,现场个人赛分两天,每天三道任务,共六题。

本篇为个人赛第3题,也是第1天的第3题Potato。程序只能通过词语间的语义远近反馈寻找隐藏词,必须在30轮内猜中确切单词。它是一道逐轮交互题,反馈会影响下一次提交的词。

01 运行环境与预算

题目描述

- Time limit:10 minutes

- Environment:one GPU (≈16 GB VRAM), no internet

- Solution size:solution.ipynb≤ 1 MB

- Storage:5 GB

解读

整个程序运行限10分钟,不能联网,使用约16 GB显存的GPU,存储5 GB,提交Notebook不超过1 MB。准备与交互都要在运行预算内完成。

02 任务:比较语义远近,猜中隐藏词

题目描述

Task

Your friend suggests playing a guessing game. He, as the judge, picks one hidden word from a fixed vocabulary, and you must find it in at most 30 turns. Each turn the judge compares two words and reports which is semantically closer to the hidden word. Every game starts from the fixed pair lamp vs potato, because they are two of your friend's favorite things. Your program then proposes one new word. The winner of the comparison is kept and compared against your next proposal. You win a game the moment you propose the hidden word exactly. Matching is case-insensitive. Every word you propose must be in dataset/vocabulary.json.

There is a full example in solution.ipynbwith protocol and data loading. You can change the PublicEmbeddingPlayer class. Your program is initialized once and plays every game in a single run; the protocol creates a fresh PublicEmbeddingPlayer at the start of each game.

解读

每局从lamp与potato的固定比较开始,程序提出新词,裁判保留更接近隐藏词的一方进入下一轮。胜利要求精确命中隐藏词,大小写不敏感;只猜到近义词不算成功。所有提议必须来自给定词表。程序在一次运行中完成所有游戏,每局创建新的PublicEmbeddingPlayer对象。

03 官方交互示例与反馈含义

题目描述

The Judge

Your program sends one JSON object to the Judge and the Judge responds with one JSON object.

A worked example, with the hidden word shown only to explain the protocol:

Hidden word: shovel          Fixed opening: lamp vs potato  <- {"turn": 1, "winner_word": "potato", "verdict": "second", "word1": "lamp",   "word2": "potato"} -> {"new_word": "rock"} <- {"turn": 2, "winner_word": "rock",   "verdict": "second", "word1": "potato", "word2": "rock"} -> {"new_word": "hammer"} <- {"turn": 3, "winner_word": "hammer", "verdict": "second", "word1": "rock",   "word2": "hammer"} -> {"new_word": "shovel"}                                    <- matches: game won -> {"status": "win"}

Turns are indexed from 1 to 30.

The verdictoptions are firstmeaning that word1 is closer, secondmeaning that word2 is closer or samemeaning that both words are equally close to the hidden word.

winner_wordis the word retained for the next comparison. On a sameverdict, the first word stays.

解读

官方示例把shovel作为隐藏词,演示从potato到rock、hammer,再猜中shovel的交互。示例展示答案是为了说明协议,正式运行不会直接告诉程序隐藏词。first、second、same是比较结果;same时保留第一个词,winner_word是下一轮继续参加比较的词。原示例末尾status行的箭头按题面保留,实际接口调用应对照官方Notebook,不自行改写协议。

04 词表、公开表示与私有裁判

题目描述

Dataset

Shared by every split:

- dataset/vocabulary.json— 1602 unique lowercase words. The hidden word is always one of these.

- dataset/public_embeddings.npy— float32, shape (1602, 2560). Row icorresponds to word iin the vocabulary. These are publicembeddings; the judge uses a different, private representation.

The splits are sets of hidden words:

Split
Words
Answers
Use it to
test_public
120
✅ dataset/test_public.json
run your solution and self-score
test_leaderboard_a
120
hidden
live leaderboard
test_leaderboard_b
120
hidden
final ranking

There is no trainsplit — nothing is fitted from labelled rows.

解读

词表有1,602个小写词,公开向量数组的行与词表逐项对应,每行2,560维。裁判使用另一套私有表示,公开相似度与真实反馈不能视为完全相同。三组各120个隐藏词;没有普通的带标签训练集,公开测试用来运行和自测。

05 允许模型与本地加载

题目描述

Provided models

Two pretrained embedding models ship with the task and may be used:

models/Qwen/Qwen3-Embedding-0.6B [Qwen Embedding model card](https://yastatic.net/s3/contest/ioai/3/01_Qwen_Qwen3-Embedding-0.6B_MODEL_CARD.html) models/BAAI/bge-m3 [bge-m3 model card](https://yastatic.net/s3/contest/ioai/3/02_BAAI_bge-m3_MODEL_CARD.html)

Both must be loaded from their local path; a Hugging Face hub id such as "BAAI/bge-m3"triggers a download and fails, because judging is offline. Each directory contains a runnable example.pyshowing the offline call.

Available libraries: numpy, torch, sentence-transformers. No internet, no downloads, no other packages.

解读

可用Qwen3-Embedding-0.6B和bge-m3,必须从题目提供的本地目录加载。使用在线模型ID可能触发下载,离线评测会失败。可用库被限定为numpy、torch和sentence-transformers,各目录有离线调用示例。

06 通过标准输入输出交互

题目描述

Output

None. This is an interactive task: your solution writes no answer file; it talks to the judge over stdin/stdout as described above.

解读

本题不生成答案文件。程序通过stdin/stdout逐轮与裁判交换JSON,这与其他题一次性写预测文件的形式不同。输出协议本身就是提交程序必须正确完成的部分。

07 成功轮次与得分

题目描述

Metric

A game found on turn tscores 1.0 - 0.02 × max(0, t - 10); a game not solved within 30 turns scores 0. So turns 1–10 score 1.00, turn 20 scores 0.80, turn 30 scores 0.60.

Your task score is the mean game score × 100, between 0.00and 100.00.

The 10-minute limit is a single budget covering start-up, preparation and all 120 games in the test set.

解读

第1—10轮猜中均得1.00,之后每轮递减0.02;官方给出的第20轮得0.80、第30轮得0.60,30轮内未猜中得0。各局平均再乘100。10分钟包括启动、准备和全部120局,不是每局各有10分钟。

08 修改玩家类并提交Notebook

题目描述

How to submit

1. Open solution.ipynb, edit PublicEmbeddingPlayer, and run all cells to make sure it's working.

2. Optionally, check it locally: python local_test.py solution.ipynb --limit 5. The local judge uses the publicembeddings, so its score is only a guide.

3. Save solution.ipynb.

4. Open the Git tab in the left sidebar of JupyterLab.

5. Stage solution.ipynb(the +icon next to it).

6. Enter a commit message and click Commit.

7. Click the cloud-with-up-arrow to push.

8. Return to this Contest page and click Submit, with the commit message matching the one you've provided.

Submit exactly one file, named solution.ipynb, covering any necessary preparations and inference.

解读

修改PublicEmbeddingPlayer并检查完整运行。官方local_test命令可做本地检查,但本地裁判使用公开向量,其分数只作参考。提交文件必须是solution.ipynb,包含需要的准备与推理步骤,按规定的Git与比赛页面流程提交。

题面来源:IOAI 2026官方个人赛 Potato。题目描述保留英文原文,公式以等价数学字符排版;插图使用本地保存的官方图片。

09 IOAI 2026个人赛金牌得主

以下为官方个人赛全部37位金牌得主,按六题总分排名排列。“本题得分”为第3题Potato的成绩,不是本题单项排名。姓名、代表队及四位小数成绩沿用官方记录。

总排名
选手 / 代表队
本题得分
六题总分
1
Artem GorokhovRussia 1 🇷🇺
85.5130
428.4866
2
Dauzhan BeketovKazakhstan 1 🇰🇿
68.0383
405.5759
3
Shao ZixiChina 1 🇨🇳
66.9955
375.4507
4
Michał MasnyPoland 2 🇵🇱
0.0000
370.0192
5
Kirill LabzinRussia 1 🇷🇺
65.0225
365.7619
6
Anango Dev PrabhatUnited Kingdom 1 🇬🇧
79.1995
350.4524
7
Ian BobrusKyrgyzstan 1 🇰🇬
45.0958
338.2170
8
Nguyen Viet Trung NhanVietnam 1 🇻🇳
22.1815
334.4282
9
Cheng Lai YinHong Kong 1 🇭🇰
66.7700
327.0589
10
Shi ZhenghangChina 1 🇨🇳
32.6099
313.5448
11
Krzysztof RojekPoland 1 🇵🇱
59.3010
312.9109
12
Luvidi Pranawa AlghariIndonesia 1 🇮🇩
0.0000
311.7292
13
Samik GoyalIndia 1 🇮🇳
39.2334
307.8219
14
Mikhail VershininRussia 2 🇷🇺
11.6685
306.8896
15
Nguyen Huu TuanVietnam 1 🇻🇳
52.3957
296.1471
16
Feodor YevtushenkoUnited States 1 🇺🇸
75.6764
295.3775
17
Karl Teo LovmarSweden 1 🇸🇪
57.2717
292.4444
18
Nikita PudovkinRussia 2 🇷🇺
50.5073
291.8335
19
Thury-Burileanu AlexandruRomania 1 🇷🇴
75.0846
287.6469
20
Petrean RolandRomania 1 🇷🇴
0.0000
282.2468
21
Semen RodionovRussia 2 🇷🇺
33.7373
278.9932
22
Hong ZeChina 1 🇨🇳
36.4149
271.3354
23
Alexander PovarovRussia 2 🇷🇺
0.0000
269.5110
24
Xi FanglongChina 2 🇨🇳
41.4882
268.9771
25
Shayan RezazadehIran 1 🇮🇷
51.4938
267.8067
26
Nikoloz GegenavaGeorgia 1 🇬🇪
11.6685
257.5146
27
Elisei KirpichenkoRussia 1 🇷🇺
0.0000
257.4105
28
Zane KumarUnited Kingdom 1 🇬🇧
0.0000
251.5938
29
Haoxuan ZhangSweden 1 🇸🇪
33.8782
251.3079
30
Zhang RenzeChina 1 🇨🇳
3.8613
250.2285
31
Zhang Jerry LiUnited States 2 🇺🇸
0.0000
249.6299
32
Mikołaj ŹrałekPoland 2 🇵🇱
17.1928
246.3154
33
Sultanbi IssatayKazakhstan 1 🇰🇿
26.9729
241.8351
34
Tomasz KuczkoPoland 1 🇵🇱
0.0000
240.2035
35
Giorgi MaisuradzeGeorgia 1 🇬🇪
31.9899
233.3991
36
Georgescu DavidRomania 1 🇷🇴
32.8072
228.9866
37
Wang Enrun MatthiasSingapore 2 🇸🇬
35.5693
222.5052

来源:IOAI 2026官方个人赛成绩。核对日期:2026年11月3日。

关注「魔都IOAI.NOAI」

专注上海青少年人工智能学习与备考项目解读 · 学习路线 · 实践训练 · 活动资讯

关注后,在公众号对话框回复 「手册」领取《IOAI中国区-NOAI-项目手册》

看清学习路径,练好AI本领。

回首页,搜索你想要的文档

在首页搜索框输入关键词,快速定位资料

文库首页搜索指引 立即搜索

相关学习资料