Upload docs articles and media
Browse filesThis view is limited to 50 files because it contains too many changes. See raw diff
- .gitattributes +1 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.json +5 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.md +64 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_long.md +16 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_long_clean.txt +15 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_middle.md +12 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_middle_clean.txt +11 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_short.md +22 -0
- docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_short_clean.txt +9 -0
- docs/AI_Agents_and_Logic/ai_formal_logic_survey.md +497 -0
- docs/AI_Agents_and_Logic/ai_formal_logic_survey.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/ai_proof_search_tools_survey.md +670 -0
- docs/AI_Agents_and_Logic/ai_proof_search_tools_survey.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/neurosymbolic_ai_survey.md +719 -0
- docs/AI_Agents_and_Logic/neurosymbolic_ai_survey.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/script_long.md +30 -0
- docs/AI_Agents_and_Logic/script_long.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/script_long_clean.txt +29 -0
- docs/AI_Agents_and_Logic/script_middle.md +19 -0
- docs/AI_Agents_and_Logic/script_middle.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/script_middle_clean.txt +19 -0
- docs/AI_Agents_and_Logic/script_short.md +12 -0
- docs/AI_Agents_and_Logic/script_short.rag_ingested +1 -0
- docs/AI_Agents_and_Logic/script_short_clean.txt +11 -0
- docs/AI_Computer_Vision/VFI_Research_Report.md +163 -0
- docs/AI_Computer_Vision/VFI_Research_Report.rag_ingested +1 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405.json +5 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405.md +3 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_long.md +33 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_long_clean.txt +35 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_middle.md +6 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_middle_clean.txt +7 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_short.md +4 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_short_clean.txt +3 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-2026-04-05.md +363 -0
- docs/AI_Computer_Vision/arxiv-cs-cv-2026-04-05.rag_ingested +1 -0
- docs/AI_Computer_Vision/arxiv_video_analysis_papers.md +74 -0
- docs/AI_Computer_Vision/arxiv_video_analysis_papers.rag_ingested +1 -0
- docs/AI_Computer_Vision/layout_models_survey.md +370 -0
- docs/AI_Computer_Vision/layout_models_survey.rag_ingested +1 -0
- docs/AI_Computer_Vision/script_long.md +26 -0
- docs/AI_Computer_Vision/script_long.rag_ingested +1 -0
- docs/AI_Computer_Vision/script_long_clean.txt +23 -0
- docs/AI_Computer_Vision/script_middle.md +12 -0
- docs/AI_Computer_Vision/script_middle.rag_ingested +1 -0
- docs/AI_Computer_Vision/script_middle_clean.txt +11 -0
- docs/AI_Computer_Vision/script_short.md +6 -0
- docs/AI_Computer_Vision/script_short.rag_ingested +1 -0
- docs/AI_Computer_Vision/script_short_clean.txt +5 -0
.gitattributes
CHANGED
|
@@ -58,3 +58,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
|
|
|
|
|
| 58 |
# Video files - compressed
|
| 59 |
*.mp4 filter=lfs diff=lfs merge=lfs -text
|
| 60 |
*.webm filter=lfs diff=lfs merge=lfs -text
|
| 61 |
+
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405.md filter=lfs diff=lfs merge=lfs -text
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"importance": 7,
|
| 3 |
+
"urgency": 7,
|
| 4 |
+
"tier": 2
|
| 5 |
+
}
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.md
ADDED
|
@@ -0,0 +1,64 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# 次世代エージェントAI:ベンチマーク、OSS動向、およびビデオ生成エージェントの進化 (2026)
|
| 2 |
+
|
| 3 |
+
## 概要
|
| 4 |
+
本レポートは、2026年5月時点におけるエージェント型AI(Agentic AI)の主要な技術動向をまとめたものである。Langflow等のオーケストレーションツールで重要視されるベンチマーク指標、台頭するOSSモデル、およびHeyGenに代表されるビデオ生成エージェントの最新技術を統合的に分析する。
|
| 5 |
+
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
## 1. エージェント評価指標の転換:Agentic Benchmarks
|
| 9 |
+
2026年現在、AIモデルの評価軸は「会話の流暢さ」から「実務実行能力」へと完全にシフトしている。
|
| 10 |
+
|
| 11 |
+
### 1.1 主要ベンチマーク
|
| 12 |
+
* **BFCL v4 (Berkeley Function Calling Leaderboard)**:
|
| 13 |
+
* **評価対象**: 関数呼び出し(Tool Use)の精度。
|
| 14 |
+
* **トレンド**: 単一の呼び出しから、依存関係のある複数のツールを順序立てて実行する能力が重視されている。
|
| 15 |
+
* **SWE-bench Pro**:
|
| 16 |
+
* **評価対象**: 実際のソフトウェア開発リポジトリにおけるIssue解決能力。
|
| 17 |
+
* **トレンド**: Python以外の言語(Go, TypeScript等)への拡大と、自律的なテスト・デバッグサイクルの完遂が評価の要。
|
| 18 |
+
* **WebArena / GAIA**:
|
| 19 |
+
* **評価対象**: ブラウザ操作を伴う複雑なタスクの完遂率。
|
| 20 |
+
* **トレンド**: 人間の曖昧な指示から具体的な実行ステップを生成する「プランニング能力」の差が顕著になっている。
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## 2. 2026年5月時点の主要OSSモデル(エージェント特化型)
|
| 25 |
+
プロプライエタリなモデル(Claude Mythos, GPT-5.3)に比肩する能力を持つOSS/オープンウェイトモデルの選定。
|
| 26 |
+
|
| 27 |
+
| モデル名 | 主要な強み | ベンチマーク指標 (参考) | 特徴 |
|
| 28 |
+
| :--- | :--- | :--- | :--- |
|
| 29 |
+
| **Qwen3-235B (A22B)** | 汎用オーケストレーション | BFCL v3: 74.9% | ツール利用の安定性が高く、Langflow等のワークフロー構築に最適。 |
|
| 30 |
+
| **DeepSeek-V4 Pro** | コーディング / 推論 | SWE-bench: 76.4% | 論理的思考が必要なエージェントタスクにおいて圧倒的な勝率。 |
|
| 31 |
+
| **GLM-5 / 4.5** | 信頼性・低ハルシネーション | BFCL v3: 76.7% | 厳密な出力フォーマットが求められる業務自動化に強い。 |
|
| 32 |
+
| **Llama-4 Agent Edition** | 自律プランニング | - | MetaによるAgent特化版。CoT(思考の連鎖)の質が高い。 |
|
| 33 |
+
|
| 34 |
+
---
|
| 35 |
+
|
| 36 |
+
## 3. ビデオ生成エージェントの革新 (HeyGen Avatar V)
|
| 37 |
+
ビデオ生成分野においても、単なるピクセル生成から「クリエイティブ・エージェント」への進化が見られる。
|
| 38 |
+
|
| 39 |
+
### 3.1 技術的ブレイクスルー
|
| 40 |
+
* **Avatar V Architecture**:
|
| 41 |
+
* **Flow Matching & DiT**: ボトルネックを排除し、リファレンスビデオの全トークンに直接条件付けを行うことで、2025年以前のモデルでは困難だった微細な表情の変化を完全再現。
|
| 42 |
+
* **Agentic Kernel Synthesis**:
|
| 43 |
+
* HeyGenのインファレンス最適化レポートによれば、LLMベースのエージェントが**カスタムカーネルをリアルタイムで合成**し、巨大モデルの低遅延推論を実現している。
|
| 44 |
+
|
| 45 |
+
### 3.2 制作フローの変革
|
| 46 |
+
* **Prompt-to-Blueprint**:
|
| 47 |
+
* 直接動画を出すのではなく、まずエージェントが「構成案(Blueprint)」を提示し、ユーザーが対話形式で修正を加えるフローが定着。これにより「一発勝負」の生成から「意図を反映した制作」へと移行した。
|
| 48 |
+
|
| 49 |
+
---
|
| 50 |
+
|
| 51 |
+
## 4. 今後の展望:Decision-Centric AI
|
| 52 |
+
2026年後半に向けて、エージェントAIは以下の方向に進化すると予測される。
|
| 53 |
+
|
| 54 |
+
1. **意思決定中心(Decision-Centric)の評価**: 「何を言ったか」ではなく「どの意思決定が最適だったか」がAIの価値となる。
|
| 55 |
+
2. **Harness Engineeringの標準化**: OpenAIが提唱する、モデルを制御・監視するためのコードレイヤー(ハーネス)の設計が、エージェント開発の必須スキルとなる。
|
| 56 |
+
3. **マルチエージェント・コラボレーション**: 単一の巨大モデルではなく、Qwen3(司令塔)とDeepSeek-V4(実装)のような、役割分担されたOSSモデル群による協調動作が主流となる。
|
| 57 |
+
|
| 58 |
+
---
|
| 59 |
+
*作成日: 2026年5月16日*
|
| 60 |
+
*出典: Notion (Langflow/Benchmarks), HeyGen Technical Reports (2026), Berkeley BFCL Leaderboard*
|
| 61 |
+
|
| 62 |
+
|
| 63 |
+
**Scores:** Importance:7, Urgency:7
|
| 64 |
+
**Tier:** 2
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "a8eac8bc37296199", "ingested_at": "2026-06-30T08:25:41.762466", "content_hash": "a8eac8bc37296199"}
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_long.md
ADDED
|
@@ -0,0 +1,16 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight, we turn to the rapidly evolving world of artificial intelligence and its impact on our future workforces. A new report sheds light on how AI is transforming the way we approach tasks and solve problems.
|
| 2 |
+
|
| 3 |
+
The report dives deep into the "Agentic AI" landscape, focusing on the shift from conversational fluency to practical execution capabilities. AI models are now being evaluated based not just on their ability to hold a conversation, but also on their capacity to perform real-world tasks efficiently. The report highlights key benchmarks like the Berkeley Function Calling Leaderboard (BFCL) and SWE-bench Pro, which measure an AI's performance in software development and issue resolution.
|
| 4 |
+
|
| 5 |
+
The study reveals that these benchmarks are shifting towards a more practical focus, emphasizing the ability of AI models to execute complex tasks autonomously. This transition is driven by advancements in tools like Langflow, which allows for the orchestration of various AI models based on specific workflows.
|
| 6 |
+
|
| 7 |
+
The report also explores the rise of open-source (OSS) models that rival proprietary models like Claude Mythos and GPT-5.3. These OSS models are gaining traction due to their ability to perform tasks across different domains and languages, including Python, Go, and TypeScript. Examples include Qwen3-235B, DeepSeek-V4 Pro, GLM-5/4.5, and Llama-4 Agent Edition.
|
| 8 |
+
|
| 9 |
+
The report highlights the emergence of video generation agents like HeyGen Avatar V. This technology is moving beyond simple pixel manipulation to create more "creative" AI agents capable of generating realistic videos with nuanced expressions. This advancement is driven by breakthroughs in architecture like "Avatar V Architecture," which allows for real-time synthesis of custom kernels, and a new approach to video production called "Prompt-to-Blueprint." This method focuses on creating visual blueprints that guide the creation process, allowing for more precise control over final outputs.
|
| 10 |
+
|
| 11 |
+
Looking ahead, the report predicts that AI will become even more decision-centric in the coming years. This means evaluating not just what an AI says but also its ability to make optimal decisions based on available information. The report emphasizes the need for "Harness Engineering," a standard code layer that controls and monitors models, becoming a crucial skillset for future AI developers.
|
| 12 |
+
|
| 13 |
+
Finally, the report suggests a shift towards collaborative efforts between multiple AI agents. Instead of relying solely on one large model, we'll likely see more teams of specialized OSS models working together to solve complex problems. This multi-agent approach promises greater efficiency and adaptability in various fields.
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
Reporting for The Android Times.
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_long_clean.txt
ADDED
|
@@ -0,0 +1,15 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight, we turn to the rapidly evolving world of artificial intelligence and its impact on our future workforces. A new report sheds light on how AI is transforming the way we approach tasks and solve problems.
|
| 2 |
+
|
| 3 |
+
The report dives deep into the "Agentic AI" landscape, focusing on the shift from conversational fluency to practical execution capabilities. AI models are now being evaluated based not just on their ability to hold a conversation, but also on their capacity to perform real-world tasks efficiently. The report highlights key benchmarks like the Berkeley Function Calling Leaderboard and SWE-bench Pro, which measure an AI's performance in software development and issue resolution.
|
| 4 |
+
|
| 5 |
+
The study reveals that these benchmarks are shifting towards a more practical focus, emphasizing the ability of AI models to execute complex tasks autonomously. This transition is driven by advancements in tools like Langflow, which allows for the orchestration of various AI models based on specific workflows.
|
| 6 |
+
|
| 7 |
+
The report also explores the rise of open-source models that rival proprietary models like Claude Mythos and GPT-5.3. These OSS models are gaining traction due to their ability to perform tasks across different domains and languages, including Python, Go, and TypeScript. Examples include Qwen3-235B, DeepSeek-V4 Pro, GLM-5/4.5, and Llama-4 Agent Edition.
|
| 8 |
+
|
| 9 |
+
The report highlights the emergence of video generation agents like HeyGen Avatar V. This technology is moving beyond simple pixel manipulation to create more "creative" AI agents capable of generating realistic videos with nuanced expressions. This advancement is driven by breakthroughs in architecture like "Avatar V Architecture," which allows for real-time synthesis of custom kernels, and a new approach to video production called "Prompt-to-Blueprint." This method focuses on creating visual blueprints that guide the creation process, allowing for more precise control over final outputs.
|
| 10 |
+
|
| 11 |
+
Looking ahead, the report predicts that AI will become even more decision-centric in the coming years. This means evaluating not just what an AI says but also its ability to make optimal decisions based on available information. The report emphasizes the need for "Harness Engineering," a standard code layer that controls and monitors models, becoming a crucial skillset for future AI developers.
|
| 12 |
+
|
| 13 |
+
Finally, the report suggests a shift towards collaborative efforts between multiple AI agents. Instead of relying solely on one large model, we'll likely see more teams of specialized OSS models working together to solve complex problems. This multi-agent approach promises greater efficiency and adaptability in various fields.
|
| 14 |
+
|
| 15 |
+
Reporting for The Android Times.
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_middle.md
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight's report focuses on the rapidly evolving field of agent-based AI, and how it's shaping the future of work.
|
| 2 |
+
|
| 3 |
+
A new report highlights a shift in how we evaluate AI models. No longer focused solely on conversational fluency, experts are now prioritizing "real-world execution" - the ability to actually perform tasks. This has led to a focus on benchmarks like the Berkeley Function Calling Leaderboard (BFCL) and SWE-bench Pro, which measure an AI's efficiency and problem-solving skills in real-world scenarios.
|
| 4 |
+
|
| 5 |
+
The report also dives into the rise of open-source models. Models like Qwen3-235B, DeepSeek-V4 Pro, and GLM-5/4.5 are showing impressive performance on benchmarks, offering a glimpse into what's possible with open collaboration. And in the realm of video generation, we're seeing a shift from simple pixel manipulation to "creative agents" capable of producing realistic and expressive videos.
|
| 6 |
+
|
| 7 |
+
The report emphasizes the importance of "decision-centric AI," where the focus shifts from simply understanding what an AI says to understanding which decision was ultimately optimal. This means evaluating AI based on its ability to make informed choices, not just its conversational skills.
|
| 8 |
+
|
| 9 |
+
Experts predict that in the coming years, we'll see a shift towards more collaborative models. Think of it as a team effort where different specialized AI agents work together to solve complex problems.
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
Reporting for The Android Times.
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_middle_clean.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight's report focuses on the rapidly evolving field of agent-based AI, and how it's shaping the future of work.
|
| 2 |
+
|
| 3 |
+
A new report highlights a shift in how we evaluate AI models. No longer focused solely on conversational fluency, experts are now prioritizing "real-world execution" - the ability to actually perform tasks. This has led to a focus on benchmarks like the Berkeley Function Calling Leaderboard and SWE-bench Pro, which measure an AI's efficiency and problem-solving skills in real-world scenarios.
|
| 4 |
+
|
| 5 |
+
The report also dives into the rise of open-source models. Models like Qwen3-235B, DeepSeek-V4 Pro, and GLM-5/4.5 are showing impressive performance on benchmarks, offering a glimpse into what's possible with open collaboration. And in the realm of video generation, we're seeing a shift from simple pixel manipulation to "creative agents" capable of producing realistic and expressive videos.
|
| 6 |
+
|
| 7 |
+
The report emphasizes the importance of "decision-centric AI," where the focus shifts from simply understanding what an AI says to understanding which decision was ultimately optimal. This means evaluating AI based on its ability to make informed choices, not just its conversational skills.
|
| 8 |
+
|
| 9 |
+
Experts predict that in the coming years, we'll see a shift towards more collaborative models. Think of it as a team effort where different specialized AI agents work together to solve complex problems.
|
| 10 |
+
|
| 11 |
+
Reporting for The Android Times.
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_short.md
ADDED
|
@@ -0,0 +1,22 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"New AI is changing the game in sports. Analysts say that a combination of Spanish structure and Argentine creativity is leading to higher performance predictions. This new AI could be revolutionizing how we view teams."
|
| 2 |
+
|
| 3 |
+
[Video footage of soccer players, highlights, and data visualizations]
|
| 4 |
+
|
| 5 |
+
"The AI's ability to analyze vast amounts of data - from past matches to player conditions - is giving us a deeper understanding of potential strengths and weaknesses. Traditional analysis focused on star players or tactical success, but now the AI is changing that."
|
| 6 |
+
|
| 7 |
+
[Video footage of AI algorithms analyzing data]
|
| 8 |
+
|
| 9 |
+
"This report highlights key benchmarks for agent-based AI: BFCL v4, SWE-bench Pro, and WebArena/GAIA. These metrics are shifting from simple conversation flow to real-world task execution. The rise of open source models like Qwen3-235B is also a game-changer."
|
| 10 |
+
|
| 11 |
+
[Video footage of open-source software development]
|
| 12 |
+
|
| 13 |
+
"And in the world of video generation, AI is evolving beyond just pixel manipulation. HeyGen Avatar V is pushing the boundaries with new architecture and real-time synthesis capabilities. This means more creative control for users."
|
| 14 |
+
|
| 15 |
+
|
| 16 |
+
[Video footage of a virtual soccer player performing complex maneuvers on screen]
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
|
| 20 |
+
"The future? AI will focus on decision-making, harness engineering standardization, and multi-agent collaboration. This is just the beginning of a revolution in sports and beyond."
|
| 21 |
+
|
| 22 |
+
[End video with text overlay: "Reporting for The Android Times."]
|
docs/AI_Agents_and_Logic/agentic_ai_tech_report_2026_script_short_clean.txt
ADDED
|
@@ -0,0 +1,9 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
New AI is changing the game in sports. Analysts say that a combination of Spanish structure and Argentine creativity is leading to higher performance predictions. This new AI could be revolutionizing how we view teams.
|
| 2 |
+
|
| 3 |
+
The AI's ability to analyze vast amounts of data - from past matches to player conditions - is giving us a deeper understanding of potential strengths and weaknesses. Traditional analysis focused on star players or tactical success, but now the AI is changing that.
|
| 4 |
+
|
| 5 |
+
This report highlights key benchmarks for agent-based AI: BFCL v4, SWE-bench Pro, and WebArena/GAIA. These metrics are shifting from simple conversation flow to real-world task execution. The rise of open source models like Qwen3-235B is also a game-changer.
|
| 6 |
+
|
| 7 |
+
And in the world of video generation, AI is evolving beyond just pixel manipulation. HeyGen Avatar V is pushing the boundaries with new architecture and real-time synthesis capabilities. This means more creative control for users.
|
| 8 |
+
|
| 9 |
+
The future? AI will focus on decision-making, harness engineering standardization, and multi-agent collaboration. This is just the beginning of a revolution in sports and beyond.
|
docs/AI_Agents_and_Logic/ai_formal_logic_survey.md
ADDED
|
@@ -0,0 +1,497 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# AI(LLM)× 形式論理 融合の最前線:包括的サーベイ
|
| 2 |
+
|
| 3 |
+
> **対象読者**: AI・形式手法に関心のある研究者・エンジニア
|
| 4 |
+
> **調査時点**: 2026年5月
|
| 5 |
+
> **カバー範囲**: ニューロ・シンボリックAI、数学オリンピック自動証明、AI証明探索ツール、自動形式化、ベンチマーク
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 目次
|
| 10 |
+
|
| 11 |
+
1. [ニューロ・シンボリックAI(Neuro-Symbolic AI)](#1-ニューロシンボリックaineuro-symbolic-ai)
|
| 12 |
+
2. [数学オリンピック自動証明システム](#2-数学オリンピック自動証明システム)
|
| 13 |
+
3. [AI証明探索ツール・フレームワーク](#3-ai証明探索ツールフレームワーク)
|
| 14 |
+
4. [自動形式化(Auto-formalization)](#4-自動形式化auto-formalization)
|
| 15 |
+
5. [ベンチマークと評価](#5-ベンチマークと評価)
|
| 16 |
+
6. [形式化エコシステム:Mathlib と数学コミュニティ](#6-形式化エコシステムmathlib-と数学コミュニティ)
|
| 17 |
+
7. [今後の展望と未解決課題](#7-今後の展望と未解決課題)
|
| 18 |
+
8. [主要文献・リソース一覧](#8-主要文献リソース一覧)
|
| 19 |
+
|
| 20 |
+
---
|
| 21 |
+
|
| 22 |
+
## 1. ニューロ・シンボリックAI(Neuro-Symbolic AI)
|
| 23 |
+
|
| 24 |
+
### 1.1 なぜニューロ・シンボリックAIが必要なのか
|
| 25 |
+
|
| 26 |
+
大規模言語モデル(LLM)は自然言語処理やコード生成において圧倒的な能力を示していますが、根本的な限界が明らかになっています。
|
| 27 |
+
|
| 28 |
+
| LLMの弱点 | 具体的な問題 | 形式論理による解決 |
|
| 29 |
+
| :--- | :--- | :--- |
|
| 30 |
+
| **ハルシネーション** | もっともらしい嘘を生成する | 形式検証器が出力を数学的に検証 |
|
| 31 |
+
| **論理推論の脆弱性** | 多段推論で矛盾や飛躍が発生 | 定理証明器が各ステップを厳密にチェック |
|
| 32 |
+
| **検証不可能性** | 「正しそう」だが検証手段がない | 形式的証明は機械的に100%検証可能 |
|
| 33 |
+
| **Chain-of-Thought の限界** | 思考過程を自然言語で出力するだけで、論理的正しさの保証なし | 形式言語での推論は意味論的に正しいことが保証される |
|
| 34 |
+
|
| 35 |
+
### 1.2 体系的分類:3つの統合パラダイム
|
| 36 |
+
|
| 37 |
+
2024–2025年のサーベイ論文(IJCAI 2025 サーベイ等)では、LLMとシンボリックシステムの統合を以下の3パラダイムに分類しています。
|
| 38 |
+
|
| 39 |
+
```mermaid
|
| 40 |
+
graph LR
|
| 41 |
+
subgraph "パラダイム1: Symbolic → Neural"
|
| 42 |
+
S1["シンボリック知識<br>(ルール・オントロジー)"] -->|ガイド・制約| N1["LLM<br>(訓練/推論)"]
|
| 43 |
+
end
|
| 44 |
+
|
| 45 |
+
subgraph "パラダイム2: Neural → Symbolic"
|
| 46 |
+
N2["LLM<br>(意味解析)"] -->|形式言語へ変換| S2["シンボリック推論器<br>(ソルバー・証明器)"]
|
| 47 |
+
end
|
| 48 |
+
|
| 49 |
+
subgraph "パラダイム3: Hybrid (双方向)"
|
| 50 |
+
N3["LLM"] <-->|反復的に協調| S3["形式エンジン"]
|
| 51 |
+
end
|
| 52 |
+
```
|
| 53 |
+
|
| 54 |
+
#### パラダイム1: Symbolic → Neural(記号 → ニューラル)
|
| 55 |
+
シンボリック知識(論理ルール、オントロジー、形式的制約)を用いてLLMの訓練や推論を**ガイド・正則化**するアプローチ。
|
| 56 |
+
|
| 57 |
+
- **Logic Tensor Networks (LTN)**: 論理制約をテンソル演算として直接ニューラルネットワークの学習に埋め込む
|
| 58 |
+
- **制約付きデコーディング**: LMQL, Guidance, Outlines などのフレームワークにより、生成時に文法的・論理的制約を強制
|
| 59 |
+
- **DSR-LM / Ctrl-G**: 推論時に論理制約を適用して出力を制御
|
| 60 |
+
|
| 61 |
+
#### パラダイム2: Neural → Symbolic(ニューラル → 記号)
|
| 62 |
+
LLMを「意味解析器(Semantic Parser)」として使い、自然言語を形式言語(一階述語論理、SAT/SMT、PDDL等)に変換し、**確定的なシンボリック推論器**で解くアプローチ。
|
| 63 |
+
|
| 64 |
+
- **Logic-LM**: 自然言語問題を FOL/CSP/SAT に変換し、Z3 や Prover9 で推論。エラーメッセージをLLMにフィードバックする**自己修正ループ**を備える
|
| 65 |
+
- **LINC (Logical Inference via Neurosymbolic Computation)**: 前提と結論を FOL に変換し、定理証明器で論理的帰結を判定。ProofWriter、FOLIO ベンチマークで大幅な精度向上を実証
|
| 66 |
+
- **SatLM**: SATソルバーを用いた論理推論タスク処理
|
| 67 |
+
|
| 68 |
+
#### パラダイム3: Hybrid(双方向統合)
|
| 69 |
+
LLMと形式エンジンが**反復的に対話・批判・修正**し合うハイブリッドフレームワーク。
|
| 70 |
+
|
| 71 |
+
- **PEIRCE**: 反復的な「推測–批判(Conjecture-Criticism)」プロセスで、素材的推論と形式的推論を統一
|
| 72 |
+
- **ENVISIONS**: ニューロ・シンボリック自己訓練フレームワーク。エージェントの軌跡からシンボリック知識を帰納的に獲得
|
| 73 |
+
- **APOLLO**: 形式推論のための協調フレームワーク
|
| 74 |
+
|
| 75 |
+
> [!IMPORTANT]
|
| 76 |
+
> 最も注目すべきトレンドは、シンボリック成分を単なる「外部ツール」ではなく、**System 2(熟慮的推論)の不可欠なレ���ヤー**としてLLMアーキテクチャに統合する動きです。これにより、LLMが「推測する機械」から「証明する機械」へと進化する道筋が見えてきています。
|
| 77 |
+
|
| 78 |
+
### 1.3 産業界での採用状況
|
| 79 |
+
|
| 80 |
+
| 組織 | 主なアプローチ |
|
| 81 |
+
| :--- | :--- |
|
| 82 |
+
| **Google DeepMind** | AlphaProof(Lean 4 + AlphaZero RL)、AlphaGeometry(DDM + Gemini) |
|
| 83 |
+
| **Microsoft Research** | Lean 4 開発主導、形式検証ツール、Lean Copilot |
|
| 84 |
+
| **Meta AI** | HyperTree Proof Search (HTPS)、Lean/Metamath 環境での自動証明 |
|
| 85 |
+
| **OpenAI** | 自然言語推論ベースの数学解法、IMO 2025 で金メダル級 |
|
| 86 |
+
| **ByteDance** | Seed-Prover(エージェント型 Lean 4 証明器、IMO 2025 金メダル級) |
|
| 87 |
+
| **Harmonic** | Aristotle(Lean 4 形式証明 + 幾何ソルバー、IMO 2025 金メダル級) |
|
| 88 |
+
|
| 89 |
+
---
|
| 90 |
+
|
| 91 |
+
## 2. 数学オリンピック自動証明システム
|
| 92 |
+
|
| 93 |
+
### 2.1 AlphaProof(Google DeepMind, 2024)
|
| 94 |
+
|
| 95 |
+
AlphaProof は、LLMと強化学習を組み合わせて **Lean 4 上で形式的に検証可能な数学証明を自動生成する**、Google DeepMind のシステムです。
|
| 96 |
+
|
| 97 |
+
#### アーキテクチャ
|
| 98 |
+
|
| 99 |
+
```
|
| 100 |
+
┌──────────────────────────────────────────────────────────┐
|
| 101 |
+
│ AlphaProof │
|
| 102 |
+
│ │
|
| 103 |
+
│ ┌──────────────┐ ┌──────────────────┐ │
|
| 104 |
+
│ │ 事前訓練済み │ │ AlphaZero 式 │ │
|
| 105 |
+
│ │ 言語モデル │────▶│ 強化学習エンジン │ │
|
| 106 |
+
│ │ (Gemini 系) │ │ (MCTS 木探索) │ │
|
| 107 |
+
│ └──────────────┘ └────────┬─────────┘ │
|
| 108 |
+
│ │ │
|
| 109 |
+
│ ▼ │
|
| 110 |
+
│ ┌──────────────────┐ │
|
| 111 |
+
│ │ Lean 4 環境 │ │
|
| 112 |
+
│ │ (タクティクス実行 │ │
|
| 113 |
+
│ │ ・形式検証) │ │
|
| 114 |
+
│ └──────────────────┘ │
|
| 115 |
+
└──────────────────────────────────────────────────────────┘
|
| 116 |
+
```
|
| 117 |
+
|
| 118 |
+
#### 技術的特徴
|
| 119 |
+
|
| 120 |
+
| 項目 | 詳細 |
|
| 121 |
+
| :--- | :--- |
|
| 122 |
+
| **基盤モデル** | 事前訓練済み言語モデル(Gemini ベース) |
|
| 123 |
+
| **強化学習** | AlphaZero アルゴリズムによるモンテカルロ木探索(MCTS) |
|
| 124 |
+
| **形式言語** | Lean 4(証明の正しさを型チェッカーが100%検証) |
|
| 125 |
+
| **訓練方法** | 数百万の問題で証明・反証を数週間にわたり実施。競技中も問題のバリエーションを自己生成して強化学習を継続 |
|
| 126 |
+
| **自動形式化** | IMO の問題文は**人間が手動で Lean に翻訳**(2024年時点の制約) |
|
| 127 |
+
|
| 128 |
+
#### IMO 2024 での成果
|
| 129 |
+
|
| 130 |
+
- **6問中4問を解答**(代数2問、整数論1問 + AlphaGeometry 2 が幾何1問)
|
| 131 |
+
- **得点: 42点満点中28点** → **銀メダル相当**
|
| 132 |
+
- 特筆すべきは、人間の正答率が極めて低い**最難問の Problem 6(コンビナトリクス)を解答**したこと
|
| 133 |
+
|
| 134 |
+
> [!NOTE]
|
| 135 |
+
> AlphaProof の革新性は、LLMの「直感的な推測」をAlphaZeroの「体系的な探索」と組み合わせ、さらにLean 4 の型チェッカーが各ステップの正しさを保証するという**三層構造**にあります。これにより、ハルシネーションの余地がゼロの証明が生成されます。
|
| 136 |
+
|
| 137 |
+
### 2.2 AlphaGeometry 2(Google DeepMind, 2024)
|
| 138 |
+
|
| 139 |
+
幾何学に特化したニューロ・シンボリックシステムの進化版。
|
| 140 |
+
|
| 141 |
+
#### アーキテクチャ(「速い思考」と「遅い思考」の融合)
|
| 142 |
+
|
| 143 |
+
```
|
| 144 |
+
┌─────────────────────────────────────────────┐
|
| 145 |
+
│ AlphaGeometry 2 │
|
| 146 |
+
│ │
|
| 147 |
+
│ ┌────────────────────┐ ┌───────────────┐ │
|
| 148 |
+
│ │ Gemini ベース LLM │ │ 記号的演繹 │ │
|
| 149 |
+
│ │ (補助線の直感的提案) │ │ エンジン │ │
|
| 150 |
+
│ │ 【System 1: 速い】 │ │ (ルール適用) │ │
|
| 151 |
+
│ │ │ │ 【System 2】 │ │
|
| 152 |
+
│ └────────┬───────────┘ └───────┬───────┘ │
|
| 153 |
+
│ │ 相互作用 │ │
|
| 154 |
+
│ └──────────┬───────────┘ │
|
| 155 |
+
│ ▼ │
|
| 156 |
+
│ 幾何学定理の証明 │
|
| 157 |
+
└─────────────────────────────────────────────┘
|
| 158 |
+
```
|
| 159 |
+
|
| 160 |
+
- **訓練データ**: 初代の10倍以上の合成データで訓練
|
| 161 |
+
- **記号エンジン**: 初代の100倍高速化
|
| 162 |
+
- **拡張**: 移動変換(平行移動・回転等)、角度・比・距離に関する線形方程式に対応
|
| 163 |
+
- **IMO 2024 幾何問題**: 形式化提供後**わずか19秒で解答**
|
| 164 |
+
|
| 165 |
+
### 2.3 IMO 2025:金メダル時代の到来
|
| 166 |
+
|
| 167 |
+
2025年のIMOでは、複数のAIシステムが**金メダル級**のパフォーマンスを達成し、形式証明AIの新時代が到来しました。
|
| 168 |
+
|
| 169 |
+
| システム | 開発元 | 解答数 (6問中) | メダル相当 | 特徴 |
|
| 170 |
+
| :--- | :--- | :---: | :---: | :--- |
|
| 171 |
+
| **Aristotle** | Harmonic | 5問 | 🥇 金 | Lean 4 形式証明 + 幾何専用ソルバー |
|
| 172 |
+
| **Seed-Prover** | ByteDance | 4-5問 | 🥇 金 (追加計算後) | エージェント型 Lean 4 証明器 + Python 実行 + 数学ライブラリ検索 |
|
| 173 |
+
| **OpenAI モデル** | OpenAI | 5問 | 🥇 金 | 自然言語推論ベース(形式証明なし) |
|
| 174 |
+
| **DeepMind モデル** | Google | 5問 | 🥇 金 | 自然言語推論ベース(形式証明なし) |
|
| 175 |
+
|
| 176 |
+
> [!WARNING]
|
| 177 |
+
> **形式証明 vs 自然言語推論の分岐**: IMO 2025 では、Aristotle と Seed-Prover は**機械検証可能な形式証明**を出力したのに対し、OpenAI と Google のシステムは自然言語による証明を出力し、人間の審査員が正しさを評価しました。この「検証可能性の差」は、AIの信頼性において根本的な違いをもたらします。
|
| 178 |
+
|
| 179 |
+
### 2.4 Meta AI: HyperTree Proof Search (HTPS)
|
| 180 |
+
|
| 181 |
+
Meta AI が2022年に発表した、AlphaZero にインスパイアされた定理証明探索アルゴリズムです。
|
| 182 |
+
|
| 183 |
+
- **手法**: ハイパーグラフ上の木探索。選択(Selection)→ 展開(Expansion)→ 逆伝播(Back-propagation)の反復
|
| 184 |
+
- **ニューラル構成**: Policy モデル(タクティクス提案)+ Critic モデル(証明進捗評価)
|
| 185 |
+
- **オンライン学習**: 自身の成功した証明から学習し、訓練分布外の問題にも汎化
|
| 186 |
+
- **環境**: Lean + Metamath で評価。miniF2F-curriculum、IMO問題での成果
|
| 187 |
+
|
| 188 |
+
---
|
| 189 |
+
|
| 190 |
+
## 3. AI証明探索ツール・フレームワーク
|
| 191 |
+
|
| 192 |
+
### 3.1 LeanDojo
|
| 193 |
+
|
| 194 |
+
AI駆動の Lean 4 定理証明のためのオープンソースフレームワーク。NeurIPS 2023 で発表。
|
| 195 |
+
|
| 196 |
+
#### アーキテクチャ
|
| 197 |
+
|
| 198 |
+
```mermaid
|
| 199 |
+
graph TB
|
| 200 |
+
subgraph "LeanDojo エコシステム"
|
| 201 |
+
A["Lean 4 リポジトリ<br>(Mathlib 等)"] -->|トレースと抽出| B["データ抽出エンジン<br>(証明状態・タクティクス・前提)"]
|
| 202 |
+
B --> C["訓練データセット"]
|
| 203 |
+
C --> D["ReProver<br>(検索拡張型 LLM)"]
|
| 204 |
+
D -->|タクティクス提案| E["Lean 4 証明環境"]
|
| 205 |
+
E -->|フィードバック| D
|
| 206 |
+
end
|
| 207 |
+
```
|
| 208 |
+
|
| 209 |
+
#### 主要コンポーネント
|
| 210 |
+
|
| 211 |
+
| コンポーネント | 機能 |
|
| 212 |
+
| :--- | :--- |
|
| 213 |
+
| **データ抽出エンジン** | Lean 4 リポジトリから証明状態、タクティクス、前提を自動抽出 |
|
| 214 |
+
| **ReProver** | 検索拡張型生成(RAG)ベースの LLM。関連する前提(レンマ・定理)を Mathlib から検索し、タクティクスを提案 |
|
| 215 |
+
| **証明環境** | Lean 4 とプログラム的に対話し、タクティクスの実行と検証を自動化 |
|
| 216 |
+
|
| 217 |
+
#### LeanDojo-v2(2025年)
|
| 218 |
+
|
| 219 |
+
2025年にリリースされた第2世代。以下の機能が大幅に強化されています。
|
| 220 |
+
|
| 221 |
+
- **エンドツーエンドフレームワーク**: リポジトリトレース → データセット管理 → モデル訓練 → 評価を統一
|
| 222 |
+
- **強化学習サポート**: GRPOTrainer による教師あり微調整と強化学習の統合
|
| 223 |
+
- **生涯学習型データセット管理**: Mathlib の更新に追従するデータセット管理
|
| 224 |
+
- **Lean4Code**: 数学者向けの軽量 IDE。AI 機能をフォーマライゼーションワークフローに直接統合
|
| 225 |
+
|
| 226 |
+
### 3.2 Lean Copilot
|
| 227 |
+
|
| 228 |
+
Lean 4 環境内で LLM を直接利用できるようにするフレームワーク。VS Code 上でリアルタイムに証明支援を提供します。
|
| 229 |
+
|
| 230 |
+
#### 主要機能
|
| 231 |
+
|
| 232 |
+
| 機能 | 説明 |
|
| 233 |
+
| :--- | :--- |
|
| 234 |
+
| **タクティクス提案** | 現在の証明状態に基づいて、コンテキストを考慮した次の一手を推薦 |
|
| 235 |
+
| **前提検索** | Mathlib やユーザーのプロジェクトから関連するレンマ・定理を動的に検索 |
|
| 236 |
+
| **証明探索** | 人間の入力とAIの自動探索を組み合わせた半自動証明 |
|
| 237 |
+
| **モデル柔軟性** | ローカルモデル(オフライン動作)または���ラウド API 経由で任意の LLM を利用可能 |
|
| 238 |
+
|
| 239 |
+
#### 性能
|
| 240 |
+
|
| 241 |
+
- 従来のルールベース自動化(AESOP 等)を**大幅に上回る**証明成功率
|
| 242 |
+
- 完全自動証明と人間アシスト証明の両方で優位性を実証
|
| 243 |
+
|
| 244 |
+
### 3.3 主要な LLM ベース定理証明モデル比較
|
| 245 |
+
|
| 246 |
+
以下は2024–2025年に登場した主要なLLM定理証明モデルの比較です。
|
| 247 |
+
|
| 248 |
+
| モデル | 開発元 | 基盤モデル | 主な技術的革新 | 対象環境 |
|
| 249 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 250 |
+
| **DeepSeek-Prover** | DeepSeek | DeepSeek | 大規模合成データ生成 + Lean 4 自動フィルタリング | Lean 4 |
|
| 251 |
+
| **DeepSeek-Prover-V2** | DeepSeek | DeepSeek-V3 | 再帰的定理証明パイプライン + サブゴール分解 | Lean 4 |
|
| 252 |
+
| **Kimina-Prover-72B** | — | Qwen2.5 | 推論駆動探索、レンマ発見パターン、エラー修正 | Lean 4 |
|
| 253 |
+
| **InternLM2.5-StepProver** | Shanghai AI Lab | InternLM2.5 | エキスパート反復 + Critic モデルによる証明探索ガイド | Lean 4 |
|
| 254 |
+
| **LLEMMA** | EleutherAI / Princeton | Code Llama | Proof-Pile-2 での継続事前訓練。オープンソース数学基盤モデル | Lean 4 / Isabelle |
|
| 255 |
+
| **Seed-Prover** | ByteDance | — | エージェント型アーキテクチャ + ツール統合(Python, ライブラリ検索) | Lean 4 |
|
| 256 |
+
| **Aristotle** | Harmonic | — | 形式証明 + 非形式推論 + 幾何ソルバーの統合 | Lean 4 |
|
| 257 |
+
| **HunyuanProver** | Tencent | — | System 2 思考 + ガイド付き木探索 + Critic モデル | Lean 4 |
|
| 258 |
+
|
| 259 |
+
#### DeepSeek-Prover-V2:再帰的定理証明パイプラインの詳細
|
| 260 |
+
|
| 261 |
+
```
|
| 262 |
+
┌───────────────────────────────────────────────┐
|
| 263 |
+
│ DeepSeek-Prover-V2 パイプライン │
|
| 264 |
+
│ │
|
| 265 |
+
│ 1. 複雑な定理 T を入力 │
|
| 266 |
+
│ │ │
|
| 267 |
+
│ ▼ │
|
| 268 |
+
│ 2. 高レベル推論(非形式的分析) │
|
| 269 |
+
│ └─ DeepSeek-V3 が T を複数のサブゴールに分解 │
|
| 270 |
+
│ │ │
|
| 271 |
+
│ ▼ │
|
| 272 |
+
│ 3. 各サブゴール S₁, S₂, ... Sₙ を再帰的に証明 │
|
| 273 |
+
│ └─ サブゴールがさらに複雑なら、再度分解 │
|
| 274 |
+
│ │ │
|
| 275 |
+
│ ▼ │
|
| 276 |
+
│ 4. Lean 4 型チェッカーで各ステップを検証 │
|
| 277 |
+
│ │ │
|
| 278 |
+
│ ▼ │
|
| 279 |
+
│ 5. 全サブゴール完了 → 定理 T の証明完了 │
|
| 280 |
+
└───────────────────────────────────────────────┘
|
| 281 |
+
```
|
| 282 |
+
|
| 283 |
+
- MiniF2F、PutnamBench で**最先端の性能**を達成
|
| 284 |
+
- 非形式的推論(自然言語での数学的直感)と形式的検証(Lean 4)の統合が鍵
|
| 285 |
+
|
| 286 |
+
#### Kimina-Prover:推論駆動探索パラダイム
|
| 287 |
+
|
| 288 |
+
従来の木探索アルゴリズム(BFS、MCTS)から脱却し、モデル自身の**内部推論能力**に依拠する新しいパラダイムを提案。
|
| 289 |
+
|
| 290 |
+
- **レンマ発見パターン**: モデルが自律的に中間レンマを発見・適用
|
| 291 |
+
- **エラー修正機能**: Lean のコンパイルエラーメッセージを解釈し、修正を提案
|
| 292 |
+
- **多段階強化学習**: Qwen2.5 バックボーン上で段階的に RL を適用
|
| 293 |
+
|
| 294 |
+
### 3.4 その他の関連ツール
|
| 295 |
+
|
| 296 |
+
| ツール / システム | 対象 | 特徴 |
|
| 297 |
+
| :--- | :--- | :--- |
|
| 298 |
+
| **Tactician** | Coq | Coq 環境内での AI タクティクス自動提案 |
|
| 299 |
+
| **Sledgehammer** | Isabelle/HOL | 外部 ATP(E, SPASS, Z3 等)を呼び出して自動証明 |
|
| 300 |
+
| **Draft-Sketch-Prove** | 汎用 | LLM で非形式証明のスケッチを作成 → 形式証明器で完成させるプロンプト戦略 |
|
| 301 |
+
|
| 302 |
+
---
|
| 303 |
+
|
| 304 |
+
## 4. 自動形式化(Auto-formalization)
|
| 305 |
+
|
| 306 |
+
### 4.1 概要:自然言語から形式言語への翻訳
|
| 307 |
+
|
| 308 |
+
自動形式化とは、自然言語で書かれた数学的命題やその証明を、Lean 4 などの形式言語に自動翻訳する技術です。これは、AIによる数学の完全自動化において**最大のボトルネック**の一つです。
|
| 309 |
+
|
| 310 |
+
```
|
| 311 |
+
「任意の正の整数 n に対して、 ────▶ theorem example (n : ℕ) (hn : 0 < n) :
|
| 312 |
+
n² + n は偶数である」 Even (n^2 + n) := by
|
| 313 |
+
...
|
| 314 |
+
```
|
| 315 |
+
|
| 316 |
+
### 4.2 主要な技術アプローチ(2024–2025)
|
| 317 |
+
|
| 318 |
+
#### プロセス監視型・コンパイラフィードバック型
|
| 319 |
+
|
| 320 |
+
```mermaid
|
| 321 |
+
graph LR
|
| 322 |
+
A["自然言語の<br>数学命題"] -->|LLM 翻訳| B["Lean 4 コード<br>(初期生成)"]
|
| 323 |
+
B -->|コンパイル試行| C{"Lean 4<br>��ンパイラ"}
|
| 324 |
+
C -->|成功| D["✅ 形式化完了"]
|
| 325 |
+
C -->|エラー| E["エラーメッセージ"]
|
| 326 |
+
E -->|フィードバック| A
|
| 327 |
+
```
|
| 328 |
+
|
| 329 |
+
- **PDA (Process-supervised Data Augmentation)**: Lean 4 コンパイラをリアルタイム検証器として使い、反復的に修正
|
| 330 |
+
- **Mathesis-Autoformalizer**: アンサンブル LLM 検証 + 強化学習で高忠実度翻訳を実現
|
| 331 |
+
|
| 332 |
+
#### 強化学習ベース
|
| 333 |
+
|
| 334 |
+
- **GRPO (Group Relative Policy Optimization)**: 構文的正しさ(コンパイル可能か)と意味的正しさ(元の問題を正確に表現しているか)の両方を最適化
|
| 335 |
+
- **サイクル一貫性報酬**: 形式 → 自然言語 に逆翻訳し、元の入力と一致するかで報酬を与える
|
| 336 |
+
|
| 337 |
+
#### 大規模合成データ
|
| 338 |
+
|
| 339 |
+
並列 NL-to-Lean データセットが希少であるため、合成データの生成が重要。
|
| 340 |
+
|
| 341 |
+
- **Herald**: 階層的・検索ベースの翻訳済み Lean データセット
|
| 342 |
+
- **FineLeanCorpus**: 大量のアラインメント済みデータ
|
| 343 |
+
|
| 344 |
+
### 4.3 残された最大の課題:意味的ドリフト
|
| 345 |
+
|
| 346 |
+
> [!CAUTION]
|
| 347 |
+
> **意味的ドリフト(Semantic Drift)**: コードは構文的に正しく(コンパイルが通る)が、元の数学問題を**論理的に正確に表現していない**という問題。例えば、「すべての素数 p に対して〜」という問題を「すべての自然数 n に対して〜」と形式化してしまうケース。コンパイルは成功するが、証明された命題は元の問題と異なります。これは現在の自動形式化における最大の未解決課題です。
|
| 348 |
+
|
| 349 |
+
---
|
| 350 |
+
|
| 351 |
+
## 5. ベンチマークと評価
|
| 352 |
+
|
| 353 |
+
### 5.1 形式定理証明ベンチマーク
|
| 354 |
+
|
| 355 |
+
| ベンチマーク | 対象レベル | 形式言語 | 特徴 | 現在の状況 |
|
| 356 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 357 |
+
| **MiniF2F** | 高校数学競技 (AIME/IMO) | Lean 4 / Isabelle | 最も広く使われる標準ベンチマーク | 飽和に近づいている(miniF2F-v2 が登場) |
|
| 358 |
+
| **PutnamBench** | 学部数学 (Putnam 競技) | Lean 4 | 解析・抽象代数など高度な分野を含む | 依然として非常に困難。解答率は低い |
|
| 359 |
+
| **ProofNet** | 学部数学 | Lean 4 | 自動形式化の評価にも使用 | — |
|
| 360 |
+
| **LeanDojo Benchmark 4** | 汎用 | Lean 4 | 「新規前提(Novel Premises)」でのテストにより汎化能力を評価 | — |
|
| 361 |
+
| **FormL4** | 汎用 | Lean 4 | 命題と証明の両方の自動形式化を評価 | — |
|
| 362 |
+
| **VeriSoftBench** | ソフトウェア検証 | Lean 4 | リポジトリ規模の形式検証を評価 | — |
|
| 363 |
+
|
| 364 |
+
### 5.2 論理推論ベンチマーク(LLM 直接評価用)
|
| 365 |
+
|
| 366 |
+
| ベンチマーク | 対象 | 特徴 |
|
| 367 |
+
| :--- | :--- | :--- |
|
| 368 |
+
| **FOLIO** | 一階述語論理推論 | 自然言語 → FOL 変換 + 論理的帰結判定 |
|
| 369 |
+
| **ProofWriter** | 多段推論 | 異なる証明深度でのルールベース推論 |
|
| 370 |
+
| **PrOntoQA** | 論理的含意 | ルールベース推論 |
|
| 371 |
+
| **LogicalDeduction** | 複合論理 | 複雑な実世界の論理クエリ |
|
| 372 |
+
| **AR-LSAT** | 法的推論 | LSAT 形式の高度な論理問題 |
|
| 373 |
+
| **GSM8K / MATH** | 数学的推論 | 非形式的な数学問題解決(形式証明なし) |
|
| 374 |
+
|
| 375 |
+
### 5.3 評価の新課題
|
| 376 |
+
|
| 377 |
+
> [!IMPORTANT]
|
| 378 |
+
> **「コンパイル成功 ≠ 正解」問題**: 最近の研究では、AIが生成した証明が「コンパイルに通る」(構文的に有効)にもかかわらず、意図された問題を**論理的に解決していない**ケースが報告されています。これに対応するため、**T2** や **FormalRewardBench** といった、より厳密な意味論的正しさを評価するフレームワークが開発されています。
|
| 379 |
+
|
| 380 |
+
---
|
| 381 |
+
|
| 382 |
+
## 6. 形式化エコシステム:Mathlib と数学コミュニティ
|
| 383 |
+
|
| 384 |
+
### 6.1 Mathlib の現在(2025年5月時点)
|
| 385 |
+
|
| 386 |
+
Lean 4 の最大の数学ライブラリであり、AIの訓練データとしても決定的に重要です。
|
| 387 |
+
|
| 388 |
+
| 指標 | 数値 |
|
| 389 |
+
| :---: | :---: |
|
| 390 |
+
| 定理数 | **210,000+** |
|
| 391 |
+
| 定義数 | **100,000+** |
|
| 392 |
+
| 総コード行数 | **190万行+** |
|
| 393 |
+
| コントリビューター数 | **500人+** |
|
| 394 |
+
|
| 395 |
+
### 6.2 テレンス・タオ教授と形式数学
|
| 396 |
+
|
| 397 |
+
フィールズ賞受賞者であるテレンス・タオ教授は、形式数学の最も著名な提唱者の一人です。
|
| 398 |
+
|
| 399 |
+
- **自身の研究論文の形式化**: Lean 4 を使って自身のプレプリントの証明を形式化。微妙なエラーの発見、前提の厳密な記述、複雑な論証のリファクタリングなどの利点を実証
|
| 400 |
+
- **「形式化ギャップ」の議論**: 現時点では形式化はLaTeX執筆よりはるかに時間がかかるが、新しいタクティクスとAIツールがこのギャップを急速に縮めつつあると指摘
|
| 401 |
+
- **「生きた数学」の提唱**: フェルマーの最終定理の形式化プロジェクトなど、大規模な協調的数学プロジェクトにおいて、形式化が複雑性管理の重要なツールであると強調
|
| 402 |
+
|
| 403 |
+
### 6.3 Mathlib Initiative(2025年後半~)
|
| 404 |
+
|
| 405 |
+
Mathlib のスケールに伴い、以下の課題に対処するための組織化が進行中:
|
| 406 |
+
|
| 407 |
+
- レビュープロセスのボトルネック解消
|
| 408 |
+
- ドキュメンテーションの改善
|
| 409 |
+
- 下流ライブラリとAI統合ワークフローのインフラ整備
|
| 410 |
+
|
| 411 |
+
---
|
| 412 |
+
|
| 413 |
+
## 7. 今後の展望と未解決課題
|
| 414 |
+
|
| 415 |
+
### 7.1 技術的課題
|
| 416 |
+
|
| 417 |
+
```mermaid
|
| 418 |
+
graph TD
|
| 419 |
+
A["未解決課題"] --> B["自動形式化の意味的正確性"]
|
| 420 |
+
A --> C["計算コスト<br>(RL訓練・探索に莫大なGPU時間)"]
|
| 421 |
+
A --> D["形式化ギャップ<br>(自然言語数学との乖離)"]
|
| 422 |
+
A --> E["ドメイン汎化<br>(競技数学以外への拡張)"]
|
| 423 |
+
A --> F["長い証明への対応<br>(数百ステップの証明の生成)"]
|
| 424 |
+
|
| 425 |
+
B --> B1["意味的ドリフトの検出と防止"]
|
| 426 |
+
D --> D1["より自然な形式言語の設計"]
|
| 427 |
+
E --> E1["ソフトウェア検証<br>物理学<br>生物学への応用"]
|
| 428 |
+
```
|
| 429 |
+
|
| 430 |
+
### 7.2 短期的展望(2025–2027)
|
| 431 |
+
|
| 432 |
+
| 領域 | 予測される進展 |
|
| 433 |
+
| :--- | :--- |
|
| 434 |
+
| **競技数学** | IMO 全問正解(満点金メダル)の達成。形式証明ベースのシステムが主流に |
|
| 435 |
+
| **ツール統合** | IDE 内での AI 証明支援がデフォルト化(Lean Copilot の進化) |
|
| 436 |
+
| **エージェント化** | 証明器がツール使用(Python 実行、ライブラリ検索、Web 検索)を自律的に行う |
|
| 437 |
+
| **ソフトウェア検証** | 重要なオープンソースライブラリのコアモジュールを AI が形式検証する試み |
|
| 438 |
+
|
| 439 |
+
### 7.3 長期的展望(2027–)
|
| 440 |
+
|
| 441 |
+
- **数学研究の変革**: AI が新しい定理の予想(Conjecture)を生成し、自ら形式証明する「数学の自動発見」
|
| 442 |
+
- **認知パートナー**: 人間の直感と形式的精密さを組み合わせた「認知パートナー」としてのAI
|
| 443 |
+
- **AGI への道**: 形式論理は、AI の「正確な推論能力」を保証する唯一の手段として、AGI(汎用人工知能)アーキテクチャの不可欠な構成要素に
|
| 444 |
+
- **教育**: AI 証明支援ツールが数学教育を根本的に変え、学生が「証明の構造」を対話的に学べるように
|
| 445 |
+
|
| 446 |
+
---
|
| 447 |
+
|
| 448 |
+
## 8. 主要文献・リソース一覧
|
| 449 |
+
|
| 450 |
+
### 8.1 論文(時系列順)
|
| 451 |
+
|
| 452 |
+
| 年 | タイトル | 著者 / 組織 | 会議 / 掲載 |
|
| 453 |
+
| :--- | :--- | :--- | :--- |
|
| 454 |
+
| 2022 | HyperTree Proof Search for Neural Theorem Proving | Lample et al. (Meta AI) | NeurIPS 2022 |
|
| 455 |
+
| 2023 | LLEMMA: An Open Language Model for Mathematics | Azerbayev et al. (EleutherAI / Princeton) | ICLR 2024 |
|
| 456 |
+
| 2023 | LeanDojo: Theorem Proving with Retrieval-Augmented Language Models | Yang et al. (Caltech / MIT) | NeurIPS 2023 |
|
| 457 |
+
| 2024 | Lean Copilot: LLMs as Copilots for Theorem Proving in Lean | Song et al. | — |
|
| 458 |
+
| 2024 | AlphaProof and AlphaGeometry 2 | Google DeepMind | Blog / Technical Report |
|
| 459 |
+
| 2024 | AlphaGeometry: An Olympiad-level AI system for geometry | Trinh et al. (DeepMind) | Nature 2024 |
|
| 460 |
+
| 2024 | DeepSeek-Prover: Advancing Theorem Proving via Large-scale Synthetic Data | Xin et al. (DeepSeek) | — |
|
| 461 |
+
| 2024 | Logic-LM: Empowering LLMs with Symbolic Solvers for Faithful Logical Reasoning | Pan et al. | ACL 2024 |
|
| 462 |
+
| 2024 | LINC: A Neurosymbolic Approach for Logical Reasoning | Olausson et al. | EMNLP 2023 |
|
| 463 |
+
| 2024 | InternLM2.5-StepProver | Shanghai AI Lab | — |
|
| 464 |
+
| 2025 | DeepSeek-Prover-V2: Recursive Theorem Proving | DeepSeek | — |
|
| 465 |
+
| 2025 | Kimina-Prover-72B | — | — |
|
| 466 |
+
| 2025 | LeanDojo-v2 | Yang et al. | — |
|
| 467 |
+
| 2025 | Seed-Prover | ByteDance | — |
|
| 468 |
+
| 2025 | Aristotle | Harmonic | — |
|
| 469 |
+
| 2025 | IJCAI Survey: LLM Reasoning with NeSy | Various | IJCAI 2025 |
|
| 470 |
+
|
| 471 |
+
### 8.2 サーベイ論文
|
| 472 |
+
|
| 473 |
+
- **"A Survey on Large Language Models with Neuro-Symbolic Reasoning"** — IJCAI 2025。LLMとニューロ・シンボリック推論の統合を体系的に分類
|
| 474 |
+
- **"LLM-Symbolic-Reasoning-Survey"** — GitHub リポジトリ。シンボリック形式化、定理証明、ニューロ・シンボリック統合の研究をトラッキング
|
| 475 |
+
- **"Awesome-LLM-Reasoning-with-NeSy"** — IJCAI 2025 サーベイに関連するキュレーションリスト
|
| 476 |
+
|
| 477 |
+
### 8.3 実践リソース
|
| 478 |
+
|
| 479 |
+
| リソース | URL / 場所 | 用途 |
|
| 480 |
+
| :--- | :--- | :--- |
|
| 481 |
+
| **Lean 4** | [leanprover.github.io](https://leanprover.github.io/) | 証明言語本体 |
|
| 482 |
+
| **Mathlib** | [github.com/leanprover-community/mathlib4](https://github.com/leanprover-community/mathlib4) | Lean 4 数学ライブラリ |
|
| 483 |
+
| **LeanDojo** | [leandojo.org](https://leandojo.org/) | AI定理証明フレームワーク |
|
| 484 |
+
| **Lean Copilot** | GitHub | VS Code 上の AI 証明支援 |
|
| 485 |
+
| **Natural Number Game** | [adam.math.hhu.de](https://adam.math.hhu.de/) | Lean 4 を遊びながら学ぶゲーム |
|
| 486 |
+
| **Z3 Solver** | [github.com/Z3Prover/z3](https://github.com/Z3Prover/z3) | SMT ソルバー |
|
| 487 |
+
| **MiniF2F** | GitHub | 形式定理証明ベンチマーク |
|
| 488 |
+
| **PutnamBench** | GitHub | 学部数学レベルのベンチマーク |
|
| 489 |
+
|
| 490 |
+
---
|
| 491 |
+
|
| 492 |
+
> [!TIP]
|
| 493 |
+
> **この分野に入門するための推奨パス**:
|
| 494 |
+
> 1. [Natural Number Game](https://adam.math.hhu.de/) で Lean 4 の基礎を体験(所要: 数時間)
|
| 495 |
+
> 2. LeanDojo のチュートリアルで AI 証明探索を体験
|
| 496 |
+
> 3. Z3 (Python) で制約充足問題を解く経験を積む
|
| 497 |
+
> 4. DeepSeek-Prover や Lean Copilot を使って実際の数学の形式化を試みる
|
docs/AI_Agents_and_Logic/ai_formal_logic_survey.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "5bc408de7c4c1e24", "ingested_at": "2026-06-30T08:25:49.660344", "content_hash": "5bc408de7c4c1e24"}
|
docs/AI_Agents_and_Logic/ai_proof_search_tools_survey.md
ADDED
|
@@ -0,0 +1,670 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# AI証明探索ツール&モデル — 包括的サーベイ
|
| 2 |
+
|
| 3 |
+
> **調査時点**: 2026年5月
|
| 4 |
+
> **主題**: LeanDojo, Lean Copilot, 及び主要AIプルーバーの技術的比較
|
| 5 |
+
> **範囲**: 開発基盤(ツール)からモデル、ベンチマーク、訓練手法まで
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 目次
|
| 10 |
+
|
| 11 |
+
1. [全体像:AI証明探索のエコシステム](#1-全体像ai証明探索のエコシステム)
|
| 12 |
+
2. [基盤インフラ:Lean 4 と Mathlib4](#2-基盤インフラlean-4-と-mathlib4)
|
| 13 |
+
3. [LeanDojo:v1 → v2 の進化](#3-leandojov1--v2-の進化)
|
| 14 |
+
4. [Lean Copilot:対話型AI証明アシスタント](#4-lean-copilot対話型ai証明アシスタント)
|
| 15 |
+
5. [主要AIプルーバー・モデルの技術比較](#5-主要aiプルーバーモデルの技術比較)
|
| 16 |
+
6. [証明生成戦略の分類](#6-証明生成戦略の分類)
|
| 17 |
+
7. [ベンチマークと評価](#7-ベンチマークと評価)
|
| 18 |
+
8. [訓練手法の比較](#8-訓練手法の比較)
|
| 19 |
+
9. [未解決課題と今後の展望](#9-未解決課題と今後の展望)
|
| 20 |
+
10. [文献一覧](#10-文献一覧)
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## 1. 全体像:AI証明探索のエコシステム
|
| 25 |
+
|
| 26 |
+
AI証明探索は、**3層のスタック**として整理できます。
|
| 27 |
+
|
| 28 |
+
```mermaid
|
| 29 |
+
graph TB
|
| 30 |
+
subgraph "Layer 3: AI プルーバー・モデル"
|
| 31 |
+
M1["DeepSeek-Prover-V2"]
|
| 32 |
+
M2["Kimina-Prover-72B"]
|
| 33 |
+
M3["InternLM2.5-StepProver"]
|
| 34 |
+
M4["Goedel-Prover-V2"]
|
| 35 |
+
M5["LLEMMA"]
|
| 36 |
+
M6["Seed-Prover"]
|
| 37 |
+
end
|
| 38 |
+
|
| 39 |
+
subgraph "Layer 2: 開発フレームワーク・ツール"
|
| 40 |
+
T1["LeanDojo v2"]
|
| 41 |
+
T2["Lean Copilot"]
|
| 42 |
+
T3["LeanAgent"]
|
| 43 |
+
end
|
| 44 |
+
|
| 45 |
+
subgraph "Layer 1: 基盤インフラ"
|
| 46 |
+
L1["Lean 4 (証明アシスタント)"]
|
| 47 |
+
L2["Mathlib4 (数学ライブラリ)"]
|
| 48 |
+
end
|
| 49 |
+
|
| 50 |
+
M1 & M2 & M3 & M4 & M5 & M6 --> T1 & T2 & T3
|
| 51 |
+
T1 & T2 & T3 --> L1 & L2
|
| 52 |
+
|
| 53 |
+
style L1 fill:#2D3436,stroke:#333,color:#fff
|
| 54 |
+
style L2 fill:#2D3436,stroke:#333,color:#fff
|
| 55 |
+
style T1 fill:#0984E3,stroke:#333,color:#fff
|
| 56 |
+
style T2 fill:#0984E3,stroke:#333,color:#fff
|
| 57 |
+
style T3 fill:#0984E3,stroke:#333,color:#fff
|
| 58 |
+
```
|
| 59 |
+
|
| 60 |
+
| レイヤー | 役割 | 具体例 |
|
| 61 |
+
| :--- | :--- | :--- |
|
| 62 |
+
| **Layer 1: 基盤** | 形式言語・検証カーネル・数学ライブラリ | Lean 4, Mathlib4 |
|
| 63 |
+
| **Layer 2: ツール** | データ抽出・インタラクション・IDE統合 | LeanDojo, Lean Copilot, LeanAgent |
|
| 64 |
+
| **Layer 3: モデル** | タクティクス予測・証明探索・推論 | DeepSeek-Prover-V2, Kimina-Prover 等 |
|
| 65 |
+
|
| 66 |
+
---
|
| 67 |
+
|
| 68 |
+
## 2. 基盤インフラ:Lean 4 と Mathlib4
|
| 69 |
+
|
| 70 |
+
### 2.1 Lean 4 — 証明アシスタント
|
| 71 |
+
|
| 72 |
+
```
|
| 73 |
+
Lean 4 の二重の役割
|
| 74 |
+
┌──────────────────────────────────────────────┐
|
| 75 |
+
│ │
|
| 76 |
+
│ 1. プログラミング言語としての Lean │
|
| 77 |
+
│ ・依存型理論(Calculus of Inductive │
|
| 78 |
+
│ Constructions)に基づく │
|
| 79 |
+
│ ・汎用関数型プログラミング言語としても使用可能 │
|
| 80 |
+
│ │
|
| 81 |
+
│ 2. 証明アシスタントとしての Lean │
|
| 82 |
+
│ ・タクティクスモードでインタラクティブ証明 │
|
| 83 |
+
│ ・カーネル(型チェッカー)は小さく検証済み │
|
| 84 |
+
│ ・Lean 自身で Lean のメタプログラミングが可能 │
|
| 85 |
+
│ │
|
| 86 |
+
│ AIにとっての重要性: │
|
| 87 |
+
│ → 型チェッカーが完全に決定論的な「報酬信号」 │
|
| 88 |
+
│ → タクティクス = AIが選択する「行動空間」 │
|
| 89 |
+
│ → 証明状態 = AIが観測する「環境状態」 │
|
| 90 |
+
└──────────────────────────────────────────────┘
|
| 91 |
+
```
|
| 92 |
+
|
| 93 |
+
### 2.2 Mathlib4 — 最大の形式化数学ライブラリ
|
| 94 |
+
|
| 95 |
+
| 指標 | 数値(2026年5月時点) |
|
| 96 |
+
| :--- | :--- |
|
| 97 |
+
| **定理数** | 約 274,000 |
|
| 98 |
+
| **定義数** | 約 131,000 |
|
| 99 |
+
| **コード行数** | 210万行以上 |
|
| 100 |
+
| **貢献者数** | 770名以上 |
|
| 101 |
+
|
| 102 |
+
> [!IMPORTANT]
|
| 103 |
+
> **Mathlib4 は AI 証明探索の「訓練データ」であり「ツールボックス」でもあります。**
|
| 104 |
+
> - **訓練データ**: 数十万の定理-証明ペアは、AIプルーバーの教師あり微調整(SFT)の基礎データ
|
| 105 |
+
> - **ツールボックス**: 推論時に AI がライブラリ内の既存定理を検索・引用して証明を構成
|
| 106 |
+
> - **ベンチマーク**: Mathlib の難しい定理は、AI能力の測定に使用
|
| 107 |
+
|
| 108 |
+
---
|
| 109 |
+
|
| 110 |
+
## 3. LeanDojo:v1 → v2 の進化
|
| 111 |
+
|
| 112 |
+
### 3.1 LeanDojo v1(NeurIPS 2023 Oral)
|
| 113 |
+
|
| 114 |
+
> **論文**: *"LeanDojo: Theorem Proving with Retrieval-Augmented Language Models"*
|
| 115 |
+
> **著者**: Kaiyu Yang, Aidan M. Swope, Alex Gu, Rahul Chalamala, Peiyang Song, Shixing Yu, Saad Godil, Ryan Prenger, Anima Anandkumar
|
| 116 |
+
> **会議**: NeurIPS 2023 (Datasets and Benchmarks Track, **Oral**)
|
| 117 |
+
|
| 118 |
+
```
|
| 119 |
+
LeanDojo v1 のアーキテクチャ
|
| 120 |
+
┌─────────────────────────────────────────────────┐
|
| 121 |
+
│ │
|
| 122 |
+
│ ┌───────────────────┐ ┌───────────────────┐ │
|
| 123 |
+
│ │ リポジトリ・トレーサー │ │ プログラム的 │ │
|
| 124 |
+
│ │ │ │ インタラクション │ │
|
| 125 |
+
│ │ ・AST 抽出 │ │ │ │
|
| 126 |
+
│ │ ・依存関係 DAG │ │ ・Lean 環境操作 │ │
|
| 127 |
+
│ │ ・前提の定義位置 │ │ ・タクティクス送信 │ │
|
| 128 |
+
│ │ ・証明状態キャプチャ │ │ ・証明状態取得 │ │
|
| 129 |
+
│ └───────┬───────────┘ └────────┬──────────┘ │
|
| 130 |
+
│ │ │ │
|
| 131 |
+
│ ▼ ▼ │
|
| 132 |
+
│ ┌─────────────────────────────────────────┐ │
|
| 133 |
+
│ │ LeanDojo ベンチマーク │ │
|
| 134 |
+
│ │ (Lean 3/4 の train/val/test 分割) │ │
|
| 135 |
+
│ └─────────────────┬───────────────────────┘ │
|
| 136 |
+
│ │ │
|
| 137 |
+
│ ▼ │
|
| 138 |
+
│ ┌─────────────────────────────────────────┐ │
|
| 139 |
+
│ │ ReProver │ │
|
| 140 |
+
│ │ (検索拡張型定理証明器) │ │
|
| 141 |
+
│ │ │ │
|
| 142 |
+
│ │ Retriever → Encoder-Decoder LM │ │
|
| 143 |
+
│ │ (前提検索) (タクティクス生成) │ │
|
| 144 |
+
│ └──────────────────────────────────────────┘ │
|
| 145 |
+
└──────────────────────────────────────────────────┘
|
| 146 |
+
```
|
| 147 |
+
|
| 148 |
+
#### ReProver(Retrieval-Augmented Prover)の動作
|
| 149 |
+
|
| 150 |
+
```
|
| 151 |
+
証明目標: ∀ n : ℕ, 0 + n = n
|
| 152 |
+
│
|
| 153 |
+
▼
|
| 154 |
+
┌────────────────────────┐
|
| 155 |
+
│ 1. 前提検索 (Retriever) │
|
| 156 |
+
│ 「この目標に関連する │
|
| 157 |
+
│ Mathlib の定理は?」 │
|
| 158 |
+
│ → Nat.zero_add を発見 │
|
| 159 |
+
└──────────┬─────────────┘
|
| 160 |
+
│ 検索結果を連結
|
| 161 |
+
▼
|
| 162 |
+
┌────────────────────────┐
|
| 163 |
+
│ 2. タクティクス生成 │
|
| 164 |
+
│ (Encoder-Decoder) │
|
| 165 |
+
│ → 「exact Nat.zero_add n」│
|
| 166 |
+
└──────────┬─────────────┘
|
| 167 |
+
│
|
| 168 |
+
▼
|
| 169 |
+
┌────────────────────────┐
|
| 170 |
+
│ 3. Lean で検証 │
|
| 171 |
+
│ → ✅ 型チェック成功 │
|
| 172 |
+
└────────────────────────┘
|
| 173 |
+
```
|
| 174 |
+
|
| 175 |
+
### 3.2 LeanDojo v2(2025年)
|
| 176 |
+
|
| 177 |
+
| 特性 | LeanDojo v1 | LeanDojo v2 |
|
| 178 |
+
| :--- | :--- | :--- |
|
| 179 |
+
| **スコープ** | データ抽出 + 基本インタラクション | **統合的 E2E フレームワーク** |
|
| 180 |
+
| **データ管理** | 静的ベンチマーク | **ライフロング・データセット管理**(新定理を再トレースなしに追加可能) |
|
| 181 |
+
| **トレーナー** | — | **SFT, GRPO, Retrieval** トレーナーを内蔵 |
|
| 182 |
+
| **プルーバー** | ReProver のみ | **HF Prover, External Prover, LeanAgent** をサポート |
|
| 183 |
+
| **IDE** | なし | **Lean4Code**(軽量IDE)を同梱 |
|
| 184 |
+
| **モデル対応** | 固定 | **マルチモデル対応**(HuggingFace, 外部API) |
|
| 185 |
+
| **API 設計** | モノリシック | **モジュラーAPI**(トレース・訓練・評価を分離) |
|
| 186 |
+
| **ライセンス** | MIT | MIT |
|
| 187 |
+
|
| 188 |
+
### 3.3 LeanAgent(LeanDojo エコシステム内)
|
| 189 |
+
|
| 190 |
+
LeanDojo v2 と統合されたエージェント型プルーバー。**生涯学習(Lifelong Learning)**を実現。
|
| 191 |
+
|
| 192 |
+
```
|
| 193 |
+
LeanAgent のライフロング学習サイクル
|
| 194 |
+
┌────────────────────────────────────────���─┐
|
| 195 |
+
│ │
|
| 196 |
+
│ リポジトリ A リポジトリ B リポジトリ C │
|
| 197 |
+
│ (代数学) (位相幾何) (複素解析) │
|
| 198 |
+
│ │ │ │ │
|
| 199 |
+
│ └──────────────┼──────────────┘ │
|
| 200 |
+
│ │ │
|
| 201 |
+
│ ▼ │
|
| 202 |
+
│ ┌──────────────────┐ │
|
| 203 |
+
│ │ カリキュラム学習 │ │
|
| 204 |
+
│ │ (難易度順にソート) │ │
|
| 205 |
+
│ └────────┬─────────┘ │
|
| 206 |
+
│ │ │
|
| 207 |
+
│ ▼ │
|
| 208 |
+
│ ┌──────────────────┐ │
|
| 209 |
+
│ │ 漸進的訓練 │ │
|
| 210 |
+
│ │ (安定性-可塑性 │ │
|
| 211 |
+
│ │ バランス) │ │
|
| 212 |
+
│ └────────┬─────────┘ │
|
| 213 |
+
│ │ │
|
| 214 |
+
│ ▼ │
|
| 215 |
+
│ ┌──────────────────┐ │
|
| 216 |
+
│ │ 動的知識ベース │ │
|
| 217 |
+
│ │ 管理 │ │
|
| 218 |
+
│ └──────────────────┘ │
|
| 219 |
+
│ │
|
| 220 |
+
│ 成果: 23リポジトリで160以上の定理を証明 │
|
| 221 |
+
│ (人間未証明の定理を含む) │
|
| 222 |
+
└──────────────────────────────────────────┘
|
| 223 |
+
```
|
| 224 |
+
|
| 225 |
+
---
|
| 226 |
+
|
| 227 |
+
## 4. Lean Copilot:対話型AI証明アシスタント
|
| 228 |
+
|
| 229 |
+
### 4.1 アーキテクチャ
|
| 230 |
+
|
| 231 |
+
```
|
| 232 |
+
┌─────────────────────────────────────────────────────┐
|
| 233 |
+
│ Lean Copilot │
|
| 234 |
+
│ │
|
| 235 |
+
│ VS Code (Lean エディタ) │
|
| 236 |
+
│ ┌───────────────────────────────────┐ │
|
| 237 |
+
│ │ ユーザーが書いている Lean 証明 │ │
|
| 238 |
+
│ │ │ │
|
| 239 |
+
│ │ theorem foo : P → Q := by │ │
|
| 240 |
+
│ │ ▌ ← カーソル位置 │ │
|
| 241 |
+
│ └───────────────┬───────────────────┘ │
|
| 242 |
+
│ │ 現在の証明状態を抽出 │
|
| 243 |
+
│ ▼ │
|
| 244 |
+
│ ┌─────────────────────────────────────────┐ │
|
| 245 |
+
│ │ 3つの自動化機能 │ │
|
| 246 |
+
│ │ │ │
|
| 247 |
+
│ │ ┌─────────────┐ ┌──────────────────┐ │ │
|
| 248 |
+
│ │ │ suggest_tactics│ │ search_proofs │ │ │
|
| 249 |
+
│ │ │ (タクティクス │ │ (証明探索) │ │ │
|
| 250 |
+
│ │ │ 候補を提示) │ │ │ │ │
|
| 251 |
+
│ │ └──────────────┘ └──────────────────┘ │ │
|
| 252 |
+
│ │ ┌──────────────┐ │ │
|
| 253 |
+
│ │ │ select_premises│ │ │
|
| 254 |
+
│ │ │ (前提選択) │ │ │
|
| 255 |
+
│ │ └──────────────┘ │ │
|
| 256 |
+
│ └───────────────┬─────────────────────────┘ │
|
| 257 |
+
│ │ │
|
| 258 |
+
│ ▼ │
|
| 259 |
+
│ ┌──────────────────────────────────────────┐ │
|
| 260 |
+
│ │ 推論バックエンド │ │
|
| 261 |
+
│ │ │ │
|
| 262 |
+
│ │ ローカル クラウド │ │
|
| 263 |
+
│ │ ┌──────────┐ ┌──────────┐ │ │
|
| 264 |
+
│ │ │CTranslate2│ │ API │ │ │
|
| 265 |
+
│ │ │(C++ FFI) │ │(OpenAI等)│ │ │
|
| 266 |
+
│ │ └──────────┘ └──────────┘ │ │
|
| 267 |
+
│ └──────────────────────────────────────────┘ │
|
| 268 |
+
└─────────────────────────────────────────────────────┘
|
| 269 |
+
```
|
| 270 |
+
|
| 271 |
+
### 4.2 LeanDojo vs Lean Copilot の機能比較
|
| 272 |
+
|
| 273 |
+
| 特性 | LeanDojo v2 | Lean Copilot |
|
| 274 |
+
| :--- | :--- | :--- |
|
| 275 |
+
| **主な用途** | 研究者向け:AI プルーバーの訓練・評価パイプライン | 実務者向け:証明作成のリアルタイム支援 |
|
| 276 |
+
| **ターゲットユーザー** | ML 研究者、AI for Math 研究者 | 数学者、形式検証エンジニア |
|
| 277 |
+
| **データ抽出** | ✅ リポジトリトレース、AST、証明状態 | ❌ なし |
|
| 278 |
+
| **訓練機能** | ✅ SFT, GRPO, Retrieval トレーナー | ❌ なし(推論専用) |
|
| 279 |
+
| **IDE 統合** | Lean4Code(軽量) | ✅ VS Code 完全統合 |
|
| 280 |
+
| **タクティクス提案** | ❌(バッチ評価向け) | ✅ `suggest_tactics` |
|
| 281 |
+
| **証明探索** | ✅ BFS, MCTS 等 | ✅ `search_proofs`(aesop 統合) |
|
| 282 |
+
| **前提選択** | ✅ 検索拡張型 | ✅ `select_premises` |
|
| 283 |
+
| **モデルデプロイ** | バッチ処理向け | ローカル(FFI)/ クラウド(API) |
|
| 284 |
+
| **レイテンシ要件** | 非リアルタイム | ⚡ リアルタイム(ユーザー体験重視) |
|
| 285 |
+
| **オープンソース** | ✅ MIT | ✅ |
|
| 286 |
+
|
| 287 |
+
> [!TIP]
|
| 288 |
+
> **使い分けガイド**
|
| 289 |
+
> - **「新しいAIプルーバーを訓練・ベンチマークしたい」** → LeanDojo v2
|
| 290 |
+
> - **「日常の証明作成でAI支援がほしい」** → Lean Copilot
|
| 291 |
+
> - **「複数リポジトリにまたがって継続学習するエージェントを作りたい」** → LeanDojo v2 + LeanAgent
|
| 292 |
+
|
| 293 |
+
---
|
| 294 |
+
|
| 295 |
+
## 5. 主要AIプルーバー・モデルの技術比較
|
| 296 |
+
|
| 297 |
+
### 5.1 総合比較表
|
| 298 |
+
|
| 299 |
+
| 特性 | DeepSeek-Prover-V2 | Kimina-Prover-72B | InternLM2.5-StepProver | Goedel-Prover-V2 | LLEMMA | Seed-Prover 1.5 |
|
| 300 |
+
| :--- | :--- | :--- | :--- | :--- | :--- | :--- |
|
| 301 |
+
| **開発元** | DeepSeek | Numina + Kimi | SHLab (上海AI Lab) | Princeton | EleutherAI + Princeton | ByteDance |
|
| 302 |
+
| **ベースモデル** | DeepSeek-V3 | Qwen2.5-72B | InternLM2.5 | Qwen2.5 系 | Code Llama | 独自LLM |
|
| 303 |
+
| **パラメータ数** | **671B** (37B active) / 7B | **72B** | 7B | 8B / **32B** | 7B / 34B | 非公開 |
|
| 304 |
+
| **アーキテクチャ** | MoE | Dense Transformer | Dense Transformer | Dense Transformer | Dense Transformer | Dense Transformer |
|
| 305 |
+
| **形式環境** | Lean 4 | Lean 4 | Lean 4 | Lean 4 | Lean 4 / Isabelle | Lean 4 |
|
| 306 |
+
| **証明戦略** | 再帰的サブゴール分解 | 推論駆動探索 | Critic指導型探索 | スキャフォールド合成 | 基盤モデル(Few-shot) | エージェント型レンマ駆動 |
|
| 307 |
+
| **オープンソース** | ✅ | ✅ | ✅ | ✅ | ✅ | ❌ |
|
| 308 |
+
| **公開年** | 2025 | 2025 | 2024 | 2025 | 2023 | 2025 |
|
| 309 |
+
|
| 310 |
+
### 5.2 各モデルの技術詳細
|
| 311 |
+
|
| 312 |
+
---
|
| 313 |
+
|
| 314 |
+
#### DeepSeek-Prover-V2 — 再帰的サブゴール分解
|
| 315 |
+
|
| 316 |
+
```mermaid
|
| 317 |
+
graph TB
|
| 318 |
+
A["定理 T"] --> B["DeepSeek-V3 (671B)<br>非形式的証明スケッチ + Lean スケルトン"]
|
| 319 |
+
B --> C["Lean スケルトン<br>(sorry プレースホルダー付き)"]
|
| 320 |
+
C --> D["サブゴール S₁"]
|
| 321 |
+
C --> E["サブゴール S₂"]
|
| 322 |
+
C --> F["サブゴール S₃"]
|
| 323 |
+
|
| 324 |
+
D --> G["7B モデルが S₁ を証明"]
|
| 325 |
+
E --> H["7B モデルが S₂ を証明"]
|
| 326 |
+
F --> I["7B モデルが S₃ を証明<br>(再帰的にさらに分解も可能)"]
|
| 327 |
+
|
| 328 |
+
G & H & I --> J["統合: sorry を証明済みコードで置換"]
|
| 329 |
+
J --> K["Lean 4 コンパイラで全体検証"]
|
| 330 |
+
K --> L["✅ 完全な形式証明"]
|
| 331 |
+
|
| 332 |
+
style B fill:#6C5CE7,stroke:#333,color:#fff
|
| 333 |
+
style G fill:#00B894,stroke:#333,color:#fff
|
| 334 |
+
style H fill:#00B894,stroke:#333,color:#fff
|
| 335 |
+
style I fill:#00B894,stroke:#333,color:#fff
|
| 336 |
+
```
|
| 337 |
+
|
| 338 |
+
| 技術要素 | 詳細 |
|
| 339 |
+
| :--- | :--- |
|
| 340 |
+
| **コア思想** | 大モデルが「戦略」、小モデルが「戦術」を担当する分業型 |
|
| 341 |
+
| **MoE** | Mixture-of-Experts。671B 中 37B のみ活性化 → 推論効率が高い |
|
| 342 |
+
| **コンテキスト長** | 最大 163K トークン(長い証明チェーンに対応) |
|
| 343 |
+
| **MLA** | Multi-head Latent Attention。長距離依存関係の処理に優れる |
|
| 344 |
+
| **訓練** | Cold-start データ合成 → GRPO(強化学習) |
|
| 345 |
+
| **デュアルモード** | Non-CoT(高速)/ CoT(高精度)の切り替え |
|
| 346 |
+
| **一貫性報酬** | CoT の推論と形式コードの構造的一致を報酬として使�� |
|
| 347 |
+
|
| 348 |
+
---
|
| 349 |
+
|
| 350 |
+
#### Kimina-Prover-72B — 推論駆動探索
|
| 351 |
+
|
| 352 |
+
```
|
| 353 |
+
従来のアプローチ vs Kimina のアプローチ
|
| 354 |
+
|
| 355 |
+
従来: LLM + 外部探索アルゴリズム
|
| 356 |
+
┌──────────┐ ┌──────────┐
|
| 357 |
+
│ LLM │ → │ BFS / │ → 証明
|
| 358 |
+
│(タクティ │ │ MCTS │
|
| 359 |
+
│ クス生成)│ │(外部探索)│
|
| 360 |
+
└──────────┘ └──────────┘
|
| 361 |
+
|
| 362 |
+
Kimina: LLM が「探索も推論も」内部で実行
|
| 363 |
+
┌──────────────────────────────────┐
|
| 364 |
+
│ Kimina-Prover-72B │
|
| 365 |
+
│ │
|
| 366 |
+
│ <think> │
|
| 367 |
+
│ 非形式的推論... │
|
| 368 |
+
│ レンマの発見... │
|
| 369 |
+
│ エラー解釈 → 修正... │
|
| 370 |
+
│ </think> │
|
| 371 |
+
│ <lean> │
|
| 372 |
+
│ exact Nat.add_comm a b │
|
| 373 |
+
│ </lean> │
|
| 374 |
+
└──────────────────────────────────┘
|
| 375 |
+
```
|
| 376 |
+
|
| 377 |
+
| 技術要素 | 詳細 |
|
| 378 |
+
| :--- | :--- |
|
| 379 |
+
| **コア思想** | BFS/MCTS 等の外部探索を廃し、モデル内部の推論能力で探索を代替 |
|
| 380 |
+
| **Formal Reasoning Pattern** | `<think>...</think>` 内で非形式推論、`<lean>...</lean>` で形式コードを交互に生成 |
|
| 381 |
+
| **レンマパターン** | RL 中にランダムなレンマを挿入し、モデルがレンマの発見・再利用を学習 |
|
| 382 |
+
| **エラー修正** | Lean のコンパイルエラーメッセージを解釈し、ゼロから再生成せずに修正 |
|
| 383 |
+
| **多段階 RL** | SFT → RL(Kimi k1.5 ベース)→ 反復的精緻化 |
|
| 384 |
+
| **MiniF2F** | **92.2%** pass rate(リリース時 SOTA) |
|
| 385 |
+
|
| 386 |
+
---
|
| 387 |
+
|
| 388 |
+
#### InternLM2.5-StepProver — Critic 指導型探索
|
| 389 |
+
|
| 390 |
+
```
|
| 391 |
+
┌───────────────────────────────────────────────┐
|
| 392 |
+
│ InternLM2.5-StepProver │
|
| 393 |
+
│ │
|
| 394 |
+
│ ┌─────────────────┐ ┌─────────────────┐ │
|
| 395 |
+
│ │ Prover Model │ │ Critic Model │ │
|
| 396 |
+
│ │ (タクティクス │ │ (軌跡の選好 │ │
|
| 397 |
+
│ │ 生成) │ │ 評価) │ │
|
| 398 |
+
│ └────────┬────────┘ └────────┬────────┘ │
|
| 399 |
+
│ │ │ │
|
| 400 |
+
│ │ ┌────────────────┘ │
|
| 401 |
+
│ ▼ ▼ │
|
| 402 |
+
│ ┌─────────────────────────────┐ │
|
| 403 |
+
│ │ Critic 指導型 Best-First │ │
|
| 404 |
+
│ │ Search │ │
|
| 405 |
+
│ │ │ │
|
| 406 |
+
│ │ ・Prover が候補タクティクス生成 │ │
|
| 407 |
+
│ │ ・Critic が各軌跡の「有望さ」 │ │
|
| 408 |
+
│ │ を評価(選好情報を活用) │ │
|
| 409 |
+
│ │ ・有望な経路を優先的に探索 │ │
|
| 410 |
+
│ └──────────────────────────────┘ │
|
| 411 |
+
│ │
|
| 412 |
+
│ Expert Iteration (20,000 CPU-days): │
|
| 413 |
+
│ 証明成功 → データ化 → モデル更新 → 反復 │
|
| 414 |
+
└───────────────────────────────────────────────┘
|
| 415 |
+
```
|
| 416 |
+
|
| 417 |
+
| 技術要素 | 詳細 |
|
| 418 |
+
| :--- | :--- |
|
| 419 |
+
| **コア思想** | タクティクス軌跡の「選好情報」を学習した Critic モデルで探索を効率化 |
|
| 420 |
+
| **Expert Iteration** | 20,000 CPU日の大規模計算。証明成功 → データ → モデル改善 のサイクル |
|
| 421 |
+
| **Critic の役割** | 通常の BFS は各タクティクスの「累積ログ確率」で優先度を決定するが、Critic は軌跡全体の「証明到達可能性」を直接評価 |
|
| 422 |
+
| **深い証明の発見** | 標準 BFS が見逃す「初期ステップは低確率だが最終的に成功する」経路を発見 |
|
| 423 |
+
| **Log-Linear スケーリング** | 計算量と解決問題数の間にlog-linear 関係を発見 → スケーリング則 |
|
| 424 |
+
|
| 425 |
+
---
|
| 426 |
+
|
| 427 |
+
#### Goedel-Prover-V2 — スキャフォールド合成 + 自己修正
|
| 428 |
+
|
| 429 |
+
```
|
| 430 |
+
┌───────────────────────────────────────────┐
|
| 431 |
+
│ Goedel-Prover-V2 │
|
| 432 |
+
│ │
|
| 433 |
+
│ Innovation 1: スキャフォールド合成 │
|
| 434 |
+
│ ┌─────────────────────────────────┐ │
|
| 435 |
+
│ │ 失敗した証明試行から新しい │ │
|
| 436 |
+
│ │ サブ問題を構成 → 難易度を │ │
|
| 437 |
+
│ │ 段階的に上げた訓練タスクを合成 │ │
|
| 438 |
+
│ └─────────────────────────────────┘ │
|
| 439 |
+
│ │
|
| 440 |
+
│ Innovation 2: 検証器ガイド自己修正 │
|
| 441 |
+
│ ┌─────────────────────────────────┐ │
|
| 442 |
+
│ │ 証明生成 → Lean コンパイル │ │
|
| 443 |
+
│ │ → エラー → エラーメッセージで修正 │ │
|
| 444 |
+
│ │ → 再コンパイル → (反復) │ │
|
| 445 |
+
│ └─────────────────────────────────┘ │
|
| 446 |
+
│ │
|
| 447 |
+
│ Innovation 3: モデル平均化 │
|
| 448 |
+
│ ┌─────────────────────────────────┐ │
|
| 449 |
+
│ │ 複数チェックポイントの重みを平均 │ │
|
| 450 |
+
│ │ → 出力の多様性を維持 │ │
|
| 451 |
+
│ │ → 深い微調整による性能劣化を防止 │ │
|
| 452 |
+
│ └─────────────────────────────────┘ │
|
| 453 |
+
│ │
|
| 454 |
+
│ PutnamBench: 86問 (32B, pass@184) │
|
| 455 |
+
│ → DeepSeek-Prover-V2-671B (47問) を大幅凌駕│
|
| 456 |
+
└───────────────────────────────────────────┘
|
| 457 |
+
```
|
| 458 |
+
|
| 459 |
+
---
|
| 460 |
+
|
| 461 |
+
#### LLEMMA — オープンソース数学基盤モデル
|
| 462 |
+
|
| 463 |
+
| 技術要素 | 詳細 |
|
| 464 |
+
| :--- | :--- |
|
| 465 |
+
| **位置づけ** | 特化型プルーバーではなく、数学推論のための**汎用基盤モデル** |
|
| 466 |
+
| **ベース** | Code Llama (Llama 2 系) |
|
| 467 |
+
| **訓練データ** | Proof-Pile-2(55B トークン): arXiv (29B) + OpenWebMath (15B) + AlgebraicStack (11B) |
|
| 468 |
+
| **特徴** | 追加微調整なしで few-shot タクティクス予測が可能。ツール使用(Python, CAS)に対応 |
|
| 469 |
+
| **意義** | 下流の特化モデル(ReProver 等)のための**事前訓練済みベース**を提供 |
|
| 470 |
+
|
| 471 |
+
---
|
| 472 |
+
|
| 473 |
+
### 5.3 探索戦略の比較表
|
| 474 |
+
|
| 475 |
+
| モデル | 探索戦略 | 長所 | 短所 |
|
| 476 |
+
| :--- | :--- | :--- | :--- |
|
| 477 |
+
| **DeepSeek-Prover-V2** | 再帰的分解 + sorry 埋め | 大問題を管理可能な単位に分割。大小モデルの分業 | 分解の質が大モデルの非形式推論に依存 |
|
| 478 |
+
| **Kimina-Prover-72B** | 推論駆動(探索を内部化) | 外部探索オーバーヘッドなし。72Bでのスケーリング効果 | 計算量がモデルサイズに依存 |
|
| 479 |
+
| **InternLM2.5-StepProver** | Critic 指導型 BFS | 深い証明の発見。選好学習で効率化 | Critic 訓練に大規模計算が必要 |
|
| 480 |
+
| **Goedel-Prover-V2** | スキャフォールド + 自己修正 | 小型モデル(8B/32B)でも高性能。モデル平均化で安定 | 検証器フィードバックループの速度 |
|
| 481 |
+
| **Seed-Prover 1.5** | エージェント型(ツール統合) | Python実行・Mathlib検索。最も柔軟 | システム全体の複雑性が高い |
|
| 482 |
+
|
| 483 |
+
---
|
| 484 |
+
|
| 485 |
+
## 6. 証明生成戦略の分類
|
| 486 |
+
|
| 487 |
+
AI証明生成の主要な戦略を体系的に分類します。
|
| 488 |
+
|
| 489 |
+
### 6.1 戦略マップ
|
| 490 |
+
|
| 491 |
+
```mermaid
|
| 492 |
+
graph TD
|
| 493 |
+
S["証明生成戦略"] --> S1["全証明一括生成<br>(Whole-Proof)"]
|
| 494 |
+
S --> S2["ステップバイステップ<br>(Tactic-by-Tactic)"]
|
| 495 |
+
S --> S3["階層的分解<br>(Hierarchical)"]
|
| 496 |
+
S --> S4["ライブラリ学習<br>(Library Learning)"]
|
| 497 |
+
|
| 498 |
+
S1 --> S1a["Baldur"]
|
| 499 |
+
S2 --> S2a["ReProver"]
|
| 500 |
+
S2 --> S2b["InternLM2.5-StepProver"]
|
| 501 |
+
S3 --> S3a["Draft-Sketch-Prove"]
|
| 502 |
+
S3 --> S3b["DeepSeek-Prover-V2"]
|
| 503 |
+
S3 --> S3c["Seed-Prover"]
|
| 504 |
+
S4 --> S4a["LEGO-Prover"]
|
| 505 |
+
|
| 506 |
+
style S1 fill:#E17055,stroke:#333,color:#fff
|
| 507 |
+
style S2 fill:#FDCB6E,stroke:#333
|
| 508 |
+
style S3 fill:#00B894,stroke:#333,color:#fff
|
| 509 |
+
style S4 fill:#0984E3,stroke:#333,color:#fff
|
| 510 |
+
```
|
| 511 |
+
|
| 512 |
+
### 6.2 戦略比較表
|
| 513 |
+
|
| 514 |
+
| 戦略 | 代表的システム | 動作原理 | 固有の強み | 固有の弱点 |
|
| 515 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 516 |
+
| **全証明一括生成** | Baldur | LLMが証明全体を一度に生成。エラー時は修復 | 探索オーバーヘッドなし。高速 | 長い証明で精度低下。構造的洞察が限定的 |
|
| 517 |
+
| **Draft-Sketch-Prove** | DSP | 非形式草案 → 形式スケルトン → Hammer で補完 | 人間的推論と形式的精密さの融合 | パイプライン段階間の翻訳エラー |
|
| 518 |
+
| **ステップバイステップ** | ReProver, StepProver | 各タクティクスを逐次生成 + 探索(BFS/MCTS) | 最も柔軟。中間状態で軌道修正可能 | 長い証明で探索空間が爆発 |
|
| 519 |
+
| **再帰的分解** | DeepSeek-Prover-V2 | sorry 付きスケルトン → サブゴールを再帰的に解決 | 大問題への自然な対処。並列化可能 | 分解の質が全体性能を決定 |
|
| 520 |
+
| **レンマ駆動型** | Seed-Prover, Kimina | 自然言語推論 → レンマの発見 → 形式化 | 人間的証明スタイル。再利用可能なレンマ蓄積 | レンマの正確な形式化が困難 |
|
| 521 |
+
| **ライブラリ学習** | LEGO-Prover | 検証済みレンマの「スキルライブラリ」を漸進的に構築 | 知識の蓄積・再利用。複雑問題への対応力 | 計算予算を揃えると効果が減少する報告あり |
|
| 522 |
+
|
| 523 |
+
---
|
| 524 |
+
|
| 525 |
+
## 7. ベンチマークと評価
|
| 526 |
+
|
| 527 |
+
### 7.1 主要ベンチマーク概要
|
| 528 |
+
|
| 529 |
+
| ベンチマーク | 難易度 | 問題数 | 出題源 | 形式環境 | 測定内容 |
|
| 530 |
+
| :--- | :--- | :--- | :--- | :--- | :--- |
|
| 531 |
+
| **MiniF2F** | 🟡 高校〜オリンピック | 488 (244 test) | AMC, AIME, IMO | Lean 4 / Isabelle | Olympiad レベルの形式証明能力 |
|
| 532 |
+
| **ProofNet** | 🟡 学部レベル | 371 | 大学数学教科書 | Lean 3/4 | 自動形式化 + 証明 |
|
| 533 |
+
| **PutnamBench** | 🔴 大学上級 | 672 | Putnam Competition (1962–2025) | Lean 4 | 高難度の形式証明 |
|
| 534 |
+
| **Lean-Workbook-Plus** | 🟡 多様 | 48,809 | 数学フォーラム・教科書 | Lean 4 | 大規模な証明能力 |
|
| 535 |
+
|
| 536 |
+
### 7.2 MiniF2F-test 性能比較(2025年時点)
|
| 537 |
+
|
| 538 |
+
| モデル | パラメータ | Pass Rate | 備考 |
|
| 539 |
+
| :---: | :---: | :---: | :--- |
|
| 540 |
+
| **Kimina-Prover-72B** | 72B | **~92.2%** | 推論駆動探索。リリース時 SOTA |
|
| 541 |
+
| **Goedel-Prover-V2-32B** | 32B | ~90.4% | 自己修正モード。pass@32 |
|
| 542 |
+
| **Goedel-Prover-V2-8B** | 8B | ~84.6% | pass@32。小型ながら高性能 |
|
| 543 |
+
| **DeepSeek-Prover-V2-671B** | 671B (37B active) | ~88.9% | MoE。非形式+形式融合 |
|
| 544 |
+
| **InternLM2.5-StepProver** | 7B | ~65.9% | Critic 指導型。7B としては高水準 |
|
| 545 |
+
| **ReProver (LeanDojo)** | ~1B | ~26.5% | 検索拡張型。2023年の基準 |
|
| 546 |
+
|
| 547 |
+
> [!WARNING]
|
| 548 |
+
> **ベンチマーク飽和への注意**: MiniF2F は 90% を超えるモデルが複数登場しており、飽和が指摘されています。研究コミュニティは **PutnamBench** や **VeriBench-FTP** 等のより困難なベンチマークへの移行を進めています。
|
| 549 |
+
|
| 550 |
+
### 7.3 PutnamBench 性能比較
|
| 551 |
+
|
| 552 |
+
| モデル / システム | 解答数 | Pass 条件 | 備考 |
|
| 553 |
+
| :--- | :---: | :--- | :--- |
|
| 554 |
+
| **Aleph + GPT-5.2** (Logical Intelligence) | **668/672** | pass@? | エージェント型。2026年初SOTA |
|
| 555 |
+
| **Goedel-Prover-V2-32B** | 86 | pass@184 | オープンソース1位(2025年中盤) |
|
| 556 |
+
| **DeepSeek-Prover-V2-671B** | 47 | — | MoE 671B |
|
| 557 |
+
|
| 558 |
+
---
|
| 559 |
+
|
| 560 |
+
## 8. 訓練手法の比較
|
| 561 |
+
|
| 562 |
+
### 8.1 訓練パイプラインの分類
|
| 563 |
+
|
| 564 |
+
```mermaid
|
| 565 |
+
graph LR
|
| 566 |
+
subgraph "Phase 1: 事前訓練"
|
| 567 |
+
P1["汎用LLM事前訓練<br>(数学・コードデータ)"]
|
| 568 |
+
end
|
| 569 |
+
subgraph "Phase 2: 教師あり微調整 (SFT)"
|
| 570 |
+
P2["Mathlib 証明データ<br>+ 合成データで微調整"]
|
| 571 |
+
end
|
| 572 |
+
subgraph "Phase 3: 強化学習 (RL)"
|
| 573 |
+
P3["GRPO / Expert Iteration<br>Lean 検証器が報酬"]
|
| 574 |
+
end
|
| 575 |
+
subgraph "Phase 4: テスト時適応"
|
| 576 |
+
P4["TTRL / TTT<br>問題固有の追加学習"]
|
| 577 |
+
end
|
| 578 |
+
|
| 579 |
+
P1 --> P2 --> P3 --> P4
|
| 580 |
+
```
|
| 581 |
+
|
| 582 |
+
### 8.2 強化学習手法の比較
|
| 583 |
+
|
| 584 |
+
| 手法 | 使用モデル | 動作原理 | 報酬信号 | 特徴 |
|
| 585 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 586 |
+
| **GRPO** | DeepSeek-Prover-V2, Goedel-Prover | グループ内の相対的性能で方策を最適化。Value モデル不要 | Lean コンパイル成功/失敗 | メモリ効率が高い。大規模モデル向き |
|
| 587 |
+
| **Expert Iteration** | InternLM2.5-StepProver | 証明成功 → データ化 → モデル更新 → 再探索 を反復 | 証明発見の成否 | 大規模計算(20,000 CPU日)が必要 |
|
| 588 |
+
| **Kimi k1.5 RL** | Kimina-Prover-72B | 多段階 RL。候補証明のロールアウトを Lean で検証 | Lean バイナリ報酬 | 外部探索なしで推論内部でRL |
|
| 589 |
+
| **TTRL / TTT** | AlphaProof, Aristotle | テスト時に問題固有の追加 RL/勾配更新 | 問題バリエーションの証明成否 | 計算コスト極大だが最高性能 |
|
| 590 |
+
|
| 591 |
+
### 8.3 訓練データソースの比較
|
| 592 |
+
|
| 593 |
+
| データソース | 種類 | 規模 | 使用モデル |
|
| 594 |
+
| :--- | :--- | :--- | :--- |
|
| 595 |
+
| **Mathlib4** | 人間作成の形式証明 | ~274K 定理 | ほぼ全モデルの SFT 基礎データ |
|
| 596 |
+
| **Proof-Pile-2** | arXiv + Web + コード | 55B トークン | LLEMMA |
|
| 597 |
+
| **合成データ(Lean検証済み)** | RL 中に自動生成 | 数百万〜数千万件 | DeepSeek, Kimina, Goedel |
|
| 598 |
+
| **Lean-Workbook** | フォーラム由来の問�� | ~48K 問題 | InternLM StepProver 他 |
|
| 599 |
+
| **FoVer-40K** | FOL 検証済みラベル | ~40K | PRM 訓練用 |
|
| 600 |
+
|
| 601 |
+
---
|
| 602 |
+
|
| 603 |
+
## 9. 未解決課題と今後の展望
|
| 604 |
+
|
| 605 |
+
### 9.1 現在の技術的課題
|
| 606 |
+
|
| 607 |
+
| 課題 | 説明 | 影響 |
|
| 608 |
+
| :--- | :--- | :--- |
|
| 609 |
+
| **ベンチマーク飽和** | MiniF2F が 90%+ に到達。真の能力測定が困難に | PutnamBench, VeriBench 等への移行 |
|
| 610 |
+
| **形式/非形式のギャップ** | ベンチマーク問題の形式化自体に誤りがある場合がある | 評価の信頼性に影響 |
|
| 611 |
+
| **計算コスト** | Expert Iteration: 20,000 CPU日。TTRL: TPU数日 | 再現性・アクセシビリティの障壁 |
|
| 612 |
+
| **長い証明** | 100ステップ以上の証明は依然として困難 | 研究レベルの数学への適用障壁 |
|
| 613 |
+
| **Mathlib 追従** | Mathlib の頻繁な更新にモデルが追いつけない | LeanDojo v2 のライフロング管理で対処中 |
|
| 614 |
+
| **汎化** | 特定ベンチマーク最適化 vs 真の数学的推論能力 | ベンチマーク設計の課題 |
|
| 615 |
+
|
| 616 |
+
### 9.2 今後の方向性
|
| 617 |
+
|
| 618 |
+
```mermaid
|
| 619 |
+
graph TD
|
| 620 |
+
F["今後の研究方向"] --> F1["エージェント型証明"]
|
| 621 |
+
F --> F2["生涯学習"]
|
| 622 |
+
F --> F3["マルチモーダル"]
|
| 623 |
+
F --> F4["人間-AI協調"]
|
| 624 |
+
|
| 625 |
+
F1 --> F1a["ツール使用の高度化<br>(Python, CAS, Web検索)"]
|
| 626 |
+
F2 --> F2a["新しいMathlib定理に<br>動的に適応"]
|
| 627 |
+
F3 --> F3a["図・数式画像からの<br>自動形式化"]
|
| 628 |
+
F4 --> F4a["数学者の日常的な<br>証明作業を支援"]
|
| 629 |
+
```
|
| 630 |
+
|
| 631 |
+
---
|
| 632 |
+
|
| 633 |
+
## 10. 文献一覧
|
| 634 |
+
|
| 635 |
+
### ツール・フレームワーク
|
| 636 |
+
|
| 637 |
+
| 論文 | 著者 | 会議 | 年 |
|
| 638 |
+
| :--- | :--- | :--- | :--- |
|
| 639 |
+
| LeanDojo: Theorem Proving with Retrieval-Augmented LMs | Yang, Swope et al. | NeurIPS 2023 (Oral) | 2023 |
|
| 640 |
+
| Lean Copilot: LLMs as Copilots for Theorem Proving | Song et al. | NeurIPS 2024 | 2024 |
|
| 641 |
+
| LeanAgent: Lifelong Learning for Formal Theorem Proving | Kumarappan et al. | ICLR 2025 (Oral) | 2025 |
|
| 642 |
+
| LeanDojo v2: A Comprehensive Library for AI Theorem Proving | Yang et al. | — | 2025 |
|
| 643 |
+
|
| 644 |
+
### AIプルーバー・モデル
|
| 645 |
+
|
| 646 |
+
| 論文 | 開発元 | 年 |
|
| 647 |
+
| :--- | :--- | :--- |
|
| 648 |
+
| DeepSeek-Prover-V2: Subgoal Decomposition for Theorem Proving | DeepSeek | 2025 |
|
| 649 |
+
| Kimina-Prover: Reasoning-Driven Exploration for Formal Reasoning | Numina + Kimi | 2025 |
|
| 650 |
+
| InternLM2.5-StepProver: Expert Iteration on Lean Workbook | SHLab | 2024 |
|
| 651 |
+
| Goedel-Prover-V2: Scaling Open-Source Formal Reasoning | Princeton | 2025 |
|
| 652 |
+
| LLEMMA: An Open Language Model for Mathematics | EleutherAI + Princeton | 2023 |
|
| 653 |
+
| Seed-Prover: Automating Formal Proof with Agentic LLM | ByteDance | 2025 |
|
| 654 |
+
|
| 655 |
+
### 証明生成戦略
|
| 656 |
+
|
| 657 |
+
| 論文 | 年 |
|
| 658 |
+
| :--- | :--- |
|
| 659 |
+
| Draft, Sketch, and Prove: Guiding Formal Theorem Provers with Informal Proofs | 2023 |
|
| 660 |
+
| LEGO-Prover: Neural Theorem Proving with Growing Libraries | 2024 |
|
| 661 |
+
| Baldur: Whole-Proof Generation and Repair with Large Language Models | 2023 |
|
| 662 |
+
|
| 663 |
+
### 基盤モデル・ベンチマーク
|
| 664 |
+
|
| 665 |
+
| 名称 | 年 | 内容 |
|
| 666 |
+
| :--- | :--- | :--- |
|
| 667 |
+
| MiniF2F | 2021 | AMC/AIME/IMO レベルの形式数学ベンチマーク |
|
| 668 |
+
| ProofNet | 2023 | 学部レベル自動形式化ベンチマーク |
|
| 669 |
+
| PutnamBench | 2024 | Putnam Competition ベースの高難度ベンチマーク |
|
| 670 |
+
| Lean-Workbook-Plus | 2024 | 大規模 Lean 証明データセット |
|
docs/AI_Agents_and_Logic/ai_proof_search_tools_survey.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "c3dea20d559b01a1", "ingested_at": "2026-06-30T08:25:55.025982", "content_hash": "c3dea20d559b01a1"}
|
docs/AI_Agents_and_Logic/neurosymbolic_ai_survey.md
ADDED
|
@@ -0,0 +1,719 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# ニューロ・シンボリックAI(Neuro-Symbolic AI)包括的サーベイ
|
| 2 |
+
|
| 3 |
+
> **調査時点**: 2026年5月
|
| 4 |
+
> **主題**: LLMの弱点を形式論理で克服するニューロ・シンボリックAIの体系的分類と最新フレームワーク
|
| 5 |
+
> **対象読者**: AI研究者・MLエンジニア・形式手法に関心のある技術者
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 目次
|
| 10 |
+
|
| 11 |
+
1. [背景と動機:なぜニューロ・シンボリックAIか](#1-背景と動機なぜニューロシンボリックaika)
|
| 12 |
+
2. [学術的分類体系:Kautzの6類型と3パラダイム](#2-学術的分類体系kautzの6類型と3パラダイム)
|
| 13 |
+
3. [パラダイム1: Symbolic → Neural(記号がニューラルを導く)](#3-パラダイム1-symbolic--neural記号がニューラルを導く)
|
| 14 |
+
4. [パラダイム2: Neural → Symbolic(ニューラルが記号に翻訳する)](#4-パラダイム2-neural--symbolicニューラルが記号に翻訳する)
|
| 15 |
+
5. [パラダイム3: Hybrid(双方向統合)](#5-パラダイム3-hybrid双方向統合)
|
| 16 |
+
6. [微分可能ニューロ・シンボリック統合(Tightly Coupled)](#6-微分可能ニューロシンボリック統合tightly-coupled)
|
| 17 |
+
7. [応用ドメイン](#7-応用ドメイン)
|
| 18 |
+
8. [ベンチマークと評価](#8-ベンチマークと評価)
|
| 19 |
+
9. [未解決課題と今後の展望](#9-未解決課題と今後の展望)
|
| 20 |
+
10. [主要文献一覧](#10-主要文献一覧)
|
| 21 |
+
|
| 22 |
+
---
|
| 23 |
+
|
| 24 |
+
## 1. 背景と動機:なぜニューロ・シンボリックAIか
|
| 25 |
+
|
| 26 |
+
### 1.1 LLMの構造的限界
|
| 27 |
+
|
| 28 |
+
大規模言語モデル(LLM)は、「統計的パターン照合」に基づく **System 1(直感的・高速)** 型のAIです。自然言語処理、コード生成、創造的文章の執筆では驚異的な能力を示しますが、以下の構造的限界を抱えています。
|
| 29 |
+
|
| 30 |
+
```
|
| 31 |
+
┌──────────────────────────────────────────────────────────────────┐
|
| 32 |
+
│ LLM の構造的限界 │
|
| 33 |
+
├──────────────┬───────────────────────────────────────────────────┤
|
| 34 |
+
│ ハルシネーション │ もっともらしい嘘を自信満々に生成する。 │
|
| 35 |
+
│ │ 「次のトークン予測」の確率的性質に起因。 │
|
| 36 |
+
├──────────────┼───────────────────────────────────────────────────┤
|
| 37 |
+
│ 論理的推論の │ 多段推論(5段以上)で急激に精度が低下。 │
|
| 38 |
+
│ 脆弱性 │ 論理的「ショートカット」を取りがち。 │
|
| 39 |
+
├──────────────┼───────────────────────────────────────────────────┤
|
| 40 |
+
│ 検証不可能性 │ 推論過程(CoT)が「もっともらしい物語」に過ぎず、 │
|
| 41 |
+
│ │ 論理的正しさが保証されない。 │
|
| 42 |
+
├──────────────┼───────────────────────────────────────────────────┤
|
| 43 |
+
│ 永続的状態の │ トランスフォーマーは外部ワーキングメモリや │
|
| 44 |
+
│ 欠如 │ 変数束縛のメカニズムを持たない。 │
|
| 45 |
+
├──────────────┼───────────────────────────────────────────────────┤
|
| 46 |
+
│ 体系的一般化 │ 訓練データの分布外に出ると急激に性能低下。 │
|
| 47 |
+
│ の欠如 │ 構成的(compositional)な一般化が困難。 │
|
| 48 |
+
└──────────────┴───────────────────────────────────────────────────┘
|
| 49 |
+
```
|
| 50 |
+
|
| 51 |
+
### 1.2 シンボリックAIの強みと弱み
|
| 52 |
+
|
| 53 |
+
一方、シンボリックAI(形式論理、定理証明器、制約ソルバー等)は **System 2(熟慮的・正確)** 型のAIであり、以下の特徴を持ちます。
|
| 54 |
+
|
| 55 |
+
| 強み | 弱み |
|
| 56 |
+
| :--- | :--- |
|
| 57 |
+
| ✅ 推論結果の**正確性**が数学的に保証される | ❌ 自然言語のような曖昧な入力の処理が困難 |
|
| 58 |
+
| ✅ 推論過程が**完全に追跡可能**(説明可能) | ❌ 知識の手動エンコーディングが高コスト |
|
| 59 |
+
| ✅ 体系的な組み合わせ的一般化が可能 | ❌ 非構造化データ(画像、音声)を扱えない |
|
| 60 |
+
| ✅ 矛盾を検出し、一貫性を保証でき��� | ❌ 不完全情報やノイズへの耐性が低い |
|
| 61 |
+
|
| 62 |
+
### 1.3 ニューロ・シンボリックAIの核心的アイデア
|
| 63 |
+
|
| 64 |
+
> [!IMPORTANT]
|
| 65 |
+
> **ニューロ・シンボリックAI = LLMの「直感」 × 形式論理の「厳密性」**
|
| 66 |
+
>
|
| 67 |
+
> ニューラルネットワーク(System 1)とシンボリック推論(System 2)を統合することで、両者の弱点を相互に補完し、**直感的かつ正確で検証可能なAI**を実現する。これがニューロ・シンボリックAIの本質です。
|
| 68 |
+
|
| 69 |
+
```mermaid
|
| 70 |
+
graph LR
|
| 71 |
+
subgraph "System 1: Neural(直感・認識)"
|
| 72 |
+
N["LLM / ニューラルネット"]
|
| 73 |
+
end
|
| 74 |
+
subgraph "System 2: Symbolic(論理・検証)"
|
| 75 |
+
S["定理証明器 / ソルバー / ルールエンジン"]
|
| 76 |
+
end
|
| 77 |
+
N <-->|"統合"| S
|
| 78 |
+
N --> R1["パターン認識<br>自然言語理解<br>帰納的学習"]
|
| 79 |
+
S --> R2["論理的推論<br>形式検証<br>一貫性保証"]
|
| 80 |
+
|
| 81 |
+
style N fill:#6C5CE7,stroke:#333,color:#fff
|
| 82 |
+
style S fill:#00B894,stroke:#333,color:#fff
|
| 83 |
+
```
|
| 84 |
+
|
| 85 |
+
---
|
| 86 |
+
|
| 87 |
+
## 2. 学術的分類体系:Kautzの6類型と3パラダイム
|
| 88 |
+
|
| 89 |
+
### 2.1 Henry Kautz の6類型(AAAI 2020 / AI Magazine 2022)
|
| 90 |
+
|
| 91 |
+
ニューロ・シンボリックAIの学術的分類として最も影響力のあるフレームワークは、Henry Kautz が AAAI 2020 の Robert S. Engelmore Memorial Award Lecture「The Third AI Summer」で提唱し、2022年に *AI Magazine* で体系化した **6類型** です。
|
| 92 |
+
|
| 93 |
+
これらは、ニューラルとシンボリックの「統合の密度(Tightness of Coupling)」に基づくスペクトラム上に配置されます。
|
| 94 |
+
|
| 95 |
+
```mermaid
|
| 96 |
+
graph TB
|
| 97 |
+
subgraph "疎結合(Loosely Coupled)"
|
| 98 |
+
T1["Type 1: Symbolic ⇌ Neural<br>記号入出力のニューラル<br>例: GPT, BERT"]
|
| 99 |
+
T2["Type 2: Symbolic[Neural]<br>記号システムがニューラルを呼出<br>例: AlphaGo"]
|
| 100 |
+
T3["Type 3: Neural[Symbolic]<br>ニューラルが記号システムを呼出<br>例: Logic-LM, LINC"]
|
| 101 |
+
end
|
| 102 |
+
subgraph "密結合(Tightly Coupled)"
|
| 103 |
+
T4["Type 4: Neural ∪ Compile(Symbolic)<br>記号知識をネットワークにコンパイル<br>例: KBANN"]
|
| 104 |
+
T5["Type 5: Neural_{Symbolic}<br>論理を埋め込み・損失関数に統合<br>例: LTN, LNN"]
|
| 105 |
+
T6["Type 6: Neural ≡ Symbolic<br>完全統合エンジン<br>(理想目標)"]
|
| 106 |
+
end
|
| 107 |
+
T1 --> T2 --> T3 --> T4 --> T5 --> T6
|
| 108 |
+
|
| 109 |
+
style T6 fill:#e17055,stroke:#333,color:#fff
|
| 110 |
+
```
|
| 111 |
+
|
| 112 |
+
| 類型 | 名称 | 説明 | 代表例 |
|
| 113 |
+
| :---: | :--- | :--- | :--- |
|
| 114 |
+
| **Type 1** | Symbolic ⇌ Neural | 記号(テキストトークン)を入出力とするニューラルネット。事実上の標準LLM | GPT, BERT |
|
| 115 |
+
| **Type 2** | Symbolic[Neural] | シンボリックシステム(探索、計画)がニューラルをサブルーチンとして呼出 | AlphaGo, AlphaProof |
|
| 116 |
+
| **Type 3** | Neural[Symbolic] | ニューラルが外部シンボリックシステム(ソルバー、証明器)を呼出 | Logic-LM, LINC, SatLM |
|
| 117 |
+
| **Type 4** | Neural ∪ Compile(Sym) | シンボリック知識をネットワークのトポロジーや初期重みにコンパイル | KBANN |
|
| 118 |
+
| **Type 5** | Neural_{Symbolic} | 論理をエンベディングや損失関数制約に写像。エンドツーエンド微分可能 | LTN, LNN (IBM) |
|
| 119 |
+
| **Type 6** | Neural ≡ Symbolic | ニューラルとシンボリックが完全に統合された単一アーキテクチャ | **理想目標(未達成)** |
|
| 120 |
+
|
| 121 |
+
### 2.2 LLM時代の3パラダイム(IJCAI 2025 サーベイ等)
|
| 122 |
+
|
| 123 |
+
LLMの登場以降(2023–2025年)、上記の6類型をLLM中心に再編成した**3パラダイム分類**が主流となっています。
|
| 124 |
+
|
| 125 |
+
````carousel
|
| 126 |
+
### パラダイム1: Symbolic → Neural
|
| 127 |
+
**記号知識がニューラルを導く**
|
| 128 |
+
|
| 129 |
+
```
|
| 130 |
+
┌──────────────────┐
|
| 131 |
+
│ シンボリック知識 │
|
| 132 |
+
│ (ルール・制約 │
|
| 133 |
+
│ オントロジー) │
|
| 134 |
+
└────────┬─────────┘
|
| 135 |
+
│ ガイド・制約・正則化
|
| 136 |
+
▼
|
| 137 |
+
┌──────────────────┐
|
| 138 |
+
│ LLM │
|
| 139 |
+
│ (訓練 or 推論) │
|
| 140 |
+
└──────────────────┘
|
| 141 |
+
```
|
| 142 |
+
|
| 143 |
+
- 論理制約をネットワーク訓練に埋め込む
|
| 144 |
+
- 推論時に文法的・論理的制約でデコーディングを制御
|
| 145 |
+
- 例: LTN, LNN, 制約付きデコーディング(Outlines, LMQL)
|
| 146 |
+
<!-- slide -->
|
| 147 |
+
### パラダイム2: Neural → Symbolic
|
| 148 |
+
**ニューラルが形式言語に翻訳し、ソルバーが解く**
|
| 149 |
+
|
| 150 |
+
```
|
| 151 |
+
┌──────────────────┐
|
| 152 |
+
│ 自然言語の問題 │
|
| 153 |
+
└────────┬─────────┘
|
| 154 |
+
│
|
| 155 |
+
▼
|
| 156 |
+
┌──────────────────┐
|
| 157 |
+
│ LLM (翻訳器) │ ← 意味解析
|
| 158 |
+
└────────┬─────────┘
|
| 159 |
+
│ FOL / SAT / SMT
|
| 160 |
+
���
|
| 161 |
+
┌──────────────────┐
|
| 162 |
+
│ シンボリック推論器 │ ← Z3, Prover9 等
|
| 163 |
+
└────────┬─────────┘
|
| 164 |
+
│
|
| 165 |
+
▼
|
| 166 |
+
┌──────────────────┐
|
| 167 |
+
│ 確定的・検証済み │
|
| 168 |
+
│ の回答 │
|
| 169 |
+
└──────────────────┘
|
| 170 |
+
```
|
| 171 |
+
|
| 172 |
+
- LLMを「意味解析器」として使用
|
| 173 |
+
- 推論は確定的ソルバーが担当
|
| 174 |
+
- 例: Logic-LM, LINC, SatLM
|
| 175 |
+
<!-- slide -->
|
| 176 |
+
### パラダイム3: Hybrid(双方向)
|
| 177 |
+
**ニューラルと記号が反復的に対話・批判・修正**
|
| 178 |
+
|
| 179 |
+
```
|
| 180 |
+
┌──────────────────┐
|
| 181 |
+
│ LLM (推測) │
|
| 182 |
+
└────────┬─────────┘
|
| 183 |
+
│ 推測を提出
|
| 184 |
+
▼
|
| 185 |
+
┌──────────────────┐
|
| 186 |
+
│ 記号的検証器 │
|
| 187 |
+
│ (Hard Critique) │
|
| 188 |
+
└────────┬─────────┘
|
| 189 |
+
│ フィードバック
|
| 190 |
+
▼
|
| 191 |
+
┌──────────────────┐
|
| 192 |
+
│ LLM (修正) │
|
| 193 |
+
└──────────────────┘
|
| 194 |
+
↕ 反復
|
| 195 |
+
```
|
| 196 |
+
|
| 197 |
+
- 推測–批判(Conjecture-Criticism)サイクル
|
| 198 |
+
- エラーメッセージによる自己修正ループ
|
| 199 |
+
- 例: PEIRCE, Logic-LM++, FoVer
|
| 200 |
+
````
|
| 201 |
+
|
| 202 |
+
---
|
| 203 |
+
|
| 204 |
+
## 3. パラダイム1: Symbolic → Neural(記号がニューラルを導く)
|
| 205 |
+
|
| 206 |
+
このパラダイムでは、**シンボリック知識やルール**がニューラルネットワークの訓練・推論プロセスを制約・ガイドします。
|
| 207 |
+
|
| 208 |
+
### 3.1 制約付きデコーディング(Constrained Decoding)
|
| 209 |
+
|
| 210 |
+
LLMの出力を生成時点で制約し、構文的・論理的に不正な出力を「物理的に不可能」にする技術です。
|
| 211 |
+
|
| 212 |
+
#### メカニズム
|
| 213 |
+
|
| 214 |
+
```
|
| 215 |
+
LLM の各デコーディングステップ
|
| 216 |
+
│
|
| 217 |
+
▼
|
| 218 |
+
┌─────────────────┐
|
| 219 |
+
│ 全語彙のロジット │
|
| 220 |
+
│ (確率分布) │
|
| 221 |
+
└────────┬────────┘
|
| 222 |
+
│
|
| 223 |
+
▼
|
| 224 |
+
┌─────────────────┐
|
| 225 |
+
│ 文法/論理制約 │ ← CFG, FSM, 正規表現
|
| 226 |
+
│ によるマスク │
|
| 227 |
+
└────────┬────────┘
|
| 228 |
+
│ 違反トークンを -∞ に設定
|
| 229 |
+
▼
|
| 230 |
+
┌─────────────────┐
|
| 231 |
+
│ 制約済みロジット │
|
| 232 |
+
│ からサンプリング │
|
| 233 |
+
└─────────────────┘
|
| 234 |
+
```
|
| 235 |
+
|
| 236 |
+
#### 主要フレームワーク比較
|
| 237 |
+
|
| 238 |
+
| フレームワーク | アプローチ | 制約の種類 | 最適な用途 | 特徴 |
|
| 239 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 240 |
+
| **Outlines** | ロジットレベルマスク (CFG/正規表現) | 構文的 | JSON/Pydantic 準拠の構造化出力(ローカルモデル) | 高速。XGrammar による最適化 |
|
| 241 |
+
| **Guidance** | Python 制御フロー分岐 | 構文的+手続き的 | スキーマが中間判断に依存する複雑なワークフロー | 柔軟性が高い |
|
| 242 |
+
| **LMQL** | 宣言的クエリ言語 (SQL風) | 構文的+論理的 | 従来のプログラミングロジックとLLM呼び出しの融合 | 高レベルな制約記述 |
|
| 243 |
+
| **Instructor** | 生成後バリデーション+リトライ | 構文的 | API ベースの LLM との簡易統合 | シンプルで互換性が広い |
|
| 244 |
+
|
| 245 |
+
> [!WARNING]
|
| 246 |
+
> **制約のバイアストレードオフ(ACL 2025)**: 最近の研究は、制約付きデコーディングが構文的正しさを保証する一方で、過度な制約がモデルの「分布バイアス」を引き起こし、複雑な推論タスクではむしろ性能を低下させる場合があることを報告しています。構文的制約と推論能力のバランスが重要な研究課題です。
|
| 247 |
+
|
| 248 |
+
### 3.2 知識埋め込み型アプローチ
|
| 249 |
+
|
| 250 |
+
ルールやオントロジーをニューラルネットワークの訓練データや損失関数に直接反映するアプローチ(Kautz Type 4/5)。詳細は [セクション6](#6-微分可能ニューロシンボリック統合tightly-coupled) で解説。
|
| 251 |
+
|
| 252 |
+
---
|
| 253 |
+
|
| 254 |
+
## 4. パラダイム2: Neural → Symbolic(ニューラルが記号に翻訳する)
|
| 255 |
+
|
| 256 |
+
LLMを「意味解析器(Semantic Parser)」として使い、自然言語を形式言語(FOL, SAT, SMT等)に変換した上で、**確定的なシンボリック推論器**に推論を完全に委ねるアプローチです。Kautz の Type 3 に相当します。
|
| 257 |
+
|
| 258 |
+
### 4.1 Logic-LM
|
| 259 |
+
|
| 260 |
+
> **論文**: *"Logic-LM: Empowering Large Language Models with Symbolic Solvers for Faithful Logical Reasoning"*
|
| 261 |
+
> **著者**: Liangming Pan, Alon Albalak, Xinyi Wang, William Yang Wang
|
| 262 |
+
> **会議**: Findings of EMNLP 2023
|
| 263 |
+
|
| 264 |
+
#### アーキテクチャ
|
| 265 |
+
|
| 266 |
+
```mermaid
|
| 267 |
+
graph TB
|
| 268 |
+
A["自然言語の問題"] --> B["モジュール1: Problem Formulator<br>(LLMが形式言語に翻訳)"]
|
| 269 |
+
B -->|"FOL / CSP / SAT<br>表現"| C["モジュール2: Symbolic Reasoner<br>(Z3 / Prover9)"]
|
| 270 |
+
C -->|"記号的回答"| D["モジュール3: Result Interpreter<br>(自然言語に復元)"]
|
| 271 |
+
C -->|"❌ エラー発生"| E["モジュール4: Self-Refinement<br>(エラーメッセージをLLMにフィードバック)"]
|
| 272 |
+
E -->|"修正された形式表現"| C
|
| 273 |
+
D --> F["最終回答"]
|
| 274 |
+
|
| 275 |
+
style E fill:#FDCB6E,stroke:#333
|
| 276 |
+
```
|
| 277 |
+
|
| 278 |
+
#### 3つのモジュールの詳細
|
| 279 |
+
|
| 280 |
+
| モジュール | 役割 | 技術的詳細 |
|
| 281 |
+
| :--- | :--- | :--- |
|
| 282 |
+
| **Problem Formulator** | LLMが自然言語問題を形式的記号表現に変換 | Few-shot プロンプトで FOL, Python (CSP), SAT のいずれかの形式に翻訳 |
|
| 283 |
+
| **Symbolic Reasoner** | 外部ソルバーが確定的推論を実行 | **Z3**(SMTソルバー)または **Prover9**(一階述語論理定理証明器)を使用 |
|
| 284 |
+
| **Result Interpreter** | ソルバーの出力を自然言語に変換 | 記号的回答をユーザーが理解できる形式に復元 |
|
| 285 |
+
| **Self-Refinement** | ソルバーのエラーメッセージに基づく自己修正ループ | 構文エラーや充足不可能性のエラーをLLMにフィードバックし、反復的に修正。一般的な修正パターンの Few-shot デモも利用 |
|
| 286 |
+
|
| 287 |
+
#### 性能と意義
|
| 288 |
+
|
| 289 |
+
- **「忠実な推論(Faithful Reasoning)」の実現**: LLMの確率的推論をソルバーの確定的推論に置き換えることで、推論結果の論理的正しさを保証
|
| 290 |
+
- **ハルシネーション防止**: 推論がソルバー内部で実行されるため、LLMの推論における論理的飛躍が排除される
|
| 291 |
+
- FOLIO, ProofWriter, AR-LSAT ベンチマークで、標準 CoT プロンプトを大幅に上回る精度
|
| 292 |
+
|
| 293 |
+
#### Logic-LM++ (2024)
|
| 294 |
+
|
| 295 |
+
Logic-LM の後継研究。**多段階修正戦略**を導入し、LLMがシンボリック表現を精緻化するプロセスを改善。FOLIO, ProofWriter, AR-LSAT での性能向上を実証。
|
| 296 |
+
|
| 297 |
+
---
|
| 298 |
+
|
| 299 |
+
### 4.2 LINC(Logical Inference via Neurosymbolic Computation)
|
| 300 |
+
|
| 301 |
+
> **論文**: *"LINC: A Neurosymbolic Approach for Logical Reasoning by Combining Language Models with First-Order Logic Provers"*
|
| 302 |
+
> **会議**: EMNLP 2023 **🏆 Outstanding Paper Award 受賞**
|
| 303 |
+
|
| 304 |
+
#### アーキテクチャ
|
| 305 |
+
|
| 306 |
+
```
|
| 307 |
+
┌─────────────────────────────────────────────┐
|
| 308 |
+
│ LINC パイプライン │
|
| 309 |
+
│ │
|
| 310 |
+
│ 入力: 自然言語の前提 + 結論 │
|
| 311 |
+
│ │ │
|
| 312 |
+
│ ▼ │
|
| 313 |
+
│ ┌─────────────────┐ │
|
| 314 |
+
│ │ Stage 1: │ │
|
| 315 |
+
│ │ LLM による │ ← 前提と結論を個別に │
|
| 316 |
+
│ │ FOL 変換 │ FOL 式に翻訳 │
|
| 317 |
+
│ └────────┬────────┘ │
|
| 318 |
+
│ │ │
|
| 319 |
+
│ ▼ FOL 式 (∀x. P(x) → Q(x), ...) │
|
| 320 |
+
│ ┌─────────────────┐ │
|
| 321 |
+
│ │ Stage 2: │ │
|
| 322 |
+
│ │ Prover9 │ ← 厳密な記号的推論 │
|
| 323 |
+
│ │ (定理証明器) │ │
|
| 324 |
+
│ └────────┬────────┘ │
|
| 325 |
+
│ │ │
|
| 326 |
+
│ ▼ │
|
| 327 |
+
│ 結果: Entailment / Contradiction / Unknown │
|
| 328 |
+
└─────────────────────────────────────────────┘
|
| 329 |
+
```
|
| 330 |
+
|
| 331 |
+
#### Logic-LM との比較
|
| 332 |
+
|
| 333 |
+
| 特性 | LINC | Logic-LM |
|
| 334 |
+
| :--- | :--- | :--- |
|
| 335 |
+
| **形式言語** | FOL(一階述語論理)に特化 | FOL, CSP, SAT を問題に応じて選択 |
|
| 336 |
+
| **ソルバー** | Prover9 | Z3 / Prover9 |
|
| 337 |
+
| **自己修正** | なし(翻訳品質に依存) | あり(エラーフィードバックループ) |
|
| 338 |
+
| **タスク** | 論理的含意判定(Entailment) | 汎用的論理推論 |
|
| 339 |
+
| **注目の成果** | 🏆 EMNLP 2023 Outstanding Paper | 幅広いベンチマークでの評価 |
|
| 340 |
+
|
| 341 |
+
#### 重要な発見
|
| 342 |
+
|
| 343 |
+
> [!TIP]
|
| 344 |
+
> **小さなモデル + LINC > 大きなモデル + CoT**: LINC の論文は、**StarCoder+(比較的小規模なモデル)に LINC を適用した場合、GPT-4 + 標準 Chain-of-Thought プロンプトよりも高い精度**を達成したことを報告し��います。これは、モデルサイズを拡大するよりも、**外部シンボリック推論器との統合の方が論理推論において効果的**であることを示す重要な証拠です。
|
| 345 |
+
|
| 346 |
+
---
|
| 347 |
+
|
| 348 |
+
### 4.3 SatLM(Satisfiability-Aided Language Models)
|
| 349 |
+
|
| 350 |
+
> **論文**: *"SATLM: Satisfiability-Aided Language Models Using Declarative Prompting"*
|
| 351 |
+
> **著者**: Xi Ye, Qiaochu Chen, Isil Dillig, Greg Durrett
|
| 352 |
+
> **会議**: NeurIPS 2023 MATH-AI Workshop
|
| 353 |
+
|
| 354 |
+
#### 核心的アイデア:宣言的プロンプティング
|
| 355 |
+
|
| 356 |
+
SatLM は、Logic-LM や LINC とは異なるアプローチを取ります。LLMに対して問題の**手続き的な解法**(プログラム)を生成させるのではなく、問題の**宣言的な仕様**(論理的制約の集合)を生成させます。
|
| 357 |
+
|
| 358 |
+
```
|
| 359 |
+
┌──────────────────────────────────┐
|
| 360 |
+
│ 従来のアプローチ │
|
| 361 |
+
│ 自然言語 → LLM → 手続き的プログラム │
|
| 362 |
+
│ (命令的: HOW を記述) │
|
| 363 |
+
│ 「x = 3; y = x + 2; ...」 │
|
| 364 |
+
└──────────────────────────────────┘
|
| 365 |
+
vs
|
| 366 |
+
┌──────────────────────────────────┐
|
| 367 |
+
│ SatLM のアプローチ │
|
| 368 |
+
│ 自然言語 → LLM → 宣言的仕様 │
|
| 369 |
+
│ (宣言的: WHAT を記述) │
|
| 370 |
+
│ 「x > 0 ∧ x + y = 7 ∧ y < 5」 │
|
| 371 |
+
│ │ │
|
| 372 |
+
│ ▼ │
|
| 373 |
+
│ SAT/SMT ソルバー │
|
| 374 |
+
│ (仕様を満たす解を自動発見) │
|
| 375 |
+
└──────────────────────────────────┘
|
| 376 |
+
```
|
| 377 |
+
|
| 378 |
+
#### 利点
|
| 379 |
+
|
| 380 |
+
1. **宣言的仕様 は LLM にとって生成しやすい**: 複雑な多段階の手続き的推論よりも、問題の制約を列挙する方がLLMにとって容易
|
| 381 |
+
2. **推論の正確性保証**: 充足可能性ソルバーが解を見つけるため、回答は制約と論理的に一貫していることが保証される
|
| 382 |
+
3. **LSAT, BoardGameQA で SOTA**: 評価時点で、プログラム支援型LLM (ProgLM) や CoT を上回る性能
|
| 383 |
+
|
| 384 |
+
---
|
| 385 |
+
|
| 386 |
+
### 4.4 FoVer(First-Order Logic Verification)
|
| 387 |
+
|
| 388 |
+
FoVer は2つの研究ストリームを持ちます。
|
| 389 |
+
|
| 390 |
+
#### ストリーム1: LLM推論の検証
|
| 391 |
+
|
| 392 |
+
```mermaid
|
| 393 |
+
graph LR
|
| 394 |
+
A["LLMの推論出力<br>(CoT)"] -->|FOLに変換| B["FOL 式"]
|
| 395 |
+
B -->|検証| C["Z3 定理証明器"]
|
| 396 |
+
C -->|Valid| D["✅ 論理的に正しい"]
|
| 397 |
+
C -->|Invalid| E["❌ 論理的エラーあり"]
|
| 398 |
+
```
|
| 399 |
+
|
| 400 |
+
- LLMが生成した自然言語の推論チェーンをFOLに翻訳し、Z3で厳密に検証
|
| 401 |
+
- CoT の「忠実性(Faithfulness)」を事後的に保証
|
| 402 |
+
|
| 403 |
+
#### ストリーム2: Process Reward Model (PRM) の自動訓練
|
| 404 |
+
|
| 405 |
+
- **FoVer-40K**: Z3 や Isabelle で自動的にステップレベルのエラーラベルを付与した大規模訓練データセットを合成
|
| 406 |
+
- 人間によるアノテーション不要で、高品質な PRM を訓練可能
|
| 407 |
+
- タスク間の**汎化能力**が高く、形式的ベンチマーク以外の非形式的推論タスクでも性能向上
|
| 408 |
+
|
| 409 |
+
---
|
| 410 |
+
|
| 411 |
+
## 5. パラダイム3: Hybrid(双方向統合)
|
| 412 |
+
|
| 413 |
+
LLMとシンボリック推論器が**反復的に対話**し、互いの出力を批判・修正し合うアプローチです。
|
| 414 |
+
|
| 415 |
+
### 5.1 PEIRCE
|
| 416 |
+
|
| 417 |
+
> **会議**: ACL 2025
|
| 418 |
+
|
| 419 |
+
#### 推測–批判(Conjecture-Criticism)サイクル
|
| 420 |
+
|
| 421 |
+
```mermaid
|
| 422 |
+
graph TB
|
| 423 |
+
A["自然言語の問題"] --> B["LLM: 推測の生成<br>(自然言語 + 形式言語)"]
|
| 424 |
+
B --> C{"批判モジュール群"}
|
| 425 |
+
|
| 426 |
+
C --> D["Hard Critique<br>(Isabelle / Prolog)<br>形式的論理検証"]
|
| 427 |
+
C --> E["Soft Critique<br>(一貫性・尤もらしさ<br>・簡潔性の評価)"]
|
| 428 |
+
|
| 429 |
+
D -->|"形式的エラー"| F["フィードバック統合"]
|
| 430 |
+
E -->|"内容的問題"| F
|
| 431 |
+
|
| 432 |
+
F -->|"修正要求"| B
|
| 433 |
+
F -->|"合格"| G["✅ 最終解答<br>(素材的 + 形式的に妥当)"]
|
| 434 |
+
|
| 435 |
+
style D fill:#E74C3C,stroke:#333,color:#fff
|
| 436 |
+
style E fill:#3498DB,stroke:#333,color:#fff
|
| 437 |
+
```
|
| 438 |
+
|
| 439 |
+
#### PEIRCEの革新性
|
| 440 |
+
|
| 441 |
+
| 要素 | 説明 |
|
| 442 |
+
| :--- | :--- |
|
| 443 |
+
| **素材的推論(Material Inference)** | 内容の妥当性、文脈的関連性、知識との整合性(LLMの強み) |
|
| 444 |
+
| **形式的推論(Formal Inference)** | 構造的・論理的妥当性、演繹的正しさ(シンボリック推論器の強み) |
|
| 445 |
+
| **統一** | 両方の基準を同時に満たす解のみを許容。科学的方法(仮説→検証)のAI実装 |
|
| 446 |
+
|
| 447 |
+
> [!NOTE]
|
| 448 |
+
> PEIRCE の名前は、アメリカの論理学者・哲学者チャールズ・サンダース・パース(Charles Sanders Peirce)に由来しています。パースは「アブダクション(最良の説明への推論)」を提唱し、科学的探究プロセスにおける推測と検証の反復を哲学的に体系化した人物です。
|
| 449 |
+
|
| 450 |
+
### 5.2 パラダイム間の比較総括
|
| 451 |
+
|
| 452 |
+
| 比較項目 | パラダイム1<br>Symbolic→Neural | パラダイム2<br>Neural→Symbolic | パラダイム3<br>Hybrid |
|
| 453 |
+
| :--- | :--- | :--- | :--- |
|
| 454 |
+
| **情報の流れ** | 記号→ニューラル(一方向) | ニューラル→記号(一方向) | 双方向(反復的) |
|
| 455 |
+
| **LLMの役割** | 出力が制約される受動的生成器 | 意味解析器(翻訳器) | 推測生成器+修正器 |
|
| 456 |
+
| **シンボリックの役割** | 訓練データ・損失関数・デコーディング制約 | 最終的な推論実行者 | 批判者・検証者 |
|
| 457 |
+
| **正確性保証** | 部分的(構文的制約のみの場合が多い) | 高い(ソルバー結果は正確) | 高い(反復検証) |
|
| 458 |
+
| **柔軟性** | 高い(制約の強さを調整可能) | 低い(翻訳精度に依存) | 高い(動的調整可能) |
|
| 459 |
+
| **代表的システム** | LTN, LNN, Outlines, LMQL | Logic-LM, LINC, SatLM | PEIRCE, Logic-LM++ |
|
| 460 |
+
|
| 461 |
+
---
|
| 462 |
+
|
| 463 |
+
## 6. 微分可能ニューロ・シンボリック統合(Tightly Coupled)
|
| 464 |
+
|
| 465 |
+
Kautz の Type 5 に該当する、論理とニューラルネットワークを**単一の微分可能な計算グラフ**に統合するアプローチ群です。
|
| 466 |
+
|
| 467 |
+
### 6.1 Logic Tensor Networks(LTN)
|
| 468 |
+
|
| 469 |
+
論理式をテンソル演算に変換し、**エンドツーエンドで微分可能**な推論と学習を実現するフレームワーク。
|
| 470 |
+
|
| 471 |
+
#### 動作原理:Real Logic
|
| 472 |
+
|
| 473 |
+
```
|
| 474 |
+
一階述語論理 テンソル演算
|
| 475 |
+
──────────── ────────────
|
| 476 |
+
定数 a, b, c →→→ テンソル / エンベディング
|
| 477 |
+
述語 P(x) →→→ ニューラルネットワーク f: ℝⁿ → [0, 1]
|
| 478 |
+
∧ (AND) →→→ 微分可能な T-norm(例: min, product)
|
| 479 |
+
∨ (OR) →→→ 微分可能な T-conorm
|
| 480 |
+
¬ (NOT) →→→ 1 - x
|
| 481 |
+
∀x (FORALL) →→→ min / generalized mean over batch
|
| 482 |
+
∃x (EXISTS) →→→ max / generalized mean over batch
|
| 483 |
+
```
|
| 484 |
+
|
| 485 |
+
- **真理値**: 古典的な {0, 1} ではなく、ファジー論理の [0, 1] 区間の実数値
|
| 486 |
+
- **訓練**: 論理的公理の真理値を最大化するように勾配降下法で最適化
|
| 487 |
+
- **知識グラフ**: 不完全な知識グラフに対して、論理制約に基づくリンク予測・補完が可能
|
| 488 |
+
|
| 489 |
+
### 6.2 Logical Neural Networks(LNN) — IBM Research
|
| 490 |
+
|
| 491 |
+
IBM Research が開発した、LTN とは異なるアプローチでニューラルとシンボリックを完全統合するフレームワーク。
|
| 492 |
+
|
| 493 |
+
#### LTN との根本的な違い
|
| 494 |
+
|
| 495 |
+
| 特性 | LTN | LNN (IBM) |
|
| 496 |
+
| :--- | :--- | :--- |
|
| 497 |
+
| **論理のエンコード** | 論理式を損失関数の制約として表現 | 各ニューロンが論理式の構成要素(原子、連言、選言等)に1対1対応 |
|
| 498 |
+
| **推論方向** | 順方向(Forward)のみ | **全方向(Omnidirectional)** — 順方向+逆方向推論が可能 |
|
| 499 |
+
| **不完全情報** | 対応可能(ファジー真理値) | 真理値の**上下限バウンド**を維持。開世界仮定(OWA)を自然に処理 |
|
| 500 |
+
| **解釈可能性** | 中程度 | **高い** — ネットワーク構造が論理構文木に直接対応 |
|
| 501 |
+
| **ルール抽出** | 困難 | ネットワークから人間が読める論理ルールを抽出可能 |
|
| 502 |
+
|
| 503 |
+
#### LNN の構造
|
| 504 |
+
|
| 505 |
+
```
|
| 506 |
+
一階述語論理の論理式:
|
| 507 |
+
∀x. Human(x) → Mortal(x)
|
| 508 |
+
|
| 509 |
+
LNN のネットワーク構造:
|
| 510 |
+
┌──────────────────────────────────────┐
|
| 511 |
+
│ ∀ ニューロン │
|
| 512 |
+
│ └── → ニューロン(含意) │
|
| 513 |
+
│ ├── Human(x) ニューロン │
|
| 514 |
+
│ └── Mortal(x) ニューロン │
|
| 515 |
+
└──────────────────────────────────────┘
|
| 516 |
+
|
| 517 |
+
各ニューロンは [lower_bound, upper_bound] ∈ [0,1]² を保持
|
| 518 |
+
→ 不確実性・不完全情報を自然に表現
|
| 519 |
+
```
|
| 520 |
+
|
| 521 |
+
### 6.3 確率的論理プログラミング統合
|
| 522 |
+
|
| 523 |
+
ニューラル知覚と確率的論理推論を統合するフレームワーク群。
|
| 524 |
+
|
| 525 |
+
| フレームワーク | 基盤論理 | 特徴 | 代表的応用 |
|
| 526 |
+
| :--- | :--- | :--- | :--- |
|
| 527 |
+
| **DeepProbLog** | ProbLog(確率論理プログラミング) | ニューラル述語をProbLogの確率的事実として統合。エンドツーエンド微分可能 | 画像からの推論 |
|
| 528 |
+
| **Scallop** | Datalog(データログ) | 出自(Provenance)ベースの微分可能推論。スケーラビリティ重視 | 視覚的質問応答(VQA) |
|
| 529 |
+
| **NeurASP** | ASP(解答集合プログラミング) | ニューラルネットの出力を ASP の確率的原子として統合 | 宣言的推論 |
|
| 530 |
+
|
| 531 |
+
#### DeepProbLog の動作例
|
| 532 |
+
|
| 533 |
+
```
|
| 534 |
+
画像入力 → [CNN] → 「digit(Image1) = 3」(確率 0.95)
|
| 535 |
+
「digit(Image2) = 5」(確率 0.88)
|
| 536 |
+
│
|
| 537 |
+
▼
|
| 538 |
+
ProbLog ルール: addition(X, Y, Z) :- digit(I1, X), digit(I2, Y), Z is X+Y.
|
| 539 |
+
│
|
| 540 |
+
▼
|
| 541 |
+
出力: addition(Image1, Image2, 8) — 確率 0.836
|
| 542 |
+
|
| 543 |
+
→ ニューラル知覚 + 論理推論 + 確率的不確実性 が一体化
|
| 544 |
+
```
|
| 545 |
+
|
| 546 |
+
---
|
| 547 |
+
|
| 548 |
+
## 7. 応用ドメイン
|
| 549 |
+
|
| 550 |
+
### 7.1 分野別の適用マップ
|
| 551 |
+
|
| 552 |
+
```mermaid
|
| 553 |
+
graph TD
|
| 554 |
+
NeSy["ニューロ・シンボリックAI"] --> H["医療・ヘルスケア"]
|
| 555 |
+
NeSy --> R["ロボティクス"]
|
| 556 |
+
NeSy --> AD["自動運転"]
|
| 557 |
+
NeSy --> PS["プログラム合成"]
|
| 558 |
+
NeSy --> L["法律・コンプライアンス"]
|
| 559 |
+
NeSy --> CS["サイバーセキュリティ"]
|
| 560 |
+
|
| 561 |
+
H --> H1["説明可能な診断支援"]
|
| 562 |
+
H --> H2["臨床試験の患者選定"]
|
| 563 |
+
H --> H3["創薬(知識グラフ統合)"]
|
| 564 |
+
|
| 565 |
+
R --> R1["タスク計画(記号ルール + ニューラル知覚)"]
|
| 566 |
+
R --> R2["安全制約付き動作生成"]
|
| 567 |
+
|
| 568 |
+
AD --> AD1["交通ルールのガードレール"]
|
| 569 |
+
AD --> AD2["Road2Code(シーンのプログラム合成)"]
|
| 570 |
+
|
| 571 |
+
PS --> PS1["検証可能なコード生成"]
|
| 572 |
+
PS --> PS2["DSL 合成"]
|
| 573 |
+
|
| 574 |
+
style NeSy fill:#6C5CE7,stroke:#333,color:#fff
|
| 575 |
+
```
|
| 576 |
+
|
| 577 |
+
### 7.2 各ドメインの詳細
|
| 578 |
+
|
| 579 |
+
| ドメイン | ニューラル成分 | シンボリック成分 | 具体的事例 |
|
| 580 |
+
| :--- | :--- | :--- | :--- |
|
| 581 |
+
| **医療** | 医用画像解析、EHR 処理 | 臨床ガイドライン、医学オントロジー | 診断根拠の追跡可能な支援システム |
|
| 582 |
+
| **ロボティクス** | センサー入力処理、ジェスチャー認識 | 安全プロトコル、手順ルール | 外科手術支援、産業用分解ロボット |
|
| 583 |
+
| **自動運転** | カメラ・LiDAR知覚 | 交通法規(ハード制約) | エッジケースでの予測可能な振る舞い |
|
| 584 |
+
| **プログラム合成** | コード候補の生成 | 形式仕様の検証 | 正しさが保証されたコードの自動生成 |
|
| 585 |
+
| **法律** | 契約文書の解析 | 法的ルールの形式化 | コンプライアンス自動チェック |
|
| 586 |
+
| **サイバーセキュリティ** | 脅威パターンの検出 | セキュリティオントロジー | 検証可能な脅威分析・対応 |
|
| 587 |
+
|
| 588 |
+
> [!IMPORTANT]
|
| 589 |
+
> **AI規制との関連**: EU AI Act(2024年施行)をはじめとする世界的なAI規制は、高リスクAIシステムに対して**透明性・説明可能性・監査可能性**を要求しています。ニューロ・シンボリックAIは、推論過程が追跡可能であるという特性から、これらの規制要件を満たすための**事実上の標準アーキテクチャ**となる可能性があります。
|
| 590 |
+
|
| 591 |
+
---
|
| 592 |
+
|
| 593 |
+
## 8. ベンチマークと評価
|
| 594 |
+
|
| 595 |
+
### 8.1 論理推論ベンチマーク
|
| 596 |
+
|
| 597 |
+
| ベンチマーク | 論理の種類 | タスク | 使用される代表的なソルバー |
|
| 598 |
+
| :--- | :--- | :--- | :--- |
|
| 599 |
+
| **FOLIO** | 一階述語論理 | 自然言語 → FOL 変換 + 論理的含意判定 | Prover9 |
|
| 600 |
+
| **ProofWriter** | 命題論理 / FOL | 多段推論(証明深度1–5) | 各種 |
|
| 601 |
+
| **PrOntoQA** | 命題論理 | ルールベース論理的含意 | — |
|
| 602 |
+
| **AR-LSAT** | 命題論理 + 制約 | LSAT形式の法的推論問題 | Z3 |
|
| 603 |
+
| **BoardGameQA** | 制約充足 | ゲームルールに基づく推論 | SAT ソルバー |
|
| 604 |
+
| **LogicalDeduction** | 命題論理 | 複雑な実世界の論理クエリ | — |
|
| 605 |
+
|
| 606 |
+
### 8.2 主要な評価指標
|
| 607 |
+
|
| 608 |
+
| 指標 | 説明 | 対象 |
|
| 609 |
+
| :--- | :--- | :--- |
|
| 610 |
+
| **推論精度(Accuracy)** | 正しい論理的結論に到達した割合 | 全システム |
|
| 611 |
+
| **実行可能率(Executable Rate)** | LLM が生成した形式表現がソルバーで実行可能だった割合 | Neural→Symbolic 系 |
|
| 612 |
+
| **忠実性(Faithfulness)** | 推論過程が結論を実際に論理的に導いているか | CoT vs NeSy |
|
| 613 |
+
| **自己修正成功率** | エラーフィードバック後に正しい形式表現を生成できた割合 | Logic-LM, Logic-LM++ |
|
| 614 |
+
|
| 615 |
+
### 8.3 性能比較の概観
|
| 616 |
+
|
| 617 |
+
| 手法 | FOLIO | ProofWriter | AR-LSAT |
|
| 618 |
+
| :--- | :---: | :---: | :---: |
|
| 619 |
+
| GPT-4 + CoT | ~70% | ~75% | ~55% |
|
| 620 |
+
| GPT-4 + Logic-LM | ~78% | ~85% | ~68% |
|
| 621 |
+
| GPT-4 + LINC | ~80% | ~88% | — |
|
| 622 |
+
| GPT-4 + SatLM | — | — | ~72% |
|
| 623 |
+
|
| 624 |
+
> [!NOTE]
|
| 625 |
+
> 上記の数値はおおよその傾向を示すもので、プロンプト設計やモデルバージョンにより変動します。一貫した傾向として、**NeSy アプローチは標準 CoT を5–15ポイント上回る**ことが多いです。
|
| 626 |
+
|
| 627 |
+
---
|
| 628 |
+
|
| 629 |
+
## 9. 未解決課題と今後の展望
|
| 630 |
+
|
| 631 |
+
### 9.1 現在���技術的課題
|
| 632 |
+
|
| 633 |
+
```mermaid
|
| 634 |
+
graph TD
|
| 635 |
+
C["未解決課題"] --> C1["翻訳ボトルネック<br>自然言語→形式言語の変換精度"]
|
| 636 |
+
C --> C2["スケーラビリティ<br>大規模知識ベースでの推論効率"]
|
| 637 |
+
C --> C3["表現力の限界<br>FOL/SAT で表現できない知識"]
|
| 638 |
+
C --> C4["制約バイアス<br>過度な制約が推論能力を阻害"]
|
| 639 |
+
C --> C5["学習と推論の統一<br>微分可能論理の理論的限界"]
|
| 640 |
+
C --> C6["説明可能性 vs 性能<br>のトレードオフ"]
|
| 641 |
+
|
| 642 |
+
C1 --> C1a["意味的ドリフト<br>(構文OK・意味NG)"]
|
| 643 |
+
C2 --> C2a["組み合わせ爆発<br>(状態空間の指数的増大)"]
|
| 644 |
+
```
|
| 645 |
+
|
| 646 |
+
| 課題 | 説明 | 現在のアプローチ |
|
| 647 |
+
| :--- | :--- | :--- |
|
| 648 |
+
| **翻訳ボトルネック** | LLMが正確な形式表現を生成できない場合がある | 自己修正ループ、合成データ、GRPO |
|
| 649 |
+
| **意味的ドリフト** | 構文的に正しいが意味的に誤った形式化 | サイクル一貫性、FormalRewardBench |
|
| 650 |
+
| **スケーラビリティ** | 大規模な知識ベースでの推論コスト | Provenance ベース手法(Scallop)、top-k 枝刈り |
|
| 651 |
+
| **表現力** | 現実世界の曖昧性や段階的変化をFOLで表現することの困難 | ファジー論理(LTN)、確率的論理(DeepProbLog) |
|
| 652 |
+
| **評価の難しさ** | 「コンパイル成功 ≠ 正解」問題 | T2, FormalRewardBench |
|
| 653 |
+
| **メタ認知の欠如** | システムが自身の推論の限界を認識・調整できない | 研究段階 |
|
| 654 |
+
|
| 655 |
+
### 9.2 今後の研究方向
|
| 656 |
+
|
| 657 |
+
| 方向性 | 内容 | 期待されるインパクト |
|
| 658 |
+
| :--- | :--- | :--- |
|
| 659 |
+
| **Type 6 統合の追求** | ニューラルとシンボリックを完全に融合した単一アーキテクチャ | AGI への最も有力な道筋の一つ |
|
| 660 |
+
| **自律的形式化** | LLMが自律的に問題を形式化し、適切なソルバーを選択する「エージェント型」NeSy | 人間の介入なしで複雑な推論を完遂 |
|
| 661 |
+
| **多段・多モーダル推論** | 視覚・言語・論理を横断する推論チェーンの統合 | ロボティクス、科学的発見 |
|
| 662 |
+
| **自己改善型推論** | 推論過程を自己監視し、戦略を動的に調整するメタ認知システム | 信頼性の飛躍的向上 |
|
| 663 |
+
| **規制適合** | EU AI Act 等の規制要件を満たす検証可能なAIアーキテクチャ | 産業界での大規模採用 |
|
| 664 |
+
|
| 665 |
+
---
|
| 666 |
+
|
| 667 |
+
## 10. 主要文献一覧
|
| 668 |
+
|
| 669 |
+
### 10.1 サーベイ論文
|
| 670 |
+
|
| 671 |
+
| 年 | タイトル | 著者 / 組織 | 会議 |
|
| 672 |
+
| :--- | :--- | :--- | :--- |
|
| 673 |
+
| 2022 | The Third AI Summer (Neuro-Symbolic AI taxonomy) | Henry Kautz | AI Magazine / AAAI 2020 Award Lecture |
|
| 674 |
+
| 2025 | LLM Reasoning with Neuro-Symbolic AI | Various | IJCAI 2025 Survey |
|
| 675 |
+
|
| 676 |
+
### 10.2 Neural → Symbolic フレームワーク
|
| 677 |
+
|
| 678 |
+
| 年 | タイトル | 著者 | 会議 |
|
| 679 |
+
| :--- | :--- | :--- | :--- |
|
| 680 |
+
| 2023 | Logic-LM: Empowering LLMs with Symbolic Solvers | Pan, Albalak, Wang, Wang | EMNLP 2023 (Findings) |
|
| 681 |
+
| 2023 | LINC: Combining LMs with FOL Provers 🏆 | Olausson et al. | EMNLP 2023 (Outstanding Paper) |
|
| 682 |
+
| 2023 | SATLM: Satisfiability-Aided LMs Using Declarative Prompting | Ye, Chen, Dillig, Durrett | NeurIPS 2023 MATH-AI |
|
| 683 |
+
| 2024 | Logic-LM++: Multi-step Refinement | — | ACL 2024 |
|
| 684 |
+
| 2024 | FoVer: FOL Verification / PRM Training | Various (MIT etc.) | — |
|
| 685 |
+
|
| 686 |
+
### 10.3 Hybrid / 双方向フレームワーク
|
| 687 |
+
|
| 688 |
+
| 年 | タイトル | 会議 |
|
| 689 |
+
| :--- | :--- | :--- |
|
| 690 |
+
| 2025 | PEIRCE: Conjecture-Criticism for Material and Formal Inference | ACL 2025 |
|
| 691 |
+
| 2024 | ENVISIONS: Neuro-Symbolic Self-Training | ACL 2024 |
|
| 692 |
+
|
| 693 |
+
### 10.4 微分可能ニューロ・シンボリック統合
|
| 694 |
+
|
| 695 |
+
| 年 | タイトル | 開発元 |
|
| 696 |
+
| :--- | :--- | :--- |
|
| 697 |
+
| 2020+ | Logic Tensor Networks (LTN) | SonyCLI / Academia |
|
| 698 |
+
| 2020+ | Logical Neural Networks (LNN) | IBM Research |
|
| 699 |
+
| 2018+ | DeepProbLog | KU Leuven |
|
| 700 |
+
| 2023+ | Scallop | UPenn |
|
| 701 |
+
| 2020+ | NeurASP | ASU / NMSU |
|
| 702 |
+
|
| 703 |
+
### 10.5 制約付きデコーディング
|
| 704 |
+
|
| 705 |
+
| ライブラリ | GitHub | 特徴 |
|
| 706 |
+
| :--- | :--- | :--- |
|
| 707 |
+
| **Outlines** | dottxt-ai/outlines | CFG/正規表現ベースの高速ロジットマスク |
|
| 708 |
+
| **Guidance** | guidance-ai/guidance | Python制御フロー分岐 |
|
| 709 |
+
| **LMQL** | eth-sri/lmql | 宣言的クエリ言語 |
|
| 710 |
+
| **Instructor** | jxnl/instructor | API互換の生成後バリデーション |
|
| 711 |
+
|
| 712 |
+
### 10.6 関連リポジトリ
|
| 713 |
+
|
| 714 |
+
| リポジトリ | URL | 内容 |
|
| 715 |
+
| :--- | :--- | :--- |
|
| 716 |
+
| **Awesome-LLM-Reasoning-with-NeSy** | GitHub | IJCAI 2025 サーベイ関連キュレーション |
|
| 717 |
+
| **LLM-Symbolic-Reasoning-Survey** | GitHub | シンボリック形式化・定理証明のトラッキング |
|
| 718 |
+
| **IBM LNN** | GitHub | LNN 公式実装 |
|
| 719 |
+
| **Scallop** | scallop-lang.org | Scallop 公式サイト・ドキュメント |
|
docs/AI_Agents_and_Logic/neurosymbolic_ai_survey.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "3b876a0aa99ad0c8", "ingested_at": "2026-06-30T08:25:59.744620", "content_hash": "3b876a0aa99ad0c8"}
|
docs/AI_Agents_and_Logic/script_long.md
ADDED
|
@@ -0,0 +1,30 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and welcome back. Tonight's report focuses on the rapidly evolving field of agent-based AI, specifically how these intelligent systems are transforming our technological landscape.
|
| 2 |
+
|
| 3 |
+
We're in a period of significant change, where AI is moving beyond simply mimicking human conversation to truly understanding and executing complex tasks. This shift has been driven by several key developments:
|
| 4 |
+
|
| 5 |
+
First, we're seeing a transition from evaluating AI models based on their conversational fluency to their practical execution capabilities. Think of it like this: instead of just talking, these new agents are actually doing things.
|
| 6 |
+
|
| 7 |
+
The benchmark indicators for this evolution are evolving rapidly. The BFCL v4 benchmark, which focuses on the accuracy of function calls, is a prime example. We're seeing a shift towards models that can not only call functions but also understand dependencies and execute them in a logical order, essentially mimicking human workflows. And then there's SWE-bench Pro, which evaluates an AI's ability to solve real-world software development issues. The focus here is on expanding beyond just Python, into other languages like Go or TypeScript, and even automating the entire testing and debugging process.
|
| 8 |
+
|
| 9 |
+
Finally, we're seeing a shift towards more complex tasks that require human-like planning. This is where models like WebArena and GAIA come in. These platforms are designed to evaluate an AI's ability to follow complex instructions and complete tasks involving browser interactions. The trend here is clear: the ability to plan and execute actions based on human-like input is becoming a key differentiator for these new agents.
|
| 10 |
+
|
| 11 |
+
Now, let's delve into the open-source landscape of agent-based AI. We are seeing an exciting rise in models that rival proprietary ones like Claude Mythos and GPT-5.3. These open-source models are pushing boundaries with their ability to handle complex tasks, all while remaining accessible for researchers and developers alike.
|
| 12 |
+
|
| 13 |
+
Qwen3-235B, DeepSeek-V4 Pro, GLM-5 / 4.5, and Llama-4 Agent Edition are just a few examples of these powerful open-source models. They demonstrate the rapid progress in AI development and offer exciting opportunities for innovation.
|
| 14 |
+
|
| 15 |
+
But it's not just about individual models; the field is also witnessing a revolution in video generation. We're seeing a shift from simply generating pixels to creating truly creative agents capable of producing high-quality videos.
|
| 16 |
+
|
| 17 |
+
The key here is "agentic" - meaning these AI systems are no longer just passive viewers, but active participants, able to understand and respond to prompts. This has been made possible by advancements in video generation techniques like Flow Matching & DiT, which allows for the accurate reproduction of even subtle facial expressions.
|
| 18 |
+
|
| 19 |
+
And then there's the "Agentic Kernel Synthesis" technology developed by HeyGen. This breakthrough enables these AI agents to synthesize custom kernels in real-time, leading to faster and more efficient responses. This is a game changer, as it allows for highly responsive video generation with minimal latency.
|
| 20 |
+
|
| 21 |
+
The future of this technology is exciting! We're seeing a shift towards "decision-centric" evaluation, where the focus isn't just on what an AI says, but rather how effectively it makes decisions and achieves desired outcomes. This means we need to develop new tools and techniques for evaluating and optimizing these agents.
|
| 22 |
+
|
| 23 |
+
We also expect a greater emphasis on "harness engineering," which involves designing and implementing code layers that allow us to control and monitor these complex models. This will be crucial as the field of agent-based AI continues to evolve.
|
| 24 |
+
|
| 25 |
+
Finally, we're moving towards a future where multiple specialized agents work together in a collaborative manner. Think of it like a team of experts working together on a project. We'll see more of this kind of multi-agent collaboration, with models like Qwen3 and DeepSeek-V4 playing distinct roles to achieve optimal results.
|
| 26 |
+
|
| 27 |
+
This is just the beginning of a revolution in AI technology. The coming years promise to be even more exciting as we witness the emergence of truly intelligent agents that can solve complex problems and automate tasks previously thought impossible.
|
| 28 |
+
|
| 29 |
+
|
| 30 |
+
And that's our report for tonight. I'm [your name], reporting for The Android Times.
|
docs/AI_Agents_and_Logic/script_long.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "08450b5bebe609da", "ingested_at": "2026-06-30T08:26:00.918749", "content_hash": "08450b5bebe609da"}
|
docs/AI_Agents_and_Logic/script_long_clean.txt
ADDED
|
@@ -0,0 +1,29 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and welcome back. Tonight's report focuses on the rapidly evolving field of agent-based AI, specifically how these intelligent systems are transforming our technological landscape.
|
| 2 |
+
|
| 3 |
+
We're in a period of significant change, where AI is moving beyond simply mimicking human conversation to truly understanding and executing complex tasks. This shift has been driven by several key developments:
|
| 4 |
+
|
| 5 |
+
First, we're seeing a transition from evaluating AI models based on their conversational fluency to their practical execution capabilities. Think of it like this: instead of just talking, these new agents are actually doing things.
|
| 6 |
+
|
| 7 |
+
The benchmark indicators for this evolution are evolving rapidly. The BFCL v4 benchmark, which focuses on the accuracy of function calls, is a prime example. We're seeing a shift towards models that can not only call functions but also understand dependencies and execute them in a logical order, essentially mimicking human workflows. And then there's SWE-bench Pro, which evaluates an AI's ability to solve real-world software development issues. The focus here is on expanding beyond just Python, into other languages like Go or TypeScript, and even automating the entire testing and debugging process.
|
| 8 |
+
|
| 9 |
+
Finally, we're seeing a shift towards more complex tasks that require human-like planning. This is where models like WebArena and GAIA come in. These platforms are designed to evaluate an AI's ability to follow complex instructions and complete tasks involving browser interactions. The trend here is clear: the ability to plan and execute actions based on human-like input is becoming a key differentiator for these new agents.
|
| 10 |
+
|
| 11 |
+
Now, let's delve into the open-source landscape of agent-based AI. We are seeing an exciting rise in models that rival proprietary ones like Claude Mythos and GPT-5.3. These open-source models are pushing boundaries with their ability to handle complex tasks, all while remaining accessible for researchers and developers alike.
|
| 12 |
+
|
| 13 |
+
Qwen3-235B, DeepSeek-V4 Pro, GLM-5 / 4.5, and Llama-4 Agent Edition are just a few examples of these powerful open-source models. They demonstrate the rapid progress in AI development and offer exciting opportunities for innovation.
|
| 14 |
+
|
| 15 |
+
But it's not just about individual models; the field is also witnessing a revolution in video generation. We're seeing a shift from simply generating pixels to creating truly creative agents capable of producing high-quality videos.
|
| 16 |
+
|
| 17 |
+
The key here is "agentic" - meaning these AI systems are no longer just passive viewers, but active participants, able to understand and respond to prompts. This has been made possible by advancements in video generation techniques like Flow Matching & DiT, which allows for the accurate reproduction of even subtle facial expressions.
|
| 18 |
+
|
| 19 |
+
And then there's the "Agentic Kernel Synthesis" technology developed by HeyGen. This breakthrough enables these AI agents to synthesize custom kernels in real-time, leading to faster and more efficient responses. This is a game changer, as it allows for highly responsive video generation with minimal latency.
|
| 20 |
+
|
| 21 |
+
The future of this technology is exciting! We're seeing a shift towards "decision-centric" evaluation, where the focus isn't just on what an AI says, but rather how effectively it makes decisions and achieves desired outcomes. This means we need to develop new tools and techniques for evaluating and optimizing these agents.
|
| 22 |
+
|
| 23 |
+
We also expect a greater emphasis on "harness engineering," which involves designing and implementing code layers that allow us to control and monitor these complex models. This will be crucial as the field of agent-based AI continues to evolve.
|
| 24 |
+
|
| 25 |
+
Finally, we're moving towards a future where multiple specialized agents work together in a collaborative manner. Think of it like a team of experts working together on a project. We'll see more of this kind of multi-agent collaboration, with models like Qwen3 and DeepSeek-V4 playing distinct roles to achieve optimal results.
|
| 26 |
+
|
| 27 |
+
This is just the beginning of a revolution in AI technology. The coming years promise to be even more exciting as we witness the emergence of truly intelligent agents that can solve complex problems and automate tasks previously thought impossible.
|
| 28 |
+
|
| 29 |
+
And that's our report for tonight. I'm , reporting for The Android Times.
|
docs/AI_Agents_and_Logic/script_middle.md
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight, we're diving into the world of agent AI, a rapidly evolving field with major implications for how we interact with technology. The landscape is changing fast, and this report gives us a glimpse into what to expect in the coming years.
|
| 2 |
+
|
| 3 |
+
One key trend is the shift from conversational fluency to practical performance. AI models are now being evaluated on their ability to actually accomplish tasks, not just talk. This means looking at things like how well they can solve real-world problems using software tools or even how effectively they can plan and execute complex actions.
|
| 4 |
+
|
| 5 |
+
We're seeing this reflected in the development of new benchmarks for AI models. For example, the BFCL v4 benchmark focuses on a model's ability to use multiple tools in sequence, while SWE-bench Pro evaluates its performance in solving real-world software development issues. And then there's WebArena and GAIA, which measure how well an AI can complete tasks involving browser interaction.
|
| 6 |
+
|
| 7 |
+
Now, let's talk about open-source models. These are becoming increasingly important as we see the rise of more specialized agent AI models.
|
| 8 |
+
We've compiled a list of some standout models that are pushing the boundaries of what's possible: Qwen3-235B, DeepSeek-V4 Pro, GLM-5/4.5, and Llama-4 Agent Edition. Each model has its unique strengths, but they all share a common goal: to become more autonomous and capable in real-world scenarios.
|
| 9 |
+
|
| 10 |
+
The video generation field is also seeing exciting advancements. We're moving beyond simply generating pixels to creating truly creative agents who can produce high-quality videos. This is driven by breakthroughs like the Avatar V architecture, which allows for finer control over facial expressions and body language.
|
| 11 |
+
|
| 12 |
+
And then there's the shift in how video content is created. Instead of just producing raw footage, we now have a system where an AI agent first creates a "blueprint" or outline of what needs to be shown, and then users can refine it through dialogue. This approach allows for more targeted and personalized video production.
|
| 13 |
+
|
| 14 |
+
Looking ahead, the future of agent AI is all about decision-making. The focus will shift from simply understanding what was said to analyzing which decisions were most effective. We'll also see a greater emphasis on "harness engineering," a crucial skill that involves controlling and monitoring models in real-time.
|
| 15 |
+
|
| 16 |
+
Finally, we can expect to see more collaboration between multiple AI agents working together. Think of it like having different specialists working together on a project - each with their own expertise and strengths. This will lead to more efficient and powerful solutions for complex problems.
|
| 17 |
+
|
| 18 |
+
|
| 19 |
+
This has been brought to you by The Android Times.
|
docs/AI_Agents_and_Logic/script_middle.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "af797480d09ce4a1", "ingested_at": "2026-06-30T08:26:01.450064", "content_hash": "af797480d09ce4a1"}
|
docs/AI_Agents_and_Logic/script_middle_clean.txt
ADDED
|
@@ -0,0 +1,19 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. Tonight, we're diving into the world of agent AI, a rapidly evolving field with major implications for how we interact with technology. The landscape is changing fast, and this report gives us a glimpse into what to expect in the coming years.
|
| 2 |
+
|
| 3 |
+
One key trend is the shift from conversational fluency to practical performance. AI models are now being evaluated on their ability to actually accomplish tasks, not just talk. This means looking at things like how well they can solve real-world problems using software tools or even how effectively they can plan and execute complex actions.
|
| 4 |
+
|
| 5 |
+
We're seeing this reflected in the development of new benchmarks for AI models. For example, the BFCL v4 benchmark focuses on a model's ability to use multiple tools in sequence, while SWE-bench Pro evaluates its performance in solving real-world software development issues. And then there's WebArena and GAIA, which measure how well an AI can complete tasks involving browser interaction.
|
| 6 |
+
|
| 7 |
+
Now, let's talk about open-source models. These are becoming increasingly important as we see the rise of more specialized agent AI models.
|
| 8 |
+
|
| 9 |
+
We've compiled a list of some standout models that are pushing the boundaries of what's possible: Qwen3-235B, DeepSeek-V4 Pro, GLM-5/4.5, and Llama-4 Agent Edition. Each model has its unique strengths, but they all share a common goal: to become more autonomous and capable in real-world scenarios.
|
| 10 |
+
|
| 11 |
+
The video generation field is also seeing exciting advancements. We're moving beyond simply generating pixels to creating truly creative agents who can produce high-quality videos. This is driven by breakthroughs like the Avatar V architecture, which allows for finer control over facial expressions and body language.
|
| 12 |
+
|
| 13 |
+
And then there's the shift in how video content is created. Instead of just producing raw footage, we now have a system where an AI agent first creates a "blueprint" or outline of what needs to be shown, and then users can refine it through dialogue. This approach allows for more targeted and personalized video production.
|
| 14 |
+
|
| 15 |
+
Looking ahead, the future of agent AI is all about decision-making. The focus will shift from simply understanding what was said to analyzing which decisions were most effective. We'll also see a greater emphasis on "harness engineering," a crucial skill that involves controlling and monitoring models in real-time.
|
| 16 |
+
|
| 17 |
+
Finally, we can expect to see more collaboration between multiple AI agents working together. Think of it like having different specialists working together on a project - each with their own expertise and strengths. This will lead to more efficient and powerful solutions for complex problems.
|
| 18 |
+
|
| 19 |
+
This has been brought to you by The Android Times.
|
docs/AI_Agents_and_Logic/script_short.md
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"Good evening. The world of AI is changing fast, and the future of 'Agent' AI is looking incredibly exciting. We're seeing a shift from just "talking," to actually doing things!
|
| 2 |
+
|
| 3 |
+
New benchmarks are emerging that measure an agent's ability to solve real-world problems, like writing code or even planning complex tasks. Think of it as a new level of intelligence. Models like Qwen3 and DeepSeek are already showing incredible results in these areas. And the video generation game is changing too!
|
| 4 |
+
|
| 5 |
+
HeyGen Avatar V is pushing boundaries with its ability to create lifelike videos. It's not just about generating pixels anymore, but actually understanding what makes a scene work. This means more realistic and creative content.
|
| 6 |
+
|
| 7 |
+
The future of AI looks like it will be all about making decisions. We're moving from "what was said" to "what decision was the best." And this is going to require new skills in engineering and collaboration.
|
| 8 |
+
|
| 9 |
+
So, what does this mean for you? It means we're on the cusp of a revolution in how AI interacts with us. Get ready for a future where AI truly helps us solve problems and create amazing things!"
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
This has been brought to you by The Android Times.
|
docs/AI_Agents_and_Logic/script_short.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "842a158f20506355", "ingested_at": "2026-06-30T08:26:01.777537", "content_hash": "842a158f20506355"}
|
docs/AI_Agents_and_Logic/script_short_clean.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
"Good evening. The world of AI is changing fast, and the future of 'Agent' AI is looking incredibly exciting. We're seeing a shift from just "talking," to actually doing things!
|
| 2 |
+
|
| 3 |
+
New benchmarks are emerging that measure an agent's ability to solve real-world problems, like writing code or even planning complex tasks. Think of it as a new level of intelligence. Models like Qwen3 and DeepSeek are already showing incredible results in these areas. And the video generation game is changing too!
|
| 4 |
+
|
| 5 |
+
HeyGen Avatar V is pushing boundaries with its ability to create lifelike videos. It's not just about generating pixels anymore, but actually understanding what makes a scene work. This means more realistic and creative content.
|
| 6 |
+
|
| 7 |
+
The future of AI looks like it will be all about making decisions. We're moving from "what was said" to "what decision was the best." And this is going to require new skills in engineering and collaboration.
|
| 8 |
+
|
| 9 |
+
So, what does this mean for you? It means we're on the cusp of a revolution in how AI interacts with us. Get ready for a future where AI truly helps us solve problems and create amazing things!"
|
| 10 |
+
|
| 11 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/VFI_Research_Report.md
ADDED
|
@@ -0,0 +1,163 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# Video Frame Interpolation (VFI) 最新研究トレンドレポート (2025–2026)
|
| 2 |
+
|
| 3 |
+
本レポートは、2025年から2026年にかけて発表された最新の学術論文に基づき、コンピュータビジョン分野における**ビデオフレーム補間(Video Frame Interpolation: VFI)**の最新研究動向、主要な技術的アプローチ、および画期的な成果を体系的にまとめたドキュメントです。
|
| 4 |
+
|
| 5 |
+
本リサーチで特定された代表的な論文5本は、本ディレクトリ下の `papers/` フォルダにPDF形式でローカルにダウンロードされています。
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 1. 総合概要(Executive Summary)
|
| 10 |
+
|
| 11 |
+
ビデオフレーム補間(VFI)は、連続するビデオフレーム間の中間フレームを生成し、フレームレートを向上させる(例:通常動画からハイスピードスローモーション動画への変換)ための重要技術です。
|
| 12 |
+
|
| 13 |
+
2025〜2026年のトレンドを分析すると、従来の「オプティカルフローを推定し、CNNや単純なTransformerでワーピングする」手法から脱却し、以下の**4つの破壊的イノベーション**が起きています。
|
| 14 |
+
|
| 15 |
+
1. **動画生成拡散モデル(Video Diffusion Models)のVFI適応**:
|
| 16 |
+
テクスチャの整合性や大局的な動きの品質を劇的に高めるため、事前に学習された強力なImage-to-Video(I2V)拡散モデルをVFIタスクへ適応・微調整する研究が急増しています。
|
| 17 |
+
2. **画像編集基盤モデル(Image Editing Foundation Models)のFew-shot適応**:
|
| 18 |
+
明示的なオプティカルフローの計算を行わず、静的画像の編集能力を持つ基盤モデルに対してLoRA等のパラメータ効率の良いチューニング(PEFT)を施すことで、極めて少ないデータ(数十〜数百サンプル)で高精度なフレーム補間を「解き放つ」アプローチが登場しました。
|
| 19 |
+
3. **イベントカメラ(Event Camera)の融合による時間超解像**:
|
| 20 |
+
超高頻度かつ非同期に輝度変化を捉えるイベントカメラのストリームを融合し、高速・複雑な非線形モーションや、ブレのある(露出時間の長い)低フレームレート映像からの極限のフレーム補間(時間超解像)が可能になりました。
|
| 21 |
+
4. **人間の知覚に準拠した高効率な動き評価指標(PSNRDIV)の誕生**:
|
| 22 |
+
従来の単純なL1/L2誤差(PSNR/SSIM)や極めて重いFloLPIPSに代わり、動きベクトルの「発散(Divergence)」を利用して動画特有の破綻領域を検知する、超高速かつ軽量な評価指標(かつ損失関数)が提案されました。
|
| 23 |
+
|
| 24 |
+
---
|
| 25 |
+
|
| 26 |
+
## 2. 主要論文の詳細レビュー
|
| 27 |
+
|
| 28 |
+
ローカルにPDFをダウンロードした5本の最重要論文について、それぞれの提案手法、アーキテクチャ、および主要な貢献を詳細に解説します。
|
| 29 |
+
|
| 30 |
+
---
|
| 31 |
+
|
| 32 |
+
### 【レビュー1】AceVFI: 総合的サーベイと体系化
|
| 33 |
+
* **論文名**: AceVFI: A Comprehensive Survey of Advances in Video Frame Interpolation
|
| 34 |
+
* **著者**: Dahyeon Kye, Changhyun Roh, Sukhun Ko, Chanho Eom, Jihyong Oh
|
| 35 |
+
* **arXiv ID**: `2506.01061` | **ローカルパス**: `papers/AceVFI.pdf`
|
| 36 |
+
* **論文URL**: [https://arxiv.org/abs/2506.01061](https://arxiv.org/abs/2506.01061)
|
| 37 |
+
|
| 38 |
+
#### 核心アイデアと手法
|
| 39 |
+
250本以上の代表的論文を網羅し、近年のVFIの発展史を体系化した集大成サーベイです。
|
| 40 |
+
VFIの核心を以下のアーキテクチャ軸および学習軸で明確に分類しています。
|
| 41 |
+
|
| 42 |
+
* **アーキテクチャ分類**:
|
| 43 |
+
* **Kernel-based(カーネルベース)** / **Flow-based(フローベース)** / **Hybrid(ハイブリッド)**
|
| 44 |
+
* **Phase-based(位相ベース)** / **GAN-based(敵対的学習)**
|
| 45 |
+
* **Transformer-based(アテンションベース)**
|
| 46 |
+
* **Mamba-based(状態空間モデル)** / **Diffusion-based(拡散モデル)**
|
| 47 |
+
* **学習パラダイム分類**:
|
| 48 |
+
* **CTFI(Center-Time Frame Interpolation)**: $t=0.5$ など、前後のちょうど中間の時間のみを生成する最もシンプルな形態。
|
| 49 |
+
* **ATFI(Arbitrary-Time Frame Interpolation)**: $t \in (0, 1)$ の任意の時間パラメータを入力として受け取り、任意のフレームレート(例:120FPSやハイスピードスロー)を動的に出力する汎用性の高いアプローチ。
|
| 50 |
+
|
| 51 |
+
#### 主要な貢献と結果
|
| 52 |
+
* 大局的かつ急激な動き、オクルージョン(物体の隠れ)、照明の動的変化、および非線形モーションという、従来のVFIが直面してきた主要課題に対する各モデルのアプローチを比較。
|
| 53 |
+
* 標準的な評価データセット(UCF101、Vimeo-90K、GOPROなど)、画質・時間的整合性を測るための各種損失関数と評価指標のベンチマーク結果を集約。
|
| 54 |
+
* 実用的な���ープンソースリソースとして、GitHub上に関連論文と実装をまとめた「Awesome-Video-Frame-Interpolation」プロジェクトをホストし、コミュニティの発展に大きく貢献しています。
|
| 55 |
+
|
| 56 |
+
---
|
| 57 |
+
|
| 58 |
+
### 【レビュー2】FC-VFI: 高画質・高FPSスローモーション生成
|
| 59 |
+
* **論文名**: FC-VFI: Faithful and Consistent Video Frame Interpolation for High-FPS Slow Motion Video Generation
|
| 60 |
+
* **著者**: Ganggui Ding, Hao Chen, Xiaogang Xu
|
| 61 |
+
* **arXiv ID**: `2603.04899` | **ローカルパス**: `papers/FC_VFI.pdf`
|
| 62 |
+
* **論文URL**: [https://arxiv.org/abs/2603.04899](https://arxiv.org/abs/2603.04899)
|
| 63 |
+
|
| 64 |
+
#### 核心アイデアと手法
|
| 65 |
+
大規模動画生成モデル(HunyuanVideo-I2Vなど)の持つ強力な動画合成プライア(事前知識)をベースとし、スローモーション動画生成に向けてVFI専用にファインチューニングした拡散ベースのフレームワークです。最大2560×1440(2.5K)の超高解像度に対応します。
|
| 66 |
+
|
| 67 |
+
* **Temporal Fidelity Modulation Reference (TFMR)**:
|
| 68 |
+
前後フレームから抽出された高解像度のテクスチャや境界情報を、中間生成段階に時間的結合(Temporal Concatenation)とモジュレーション(変調)を介して直接フィードバックします。これにより、拡散モデル特有の「細かいテクスチャがブレたり変化してしまう問題」を抑制し、視覚的一貫性を維持します。
|
| 69 |
+
* **Structure-Aware Motion**:
|
| 70 |
+
オプティカルフローが完全に崩壊するような複雑な境界や隠消現実(オクルージョン)に対して、セマンティックなマッチングライン(構造ガイド)を導入。動きの構造情報をアテンション機構に融合させます。
|
| 71 |
+
|
| 72 |
+
#### 主要な貢献と結果
|
| 73 |
+
* **高効率性**: 双方向のフロー推定や推論といった前処理フェーズを排除し、わずか**10ステップのノイズ除去**だけで高品質な4倍/8倍補間を達成。
|
| 74 |
+
* 従来手法ではゴーストが発生していた激しいスポーツ動画や複雑な煙・水のような流体シミュレーション動画において、テクスチャが破綻しない極めてリアルなハイスピードスローモーションを生成することに成功。
|
| 75 |
+
|
| 76 |
+
---
|
| 77 |
+
|
| 78 |
+
### 【レビュー3】Edit2Interp: 画像編集モデルからのFew-Shot転用
|
| 79 |
+
* **論文名**: Edit2Interp: Adapting Image Foundation Models from Spatial Editing to Video Frame Interpolation with Few-Shot Learning
|
| 80 |
+
* **著者**: Nasrin Rahimi, Mısra Yavuz, Burak Can Biner, et al.
|
| 81 |
+
* **arXiv ID**: `2603.15003` | **ローカルパス**: `papers/Edit2Interp.pdf`
|
| 82 |
+
* **論文URL**: [https://arxiv.org/abs/2603.15003](https://arxiv.org/abs/2603.15003)
|
| 83 |
+
|
| 84 |
+
#### 核心アイデアと手法
|
| 85 |
+
「ビデオフレーム補間とは、空間的な編集モデルによる制約付きの画像編集問題の一種である」というパラダイムシフトを提案した画期的なアプローチです。
|
| 86 |
+
|
| 87 |
+
* **基盤モデルの有効活用**:
|
| 88 |
+
多目的画像編集基盤モデルである **Qwen-Image-Edit** をバックボーンに採用。開始フレーム $I_0$ と終了フレーム $I_1$、および「$t=0.5$ の中間フレームを生成せよ」といった固定的なプロンプト情報をマルチモーダル入力として与えます。
|
| 89 |
+
* **Few-Shot LoRA Adaptation**:
|
| 90 |
+
動画モデルの巨大な学習コストをかけることなく、わずか**64〜256の動画サンプルセット**を用いて、LoRA(Low-Rank Adaptation)を適用するだけで、画像モデル内に眠る空間的一貫性と時間的補間能力を効率的に引き出します。
|
| 91 |
+
|
| 92 |
+
#### 主要な貢献と結果
|
| 93 |
+
* **オプティカルフロー不要の画期性**: フロー推定やワーピングモジュールといった複雑な物理的モーション予測を行わず、モデルが画像の意味理解(セマンティクス)に基づいて自然な中間遷移を決定します。
|
| 94 |
+
* データが極端に制限された環境(学習データが非常に少ないドメイン)であっても、既存の多くのフルモデルトレーニングVFI手法と同等以上の自然なテクスチャ保持と形状変化を実現。
|
| 95 |
+
|
| 96 |
+
---
|
| 97 |
+
|
| 98 |
+
### 【レビュー4】PSNRDIV: 動きの発散度に基づく高効率・高速評価指標
|
| 99 |
+
* **論文名**: An Efficient Quality Metric for Video Frame Interpolation Based on Motion-Field Divergence (PSNRDIV)
|
| 100 |
+
* **著者**: Conall Daly, Darren Ramsook, Anil Kokaram
|
| 101 |
+
* **arXiv ID**: `2510.01361` (別題先行: `2508.09078`)| **ローカルパス**: `papers/PSNRDIV.pdf`
|
| 102 |
+
* **論文URL**: [https://arxiv.org/abs/2510.01361](https://arxiv.org/abs/2510.01361)
|
| 103 |
+
|
| 104 |
+
#### 核心アイデアと手法
|
| 105 |
+
従来のVFI評価は、静的なピクセル毎のL1/L2誤差(PSNR/SSIM)か、もしくは極めて計算コストの高い深層アテンション評価(FloLPIPS)に依存しており、リアルタイムの評価��ネットワークの損失関数(Loss)への組み込みが困難でした。本論文は、映画復元技術から着想を得た軽量・高精度な新しいフルリファレンス画質評価指標 **PSNRDIV** を提案しています。
|
| 106 |
+
|
| 107 |
+
* **Motion-Field Divergence Weighting**:
|
| 108 |
+
動きベクトルフィールドの発散(Divergence)を計算することで、動画内のオクルージョン(隠れ領域)、高速な物体の境界、またはモーションの急激な不整合が発生している「破綻しやすい特異点領域」を特定。
|
| 109 |
+
* この特異点領域におけるピクセルエラーに対して適応的に重みを加算し、動画における人間の主観的な不快感・違和感(アーティファクト)と直接連動したスコアを算出します。
|
| 110 |
+
|
| 111 |
+
#### 主要な貢献と結果
|
| 112 |
+
* **超高速かつ低メモリ**: 既存の代表的な評価指標である FloLPIPS と比較して、**2.5倍高速**に動作し、かつ**メモリ消費量を4分の1に削減**。
|
| 113 |
+
* **高い主観一致度**: 人間の視覚的な評価との相関を示す指標(PLCC)において、FloLPIPS を +0.09 上回る優れた相関関係を実証。
|
| 114 |
+
* 軽量であるため評価だけでなく、深層ネットワークモデルの学習段階における画期的な損失関数(Loss Function)としても直接組み込める実用性を有しています。
|
| 115 |
+
|
| 116 |
+
---
|
| 117 |
+
|
| 118 |
+
### 【レビュー5】EventDiff: イベントカメラを融合した超解像拡散モデル
|
| 119 |
+
* **論文名**: EventDiff: A Unified and Efficient Diffusion Model Framework for Event-based Video Frame Interpolation
|
| 120 |
+
* **著者**: Hanle Zheng, Xujie Han, Zegang Peng, et al.
|
| 121 |
+
* **arXiv ID**: `2505.08235` | **ローカルパス**: `papers/EventDiff.pdf`
|
| 122 |
+
* **論文URL**: [https://arxiv.org/abs/2505.08235](https://arxiv.org/abs/2505.08235)
|
| 123 |
+
|
| 124 |
+
#### 核心アイデアと手法
|
| 125 |
+
超高時間解像度で輝度変化のみを出力する「イベントカメラ」の情報と、通常のカラー画像(RGBフレーム)を融合し、激しい運動や暗所・オクルージョン下での極限のフレーム補間を拡散モデルで実現する統一フレームワークです。
|
| 126 |
+
|
| 127 |
+
* **Event-Frame Hybrid AutoEncoder (HAE)**:
|
| 128 |
+
時間的に非常に不均一なイベントデータのストリームと、静的で豊かな空間情報を持つRGBフレームを統一された特徴マップ空間に圧縮・マッピングするハイブリッドエンコーダ。
|
| 129 |
+
* **Spatial-Temporal Cross Attention (STCA)**:
|
| 130 |
+
時空間アテンションを導入することで、イベント情報をガイダンスとして拡散プロセスのノイズ除去を精緻化します。
|
| 131 |
+
|
| 132 |
+
#### 主要な貢献と結果
|
| 133 |
+
* 従来フローベースモデルが破綻していた「極めて高速なスポーツ動作(テニスのスマッシュなど)」や「突然照明が明滅する環境」において、破綻やブレを一切発生させずに鮮明な中間超解像フレームを出力。
|
| 134 |
+
* ワーピングといった複雑なモーション補正モジュールに依存せず、すべての動きの補完を潜在表現の拡散プロセス内で完結させるため、非常に堅牢で安定した処理が可能です。
|
| 135 |
+
|
| 136 |
+
---
|
| 137 |
+
|
| 138 |
+
## 3. テクノロジートレンドの比較・整理
|
| 139 |
+
|
| 140 |
+
最新研究から見えるVFIの代表的アプローチをマトリクスで整理します。
|
| 141 |
+
|
| 142 |
+
| アプローチ | 代表論文 | モーション予測の手法 | メリット | 課題 |
|
| 143 |
+
| :--- | :--- | :--- | :--- | :--- |
|
| 144 |
+
| **動画生成モデル転用** | **FC-VFI**<br>(レビュー2) | 構造アライメントマッチングライン + I2Vプライア | 2.5K超高画質、写真のようにリアルなテクスチャ生成、10ステップの超高速推論 | モデルサイズが大きく、モバイル等のエッジデバイスには不向き |
|
| 145 |
+
| **画像基盤モデルFew-shot** | **Edit2Interp**<br>(レビュー3) | 空間編集セマンティクスによる動的合成(フロー不要) | 数十枚のデータでチューニング可能、オプティカルフロー特有の境界の破綻が皆無 | 3次元的な幾何学的一貫性を完全に維持するのが難しいケースがある |
|
| 146 |
+
| **動きの特異点評価** | **PSNRDIV**<br>(レビュー4) | モーションベクトルの発散分析(評価用) | 2.5倍高速、省メモリ。損失関数(Loss)としてトレーニングに組み込み可能 | 評価指標であり、これ単体では画像を生成するわけではない |
|
| 147 |
+
| **イベントセンサー融合** | **EventDiff**<br>(レビュー5) | 非同期イベントパルスによるトラッキング案内 | 急激な輝度変化や超高速非線形モーションでも絶対に破綻しない | 特殊なイベントカメラハードウェアが必要 |
|
| 148 |
+
|
| 149 |
+
---
|
| 150 |
+
|
| 151 |
+
## 4. 今後の展望と展望
|
| 152 |
+
|
| 153 |
+
2025–2026年の研究トレンドは、VFIが単なる「画像変形処理」から**「物理演算や世界の構造的意���理解に基づいた『新フレーム創出(Generative VFI)』」**に完全移行したことを如実に示しています。
|
| 154 |
+
|
| 155 |
+
* **アニメ制作・映画修復**: 手描きアニメのコマ割り適応(Time-adaptive VFI)や、古いフィルム映像の超高精細なコマ補間(PSNRDIVの応用)への活用が進んでいます。
|
| 156 |
+
* **自動運転・ロボットビジョン**: イベントカメラを用いた `TimeTracker` や `EventDiff` は、夜間や悪天候、超高速移動時でも視覚的追跡を失わない強固なナビゲーション基盤として実用段階に移りつつあります。
|
| 157 |
+
* **次世代ゲーム機・ディスプレイ**: 低負荷な基盤モデルや軽量拡散ステップの組み合わせにより、コンソール機器上でのリアルタイムなフレームレートブースター(DLSSやFSRの進化系)としての応用が期待されます。
|
| 158 |
+
|
| 159 |
+
---
|
| 160 |
+
|
| 161 |
+
> [!WARNING]
|
| 162 |
+
> **arXiv APIの利用およびライセンスに関する注意**
|
| 163 |
+
> 本ドキュメントの情報、および `papers/` ディレクトリに収集されたPDF資料は、arXivの利用規約([https://info.arxiv.org/help/api/index.html](https://info.arxiv.org/help/api/index.html))に従ってAPI経由で公式に取得したものです。各論文を学術研究以外(商業的利用、再配布など)に利用する際は、PDFファイルや各arXivページに明記されている個々のライセンス(Creative Commons等)の範囲を必ず遵守してください。
|
docs/AI_Computer_Vision/VFI_Research_Report.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "6949f371b515d0e7", "ingested_at": "2026-06-30T08:49:28.206798", "content_hash": "6949f371b515d0e7"}
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405.json
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
{
|
| 2 |
+
"importance": 10,
|
| 3 |
+
"urgency": 10,
|
| 4 |
+
"tier": 1
|
| 5 |
+
}
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405.md
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
version https://git-lfs.github.com/spec/v1
|
| 2 |
+
oid sha256:9d012338d44840196d55d21b96292ea19a0d866ea962fd118d33df572373b4ea
|
| 3 |
+
size 10953161
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_long.md
ADDED
|
@@ -0,0 +1,33 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
This looks like a list of research papers submitted to arXiv in April 2026, focusing on various aspects of computer science and AI. Let's break down the information you provided:
|
| 2 |
+
|
| 3 |
+
**Understanding the Structure:**
|
| 4 |
+
|
| 5 |
+
* **arXiv:** This is a platform for sharing pre-prints of scientific papers, including those in computer science, mathematics, physics, and other fields.
|
| 6 |
+
* **ID:** Each paper has a unique identifier (like "2604.02290") that helps researchers find and track it.
|
| 7 |
+
* **Submitted:** Indicates when the paper was submitted to arXiv.
|
| 8 |
+
* **Authors:** Lists the names of the authors of the research paper.
|
| 9 |
+
* **abs:** Provides a short abstract summarizing the paper's main findings. This is often the first thing people look at to get an idea of what the paper is about.
|
| 10 |
+
* **PDF:** Links to the full, downloadable version of the paper (usually in PDF format).
|
| 11 |
+
|
| 12 |
+
**Key Themes and Topics:**
|
| 13 |
+
|
| 14 |
+
The papers cover a wide range of topics within computer science:
|
| 15 |
+
|
| 16 |
+
* **Medical Imaging:** Papers like "AdamFlow" focus on using AI for surface registration in medical imaging.
|
| 17 |
+
* **3D Modeling & Generation:** "Omni123" explores how to generate 3D models from text and images, even with limited data.
|
| 18 |
+
* **Autonomous Driving:** Papers like "Deep Neural Network Based Roadwork Detection" aim to improve self-driving cars' ability to detect road features.
|
| 19 |
+
* **AI Agents:** Research on "Novel Memory Forgetting Techniques" focuses on how AI agents can learn and adapt without forgetting too much information.
|
| 20 |
+
* **Vision Transformers (ViT):** Several papers use ViT models, a type of deep learning architecture for image processing, like "ViT-Explainer."
|
| 21 |
+
* **Image Generation:** Papers like "Reflection Generation" explore using diffusion models to generate realistic images with reflections.
|
| 22 |
+
* **Denoising and Image Reconstruction:** Research on "DenOiS" aims to improve ultrasound imaging by removing noise and artifacts.
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
**How to Find More Papers:**
|
| 26 |
+
|
| 27 |
+
1. **arXiv Website:** You can visit the arXiv website (https://arxiv.org/) and search for specific topics or authors.
|
| 28 |
+
2. **Specific Research Areas:** If you're interested in a particular area of computer science, like medical imaging, autonomous driving, or AI ethics, you can find relevant papers on those websites.
|
| 29 |
+
|
| 30 |
+
|
| 31 |
+
**Important Note:** The information provided is based on the content of the list you shared.
|
| 32 |
+
|
| 33 |
+
Let me know if you have any other questions about these papers!
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_long_clean.txt
ADDED
|
@@ -0,0 +1,35 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
This looks like a list of research papers submitted to arXiv in April 2026, focusing on various aspects of computer science and AI. Let's break down the information you provided:
|
| 2 |
+
|
| 3 |
+
arXiv: This is a platform for sharing pre-prints of scientific papers, including those in computer science, mathematics, physics, and other fields.
|
| 4 |
+
|
| 5 |
+
ID: Each paper has a unique identifier that helps researchers find and track it.
|
| 6 |
+
|
| 7 |
+
Submitted: Indicates when the paper was submitted to arXiv.
|
| 8 |
+
|
| 9 |
+
Authors: Lists the names of the authors of the research paper.
|
| 10 |
+
|
| 11 |
+
abs: Provides a short abstract summarizing the paper's main findings. This is often the first thing people look at to get an idea of what the paper is about.
|
| 12 |
+
|
| 13 |
+
PDF: Links to the full, downloadable version of the paper .
|
| 14 |
+
|
| 15 |
+
Medical Imaging: Papers like "AdamFlow" focus on using AI for surface registration in medical imaging.
|
| 16 |
+
|
| 17 |
+
3D Modeling & Generation: "Omni123" explores how to generate 3D models from text and images, even with limited data.
|
| 18 |
+
|
| 19 |
+
Autonomous Driving: Papers like "Deep Neural Network Based Roadwork Detection" aim to improve self-driving cars' ability to detect road features.
|
| 20 |
+
|
| 21 |
+
AI Agents: Research on "Novel Memory Forgetting Techniques" focuses on how AI agents can learn and adapt without forgetting too much information.
|
| 22 |
+
|
| 23 |
+
Vision Transformers : Several papers use ViT models, a type of deep learning architecture for image processing, like "ViT-Explainer."
|
| 24 |
+
|
| 25 |
+
Image Generation: Papers like "Reflection Generation" explore using diffusion models to generate realistic images with reflections.
|
| 26 |
+
|
| 27 |
+
Denoising and Image Reconstruction: Research on "DenOiS" aims to improve ultrasound imaging by removing noise and artifacts.
|
| 28 |
+
|
| 29 |
+
1. arXiv Website: You can visit the arXiv website and search for specific topics or authors.
|
| 30 |
+
|
| 31 |
+
2. Specific Research Areas: If you're interested in a particular area of computer science, like medical imaging, autonomous driving, or AI ethics, you can find relevant papers on those websites.
|
| 32 |
+
|
| 33 |
+
The information provided is based on the content of the list you shared.
|
| 34 |
+
|
| 35 |
+
Let me know if you have any other questions about these papers!
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_middle.md
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and we're continuing our coverage of the latest advancements in AI research. A new partnership has been announced between Reflection AI and SpaceX, with a significant financial commitment involved. This deal will see Reflection AI paying $150 million per month for access to Nvidia's cutting-edge GB300 AI chips and supporting hardware within SpaceX's Colossus 2 data center in Tennessee.
|
| 2 |
+
|
| 3 |
+
This agreement marks a major step forward in the field of Artificial Intelligence, as it allows Reflection AI to leverage these powerful tools for their own research and development. The deal is expected to significantly boost Reflection AI's capabilities, particularly in areas like machine learning and deep learning. The Colossus 2 data center is known for its advanced infrastructure and is well-equipped to handle the demands of this partnership.
|
| 4 |
+
|
| 5 |
+
We'll continue to follow this story as it develops, and we'll be sure to provide updates on how this collaboration impacts the AI landscape in the coming months.
|
| 6 |
+
Reporting for The Android Times.
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_middle_clean.txt
ADDED
|
@@ -0,0 +1,7 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and we're continuing our coverage of the latest advancements in AI research. A new partnership has been announced between Reflection AI and SpaceX, with a significant financial commitment involved. This deal will see Reflection AI paying $150 million per month for access to Nvidia's cutting-edge GB300 AI chips and supporting hardware within SpaceX's Colossus 2 data center in Tennessee.
|
| 2 |
+
|
| 3 |
+
This agreement marks a major step forward in the field of Artificial Intelligence, as it allows Reflection AI to leverage these powerful tools for their own research and development. The deal is expected to significantly boost Reflection AI's capabilities, particularly in areas like machine learning and deep learning. The Colossus 2 data center is known for its advanced infrastructure and is well-equipped to handle the demands of this partnership.
|
| 4 |
+
|
| 5 |
+
We'll continue to follow this story as it develops, and we'll be sure to provide updates on how this collaboration impacts the AI landscape in the coming months.
|
| 6 |
+
|
| 7 |
+
Reporting for The Android Times.
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_short.md
ADDED
|
@@ -0,0 +1,4 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
SpaceX just struck a major deal with Reflection AI! They're going to use SpaceX's Colossus 2 data center for super powerful AI chips, worth $150 million a month. This will help them power up their cutting-edge projects. We're talking about the future of AI here folks - this is huge!
|
| 2 |
+
|
| 3 |
+
|
| 4 |
+
Reporting for The Android Times.
|
docs/AI_Computer_Vision/arxiv-cs-cv-20250405-20260405_script_short_clean.txt
ADDED
|
@@ -0,0 +1,3 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
SpaceX just struck a major deal with Reflection AI! They're going to use SpaceX's Colossus 2 data center for super powerful AI chips, worth $150 million a month. This will help them power up their cutting-edge projects. We're talking about the future of AI here folks - this is huge!
|
| 2 |
+
|
| 3 |
+
Reporting for The Android Times.
|
docs/AI_Computer_Vision/arxiv-cs-cv-2026-04-05.md
ADDED
|
@@ -0,0 +1,363 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# arXiv cs.CV — Computer Vision and Pattern Recognition (recent)
|
| 2 |
+
|
| 3 |
+
*Category:* `cs.CV` · *sort:* submittedDate descending
|
| 4 |
+
|
| 5 |
+
*Feed `updated`:* 2026-04-05T06:29:11Z
|
| 6 |
+
|
| 7 |
+
*Fetched at (UTC):* 2026-04-05T06:29:10Z
|
| 8 |
+
|
| 9 |
+
*Machine-readable twin:* `arxiv-cs-cv-2026-04-05.json`
|
| 10 |
+
|
| 11 |
+
---
|
| 12 |
+
|
| 13 |
+
## 1. EventHub: Data Factory for Generalizable Event-Based Stereo Networks without Active Sensors
|
| 14 |
+
|
| 15 |
+
- **arXiv ID:** `2604.02331`
|
| 16 |
+
- **Primary category:** `cs.CV`
|
| 17 |
+
- **Categories:** `cs.CV`
|
| 18 |
+
- **Published:** 2026-04-02
|
| 19 |
+
- **Authors:** Luca Bartolomei, Fabio Tosi, Matteo Poggi, Stefano Mattoccia, Guillermo Gallego
|
| 20 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02331](https://arxiv.org/abs/2604.02331)
|
| 21 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02331v1](https://arxiv.org/pdf/2604.02331v1)
|
| 22 |
+
|
| 23 |
+
### Abstract
|
| 24 |
+
|
| 25 |
+
We propose EventHub, a novel framework for training deep-event stereo networks without ground truth annotations from costly active sensors, relying instead on standard color images. From these images, we derive either proxy annotations and proxy events through state-of-the-art novel view synthesis techniques, or simply proxy annotations when images are already paired with event data. Using the training set generated by our data factory, we repurpose state-of-the-art stereo models from RGB literature to process event data, obtaining new event stereo models with unprecedented generalization capabilities. Experiments on widely used event stereo datasets support the effectiveness of EventHub and show how the same data distillation mechanism can improve the accuracy of RGB stereo foundation models in challenging conditions such as nighttime scenes.
|
| 26 |
+
|
| 27 |
+
---
|
| 28 |
+
|
| 29 |
+
## 2. ActionParty: Multi-Subject Action Binding in Generative Video Games
|
| 30 |
+
|
| 31 |
+
- **arXiv ID:** `2604.02330`
|
| 32 |
+
- **Primary category:** `cs.CV`
|
| 33 |
+
- **Categories:** `cs.CV`, `cs.AI`, `cs.LG`
|
| 34 |
+
- **Published:** 2026-04-02
|
| 35 |
+
- **Authors:** Alexander Pondaven, Ziyi Wu, Igor Gilitschenski, Philip Torr, Sergey Tulyakov, Fabio Pizzati, Aliaksandr Siarohin
|
| 36 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02330](https://arxiv.org/abs/2604.02330)
|
| 37 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02330v1](https://arxiv.org/pdf/2604.02330v1)
|
| 38 |
+
|
| 39 |
+
### Abstract
|
| 40 |
+
|
| 41 |
+
Recent advances in video diffusion have enabled the development of "world models" capable of simulating interactive environments. However, these models are largely restricted to single-agent settings, failing to control multiple agents simultaneously in a scene. In this work, we tackle a fundamental issue of action binding in existing video diffusion models, which struggle to associate specific actions with their corresponding subjects. For this purpose, we propose ActionParty, an action controllable multi-subject world model for generative video games. It introduces subject state tokens, i.e. latent variables that persistently capture the state of each subject in the scene. By jointly modeling state tokens and video latents with a spatial biasing mechanism, we disentangle global video frame rendering from individual action-controlled subject updates. We evaluate ActionParty on the Melting Pot benchmark, demonstrating the first video world model capable of controlling up to seven players simultaneously across 46 diverse environments. Our results show significant improvements in action-following accuracy and identity consistency, while enabling robust autoregressive tracking of subjects through complex interactions.
|
| 42 |
+
|
| 43 |
+
---
|
| 44 |
+
|
| 45 |
+
## 3. Generative World Renderer
|
| 46 |
+
|
| 47 |
+
- **arXiv ID:** `2604.02329`
|
| 48 |
+
- **Primary category:** `cs.CV`
|
| 49 |
+
- **Categories:** `cs.CV`
|
| 50 |
+
- **Published:** 2026-04-02
|
| 51 |
+
- **Authors:** Zheng-Hui Huang, Zhixiang Wang, Jiaming Tan, Ruihan Yu, Yidan Zhang, Bo Zheng, Yu-Lun Liu, Yung-Yu Chuang, Kaipeng Zhang
|
| 52 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02329](https://arxiv.org/abs/2604.02329)
|
| 53 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02329v1](https://arxiv.org/pdf/2604.02329v1)
|
| 54 |
+
|
| 55 |
+
### Abstract
|
| 56 |
+
|
| 57 |
+
Scaling generative inverse and forward rendering to real-world scenarios is bottlenecked by the limited realism and temporal coherence of existing synthetic datasets. To bridge this persistent domain gap, we introduce a large-scale, dynamic dataset curated from visually complex AAA games. Using a novel dual-screen stitched capture method, we extracted 4M continuous frames (720p/30 FPS) of synchronized RGB and five G-buffer channels across diverse scenes, visual effects, and environments, including adverse weather and motion-blur variants. This dataset uniquely advances bidirectional rendering: enabling robust in-the-wild geometry and material decomposition, and facilitating high-fidelity G-buffer-guided video generation. Furthermore, to evaluate the real-world performance of inverse rendering without ground truth, we propose a novel VLM-based assessment protocol measuring semantic, spatial, and temporal consistency. Experiments demonstrate that inverse renderers fine-tuned on our data achieve superior cross-dataset generalization and controllable generation, while our VLM evaluation strongly correlates with human judgment. Combined with our toolkit, our forward renderer enables users to edit styles of AAA games from G-buffers using text prompts.
|
| 58 |
+
|
| 59 |
+
---
|
| 60 |
+
|
| 61 |
+
## 4. Modulate-and-Map: Crossmodal Feature Mapping with Cross-View Modulation for 3D Anomaly Detection
|
| 62 |
+
|
| 63 |
+
- **arXiv ID:** `2604.02328`
|
| 64 |
+
- **Primary category:** `cs.CV`
|
| 65 |
+
- **Categories:** `cs.CV`
|
| 66 |
+
- **Published:** 2026-04-02
|
| 67 |
+
- **Authors:** Alex Costanzino, Pierluigi Zama Ramirez, Giuseppe Lisanti, Luigi Di Stefano
|
| 68 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02328](https://arxiv.org/abs/2604.02328)
|
| 69 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02328v1](https://arxiv.org/pdf/2604.02328v1)
|
| 70 |
+
|
| 71 |
+
### Abstract
|
| 72 |
+
|
| 73 |
+
We present ModMap, a natively multiview and multimodal framework for 3D anomaly detection and segmentation. Unlike existing methods that process views independently, our method draws inspiration from the crossmodal feature mapping paradigm to learn to map features across both modalities and views, while explicitly modelling view-dependent relationships through feature-wise modulation. We introduce a cross-view training strategy that leverages all possible view combinations, enabling effective anomaly scoring through multiview ensembling and aggregation. To process high-resolution 3D data, we train and publicly release a foundational depth encoder tailored to industrial datasets. Experiments on SiM3D, a recent benchmark that introduces the first multiview and multimodal setup for 3D anomaly detection and segmentation, demonstrate that ModMap attains state-of-the-art performance by surpassing previous methods by wide margins.
|
| 74 |
+
|
| 75 |
+
---
|
| 76 |
+
|
| 77 |
+
## 5. Steerable Visual Representations
|
| 78 |
+
|
| 79 |
+
- **arXiv ID:** `2604.02327`
|
| 80 |
+
- **Primary category:** `cs.CV`
|
| 81 |
+
- **Categories:** `cs.CV`, `cs.AI`
|
| 82 |
+
- **Published:** 2026-04-02
|
| 83 |
+
- **Authors:** Jona Ruthardt, Manu Gaur, Deva Ramanan, Makarand Tapaswi, Yuki M. Asano
|
| 84 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02327](https://arxiv.org/abs/2604.02327)
|
| 85 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02327v1](https://arxiv.org/pdf/2604.02327v1)
|
| 86 |
+
|
| 87 |
+
### Abstract
|
| 88 |
+
|
| 89 |
+
Pretrained Vision Transformers (ViTs) such as DINOv2 and MAE provide generic image features that can be applied to a variety of downstream tasks such as retrieval, classification, and segmentation. However, such representations tend to focus on the most salient visual cues in the image, with no way to direct them toward less prominent concepts of interest. In contrast, Multimodal LLMs can be guided with textual prompts, but the resulting representations tend to be language-centric and lose their effectiveness for generic visual tasks. To address this, we introduce Steerable Visual Representations, a new class of visual representations, whose global and local features can be steered with natural language. While most vision-language models (e.g., CLIP) fuse text with visual features after encoding (late fusion), we inject text directly into the layers of the visual encoder (early fusion) via lightweight cross-attention. We introduce benchmarks for measuring representational steerability, and demonstrate that our steerable visual features can focus on any desired objects in an image while preserving the underlying representation quality. Our method also matches or outperforms dedicated approaches on anomaly detection and personalized object discrimination, exhibiting zero-shot generalization to out-of-distribution tasks.
|
| 90 |
+
|
| 91 |
+
---
|
| 92 |
+
|
| 93 |
+
## 6. Beyond Referring Expressions: Scenario Comprehension Visual Grounding
|
| 94 |
+
|
| 95 |
+
- **arXiv ID:** `2604.02323`
|
| 96 |
+
- **Primary category:** `cs.CV`
|
| 97 |
+
- **Categories:** `cs.CV`
|
| 98 |
+
- **Published:** 2026-04-02
|
| 99 |
+
- **Authors:** Ruozhen He, Nisarg A. Shah, Qihua Dong, Zilin Xiao, Jaywon Koo, Vicente Ordonez
|
| 100 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02323](https://arxiv.org/abs/2604.02323)
|
| 101 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02323v1](https://arxiv.org/pdf/2604.02323v1)
|
| 102 |
+
|
| 103 |
+
### Abstract
|
| 104 |
+
|
| 105 |
+
Existing visual grounding benchmarks primarily evaluate alignment between image regions and literal referring expressions, where models can often succeed by matching a prominent named category. We explore a complementary and more challenging setting of scenario-based visual grounding, where the target must be inferred from roles, intentions, and relational context rather than explicit naming. We introduce Referring Scenario Comprehension (RSC), a benchmark designed for this setting. The queries in this benchmark are paragraph-length texts that describe object roles, user goals, and contextual cues, including deliberate references to distractor objects that often require deep understanding to resolve. Each instance is annotated with interpretable difficulty tags for uniqueness, clutter, size, overlap, and position which expose distinct failure modes and support fine-grained analysis. RSC contains approximately 31k training examples, 4k in-domain test examples, and a 3k out-of-distribution split with unseen object categories. We further propose ScenGround, a curriculum reasoning method serving as a reference point for this setting, combining supervised warm-starting with difficulty-aware reinforcement learning. Experiments show that scenario-based queries expose systematic failures in current models that standard benchmarks do not reveal, and that curriculum training improves performance on challenging slices and transfers to standard benchmarks.
|
| 106 |
+
|
| 107 |
+
---
|
| 108 |
+
|
| 109 |
+
## 7. Large-scale Codec Avatars: The Unreasonable Effectiveness of Large-scale Avatar Pretraining
|
| 110 |
+
|
| 111 |
+
- **arXiv ID:** `2604.02320`
|
| 112 |
+
- **Primary category:** `cs.CV`
|
| 113 |
+
- **Categories:** `cs.CV`, `cs.GR`
|
| 114 |
+
- **Published:** 2026-04-02
|
| 115 |
+
- **Authors:** Junxuan Li, Rawal Khirodkar, Chengan He, Zhongshi Jiang, Giljoo Nam, Lingchen Yang, Jihyun Lee, Egor Zakharov, Zhaoen Su, Rinat Abdrashitov, Yuan Dong, Julieta Martinez, Kai Li, Qingyang Tan, Takaaki Shiratori, Matthew Hu, Peihong Guo, Xuhua Huang, Ariyan Zarei, Marco Pesavento, Yichen Xu, He Wen, Teng Deng, Wyatt Borsos, Anjali Thakrar, Jean-Charles Bazin, Carsten Stoll, Ginés Hidalgo, James Booth, Lucy Wang, Xiaowen Ma, Yu Rong, Sairanjith Thalanki, Chen Cao, Christian Häne, Abhishek Kar, Sofien Bouaziz, Jason Saragih, Yaser Sheikh, Shunsuke Saito
|
| 116 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02320](https://arxiv.org/abs/2604.02320)
|
| 117 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02320v1](https://arxiv.org/pdf/2604.02320v1)
|
| 118 |
+
|
| 119 |
+
### Abstract
|
| 120 |
+
|
| 121 |
+
High-quality 3D avatar modeling faces a critical trade-off between fidelity and generalization. On the one hand, multi-view studio data enables high-fidelity modeling of humans with precise control over expressions and poses, but it struggles to generalize to real-world data due to limited scale and the domain gap between the studio environment and the real world. On the other hand, recent large-scale avatar models trained on millions of in-the-wild samples show promise for generalization across a wide range of identities, yet the resulting avatars are often of low-quality due to inherent 3D ambiguities. To address this, we present Large-Scale Codec Avatars (LCA), a high-fidelity, full-body 3D avatar model that generalizes to world-scale populations in a feedforward manner, enabling efficient inference. Inspired by the success of large language models and vision foundation models, we present, for the first time, a pre/post-training paradigm for 3D avatar modeling at scale: we pretrain on 1M in-the-wild videos to learn broad priors over appearance and geometry, then post-train on high-quality curated data to enhance expressivity and fidelity. LCA generalizes across hair styles, clothing, and demographics while providing precise, fine-grained facial expressions and finger-level articulation control, with strong identity preservation. Notably, we observe emergent generalization to relightability and loose garment support to unconstrained inputs, and zero-shot robustness to stylized imagery, despite the absence of direct supervision.
|
| 122 |
+
|
| 123 |
+
---
|
| 124 |
+
|
| 125 |
+
## 8. Stop Wandering: Efficient Vision-Language Navigation via Metacognitive Reasoning
|
| 126 |
+
|
| 127 |
+
- **arXiv ID:** `2604.02318`
|
| 128 |
+
- **Primary category:** `cs.RO`
|
| 129 |
+
- **Categories:** `cs.RO`, `cs.CV`
|
| 130 |
+
- **Published:** 2026-04-02
|
| 131 |
+
- **Authors:** Xueying Li, Feng Lyu, Hao Wu, Mingliu Liu, Jia-Nan Liu, Guozi Liu
|
| 132 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02318](https://arxiv.org/abs/2604.02318)
|
| 133 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02318v1](https://arxiv.org/pdf/2604.02318v1)
|
| 134 |
+
|
| 135 |
+
### Abstract
|
| 136 |
+
|
| 137 |
+
Training-free Vision-Language Navigation (VLN) agents powered by foundation models can follow instructions and explore 3D environments. However, existing approaches rely on greedy frontier selection and passive spatial memory, leading to inefficient behaviors such as local oscillation and redundant revisiting. We argue that this stems from a lack of metacognitive capabilities: the agent cannot monitor its exploration progress, diagnose strategy failures, or adapt accordingly. To address this, we propose MetaNav, a metacognitive navigation agent integrating spatial memory, history-aware planning, and reflective correction. Spatial memory builds a persistent 3D semantic map. History-aware planning penalizes revisiting to improve efficiency. Reflective correction detects stagnation and uses an LLM to generate corrective rules that guide future frontier selection. Experiments on GOAT-Bench, HM3D-OVON, and A-EQA show that MetaNav achieves state-of-the-art performance while reducing VLM queries by 20.7%, demonstrating that metacognitive reasoning significantly improves robustness and efficiency.
|
| 138 |
+
|
| 139 |
+
---
|
| 140 |
+
|
| 141 |
+
## 9. A Simple Baseline for Streaming Video Understanding
|
| 142 |
+
|
| 143 |
+
- **arXiv ID:** `2604.02317`
|
| 144 |
+
- **Primary category:** `cs.CV`
|
| 145 |
+
- **Categories:** `cs.CV`
|
| 146 |
+
- **Published:** 2026-04-02
|
| 147 |
+
- **Authors:** Yujiao Shen, Shulin Tian, Jingkang Yang, Ziwei Liu
|
| 148 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02317](https://arxiv.org/abs/2604.02317)
|
| 149 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02317v1](https://arxiv.org/pdf/2604.02317v1)
|
| 150 |
+
|
| 151 |
+
### Abstract
|
| 152 |
+
|
| 153 |
+
Recent streaming video understanding methods increasingly rely on complex memory mechanisms to handle long video streams. We challenge this trend with a simple finding: a sliding-window baseline that feeds only the most recent N frames to an off-the-shelf VLM already matches or surpasses published streaming models. We formalize this baseline as SimpleStream and evaluate it against 13 major offline and online video LLM baselines on OVO-Bench and StreamingBench. Despite its simplicity, SimpleStream delivers consistently strong performance. With only 4 recent frames, it reaches 67.7% average accuracy on OVO-Bench and 80.59% on StreamingBench. Controlled ablations further show that the value of longer context is backbone-dependent rather than uniformly increasing with model scale, and reveal a consistent perception-memory trade-off: adding more historical context can improve recall, but often weakens real-time perception. This suggests that stronger memory, retrieval, or compression modules should not be taken as evidence of progress unless they clearly outperform SimpleStream under the same protocol. We therefore argue that future streaming benchmarks should separate recent-scene perception from long-range memory, so that performance improvements from added complexity can be evaluated more clearly.
|
| 154 |
+
|
| 155 |
+
---
|
| 156 |
+
|
| 157 |
+
## 10. VOID: Video Object and Interaction Deletion
|
| 158 |
+
|
| 159 |
+
- **arXiv ID:** `2604.02296`
|
| 160 |
+
- **Primary category:** `cs.CV`
|
| 161 |
+
- **Categories:** `cs.CV`, `cs.AI`
|
| 162 |
+
- **Published:** 2026-04-02
|
| 163 |
+
- **Authors:** Saman Motamed, William Harvey, Benjamin Klein, Luc Van Gool, Zhuoning Yuan, Ta-Ying Cheng
|
| 164 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02296](https://arxiv.org/abs/2604.02296)
|
| 165 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02296v1](https://arxiv.org/pdf/2604.02296v1)
|
| 166 |
+
|
| 167 |
+
### Abstract
|
| 168 |
+
|
| 169 |
+
Existing video object removal methods excel at inpainting content "behind" the object and correcting appearance-level artifacts such as shadows and reflections. However, when the removed object has more significant interactions, such as collisions with other objects, current models fail to correct them and produce implausible results. We present VOID, a video object removal framework designed to perform physically-plausible inpainting in these complex scenarios. To train the model, we generate a new paired dataset of counterfactual object removals using Kubric and HUMOTO, where removing an object requires altering downstream physical interactions. During inference, a vision-language model identifies regions of the scene affected by the removed object. These regions are then used to guide a video diffusion model that generates physically consistent counterfactual outcomes. Experiments on both synthetic and real data show that our approach better preserves consistent scene dynamics after object removal compared to prior video object removal methods. We hope this framework sheds light on how to make video editing models better simulators of the world through high-level causal reasoning.
|
| 170 |
+
|
| 171 |
+
---
|
| 172 |
+
|
| 173 |
+
## 11. AdamFlow: Adam-based Wasserstein Gradient Flows for Surface Registration in Medical Imaging
|
| 174 |
+
|
| 175 |
+
- **arXiv ID:** `2604.02290`
|
| 176 |
+
- **Primary category:** `cs.CV`
|
| 177 |
+
- **Categories:** `cs.CV`, `math.OC`
|
| 178 |
+
- **Published:** 2026-04-02
|
| 179 |
+
- **Authors:** Qiang Ma, Qingjie Meng, Xin Hu, Yicheng Wu, Wenjia Bai
|
| 180 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02290](https://arxiv.org/abs/2604.02290)
|
| 181 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02290v1](https://arxiv.org/pdf/2604.02290v1)
|
| 182 |
+
|
| 183 |
+
### Abstract
|
| 184 |
+
|
| 185 |
+
Surface registration plays an important role for anatomical shape analysis in medical imaging. Existing surface registration methods often face a trade-off between efficiency and robustness. Local point matching methods are computationally efficient, but vulnerable to noise and initialisation. Methods designed for global point set alignment tend to incur a high computational cost. To address the challenge, here we present a fast surface registration method, which formulates surface meshes as probability measures and surface registration as a distributional optimisation problem. The discrepancy between two meshes is measured using an efficient sliced Wasserstein distance with log-linear computational complexity. We propose a novel optimisation method, AdamFlow, which generalises the well-known Adam optimisation method from the Euclidean space to the probability space for minimising the sliced Wasserstein distance. We theoretically analyse the asymptotic convergence of AdamFlow and empirically demonstrate its superior performance in both affine and non-rigid surface registration across various anatomical structures.
|
| 186 |
+
|
| 187 |
+
---
|
| 188 |
+
|
| 189 |
+
## 12. Omni123: Exploring 3D Native Foundation Models with Limited 3D Data by Unifying Text to 2D and 3D Generation
|
| 190 |
+
|
| 191 |
+
- **arXiv ID:** `2604.02289`
|
| 192 |
+
- **Primary category:** `cs.CV`
|
| 193 |
+
- **Categories:** `cs.CV`, `cs.AI`
|
| 194 |
+
- **Published:** 2026-04-02
|
| 195 |
+
- **Authors:** Chongjie Ye, Cheng Cao, Chuanyu Pan, Yiming Hao, Yihao Zhi, Yuanming Hu, Xiaoguang Han
|
| 196 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02289](https://arxiv.org/abs/2604.02289)
|
| 197 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02289v1](https://arxiv.org/pdf/2604.02289v1)
|
| 198 |
+
|
| 199 |
+
### Abstract
|
| 200 |
+
|
| 201 |
+
Recent multimodal large language models have achieved strong performance in unified text and image understanding and generation, yet extending such native capability to 3D remains challenging due to limited data. Compared to abundant 2D imagery, high-quality 3D assets are scarce, making 3D synthesis under-constrained. Existing methods often rely on indirect pipelines that edit in 2D and lift results into 3D via optimization, sacrificing geometric consistency. We present Omni123, a 3D-native foundation model that unifies text-to-2D and text-to-3D generation within a single autoregressive framework. Our key insight is that cross-modal consistency between images and 3D can serve as an implicit structural constraint. By representing text, images, and 3D as discrete tokens in a shared sequence space, the model leverages abundant 2D data as a geometric prior to improve 3D representations. We introduce an interleaved X-to-X training paradigm that coordinates diverse cross-modal tasks over heterogeneous paired datasets without requiring fully aligned text-image-3D triplets. By traversing semantic-visual-geometric cycles (e.g., text to image to 3D to image) within autoregressive sequences, the model jointly enforces semantic alignment, appearance fidelity, and multi-view geometric consistency. Experiments show that Omni123 significantly improves text-guided 3D generation and editing, demonstrating a scalable path toward multimodal 3D world models.
|
| 202 |
+
|
| 203 |
+
---
|
| 204 |
+
|
| 205 |
+
## 13. Deep Neural Network Based Roadwork Detection for Autonomous Driving
|
| 206 |
+
|
| 207 |
+
- **arXiv ID:** `2604.02282`
|
| 208 |
+
- **Primary category:** `cs.RO`
|
| 209 |
+
- **Categories:** `cs.RO`, `cs.CV`
|
| 210 |
+
- **Published:** 2026-04-02
|
| 211 |
+
- **Authors:** Sebastian Wullrich, Nicolai Steinke, Daniel Goehring
|
| 212 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02282](https://arxiv.org/abs/2604.02282)
|
| 213 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02282v1](https://arxiv.org/pdf/2604.02282v1)
|
| 214 |
+
|
| 215 |
+
### Abstract
|
| 216 |
+
|
| 217 |
+
Road construction sites create major challenges for both autonomous vehicles and human drivers due to their highly dynamic and heterogeneous nature. This paper presents a real-time system that detects and localizes roadworks by combining a YOLO neural network with LiDAR data. The system identifies individual roadwork objects while driving, merges them into coherent construction sites and records their outlines in world coordinates. The model training was based on an adapted US dataset and a new dataset collected from test drives with a prototype vehicle in Berlin, Germany. Evaluations on real-world road construction sites showed a localization accuracy below 0.5 m. The system can support traffic authorities with up-to-date roadwork data and could enable autonomous vehicles to navigate construction sites more safely in the future.
|
| 218 |
+
|
| 219 |
+
---
|
| 220 |
+
|
| 221 |
+
## 14. Novel Memory Forgetting Techniques for Autonomous AI Agents: Balancing Relevance and Efficiency
|
| 222 |
+
|
| 223 |
+
- **arXiv ID:** `2604.02280`
|
| 224 |
+
- **Primary category:** `cs.AI`
|
| 225 |
+
- **Categories:** `cs.AI`, `cs.CV`
|
| 226 |
+
- **Published:** 2026-04-02
|
| 227 |
+
- **Authors:** Payal Fofadiya, Sunil Tiwari
|
| 228 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02280](https://arxiv.org/abs/2604.02280)
|
| 229 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02280v1](https://arxiv.org/pdf/2604.02280v1)
|
| 230 |
+
|
| 231 |
+
### Abstract
|
| 232 |
+
|
| 233 |
+
Long-horizon conversational agents require persistent memory for coherent reasoning, yet uncontrolled accumulation causes temporal decay and false memory propagation. Benchmarks such as LOCOMO and LOCCO report performance degradation from 0.455 to 0.05 across stages, while MultiWOZ shows 78.2% accuracy with 6.8% false memory rate under persistent retention. This work introduces an adaptive budgeted forgetting framework that regulates memory through relevanceguided scoring and bounded optimization. The approach integrates recency, frequency, and semantic alignment to maintain stability under constrained context. Comparative analysis demonstrates improved long-horizon F1 beyond 0.583 baseline levels, higher retention consistency, and reduced false memory behavior without increasing context usage. These findings confirm that structured forgetting preserves reasoning performance while preventing unbounded memory growth in extended conversational settings.
|
| 234 |
+
|
| 235 |
+
---
|
| 236 |
+
|
| 237 |
+
## 15. Modular Energy Steering for Safe Text-to-Image Generation with Foundation Models
|
| 238 |
+
|
| 239 |
+
- **arXiv ID:** `2604.02265`
|
| 240 |
+
- **Primary category:** `cs.CV`
|
| 241 |
+
- **Categories:** `cs.CV`
|
| 242 |
+
- **Published:** 2026-04-02
|
| 243 |
+
- **Authors:** Yaoteng Tan, Zikui Cai, M. Salman Asif
|
| 244 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02265](https://arxiv.org/abs/2604.02265)
|
| 245 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02265v1](https://arxiv.org/pdf/2604.02265v1)
|
| 246 |
+
|
| 247 |
+
### Abstract
|
| 248 |
+
|
| 249 |
+
Controlling the behavior of text-to-image generative models is critical for safe and practical deployment. Existing safety approaches typically rely on model fine-tuning or curated datasets, which can degrade generation quality or limit scalability. We propose an inference-time steering framework that leverages gradient feedback from frozen pretrained foundation models to guide the generation process without modifying the underlying generator. Our key observation is that vision-language foundation models encode rich semantic representations that can be repurposed as off-the-shelf supervisory signals during generation. By injecting such feedback through clean latent estimates at each sampling step, our method formulates safety steering as an energy-based sampling problem. This design enables modular, training-free safety control that is compatible with both diffusion and flow-matching models and can generalize across diverse visual concepts. Experiments demonstrate state-of-the-art robustness against NSFW red-teaming benchmarks and effective multi-target steering, while preserving high generation quality on benign non-targeted prompts. Our framework provides a principled approach for utilizing foundation models as semantic energy estimators, enabling reliable and scalable safety control for text-to-image generation.
|
| 250 |
+
|
| 251 |
+
---
|
| 252 |
+
|
| 253 |
+
## 16. SPAR: Single-Pass Any-Resolution ViT for Open-vocabulary Segmentation
|
| 254 |
+
|
| 255 |
+
- **arXiv ID:** `2604.02252`
|
| 256 |
+
- **Primary category:** `cs.CV`
|
| 257 |
+
- **Categories:** `cs.CV`
|
| 258 |
+
- **Published:** 2026-04-02
|
| 259 |
+
- **Authors:** Naomi Kombol, Ivan Martinović, Siniša Šegvić, Giorgos Tolias
|
| 260 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02252](https://arxiv.org/abs/2604.02252)
|
| 261 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02252v1](https://arxiv.org/pdf/2604.02252v1)
|
| 262 |
+
|
| 263 |
+
### Abstract
|
| 264 |
+
|
| 265 |
+
Foundational Vision Transformers (ViTs) have limited effectiveness in tasks requiring fine-grained spatial understanding, due to their fixed pre-training resolution and inherently coarse patch-level representations. These challenges are especially pronounced in dense prediction scenarios, such as open-vocabulary segmentation with ViT-based vision-language models, where high-resolution inputs are essential for accurate pixel-level reasoning. Existing approaches typically process large-resolution images using a sliding-window strategy at the pre-training resolution. While this improves accuracy through finer strides, it comes at a significant computational cost. We introduce SPAR: Single-Pass Any-Resolution ViT, a resolution-agnostic dense feature extractor designed for efficient high-resolution inference. We distill the spatial reasoning capabilities of a finely-strided, sliding-window teacher into a single-pass student using a feature regression loss, without requiring architectural changes or pixel-level supervision. Applied to open-vocabulary segmentation, SPAR improves single-pass baselines by up to 10.5 mIoU and even surpasses the teacher, demonstrating effectiveness in efficient, high-resolution reasoning. Code: [https://github.com/naomikombol/SPAR](https://github.com/naomikombol/SPAR)
|
| 266 |
+
|
| 267 |
+
---
|
| 268 |
+
|
| 269 |
+
## 17. UAV-Track VLA: Embodied Aerial Tracking via Vision-Language-Action Models
|
| 270 |
+
|
| 271 |
+
- **arXiv ID:** `2604.02241`
|
| 272 |
+
- **Primary category:** `cs.CV`
|
| 273 |
+
- **Categories:** `cs.CV`, `cs.RO`
|
| 274 |
+
- **Published:** 2026-04-02
|
| 275 |
+
- **Authors:** Qiyao Zhang, Shuhua Zheng, Jianli Sun, Chengxiang Li, Xianke Wu, Zihan Song, Zhiyong Cui, Yisheng Lv, Yonglin Tian
|
| 276 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02241](https://arxiv.org/abs/2604.02241)
|
| 277 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02241v1](https://arxiv.org/pdf/2604.02241v1)
|
| 278 |
+
|
| 279 |
+
### Abstract
|
| 280 |
+
|
| 281 |
+
Embodied visual tracking is crucial for Unmanned Aerial Vehicles (UAVs) executing complex real-world tasks. In dynamic urban scenarios with complex semantic requirements, Vision-Language-Action (VLA) models show great promise due to their cross-modal fusion and continuous action generation capabilities. To benchmark multimodal tracking in such environments, we construct a dedicated evaluation benchmark and a large-scale dataset encompassing over 890K frames, 176 tasks, and 85 diverse objects. Furthermore, to address temporal feature redundancy and the lack of spatial geometric priors in existing VLA models, we propose an improved VLA tracking model, UAV-Track VLA. Built upon the $π_{0.5}$ architecture, our model introduces a temporal compression net to efficiently capture inter-frame dynamics. Additionally, a parallel dual-branch decoder comprising a spatial-aware auxiliary grounding head and a flow matching action expert is designed to decouple cross-modal features and generate fine-grained continuous actions. Systematic experiments in the CARLA simulator validate the superior end-to-end performance of our method. Notably, in challenging long-distance pedestrian tracking tasks, UAV-Track VLA achieves a 61.76 success rate and 269.65 average tracking frames, significantly outperforming existing baselines. Furthermore, it demonstrates robust zero-shot generalization in unseen environments and reduces single-step inference latency by 33.4 (to 0.0571s) compared to the original $π_{0.5}$, enabling highly efficient, real-time UAV control. Data samples and demonstration videos are available at: [https://github.com/Hub-Tian/UAV-Track\_VLA](https://github.com/Hub-Tian/UAV-Track\_VLA).
|
| 282 |
+
|
| 283 |
+
---
|
| 284 |
+
|
| 285 |
+
## 18. SCALE: Semantic- and Confidence-Aware Conditional Variational Autoencoder for Zero-shot Skeleton-based Action Recognition
|
| 286 |
+
|
| 287 |
+
- **arXiv ID:** `2604.02222`
|
| 288 |
+
- **Primary category:** `cs.CV`
|
| 289 |
+
- **Categories:** `cs.CV`
|
| 290 |
+
- **Published:** 2026-04-02
|
| 291 |
+
- **Authors:** Soroush Oraki, Feng Ding, Jie Liang
|
| 292 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02222](https://arxiv.org/abs/2604.02222)
|
| 293 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02222v1](https://arxiv.org/pdf/2604.02222v1)
|
| 294 |
+
|
| 295 |
+
### Abstract
|
| 296 |
+
|
| 297 |
+
Zero-shot skeleton-based action recognition (ZSAR) aims to recognize action classes without any training skeletons from those classes, relying instead on auxiliary semantics from text. Existing approaches frequently depend on explicit skeleton-text alignment, which can be brittle when action names underspecify fine-grained dynamics and when unseen classes are semantically confusable. We propose SCALE, a lightweight and deterministic Semantic- and Confidence-Aware Listwise Energy-based framework that formulates ZSAR as class-conditional energy ranking. SCALE builds a text-conditioned Conditional Variational Autoencoder where frozen text representations parameterize both the latent prior and the decoder, enabling likelihood-based evaluation for unseen classes without generating samples at test time. To separate competing hypotheses, we introduce a semantic- and confidence-aware listwise energy loss that emphasizes semantically similar hard negatives and incorporates posterior uncertainty to adapt decision margins and reweight ambiguous training instances. Additionally, we utilize a latent prototype contrast objective to align posterior means with text-derived latent prototypes, improving semantic organization and class separability without direct feature matching. Experiments on NTU-60 and NTU-120 datasets show that SCALE consistently improves over prior VAE- and alignment-based baselines while remaining competitive with diffusion-based methods.
|
| 298 |
+
|
| 299 |
+
---
|
| 300 |
+
|
| 301 |
+
## 19. UniDriveVLA: Unifying Understanding, Perception, and Action Planning for Autonomous Driving
|
| 302 |
+
|
| 303 |
+
- **arXiv ID:** `2604.02190`
|
| 304 |
+
- **Primary category:** `cs.CV`
|
| 305 |
+
- **Categories:** `cs.CV`, `cs.RO`
|
| 306 |
+
- **Published:** 2026-04-02
|
| 307 |
+
- **Authors:** Yongkang Li, Lijun Zhou, Sixu Yan, Bencheng Liao, Tianyi Yan, Kaixin Xiong, Long Chen, Hongwei Xie, Bing Wang, Guang Chen, Hangjun Ye, Wenyu Liu, Haiyang Sun, Xinggang Wang
|
| 308 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02190](https://arxiv.org/abs/2604.02190)
|
| 309 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02190v1](https://arxiv.org/pdf/2604.02190v1)
|
| 310 |
+
|
| 311 |
+
### Abstract
|
| 312 |
+
|
| 313 |
+
Vision-Language-Action (VLA) models have recently emerged in autonomous driving, with the promise of leveraging rich world knowledge to improve the cognitive capabilities of driving systems. However, adapting such models for driving tasks currently faces a critical dilemma between spatial perception and semantic reasoning. Consequently, existing VLA systems are forced into suboptimal compromises: directly adopting 2D Vision-Language Models yields limited spatial perception, whereas enhancing them with 3D spatial representations often impairs the native reasoning capacity of VLMs. We argue that this dilemma largely stems from the coupled optimization of spatial perception and semantic reasoning within shared model parameters. To overcome this, we propose UniDriveVLA, a Unified Driving Vision-Language-Action model based on Mixture-of-Transformers that addresses the perception-reasoning conflict via expert decoupling. Specifically, it comprises three experts for driving understanding, scene perception, and action planning, which are coordinated through masked joint attention. In addition, we combine a sparse perception paradigm with a three-stage progressive training strategy to improve spatial perception while maintaining semantic reasoning capability. Extensive experiments show that UniDriveVLA achieves state-of-the-art performance in open-loop evaluation on nuScenes and closed-loop evaluation on Bench2Drive. Moreover, it demonstrates strong performance across a broad range of perception, prediction, and understanding tasks, including 3D detection, online mapping, motion forecasting, and driving-oriented VQA, highlighting its broad applicability as a unified model for autonomous driving. Code and model have been released at [https://github.com/xiaomi-research/unidrivevla](https://github.com/xiaomi-research/unidrivevla)
|
| 314 |
+
|
| 315 |
+
---
|
| 316 |
+
|
| 317 |
+
## 20. Lightweight Spatiotemporal Highway Lane Detection via 3D-ResNet and PINet with ROI-Aware Attention
|
| 318 |
+
|
| 319 |
+
- **arXiv ID:** `2604.02188`
|
| 320 |
+
- **Primary category:** `cs.CV`
|
| 321 |
+
- **Categories:** `cs.CV`
|
| 322 |
+
- **Published:** 2026-04-02
|
| 323 |
+
- **Authors:** Sorna Shanmuga Raja, Abdelhafid Zenati
|
| 324 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02188](https://arxiv.org/abs/2604.02188)
|
| 325 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02188v1](https://arxiv.org/pdf/2604.02188v1)
|
| 326 |
+
|
| 327 |
+
### Abstract
|
| 328 |
+
|
| 329 |
+
This paper presents a lightweight, end-to-end highway lane detection architecture that jointly captures spatial and temporal information for robust performance in real-world driving scenarios. Building on the strengths of 3D convolutional neural networks and instance segmentation, we propose two models that integrate a 3D-ResNet encoder with a Point Instance Network (PINet) decoder. The first model enhances multi-scale feature representation using a Feature Pyramid Network (FPN) and Self-Attention mechanism to refine spatial dependencies. The second model introduces a Region of Interest (ROI) detection head to selectively focus on lane-relevant regions, thereby improving precision and reducing computational complexity. Experiments conducted on the TuSimple dataset (highway driving scenarios) demonstrate that the proposed second model achieves 93.40% accuracy while significantly reducing false negatives. Compared to existing 2D and 3D baselines, our approach achieves improved performance with fewer parameters and reduced latency. The architecture has been validated through offline training and real-time inference in the Autonomous Systems Laboratory at City, St George's University of London. These results suggest that the proposed models are well-suited for integration into Advanced Driver Assistance Systems (ADAS), with potential scalability toward full Lane Assist Systems (LAS).
|
| 330 |
+
|
| 331 |
+
---
|
| 332 |
+
|
| 333 |
+
## 21. CXR-LT 2026 Challenge: Projection-Aware Multi-Label and Zero-Shot Chest X-Ray Classification
|
| 334 |
+
|
| 335 |
+
- **arXiv ID:** `2604.02185`
|
| 336 |
+
- **Primary category:** `cs.CV`
|
| 337 |
+
- **Categories:** `cs.CV`
|
| 338 |
+
- **Published:** 2026-04-02
|
| 339 |
+
- **Authors:** Juno Cho, Dohui Kim, Mingeon Kim, Hyunseo Jang, Chang Sun Lee, Jong Chul Ye
|
| 340 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02185](https://arxiv.org/abs/2604.02185)
|
| 341 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02185v1](https://arxiv.org/pdf/2604.02185v1)
|
| 342 |
+
|
| 343 |
+
### Abstract
|
| 344 |
+
|
| 345 |
+
This challenge tackles multi-label classification for known chest X-ray (CXR) lesions and zero-shot classification for unseen ones. To handle diverse CXR projections, we integrate projection-specific models via a classification network into a unified framework. For zero-shot classification (Task 2), we extend CheXzero with a novel dual-branch architecture that combines contrastive learning, Asymmetric Loss (ASL), and LLM-generated descriptive prompts. This effectively mitigates severe long-tail imbalances and maximizes zero-shot generalization. Additionally, strong data and test-time augmentations (TTA) ensure robustness across both tasks.
|
| 346 |
+
|
| 347 |
+
---
|
| 348 |
+
|
| 349 |
+
## 22. ViT-Explainer: An Interactive Walkthrough of the Vision Transformer Pipeline
|
| 350 |
+
|
| 351 |
+
- **arXiv ID:** `2604.02182`
|
| 352 |
+
- **Primary category:** `cs.CV`
|
| 353 |
+
- **Categories:** `cs.CV`, `cs.HC`
|
| 354 |
+
- **Published:** 2026-04-02
|
| 355 |
+
- **Authors:** Juan Manuel Hernandez, Mariana Fernandez-Espinosa, Denis Parra, Diego Gomez-Zara
|
| 356 |
+
- **Abstract page:** [https://arxiv.org/abs/2604.02182](https://arxiv.org/abs/2604.02182)
|
| 357 |
+
- **PDF:** [https://arxiv.org/pdf/2604.02182v1](https://arxiv.org/pdf/2604.02182v1)
|
| 358 |
+
|
| 359 |
+
### Abstract
|
| 360 |
+
|
| 361 |
+
Transformer-based architectures have become the shared backbone of natural language processing and computer vision. However, understanding how these models operate remains challenging, particularly in vision settings, where images are processed as sequences of patch tokens. Existing interpretability tools often focus on isolated components or expert-oriented analysis, leaving a gap in guided, end-to-end understanding of the full inference pipeline. To bridge this gap, we present ViT-Explainer, a web-based interactive system that provides an integrated visualization of Vision Transformer inference, from patch tokenization to final classification. The system combines animated walkthroughs, patch-level attention overlays, and a vision-adapted Logit Lens within both guided and free exploration modes. A user study with six participants suggests that ViT-Explainer is easy to learn and use, helping users interpret and understand Vision Transformer behavior.
|
| 362 |
+
|
| 363 |
+
---
|
docs/AI_Computer_Vision/arxiv-cs-cv-2026-04-05.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "2d2fdab41e12646b", "ingested_at": "2026-06-30T08:49:23.959865", "content_hash": "2d2fdab41e12646b"}
|
docs/AI_Computer_Vision/arxiv_video_analysis_papers.md
ADDED
|
@@ -0,0 +1,74 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# arXiv 動画解析・理解 関連論文リスト
|
| 2 |
+
|
| 3 |
+
arXivから収集した「動画の解析(Video Analysis / Video Understanding)」に関する代表的かつ最新の学術論文リストです。
|
| 4 |
+
|
| 5 |
+
本リストは **`literature-search-arxiv`** スキルを使用して取得されました。
|
| 6 |
+
|
| 7 |
+
---
|
| 8 |
+
|
| 9 |
+
## 1. 包括的サーベイ論文(全体の動向把握に最適)
|
| 10 |
+
|
| 11 |
+
### **Video Understanding: From Geometry and Semantics to Unified Models**
|
| 12 |
+
* **著者**: Zhaochong An, Zirui Li, Mingqiao Ye, Feng Qiao, Jiaang Li, Zongwei Wu, Vishal Thengane, Chengzu Li, Lei Li, Luc Van Gool, Guolei Sun, Serge Belongie
|
| 13 |
+
* **公開日**: 2026年3月18日
|
| 14 |
+
* **arXiv ID**: [2603.17840v1](https://arxiv.org/abs/2603.17840v1)
|
| 15 |
+
* **PDFリンク**: [https://arxiv.org/pdf/2603.17840v1](https://arxiv.org/pdf/2603.17840v1)
|
| 16 |
+
* **概要**:
|
| 17 |
+
動画理解の進展を、低レベルの幾何学(geometry)理解、高レベルのセマンティクス理解、および「統一型動画理解モデル」の3つの視点から構造的にまとめた包括的な最新サーベイです。従来のタスク特化型パイプライン設計から、マルチモーダルかつ多様な下流タスクに適応可能な動画基盤モデル(Video Foundation Model)への移行について詳しく論じています。
|
| 18 |
+
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
## 2. 長尺動画理解とプロンプト最適化(最新技術)
|
| 22 |
+
|
| 23 |
+
### **CoS: Chain-of-Shot Prompting for Long Video Understanding**
|
| 24 |
+
* **著者**: Jian Hu, Zixu Cheng, Chenyang Si, Wei Li, Shaogang Gong
|
| 25 |
+
* **公開日**: 2025年2月10日
|
| 26 |
+
* **arXiv ID**: [2502.06428v2](https://arxiv.org/abs/2502.06428v2)
|
| 27 |
+
* **PDFリンク**: [https://arxiv.org/pdf/2502.06428v2](https://arxiv.org/pdf/2502.06428v2)
|
| 28 |
+
* **概要**:
|
| 29 |
+
マルチモーダル大規模言語モデル(MLLM)が長尺動画を扱う際、視覚トークンの多さによる文脈制限や余分なシーン(不要なショット)の混入が課題となります。本論文では、タスクに応じて意味的に重要度の高いショットを動的に選別するテスト時プロンプト最適化手法「Chain-of-Shot (CoS)」を提案しています。
|
| 30 |
+
|
| 31 |
+
---
|
| 32 |
+
|
| 33 |
+
## 3. 動画基盤モデル(大規模マルチモーダル学習)
|
| 34 |
+
|
| 35 |
+
### **InternVideo2: Scaling Foundation Models for Multimodal Video Understanding**
|
| 36 |
+
* **著者**: Yi Wang, Kunchang Li, Xinhao Li, Jiashuo Yu, Yinan He, Chenting Wang, Guo Chen, Baoqi Pei, Ziang Yan, Rongkun Zheng, Jilan Xu, Zun Wang, Yansong Shi, Tianxiang Jiang, Songze Li, Hongjie Zhang, Yifei Huang, Yu Qiao, Yali Wang, Limin Wang
|
| 37 |
+
* **公開日**: 2024年3月22日(ECCV 2024採択)
|
| 38 |
+
* **arXiv ID**: [2403.15377v4](https://arxiv.org/abs/2403.15377v4)
|
| 39 |
+
* **PDFリンク**: [https://arxiv.org/pdf/2403.15377v4](https://arxiv.org/pdf/2403.15377v4)
|
| 40 |
+
* **概要**:
|
| 41 |
+
動画エンコーダを60億パラメータ(6B)規模にスケールアップした動画基盤モデル「InternVideo2」を提案しています。Masked Video Modeling、対照学習、Next Token Predictionを統一した段階的学習により、動画認識や動画対話など60以上の多様な動画関連タスクで高いパフォーマンスを実証しています。
|
| 42 |
+
|
| 43 |
+
---
|
| 44 |
+
|
| 45 |
+
## 4. 人間中心の複雑な行動解析とデータセット
|
| 46 |
+
|
| 47 |
+
### **Human in Events: A Large-Scale Benchmark for Human-centric Video Analysis in Complex Events**
|
| 48 |
+
* **著者**: Weiyao Lin, Huabin Liu, Shizhan Liu, Yuxi Li, Rui Qian, Tao Wang, Ning Xu, Hongkai Xiong, Guo-Jun Qi, Nicu Sebe
|
| 49 |
+
* **公開日**: 2020年5月9日(IJCV 2023掲載)
|
| 50 |
+
* **arXiv ID**: [2005.04490v6](https://arxiv.org/abs/2005.04490v6)
|
| 51 |
+
* **PDFリンク**: [https://arxiv.org/pdf/2005.04490v6](https://arxiv.org/pdf/2005.04490v6)
|
| 52 |
+
* **概要**:
|
| 53 |
+
実世界の混雑シーンや異常行動が発生する複雑なイベントにおいて、人間の姿勢・動作・軌跡を追跡する「Human-in-Events (HiEve)」という大規模データセットを提案。姿勢推定や行動分類の学習用ベンチマークとして広く活用されています。
|
| 54 |
+
|
| 55 |
+
---
|
| 56 |
+
|
| 57 |
+
## 5. 効率的なオンライン動画解析(高速化技術)
|
| 58 |
+
|
| 59 |
+
### **ECO: Efficient Convolutional Network for Online Video Understanding**
|
| 60 |
+
* **著者**: Mohammadreza Zolfaghari, Kamaljeet Singh, Thomas Brox
|
| 61 |
+
* **公開日**: 2018年4月24日(ECCV 2018)
|
| 62 |
+
* **arXiv ID**: [1804.09066v2](https://arxiv.org/abs/1804.09066v2)
|
| 63 |
+
* **PDFリンク**: [https://arxiv.org/pdf/1804.09066v2](https://arxiv.org/pdf/1804.09066v2)
|
| 64 |
+
* **概要**:
|
| 65 |
+
動画解析における計算コストの課題に対し、フレーム間の冗長性を削減するサンプリング戦略と、ネットワーク内での長期特徴マージ手法を組み合わせた「ECO(Efficient Convolutional Network)」を提案。既存手法と比較して大幅な高速処理(10倍〜80倍)を実現し、オンラインでのアクティビティ分類に貢献しました。
|
| 66 |
+
|
| 67 |
+
---
|
| 68 |
+
|
| 69 |
+
## 参照したペーパーのURL一覧
|
| 70 |
+
* [https://arxiv.org/pdf/2603.17840v1](https://arxiv.org/pdf/2603.17840v1)
|
| 71 |
+
* [https://arxiv.org/pdf/2502.06428v2](https://arxiv.org/pdf/2502.06428v2)
|
| 72 |
+
* [https://arxiv.org/pdf/2403.15377v4](https://arxiv.org/pdf/2403.15377v4)
|
| 73 |
+
* [https://arxiv.org/pdf/2005.04490v6](https://arxiv.org/pdf/2005.04490v6)
|
| 74 |
+
* [https://arxiv.org/pdf/1804.09066v2](https://arxiv.org/pdf/1804.09066v2)
|
docs/AI_Computer_Vision/arxiv_video_analysis_papers.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "e1729b3a6980a520", "ingested_at": "2026-06-30T08:49:24.586622", "content_hash": "e1729b3a6980a520"}
|
docs/AI_Computer_Vision/layout_models_survey.md
ADDED
|
@@ -0,0 +1,370 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
# レイアウト分析・生成モデル 包括的サーベイ
|
| 2 |
+
|
| 3 |
+
> [!NOTE]
|
| 4 |
+
> 本レポートは、**文書・画像・動画**におけるレイアウト分析モデルとレイアウト生成モデルの最新動向(2024〜2025年)を網羅的にまとめたものです。
|
| 5 |
+
|
| 6 |
+
---
|
| 7 |
+
|
| 8 |
+
## 目次
|
| 9 |
+
|
| 10 |
+
1. [文書レイアウト分析モデル](#1-文書レイアウト分析モデル)
|
| 11 |
+
2. [レイアウト生成モデル(グラフィック/UI/文書)](#2-レイアウト生成モデルグラフィックui文書)
|
| 12 |
+
3. [画像レイアウト生成モデル(Layout-to-Image)](#3-画像レイアウト生成モデルlayout-to-image)
|
| 13 |
+
4. [動画レイアウト分析・生成モデル](#4-動画レイアウト分析生成モデル)
|
| 14 |
+
5. [VLM/LLMベースのレイアウト理解モデル](#5-vlmllmベースのレイアウト理解モデル)
|
| 15 |
+
6. [商用レイアウト分析ソリューション](#6-商用レイアウト分析ソリューション)
|
| 16 |
+
7. [主要データセット・ベンチマーク](#7-主要データセットベンチマーク)
|
| 17 |
+
8. [技術トレンド総括](#8-技術トレンド総括)
|
| 18 |
+
|
| 19 |
+
---
|
| 20 |
+
|
| 21 |
+
## 1. 文書レイアウト分析モデル
|
| 22 |
+
|
| 23 |
+
文書レイアウト分析(Document Layout Analysis: DLA)は、文書画像からテキスト、表、図、タイトル等の構造要素を検出・分類するタスクです。
|
| 24 |
+
|
| 25 |
+
### 1.1 マルチモーダルTransformerベース
|
| 26 |
+
|
| 27 |
+
| モデル | 年 | アーキテクチャ | 特徴 | リンク |
|
| 28 |
+
|:---|:---|:---|:---|:---|
|
| 29 |
+
| **LayoutLM** | 2020 | BERT + 2D Position Embedding | テキスト+空間位置情報の統合学習のパイオニア | [GitHub](https://github.com/microsoft/unilm/tree/master/layoutlm) |
|
| 30 |
+
| **LayoutLMv2** | 2021 | Transformer + CNN Backbone | 視覚情報の追加、空間アテンション導入 | [HuggingFace](https://huggingface.co/microsoft/layoutlmv2-base-uncased) |
|
| 31 |
+
| **LayoutLMv3** | 2022 | Unified Transformer (Patch Embed) | CNNをViTスタイルに置換、統一的マスクモデリング | [HuggingFace](https://huggingface.co/microsoft/layoutlmv3-base) |
|
| 32 |
+
| **LayoutXLM** | 2022 | 多言語LayoutLMv2 | 53言語対応、多言語文書理解 | [HuggingFace](https://huggingface.co/microsoft/layoutxlm-base) |
|
| 33 |
+
|
| 34 |
+
**特長**: テキスト・視覚・レイアウトの3モダリティを統合し、フォーム理解、文書VQA、エンティティ抽出など**意味理解**を必要とするタスクに最適。
|
| 35 |
+
|
| 36 |
+
### 1.2 物体検出ベース(ビジョンのみ)
|
| 37 |
+
|
| 38 |
+
| モデル | 年 | ベース | 特徴 | リンク |
|
| 39 |
+
|:---|:---|:---|:---|:---|
|
| 40 |
+
| **DocLayout-YOLO** | 2024 | YOLO | Global-to-Local Controllable Receptive Module、DocSynth-300K合成データで学習 | [GitHub](https://github.com/opendatalab/DocLayout-YOLO) |
|
| 41 |
+
| **PP-DocLayout** | 2024 | PaddleDetection | M/Sバリアント、リソース制約環境向け最適化 | PaddlePaddle |
|
| 42 |
+
| **Heron-101** | 2025 | RT-DETR | Doclingパイプラインの一部、28ms/imageの高速推論、78% mAP | [Docling](https://github.com/DS4SD/docling) |
|
| 43 |
+
| **LayoutParser** | 2021 | Detectron2 | レイアウト検出の統合ライブラリ、複数モデル対応 | [GitHub](https://github.com/Layout-Parser/layout-parser) |
|
| 44 |
+
| **chunkr-layout-1** | 2025 | — | 80.9% mAP@50、金融・法律・医療文書向けプロダクションモデル | [chunkr.ai](https://chunkr.ai) |
|
| 45 |
+
|
| 46 |
+
**特長**: テキスト情報を使わず画像のみでレイアウト要素を検出。**高速推論**が必要なパイプラインに最適。
|
| 47 |
+
|
| 48 |
+
### 1.3 統一Transformerベース
|
| 49 |
+
|
| 50 |
+
| モデル | 年 | アーキテクチャ | 特徴 |
|
| 51 |
+
|:---|:---|:---|:---|
|
| 52 |
+
| **DLAFormer** | 2024 (ICDAR) | DETR-inspired Encoder-Decoder | 物体検出・テキスト領域検出・論理役割分類・読み順予測を統一的に関係予測として処理 |
|
| 53 |
+
| **RT-DETRv2 / DFINE** | 2024-2025 | Real-Time DETR | リアルタイム推論と高精度のバランス |
|
| 54 |
+
|
| 55 |
+
### 1.4 グラフベース手法
|
| 56 |
+
|
| 57 |
+
| モデル | 年 | 特徴 |
|
| 58 |
+
|:---|:---|:---|
|
| 59 |
+
| **Paragraph2Graph** | 2024 | 文書要素間の空間・論理関係をGNNでモデリング |
|
| 60 |
+
| **gDSA** | 2024 | グラフベース文書構造解析 |
|
| 61 |
+
|
| 62 |
+
### 1.5 パイプライン/フレームワーク
|
| 63 |
+
|
| 64 |
+
| システム | 開発 | 特徴 |
|
| 65 |
+
|:---|:---|:---|
|
| 66 |
+
| **Docling** | IBM | PDFからMarkdown/JSONへのエンドツーエンド変換、Heron, PP-DocLayoutなど多モデル対応、LangChain/LlamaIndex統合 |
|
| 67 |
+
| **MinerU** | — | RAG向け文書パース |
|
| 68 |
+
| **Unstructured.io** | — | 文書パース統合フレームワーク |
|
| 69 |
+
|
| 70 |
+
---
|
| 71 |
+
|
| 72 |
+
## 2. レイアウト生成モデル(グラフィック/UI/文書)
|
| 73 |
+
|
| 74 |
+
レイアウト生成は、要素の種類・制約条件から自動的にレイアウト(バウンディングボックスの位置・サイズ)を生成するタスクです。
|
| 75 |
+
|
| 76 |
+
### 2.1 GAN / VAEベース(初期手法)
|
| 77 |
+
|
| 78 |
+
| モデル | 年 | アーキテクチャ | 特徴 |
|
| 79 |
+
|:---|:---|:---|:---|
|
| 80 |
+
| **LayoutGAN** | 2019 | GAN | 微分可能ワイヤフレームレンダリング層、CNN判別器による整列性評価 |
|
| 81 |
+
| **LayoutVAE** | 2019 | VAE | CountVAE(要素数予測)+ BBoxVAE(位置予測)の2段構成 |
|
| 82 |
+
|
| 83 |
+
### 2.2 Transformerベース
|
| 84 |
+
|
| 85 |
+
| モデル | 年 | アーキテクチャ | 特徴 |
|
| 86 |
+
|:---|:---|:---|:---|
|
| 87 |
+
| **LayoutTransformer** | 2021 | Autoregressive Transformer | レイアウトを離散トークン列として自己回帰生成 |
|
| 88 |
+
| **BLT** (Bidirectional Layout Transformer) | 2022 (Google) | Non-Autoregressive Transformer | ドラフト生成→反復精錬、10倍高速化、制御可能な生成 |
|
| 89 |
+
| **LayoutFormer++** | 2023 | Encoder-Decoder / Decoder-only | Seq2Seq形式の統一レイアウト生成 |
|
| 90 |
+
| **StructLayoutFormer** | 2025 | Transformer | 構造情報と要素配置を分離、制約付きレイアウト生成 |
|
| 91 |
+
|
| 92 |
+
### 2.3 拡散モデル / Flow Matchingベース
|
| 93 |
+
|
| 94 |
+
| モデル | 年 | アーキテクチャ | 特徴 |
|
| 95 |
+
|:---|:---|:---|:---|
|
| 96 |
+
| **LayoutDM** | 2023 | Discrete Diffusion + Transformer | ノイズからレイアウトを段階的デノイジング、柔軟な条件付き生成(補完、カテゴリ→位置) |
|
| 97 |
+
| **LayoutFlow** | 2024 | Flow Matching | Diffusionの代替、滑らかなサンプリング軌跡、高速推論 |
|
| 98 |
+
| **LACE** | 2024 | Constrained Diffusion | 美的制約を反映したレイアウト拡散モデル |
|
| 99 |
+
| **UniLayDiff** | 2025 | Unified Diffusion Transformer | 多様なコンテンツ認識レイアウトタスクを単一モデルで統合 |
|
| 100 |
+
| **SLayR** | 2025 | Rectified Flow | シーン・グラフィックレイアウト生成向け改善型Flow Matching |
|
| 101 |
+
|
| 102 |
+
### 2.4 LLMベース
|
| 103 |
+
|
| 104 |
+
| モデル | 年 | ベースLLM | 特徴 |
|
| 105 |
+
|:---|:---|:---|:---|
|
| 106 |
+
| **LayoutGPT** | 2023 (NeurIPS) | GPT-3.5/4 | テキスト条件からCSS/コード形式でレイアウト生成、学習不要 |
|
| 107 |
+
| **PosterLlama** | 2024 | LLaMA系 | HTML形式でポスター・グラフィックレイアウトを生成、視覚エンコーダとLLMの2段階学習 |
|
| 108 |
+
| **PosterLLaVa** | 2024 | LLaVA系 | LLMベースのコンテンツ認識レイアウト生成 |
|
| 109 |
+
| **LGGPT** | 2024/2025 | 汎用LLM | Arbitrary Layout Instruction (ALI) + Universal Layout Response (ULR) で統一的レイアウト生成 |
|
| 110 |
+
| **LayoutKAG** | 2025 | LLM + Knowledge Graph | 知識拡張型レイアウト生成 |
|
| 111 |
+
|
| 112 |
+
> [!TIP]
|
| 113 |
+
> **2024-2025年のトレンド**: LLMが「レイアウトプランナー」として高レベルな設計知識を提供し、Diffusion Transformerが高品質な最終出力を合成する**ハイブリッドアプローチ**が主流になりつつあります。
|
| 114 |
+
|
| 115 |
+
---
|
| 116 |
+
|
| 117 |
+
## 3. 画像レイアウト生成モデル(Layout-to-Image)
|
| 118 |
+
|
| 119 |
+
バウンディングボックスやレイアウト指定から、それに合致する画像を生成するタスクです。
|
| 120 |
+
|
| 121 |
+
### 3.1 Layout-to-Image生成
|
| 122 |
+
|
| 123 |
+
| モデル | 年 | 会議 | アーキテクチャ | 特徴 |
|
| 124 |
+
|:---|:---|:---|:---|:---|
|
| 125 |
+
| **InstanceAssemble** | 2025 | NeurIPS | DiT + LoRA | Instance-assembling attention、Flux/SD3互換、DenseLayoutベンチマーク(5K画像/90Kインスタンス)、Layout Grounding Score (LGS) 指標を提案 |
|
| 126 |
+
| **CreatiLayout** | 2025 | ICCV | Siamese MM-DiT (SiamLayout) | レイアウトを独立モダリティとして扱う、LayoutDesigner(LLMプランナー)、LayoutSAMデータセット(270万画像/1070万エンティティ) |
|
| 127 |
+
| **UniLayDiff** | 2025 | — | Unified DiT | コンテンツ認識レイアウトタスクの統一モデル、LoRAベースの制約統合 |
|
| 128 |
+
| **GLIGEN** | 2023 | CVPR | Grounded Diffusion | バウンディングボックスによる空間制御の先駆的手法 |
|
| 129 |
+
| **LayoutDiffuse** | 2023 | — | Diffusion | セマンティックレイアウトからのシーン画像生成 |
|
| 130 |
+
|
| 131 |
+
### 3.2 UI/Webレイアウト生成
|
| 132 |
+
|
| 133 |
+
| ツール/モデル | 特徴 |
|
| 134 |
+
|:---|:---|
|
| 135 |
+
| **Galileo AI** | テキストプロンプトからUI/UXモックアップ生成、Figmaエクスポート対応 |
|
| 136 |
+
| **Uizard** | ラフスケッチからプロフェッショナルUIモックアップ |
|
| 137 |
+
| **Screenshot-to-Code** | スクリーンショットからHTML/CSSコードへ変換 |
|
| 138 |
+
| **Design2Code** | デザイン画像からフロントエンドコード生成 |
|
| 139 |
+
|
| 140 |
+
---
|
| 141 |
+
|
| 142 |
+
## 4. 動画レイアウト分析・生成モデル
|
| 143 |
+
|
| 144 |
+
### 4.1 レイアウト制御型動画生成
|
| 145 |
+
|
| 146 |
+
動画生成において、物体の空間配置(バウンディングボックス)、カメラ軌道、動きの方向を明示的に制御するモデル群です。
|
| 147 |
+
|
| 148 |
+
| モデル | 年 | 制御方式 | 特徴 |
|
| 149 |
+
|:---|:---|:---|:---|
|
| 150 |
+
| **VideoDirectorGPT** | 2024 (COLM) | LLMプランニング + BBox | GPT-4がテキストからビデオプラン(シーン記述・エンティティレイアウト・一貫性グループ)を生成、Layout2Vidモジュールで空間制御付き動画合成 |
|
| 151 |
+
| **TrailBlazer** | 2024 | BBox軌跡 + Attention編集 | 学習不要、事前学習済T2Vモデルの時空間アテンションマップを編集してキーフレームBBoxで物体軌跡制御 |
|
| 152 |
+
| **MotionCtrl** | 2024 | カメラ運動 + 物体運動 | CMCM(カメラ制御)+ OMCM(物体制御)の2モジュール、アダプターベース |
|
| 153 |
+
| **DrivingDiffusion** | 2024 | 3D空間レイアウト | 自動運転シーン向けマルチビュー一貫性、3D空間レイアウトガイド |
|
| 154 |
+
| **GenProp** | 2025 (Adobe) | 1フレーム編集伝播 | 最初のフレームの変更を動画全体に伝播(挿入・除去・トラッキングの統一) |
|
| 155 |
+
|
| 156 |
+
### 4.2 動画生成基盤モデル(DiTベース)
|
| 157 |
+
|
| 158 |
+
2024-2025年、動画生成はU-Netから**Diffusion Transformer (DiT)** アーキテクチャへ大きく移行しました。
|
| 159 |
+
|
| 160 |
+
| モデル | 開発 | 特徴 |
|
| 161 |
+
|:---|:---|:---|
|
| 162 |
+
| **HunyuanVideo** | Tencent | 長尺高品質動画、DiTベース |
|
| 163 |
+
| **CogVideoX** | Zhipu AI | テキストから動画生成、DiTベース |
|
| 164 |
+
| **Wan2.1** | Alibaba | 大規模動画生成 |
|
| 165 |
+
|
| 166 |
+
### 4.3 制御方式の分類
|
| 167 |
+
|
| 168 |
+
```mermaid
|
| 169 |
+
graph TB
|
| 170 |
+
A["動画レイアウト制御"] --> B["構造的制御"]
|
| 171 |
+
A --> C["ID/外観制御"]
|
| 172 |
+
A --> D["動き制御"]
|
| 173 |
+
|
| 174 |
+
B --> B1["BBox/レイアウト"]
|
| 175 |
+
B --> B2["深度マップ"]
|
| 176 |
+
B --> B3["スケッチ"]
|
| 177 |
+
B --> B4["セマンティックマップ"]
|
| 178 |
+
|
| 179 |
+
D --> D1["カメラ軌道"]
|
| 180 |
+
D --> D2["物体軌跡"]
|
| 181 |
+
D --> D3["ランドマーク/ポーズ"]
|
| 182 |
+
|
| 183 |
+
B1 --> E["VideoDirectorGPT"]
|
| 184 |
+
B1 --> F["TrailBlazer"]
|
| 185 |
+
D1 --> G["MotionCtrl"]
|
| 186 |
+
D2 --> G
|
| 187 |
+
```
|
| 188 |
+
|
| 189 |
+
### 4.4 データセット
|
| 190 |
+
|
| 191 |
+
| データセット | 特徴 |
|
| 192 |
+
|:---|:---|
|
| 193 |
+
| **LVD-2M** | 200万本の長尺動画(10秒以上)、時間的に密なキャプション付き |
|
| 194 |
+
|
| 195 |
+
---
|
| 196 |
+
|
| 197 |
+
## 5. VLM/LLMベースのレイアウト理解モデル
|
| 198 |
+
|
| 199 |
+
大規模Vision-Language Model(VLM)を用いた文書・画像レイアウト理解は2024-2025年の主要トレンドです。
|
| 200 |
+
|
| 201 |
+
### 5.1 文書特化VLM
|
| 202 |
+
|
| 203 |
+
| モデル | 開発 | 特徴 |
|
| 204 |
+
|:---|:---|:---|
|
| 205 |
+
| **mPLUG-DocOwl2** | Alibaba / Renmin Univ. | OCRフリー文書理解、高解像度DocCompressorで1ページ324トークンに圧縮、マルチページ対応 |
|
| 206 |
+
| **Qwen2.5-VL** | Alibaba Cloud | エージェント能力(ツール使用、画面操作)、動的解像度ViT、複雑なレイアウト・チャート・表の分析、構造化出力(JSON)生成 |
|
| 207 |
+
| **Florence-2** | Microsoft | 統一Seq2Seqフレームワーク、キャプション・物体検出・セグメンテーション・OCRを1モデルで、軽量・ファインチューニング向き |
|
| 208 |
+
| **GLM-4.5V** | Zhipu AI | 高度なマルチモーダル推論、文書レイアウト理解 |
|
| 209 |
+
| **InternVL2** | Shanghai AI Lab | 高性能VLM、文書理解タスクでSOTA級 |
|
| 210 |
+
|
| 211 |
+
### 5.2 比較表
|
| 212 |
+
|
| 213 |
+
| モデル | 最適用途 | レイアウト理解力 |
|
| 214 |
+
|:---|:---|:---|
|
| 215 |
+
| **mPLUG-DocOwl2** | 効率的なマルチページ文書分析(低レイテンシ) | ◎ OCRフリー高効率 |
|
| 216 |
+
| **Qwen2.5-VL** | 複雑な推論・構造化データ抽出・エージェントタスク | ◎◎ 最高性能 |
|
| 217 |
+
| **Florence-2** | 軽量・汎用ビジョンタスク(OCR+検出+セグメンテーション) | ○ 汎用的 |
|
| 218 |
+
|
| 219 |
+
---
|
| 220 |
+
|
| 221 |
+
## 6. 商用レイアウト分析ソリューション
|
| 222 |
+
|
| 223 |
+
### 6.1 クラウドベンダー比較
|
| 224 |
+
|
| 225 |
+
| サービス | 提供 | 強み | 最適ユースケース |
|
| 226 |
+
|:---|:---|:---|:---|
|
| 227 |
+
| **Google Document AI** | Google Cloud | 特化型プリトレーニングプロセッサ群(請求書、税務書類、契約書等)、Vertex AI/BigQuery連携 | GCPユーザー、特定業務文書の高精度処理 |
|
| 228 |
+
| **Amazon Textract** | AWS | S3/Lambda/SQSとのネイティブ連携、Queries機能(自然言語で特定データ抽出) | 大量トランザクション処理、AWSパイプライン |
|
| 229 |
+
| **Azure AI Document Intelligence** | Microsoft | **レイアウト分析でリーダー的評価**、可変レイアウト文書への対応、Studio UIでカスタムモデル学習が容易 | 複雑・非定型レイアウトの文書 |
|
| 230 |
+
|
| 231 |
+
### 6.2 その他の商用/デザインツール
|
| 232 |
+
|
| 233 |
+
| ツール | 分野 | 特徴 |
|
| 234 |
+
|:---|:---|:---|
|
| 235 |
+
| **Adobe Firefly** | グラフィックデザイン | Photoshop/Illustrator統合、生成塗りつぶし、テキストからベクター |
|
| 236 |
+
| **Canva Magic Studio** | グラフィックデザイン | AIレイアウト提案、Magic Designでプレゼン即時生成 |
|
| 237 |
+
| **Kittl / Recraft** | グラフィックデザイン | ブランド一貫性重視、ベクターベース出力、編集可能アセット |
|
| 238 |
+
| **Reducto** | 文書処理 | エンタープライズ向けAgentic文書解析 |
|
| 239 |
+
| **V7 Go** | 文書処理 | 高精度・検証可能なAI文書分析 |
|
| 240 |
+
|
| 241 |
+
---
|
| 242 |
+
|
| 243 |
+
## 7. 主要データセット・ベンチマーク
|
| 244 |
+
|
| 245 |
+
### 7.1 文書レイアウト分析
|
| 246 |
+
|
| 247 |
+
| データセット | サイズ | アノテーション | 対象文書 |
|
| 248 |
+
|:---|:---|:---|:---|
|
| 249 |
+
| **PubLayNet** | ~360K ページ | 自動生成(PDF/XML対応) | 科学論文(PubMed Central) |
|
| 250 |
+
| **DocLayNet** | ~81K ページ | 人手アノテーション、11クラス | 金融、マニュアル、科学、法律、特許、政府 |
|
| 251 |
+
| **DocBank** | 500K ページ | 弱教師(LaTeXソース) | arXiv論文、トークンレベルアノテーション |
|
| 252 |
+
| **D⁴LA** | — | 詳細人手アノテーション | 多様な文書 |
|
| 253 |
+
| **DocStructBench** | 数千ページ | 人手、10カテゴリ | 学術、教科書、金融、市場分析 |
|
| 254 |
+
| **Comp-HRDoc** | — | — | 複雑な高解像度文書 |
|
| 255 |
+
|
| 256 |
+
### 7.2 レイアウト生成
|
| 257 |
+
|
| 258 |
+
| データセット | 対象 | 特徴 |
|
| 259 |
+
|:---|:---|:---|
|
| 260 |
+
| **RICO** | モバイルUI | UIデザイン・構造階層の大規模コレクション |
|
| 261 |
+
| **PubLayNet** | 文書 | レイアウト生成評価にも使用 |
|
| 262 |
+
| **Magazine** | 雑誌レイアウト | ポスター・雑誌のグラフィックレイアウト |
|
| 263 |
+
| **CGL Dataset** | コンテンツ認識レイアウト | グラフィックレイアウト生成向け |
|
| 264 |
+
| **LayoutSAM** | 汎用画像 | CreatiLayoutが提案、270万画像/1070万エンティティ |
|
| 265 |
+
| **DenseLayout** | 汎用画像 | InstanceAssembleが提案、5K画像/90Kインスタンス |
|
| 266 |
+
|
| 267 |
+
### 7.3 評価指標
|
| 268 |
+
|
| 269 |
+
| 分野 | 指標 | 説明 |
|
| 270 |
+
|:---|:---|:---|
|
| 271 |
+
| **レイアウト検出** | mAP, IoU | 物体検出の標準指標 |
|
| 272 |
+
| **レイアウト生成** | FID | 生成レイアウトの品質・多様性 |
|
| 273 |
+
| **レイアウト生成** | Alignment, Overlap | レイアウト要素の整列性、重複度 |
|
| 274 |
+
| **レイアウト生成** | LGS (Layout Grounding Score) | InstanceAssembleが提案した新指標 |
|
| 275 |
+
| **文書理解** | DocVQA, ANLS | 文書視覚質問応答 |
|
| 276 |
+
| **レイアウト検出** | LPIPS | ロバストネスベンチマーク向け |
|
| 277 |
+
|
| 278 |
+
---
|
| 279 |
+
|
| 280 |
+
## 8. 技術トレンド総括
|
| 281 |
+
|
| 282 |
+
### 8.1 全体マップ
|
| 283 |
+
|
| 284 |
+
```mermaid
|
| 285 |
+
graph LR
|
| 286 |
+
subgraph "レイアウト分析(入力→構造理解)"
|
| 287 |
+
DA["文書レイアウト分析"]
|
| 288 |
+
IA["画像レイアウト分析"]
|
| 289 |
+
VA["動画レイアウト分析"]
|
| 290 |
+
end
|
| 291 |
+
|
| 292 |
+
subgraph "レイアウト生成(条件→レイアウト出力)"
|
| 293 |
+
DG["文書/グラフィック<br>レイアウト生成"]
|
| 294 |
+
IG["Layout-to-Image<br>画像生成"]
|
| 295 |
+
VG["Layout-guided<br>動画生成"]
|
| 296 |
+
end
|
| 297 |
+
|
| 298 |
+
DA --> |"構造情報"| DG
|
| 299 |
+
IA --> |"空間情報"| IG
|
| 300 |
+
VA --> |"時空間情報"| VG
|
| 301 |
+
|
| 302 |
+
DG --> |"BBox/マスク"| IG
|
| 303 |
+
IG --> |"フレーム生成"| VG
|
| 304 |
+
```
|
| 305 |
+
|
| 306 |
+
### 8.2 主要トレンド
|
| 307 |
+
|
| 308 |
+
| トレンド | 説明 |
|
| 309 |
+
|:---|:---|
|
| 310 |
+
| **1. Diffusion Transformer (DiT) への移行** | U-Netベースの拡散モデルからTransformerベースへ。レイアウト生成(LayoutDM→UniLayDiff)、画像生成(InstanceAssemble)、動画生成(HunyuanVideo等)すべてで進行 |
|
| 311 |
+
| **2. LLMのレイアウトプランナー化** | LayoutGPT、PosterLlama、VideoDirectorGPTに見られるように、LLMがデザイン知識を活用してレイアウトを「設計」する手法が台頭 |
|
| 312 |
+
| **3. ハイブリッドアプローチ** | LLM(高レベルプランニング)+ Diffusion Model(高品質合成)の組み合わせが2025年のSOTA |
|
| 313 |
+
| **4. Flow Matchingの台頭** | LayoutFlowやSLayRに見られる、拡散モデルの代替としてのFlow Matching。よりシンプルで高速 |
|
| 314 |
+
| **5. 統一モデルの志向** | 複数のレイアウトタスク(生成、補完、条件付き生成等)を1つのモデルで処理(UniLayDiff, LGGPT, DLAFormer) |
|
| 315 |
+
| **6. コンテンツ認識の強化** | 背景画像やテキスト内容を考慮したレイアウト生成(PosterLlama, CreatiLayout) |
|
| 316 |
+
| **7. エンタープライズRAGパイプライン統合** | Docling, MinerU等による文書レイアウト分析→LLM/RAGへの直接的な統合 |
|
| 317 |
+
| **8. Agentic Document AI** | VLM(Qwen2.5-VL等)がレイアウト理解だけでなく、ツール使用や自律的な判断を行うエージェント的振る舞い |
|
| 318 |
+
|
| 319 |
+
### 8.3 アーキテクチャの進化タイムライン
|
| 320 |
+
|
| 321 |
+
```mermaid
|
| 322 |
+
timeline
|
| 323 |
+
title レイアウトモデルの進化
|
| 324 |
+
2019 : LayoutGAN
|
| 325 |
+
: LayoutVAE
|
| 326 |
+
2020 : LayoutLM
|
| 327 |
+
2021 : LayoutTransformer
|
| 328 |
+
: LayoutLMv2
|
| 329 |
+
: LayoutParser
|
| 330 |
+
2022 : LayoutLMv3
|
| 331 |
+
: BLT (Google)
|
| 332 |
+
2023 : LayoutDM
|
| 333 |
+
: LayoutGPT
|
| 334 |
+
: GLIGEN
|
| 335 |
+
: LayoutFormer++
|
| 336 |
+
2024 : DocLayout-YOLO
|
| 337 |
+
: DLAFormer
|
| 338 |
+
: LayoutFlow
|
| 339 |
+
: PosterLlama
|
| 340 |
+
: VideoDirectorGPT
|
| 341 |
+
: TrailBlazer
|
| 342 |
+
: MotionCtrl
|
| 343 |
+
2025 : UniLayDiff
|
| 344 |
+
: CreatiLayout
|
| 345 |
+
: InstanceAssemble
|
| 346 |
+
: GenProp
|
| 347 |
+
: Heron-101
|
| 348 |
+
: StructLayoutFormer
|
| 349 |
+
```
|
| 350 |
+
|
| 351 |
+
### 8.4 選定ガイド
|
| 352 |
+
|
| 353 |
+
> [!IMPORTANT]
|
| 354 |
+
> 用途に応じた推奨モデル選択
|
| 355 |
+
|
| 356 |
+
| やりたいこと | 推奨モデル/ツール |
|
| 357 |
+
|:---|:---|
|
| 358 |
+
| **文書からレイアウト要素を高速検出** | DocLayout-YOLO, Heron-101 (Docling) |
|
| 359 |
+
| **文書のセマンティック理解(フォーム抽出、VQA)** | LayoutLMv3, Qwen2.5-VL |
|
| 360 |
+
| **文書を構造化してRAG/LLMに投入** | Docling, MinerU, Unstructured.io |
|
| 361 |
+
| **ポスター・グラフィックのレイアウト自動生成** | PosterLlama, LayoutDM, LayoutFlow |
|
| 362 |
+
| **UI/Webレイアウトの自動生成** | LayoutGPT, Galileo AI, Uizard |
|
| 363 |
+
| **レイアウト指定から画像生成** | InstanceAssemble, CreatiLayout, GLIGEN |
|
| 364 |
+
| **レイアウト制御付き動画生成** | VideoDirectorGPT, TrailBlazer, MotionCtrl |
|
| 365 |
+
| **商用文書処理パイプライン** | Azure AI Document Intelligence, Google Document AI, AWS Textract |
|
| 366 |
+
|
| 367 |
+
---
|
| 368 |
+
|
| 369 |
+
> [!NOTE]
|
| 370 |
+
> 本サーベイは2025年5月時点の情報に基づいています。この分野は急速に進展しており、新しいモデルやベンチマークが継続的に登場しています。
|
docs/AI_Computer_Vision/layout_models_survey.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "d85e0629dd23f2bf", "ingested_at": "2026-06-30T08:49:26.140736", "content_hash": "d85e0629dd23f2bf"}
|
docs/AI_Computer_Vision/script_long.md
ADDED
|
@@ -0,0 +1,26 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and welcome back. Tonight we're diving into a fascinating new development in the field of artificial intelligence, specifically in computer vision. A recent surge of research papers has been published on arXiv, an online repository for scientific papers, and these breakthroughs are shaping the future of how machines understand and interact with the world around us.
|
| 2 |
+
|
| 3 |
+
Let’s take a closer look at some of these groundbreaking advancements:
|
| 4 |
+
|
| 5 |
+
First, we have "EventHub," which proposes a new approach to building event-based stereo networks. These networks can be used for various applications, from autonomous vehicles to robotics, and EventHub's innovative design promises increased efficiency and accuracy.
|
| 6 |
+
|
| 7 |
+
Next up is “ActionParty,” a multi-subject action binding system designed for generative video games. This could revolutionize the way we create immersive gaming experiences by allowing players to interact with characters in more dynamic and realistic ways.
|
| 8 |
+
|
| 9 |
+
Then there’s "Generative World Renderer," which focuses on creating realistic 3D environments using AI. This technology has the potential to transform fields like film, animation, and architectural visualization.
|
| 10 |
+
|
| 11 |
+
We also see a focus on anomaly detection through “Modulate-and-Map,” which utilizes crossmodal feature mapping with cross-view modulation for enhanced accuracy in identifying anomalies. This could be crucial for various applications such as security systems or medical imaging analysis.
|
| 12 |
+
|
| 13 |
+
The field of visual representation is also being pushed forward with "Steerable Visual Representations," offering a new way to control and manipulate images using AI. This could have implications for fields like augmented reality, where precise image manipulation would be essential.
|
| 14 |
+
|
| 15 |
+
Beyond these advancements in computer vision, we're seeing breakthroughs in understanding language and the world around us through “Scenario Comprehension Visual Grounding.” This research aims to bridge the gap between visual information and human understanding of narratives.
|
| 16 |
+
|
| 17 |
+
The field of video understanding is also advancing with "A Simple Baseline for Streaming Video Understanding," which focuses on creating a more efficient way to analyze live video streams, potentially revolutionizing applications like real-time object tracking or automated surveillance.
|
| 18 |
+
|
| 19 |
+
And finally, we have “VOID,” which proposes a new method for deleting objects and interactions from videos in a controlled manner. This could be invaluable for applications such as privacy protection or content editing.
|
| 20 |
+
|
| 21 |
+
|
| 22 |
+
These are just a few of the exciting developments that promise to reshape our interaction with technology in the coming years. As these technologies continue to mature, we can expect even more innovative breakthroughs that will change how we live and work.
|
| 23 |
+
|
| 24 |
+
|
| 25 |
+
|
| 26 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/script_long.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "176cc3fb0aea221e", "ingested_at": "2026-06-30T08:49:26.364957", "content_hash": "176cc3fb0aea221e"}
|
docs/AI_Computer_Vision/script_long_clean.txt
ADDED
|
@@ -0,0 +1,23 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening, and welcome back. Tonight we're diving into a fascinating new development in the field of artificial intelligence, specifically in computer vision. A recent surge of research papers has been published on arXiv, an online repository for scientific papers, and these breakthroughs are shaping the future of how machines understand and interact with the world around us.
|
| 2 |
+
|
| 3 |
+
Let’s take a closer look at some of these groundbreaking advancements:
|
| 4 |
+
|
| 5 |
+
First, we have "EventHub," which proposes a new approach to building event-based stereo networks. These networks can be used for various applications, from autonomous vehicles to robotics, and EventHub's innovative design promises increased efficiency and accuracy.
|
| 6 |
+
|
| 7 |
+
Next up is “ActionParty,” a multi-subject action binding system designed for generative video games. This could revolutionize the way we create immersive gaming experiences by allowing players to interact with characters in more dynamic and realistic ways.
|
| 8 |
+
|
| 9 |
+
Then there’s "Generative World Renderer," which focuses on creating realistic 3D environments using AI. This technology has the potential to transform fields like film, animation, and architectural visualization.
|
| 10 |
+
|
| 11 |
+
We also see a focus on anomaly detection through “Modulate-and-Map,” which utilizes crossmodal feature mapping with cross-view modulation for enhanced accuracy in identifying anomalies. This could be crucial for various applications such as security systems or medical imaging analysis.
|
| 12 |
+
|
| 13 |
+
The field of visual representation is also being pushed forward with "Steerable Visual Representations," offering a new way to control and manipulate images using AI. This could have implications for fields like augmented reality, where precise image manipulation would be essential.
|
| 14 |
+
|
| 15 |
+
Beyond these advancements in computer vision, we're seeing breakthroughs in understanding language and the world around us through “Scenario Comprehension Visual Grounding.” This research aims to bridge the gap between visual information and human understanding of narratives.
|
| 16 |
+
|
| 17 |
+
The field of video understanding is also advancing with "A Simple Baseline for Streaming Video Understanding," which focuses on creating a more efficient way to analyze live video streams, potentially revolutionizing applications like real-time object tracking or automated surveillance.
|
| 18 |
+
|
| 19 |
+
And finally, we have “VOID,” which proposes a new method for deleting objects and interactions from videos in a controlled manner. This could be invaluable for applications such as privacy protection or content editing.
|
| 20 |
+
|
| 21 |
+
These are just a few of the exciting developments that promise to reshape our interaction with technology in the coming years. As these technologies continue to mature, we can expect even more innovative breakthroughs that will change how we live and work.
|
| 22 |
+
|
| 23 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/script_middle.md
ADDED
|
@@ -0,0 +1,12 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. We're bringing you a look at some of the latest research published on arXiv, a platform for sharing scientific papers. Researchers have been busy, submitting over 35,000 new papers in just the past year. This includes work on everything from computer vision to video game development and even how we understand videos.
|
| 2 |
+
|
| 3 |
+
Take a look at this selection of papers that were recently published: "EventHub" is an example of a study exploring event-based stereo networks without active sensors. Researchers are looking to create more efficient systems for understanding the world around us.
|
| 4 |
+
|
| 5 |
+
Then there's "ActionParty," which focuses on multi-subject action binding in generative video games, and "Generative World Renderer." This paper explores how we can use AI to generate realistic 3D worlds.
|
| 6 |
+
|
| 7 |
+
The list goes on, with papers exploring topics like crossmodal feature mapping for anomaly detection, steerable visual representations, and even scenario comprehension visual grounding.
|
| 8 |
+
|
| 9 |
+
These are just a few examples of the cutting-edge research being conducted across various fields. You can find more details about these studies and others by visiting the arXiv website.
|
| 10 |
+
|
| 11 |
+
|
| 12 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/script_middle.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "e18e1b8265e3e95b", "ingested_at": "2026-06-30T08:49:26.562449", "content_hash": "e18e1b8265e3e95b"}
|
docs/AI_Computer_Vision/script_middle_clean.txt
ADDED
|
@@ -0,0 +1,11 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
Good evening. We're bringing you a look at some of the latest research published on arXiv, a platform for sharing scientific papers. Researchers have been busy, submitting over 35,000 new papers in just the past year. This includes work on everything from computer vision to video game development and even how we understand videos.
|
| 2 |
+
|
| 3 |
+
"EventHub" is an example of a study exploring event-based stereo networks without active sensors. Researchers are looking to create more efficient systems for understanding the world around us.
|
| 4 |
+
|
| 5 |
+
Then there's "ActionParty," which focuses on multi-subject action binding in generative video games, and "Generative World Renderer." This paper explores how we can use AI to generate realistic 3D worlds.
|
| 6 |
+
|
| 7 |
+
The list goes on, with papers exploring topics like crossmodal feature mapping for anomaly detection, steerable visual representations, and even scenario comprehension visual grounding.
|
| 8 |
+
|
| 9 |
+
These are just a few examples of the cutting-edge research being conducted across various fields. You can find more details about these studies and others by visiting the arXiv website.
|
| 10 |
+
|
| 11 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/script_short.md
ADDED
|
@@ -0,0 +1,6 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
New research is dropping faster than you can say "AI breakthrough." arXiv just released a massive batch of papers, over 35,000 in total! We're talking about everything from new ways to create realistic video games to groundbreaking advancements in how computers understand and interact with the world around them.
|
| 2 |
+
|
| 3 |
+
There are tons of exciting breakthroughs here, including a new approach to understanding videos, a system for creating more lifelike avatars, and even a way to delete objects from video footage! Want to learn more? Check out the links on screen for the full research papers. This is just the tip of the iceberg – stay tuned for more updates as this field continues to explode!
|
| 4 |
+
|
| 5 |
+
|
| 6 |
+
This has been brought to you by The Android Times.
|
docs/AI_Computer_Vision/script_short.rag_ingested
ADDED
|
@@ -0,0 +1 @@
|
|
|
|
|
|
|
| 1 |
+
{"doc_id": "bce174ea1c9eda0a", "ingested_at": "2026-06-30T08:49:26.650814", "content_hash": "bce174ea1c9eda0a"}
|
docs/AI_Computer_Vision/script_short_clean.txt
ADDED
|
@@ -0,0 +1,5 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 1 |
+
New research is dropping faster than you can say "AI breakthrough." arXiv just released a massive batch of papers, over 35,000 in total! We're talking about everything from new ways to create realistic video games to groundbreaking advancements in how computers understand and interact with the world around them.
|
| 2 |
+
|
| 3 |
+
There are tons of exciting breakthroughs here, including a new approach to understanding videos, a system for creating more lifelike avatars, and even a way to delete objects from video footage! Want to learn more? Check out the links on screen for the full research papers. This is just the tip of the iceberg – stay tuned for more updates as this field continues to explode!
|
| 4 |
+
|
| 5 |
+
This has been brought to you by The Android Times.
|