Monday, March 3, 2025

Europe: The Revival of Cross-Border Patent Litigation (유럽, 국경을 넘는 특허 소송 부활 )

 1. 서론(Introduction)

2025 2 25, 유럽사법재판소(ECJ) 유럽 법원 관할권에 대한 중대한 판결을 내렸습니다. 유럽사법재판소는 유럽연합(EU) 최고 사법기관으로, EU 법률의 해석과 유권해석을 담당하며 사실상 EU 대법원 역할을 수행합니다.

On February 25, 2025, the Court of Justice of the European Union (CJEU) issued a landmark ruling on the jurisdiction of European courts. As the highest judicial authority in the European Union (EU), the CJEU plays a crucial role in interpreting EU law and providing binding legal decisions, effectively functioning as the EU’s Supreme Court.

 

이번 판결의 핵심 쟁점은 브뤼셀 I bis 규정(EU 의회 이사회 민사소송 규정 No. 1215/2012) 24(4) 해석이었습니다.

The central issue in this case revolved around the interpretation of Article 24(4) of the Brussels I bis Regulation (Regulation (EU) No. 1215/2012 on jurisdiction and the recognition and enforcement of judgments in civil and commercial matters).

 

유럽사법재판소(ECJ) 피고가 EU 회원국 법원의 관할권(또는 UPC 관할권) 내에 거주하는 경우, 해당 피고( 침해를 방조하거나 공모하여 적법하게 공동피고로 추가된 ) 대해 법원이 관할하는 영토를 넘어 발생한 특허 침해 행위 대해서도 손해배상 지급 또는 침해 행위 중지 명령을 내릴 있다 판단하였습니다. 이는 피고가 특허 무효 항변을 제기한 경우에도 동일하게 적용됩니다.

다만, 판단의 효력은 당사자 (inter partes)으로 한정되며, 이를 근거로 다른 국가에서 등록된 특허가 취소되거나 청구범위가 수정되는 결과를 초래하지는 않는다 판시하였습니다.

The Court of Justice of the European Union (CJEU) ruled that if a defendant resides within the jurisdiction of an EU Member State court (or under UPC jurisdiction), the court may order the defendant—including any legally joined co-defendants who aided or conspired in the infringement—to pay damages or cease infringing activities, even if the infringement occurred outside the court’s territorial jurisdiction. This ruling applies even if the defendant raises a patent invalidity defense.

However, the court clarified that its ruling is limited to the parties involved (inter partes) and does not result in the revocation or amendment of patent registrations in other jurisdictions.

 

유럽통합특허법원(UPC) 출범 이전에도, 브뤼셀 I bis 규정 4(1) (일반 재판적) 따라 EU 회원국에 거주하는 당사자는 국적과 관계없이 해당 회원국 법원에서 민사 소송의 당사자(원고 또는 피고) 있었습니다.

그러나 브뤼셀 I bis 규정 24(특별 재판적) 따르면, 특허 등록 또는 유효성과 관련된 소송은 해당 특허가 등록된 회원국 법원이 전속 관할권을 가진다고 규정되어 있어, EU 회원국 특허 침해 소송에서 관할권 분쟁이 빈번하게 발생해 왔습니다.

이는 특허 침해 소송이 필연적으로 해당 특허의 유효성에 대한 다툼을 수반하거나 선행될 가능성이 높기 때문이며, 또한 유럽 특허(EP) 개별 국가별 특허가 하나로 묶인 형태(a bundle of national patents) 인정되지만, 국가에서 등록된 유럽 특허는 독립적으로 존재한다는 구조적 특성 때문입니다.

Before the Unified Patent Court (UPC) was established, Article 4(1) of the Brussels I bis Regulation (general jurisdiction) allowed parties residing in an EU Member State to be involved in civil litigation—either as plaintiffs or defendants—before the courts of that Member State, regardless of their nationality.

However, Article 24 (exclusive jurisdiction) of the Brussels I bis Regulation stipulates that disputes concerning the registration or validity of a patent fall under the exclusive jurisdiction of the courts of the Member State where the patent is registered. As a result, jurisdictional disputes frequently arose in cross-border patent infringement cases within the EU.

This issue stems from the fact that patent infringement litigation often involves or is preceded by disputes over the validity of the patent in question. Additionally, a European patent (EP) is recognized as a bundle of national patents, meaning that while it is granted as a unitary right, each national designation exists independently within its respective jurisdiction.

 

따라서, 동일한 민사 법원에서 특허 침해 여부와 특허의 유효성을 함께 심리할 경우, 외국에서 등록된 특허의 유효성이 쟁점이 된다면, 해당 특허의 무효 여부에 대한 판단은 전속 관할권을 가진 등록국 법원에서 이루어져야 합니다. 이에 따라, 특허 침해 소송이 제기된 법원은 해당 특허가 등록된 국가의 법원이 무효성 여부를 판단할 때까지 침해 소송을 중지하는 경우가 많았습니다.

이러한 원칙은 Case C-4/03 Gesellschaft für Antriebstechnik mbH & Co KG (GAT) v. Lamellen und Kupplungsbau Beteiligungs KG (LuK) [2006] ECR I-6509 (독일 GAT 사건)에서도 확인된 있습니다.

Therefore, when a civil court hears both a patent infringement claim and a validity challenge, and the validity of a foreign-registered patent is at issue, the determination of invalidity must fall under the exclusive jurisdiction of the court in the country where the patent is registered. As a result, courts hearing infringement cases often stay proceedings until the competent court in the patent’s country of registration rules on its validity.

This principle was affirmed in Case C-4/03 Gesellschaft für Antriebstechnik mbH & Co KG (GAT) v. Lamellen und Kupplungsbau Beteiligungs KG (LuK) [2006] ECR I-6509 (the German GAT case).

 

그러나 유럽연합(EU) 모든 국가가 동일한 입장을 취한 것은 아닙니다. 예를 들어, 네덜란드 법원은 2006 GAT v. LuK 사건 판결 이후에도 브뤼셀 I bis 규정 35(Provisional, including protective, measures) 근거로 임시 금지명령(Preliminary Injunction) 절차에서 국경 금지명령(cross-border injunctions) 내릴 있다는 입장을 유지하고 있습니다.

이러한 입장은 Bettacare v. H3 Products 사건(헤이그 임시 구제 판사, 2006 9 21), PTC v. APE Holland 사건(헤이그 임시 구제 판사, 2011 10 4), YPM v. Yell 사건(헤이그 항소법원, 2011 7 12)에서도 확인된 있습니다.

However, not all EU Member States adopted the same approach. For instance, despite the ruling in GAT v. LuK (2006), Dutch courts have continued to rely on Article 35 of the Brussels I bis Regulation (Regulation (EU) No 1215/2012) (Provisional, including protective, measures) to issue cross-border injunctions in preliminary injunction proceedings.

This stance has been reaffirmed in cases such as Bettacare v. H3 Products (Preliminary Relief Judge of The Hague, September 21, 2006), PTC v. APE Holland (Preliminary Relief Judge of The Hague, October 4, 2011), and YPM v. Yell (The Hague Court of Appeal, July 12, 2011).

 

브뤼셀 I bis 규정 35(Provisional, including protective, measures) 본안 사건에 대한 관할권이 다른 회원국 법원에 있더라도, 개별 회원국 법에 따라 해당 국가의 법원에서 임시 보호 조치를 신청할 있도록 규정하고 있습니다.

이를 근거로, 네덜란드 법원은 특정 회원국 법원이 특허 유효성(본안) 대한 관할권을 가지더라도, 임시 조치를 위해 다른 회원국 법원에서 절차를 개시할 있다고 해석하였습니다. 또한, 다른 EU 회원국에서 등록된 특허의 유효성이 다투어지는 경우에도, 네덜란드 법원은 해당 특허의 ·무효에 대한 예비적 판단을 내릴 있으며, 임시 국경 금지명령(preliminary cross-border injunction) 발령할 있는 관할권을 가진다고 판단하였습니다.

Article 35 of the Brussels I bis Regulation (Regulation (EU) No 1215/2012) (Provisional, including protective, measures) states that provisional and protective measures may be requested from the courts of a Member State, even if another Member State has jurisdiction over the substance of the case.

Based on this provision, Dutch courts have interpreted that even if a court in another Member State has jurisdiction over the validity of a patent (the merits of the case), proceedings for provisional measures may still be initiated in a different Member State. Furthermore, even when the validity of a patent registered in another EU Member State is challenged, Dutch courts maintain that they have jurisdiction to make a preliminary assessment of patent validity and to issue preliminary cross-border injunctions.

 

이러한 선례를 바탕으로 네덜란드 법원의 실무적 관행은 2012 Solvay v. Honeywell 사건에서 유럽사법재판소(ECJ) 이를 승인하면서 더욱 확립되었습니다 (Solvay v. Honeywell, C-616/10, 2012 7 12).

다만, 브뤼셀 I bis 규정 24(4) 따라 해당 외국 특허의 등록국 법원이 전속 관할권을 가지며, 외국 법원이 해당 특허를 무효로 판결할 합리적이고 무시할 없는 가능성이 있다고 판단할 경우, 국경 금지명령(cross-border injunction) 대한 요청이 거부될 있다고 판시하였습니다.

This established practice of Dutch courts was further solidified when the Court of Justice of the European Union (CJEU) endorsed it in the 2012 case of Solvay v. Honeywell (C-616/10, July 12, 2012).

However, the CJEU also ruled that under Article 24(4) of the Brussels I bis Regulation, if the court of the Member State with exclusive jurisdiction over the foreign patent reaches the conclusion that there is a reasonable and non-negligible possibility that the patent will be declared invalid, a request for a cross-border injunction may be denied.

 

유럽사법재판소(ECJ) 또한 여러 개의 기업이 EU 회원국 여러 곳에서 동일한 제품을 대상으로 동일한 유럽 특허를 침해하는 경우(기업 연계(Link) 요건), 특허권자는 특정 회원국 법원에서 국경을 넘는 임시 금지명령(preliminary cross-border injunction) 청구하고 이를 받을 권리가 있다 판결하였습니다.

이로 인해 특허권자에게 유리하고 적극적인 소송 전략이 가능한 네덜란드 법원의 관할권이 매력적인 포럼으로 주목받기 시작했습니다.

The Court of Justice of the European Union (CJEU) also ruled that when multiple companies infringe the same European patent with the same product across multiple EU Member States (meeting the link requirement between defendants), the patent holder has the right to request and obtain a preliminary cross-border injunction from a single national court with jurisdiction.

As a result, the pro-patentee and assertive approach of Dutch courts gained attention as an attractive forum for patent litigation.

 

2. 유럽의 민사소송법: 브뤼셀 협약(Brussels Convention)과 루가노 협약(Lugano Convention)의 연혁 (European Civil Procedure Law: The History of the Brussels Convention and the Lugano Convention)

 

브뤼셀 규정(Brussels Regulation) 유럽 연합 의회 이사회가 제정한 민사소송법(EU 규정)입니다. 규정을 정확히 이해하려면, 유럽연합(EU) 민사소송 규정이 어떻게 발전해 왔는지, 그리고 법원이 이를 어떻게 해석해왔는지에 대한 역사적 배경을 먼저 살펴볼 필요가 있습니다.

The Brussels Regulation is a civil procedure law enacted by the European Parliament and the Council of the European Union. To fully understand this regulation, it is essential to first examine the historical background of the EU’s civil procedure rules and how the courts have interpreted them over time.

 

유럽연합(European Union, EU) 벨기에, 프랑스, 독일(당시 서독), 이탈리아, 룩셈부르크, 네덜란드 6개국이 경제 통합(Common Market) 목표로 설립한 유럽경제공동체(European Economic Community, EEC)에서 발전한 유럽공동체(EC) 연합체입니다. 현재 27개국이 가입되어 있으며, 영국은 2020 1 31일을 기해 EU 탈퇴(Brexit)하였습니다.

한편, 유럽자유무역연합(European Free Trade Association, EFTA) 1960, 유럽경제공동체(EEC) 가입하지 않은 서유럽 7개국이 이에 대응하여 결성한 경제 연합으로, 현재는 스위스, 노르웨이, 아이슬란드, 리히텐슈타인 4개국으로 구성되어 있습니다.

The European Union (EU) evolved from the European Economic Community (EEC), which was established by six founding countries: Belgium, France, Germany (then West Germany), Italy, Luxembourg, and the Netherlands. The goal of the EEC was economic integration and the creation of a Common Market. Today, the EU consists of 27 member states, and the United Kingdom formally left the EU (Brexit) on January 31, 2020.

Meanwhile, the European Free Trade Association (EFTA) was established in 1960 as an economic alliance by seven Western European countries that had not joined the EEC. Currently, EFTA consists of four member states: Switzerland, Norway, Iceland, and Liechtenstein.

 

1994 냉전시대가 끝나자 유럽 국가의 유럽 연합(EU) 유럽자유무역연합(EFTA) 유럽 확대 통일 시장을 만들기 위해 유럽 경제 지역(EEA) 설립하고, 유럽 전역을 역내 자유 무역 지역으로 선언하였습니다.

Following the end of the Cold War, in 1994, the European Union (EU) and European Free Trade Association (EFTA) member states established the European Economic Area (EEA) to create an expanded, unified European market. This agreement effectively declared Europe a single internal free trade area.

 

유럽경제공동체(EEC) 1968 브뤼셀 협약(Brussels Convention, 1968) 체결하여 회원국 민사 상사 사건에서의 관할권 판결 집행에 대한 통일된 규칙을 마련하였습니다. 이후, 1988 유럽자유무역연합(EFTA) 루가노 협약(Lugano Convention, 1988) 체결하여, 브뤼셀 협약(1968) 원칙을 EFTA 회원국에도 적용할 있게 되었습니다.

The European Economic Community (EEC) established the Brussels Convention (1968) to create a unified framework for jurisdiction and the enforcement of judgments in civil and commercial matters among its member states. Later, in 1988, the EEC signed the Lugano Convention with the European Free Trade Association (EFTA), extending the principles of the Brussels Convention (1968) to EFTA member states.

 

브뤼셀 협약(1968) 2001 브뤼셀 I 규정(EC No 44/2001)으로 대체되면서, 관할권 판결 집행 절차를 더욱 단순화하여 유럽공동체(EC) 내에서 적용될 있도록 개선되었습니다.

이후 유럽연합(EU) 출범과 함께, EU 회원국에서 내려진 판결이 별도의 승인 절차 없이 다른 회원국에서 즉시 집행될 있도록 Exequatur 절차가 완전히 폐지되었습니다. 또한, () EU 국가 피고에 대한 관할권이 확대되었으며, 국제적으로 동일 소송이 중복 진행되는 것을 방지하는 원칙이 강화되었습니다.

아울러, 지식재산권, 계약, 소비자 보호 등과 관련된 관할권 규정이 명확히 정리되면서, 브뤼셀 I 규정은 유럽연합(EU) 통합 민사소송법으로 자리 잡았고, 최종적으로 유럽 의회 이사회 민사소송 규정(EU No 1215/2012) (브뤼셀 I bis 규정)으로 대체되었습니다.

The Brussels Convention (1968) was replaced by the Brussels I Regulation (EC No 44/2001) in 2001, simplifying and streamlining rules on jurisdiction and the enforcement of judgments to ensure consistent application within the European Community (EC).

With the establishment of the European Union (EU), the exequatur procedure was fully abolished, allowing judgments rendered in one EU Member State to be enforced in another without the need for separate approval. Additionally, jurisdiction over defendants from non-EU countries was expanded, and the principle of preventing parallel proceedings in identical cases across multiple jurisdictions was strengthened.

Furthermore, as jurisdictional rules regarding intellectual property, contracts, and consumer protection were clarified, the Brussels I Regulation became the cornerstone of the EU’s unified civil procedure law. Eventually, it was replaced by the Regulation (EU) No 1215/2012 (Brussels I bis Regulation), adopted by the European Parliament and the Council.

 

한편, 루가노 협약(Lugano Convention, 1988) 2007 개정을 통해 브뤼셀 I 규정(EC No 44/2001) 일치하도록 조정되었으며, 이를 통해 EU EFTA 간의 법적 협력이 강화되었습니다 (루가노 협약 2007).

그러나 루가노 협약(2007) () EU 국가 피고에 대한 관할권 확대, 집행 승인 절차 폐지 등의 측면에서 이후 도입된 브뤼셀 I bis 규정(1215/2012) 내용이 반영되지 않아 일부 차이가 존재합니다.

Meanwhile, the Lugano Convention (1988) was revised in 2007 to align with the Brussels I Regulation (EC No 44/2001), thereby strengthening legal cooperation between the EU and EFTA (Lugano Convention 2007).

However, the 2007 Lugano Convention differs in certain aspects from the Brussels I bis Regulation (Regulation (EU) No 1215/2012), as it does not incorporate subsequent updates, such as expanded jurisdiction over non-EU defendants and the abolition of the exequatur procedure.

  

3. BSH Hausgeräte v. Electrolux 사건 이전 유럽 법원의 태도와 법규정 해석 (The European Courts' Stance and Interpretation of Legal Provisions Before the BSH Hausgeräte v. Electrolux Case)

 

3.1. 브뤼셀 I bis 규정 (Regulation (EU) No 1215/2012)의 일반관할규정 (General Jurisdiction Provisions Under the Brussels I bis Regulation (Regulation (EU) No 1215/2012))

 

1) 일반 재판적: 4 (General Jurisdiction: Article 4)

Article 4(1)
Subject to this Regulation, persons domiciled in a Member State shall, whatever their nationality, be sued in the courts of that Member State.

Article 4(2)
Persons who are not nationals of the Member State in which they are domiciled shall be subject to the rules of jurisdiction applicable to nationals of that Member State.

조항은 브뤼셀 I 규정(EC No 44/2001) 2조를 대체하며, 피고의 주소지를 관할권의 기준으로 삼는 일반 관할권 원칙을 규정하고 있습니다.

이에 따라, EU 회원국에 거주하는 당사자는 국적에 관계없이 해당 회원국 법원의 관할을 따르게 됩니다.

These provisions replace Article 2 of the Brussels I Regulation (EC No 44/2001) and establish the general jurisdiction principle, which is based on the defendant's domicile.

Accordingly, any individual or entity domiciled in an EU Member State is subject to the jurisdiction of that Member State’s courts, regardless of nationality.

 

2) 불법행위 관련 관할권: 7(2) (Jurisdiction Over Tort Claims: Article 7(2))

Article 7(2)
A person domiciled in a Member State may be sued in another Member State, in matters relating to tort, delict or quasi-delict, in the courts for the place where the harmful event occurred or may occur.

규정은 브뤼셀 I 규정(EC No 44/2001) 5(3) 대체하는 조항입니다.

불법행위(tort, delict, quasi-delict) 대한 소송은 피고의 주소지가 아닌, 피해가 발생하거나 발생할 가능성이 있는 국가에서 제기할 있도록 규정하고 있습니다.

예를 들어, 온라인 명예훼손 사건의 경우, 피해자는 자신이 피해를 입은 국가에서 소송을 제기할 있습니다.

This provision replaces Article 5(3) of the Brussels I Regulation (EC No 44/2001).

Under this rule, lawsuits concerning torts, delicts, or quasi-delicts can be filed not in the defendant’s country of domicile, but in the country where the harm occurred or is likely to occur.

For instance, in cases of online defamation, the injured party may bring a lawsuit in the country where the harm was suffered.

 

3) 복수 피고에 대한 소송: 8(1) (Lawsuits Involving Multiple Defendants: Article 8(1))

Article 8(1)
A person domiciled in a Member State may also be sued:
(1) where he is one of a number of defendants, in the courts for the place where any one of them is domiciled, provided the claims are so closely connected that it is expedient to hear and determine them together to avoid the risk of irreconcilable judgments resulting from separate proceedings.

규정은 브뤼셀 I 규정(EC No 44/2001) 6(1) 대체하는 조항입니다.

특허 침해나 상표권 침해 여러 국가에 걸쳐 발생하는 침해 사건에서, 다수의 피고가 포함된 경우, 국가의 법원에서 개별적으로 소송이 진행될 경우 모순된 판결이 나올 위험이 있습니다. 이에 따라, 피고 명의 주소지가 있는 회원국 법원에서 모든 피고를 대상으로 소송을 제기할 있도록 규정하고 있으며, 이는 다국적 분쟁에서 중요한 역할을 합니다.

This provision replaces Article 6(1) of the Brussels I Regulation (EC No 44/2001).

In cross-border infringement cases—such as patent or trademark infringement—where multiple defendants are involved, there is a risk of conflicting judgments if separate lawsuits are conducted in different national courts.

To prevent this, the regulation allows all defendants to be sued in the courts of the Member State where at least one of them is domiciled, provided that the claims are closely connected.

This provision plays a crucial role in multinational disputes, ensuring procedural efficiency and consistency in rulings.

 

4) 법인의 주소지: 63(1) (Corporate Domicile: Article 63(1))

Article 63(1)
For the purposes of this Regulation, a company or other legal person or association of natural or legal persons is domiciled at the place where it has its:
(a) statutory seat,
(b) central administration, or
(c) principal place of business.

규정은 브뤼셀 협약(Brussels Convention) 60조를 대체하는 조항입니다.

법인의 주소지(domicile) 법적 등록지(statutorily registered office), 본사 소재지(central administration), 또는 주된 영업소 소재지(principal place of business) 하나로 간주됩니다.

이는 기업 소송에서 관할권을 결정하는 중요한 기준으로 작용합니다.

This provision replaces Article 60 of the Brussels Convention.

Under this rule, a company’s domicile is determined based on one of three factors:

                       i.            Statutory seat (the legally registered office),

                    ii.            Central administration (the headquarters or place of management), or

                  iii.            Principal place of business (the main business location).

This definition serves as a key factor in determining jurisdiction for corporate litigation within the EU.

 

5) 종합해설 (Comprehensive Interpretation of Relevant Provisions)

유럽연합(EU) 회원국 내에 주소를 두고 있는 자는 국적에 관계없이 해당 주소지 국가의 법원 관할권에 따릅니다 (브뤼셀 I bis 규정 4(1)). 따라서, 유럽연합 피고의 주소지국 법원은 특허 침해와 관련된 모든 소송에 대해 재판 관할권을 가집니다.

예를 들어, 프랑스 법원은 프랑스에 거주하는 피고를 상대로 특허 침해 소송에 대한 관할권을 가지며, 침해 대상이 독일 특허이든, EU 회원국 또는 기타 EU 회원국의 특허이든 관계없이 소송을 진행할 있습니다.

따라서, 프랑스, 독일, 스위스에서 등록된 유럽 특허를 보유한 특허권자가 본점을 프랑스에 기업을 상대로 프랑스 특허의 침해를 이유로 프랑스 법원에 소송을 제기할 있음은 물론, 독일 특허나 스위스 특허의 침해를 이유로도 프랑스 법원에 소송을 제기할 있습니다.

Under Article 4(1) of the Brussels I bis Regulation, any individual or entity domiciled in an EU Member State is subject to the jurisdiction of that Member State’s courts, regardless of nationality. Consequently, the courts of the defendant’s domicile within the EU have jurisdiction over all patent infringement cases involving that defendant.

For instance, a French court has jurisdiction over a patent infringement case against a defendant domiciled in France, regardless of whether the alleged infringement concerns a German patent, an EU Member State patent, or even a non-EU patent.

Accordingly, a patent holder with European patents registered in France, Germany, and Switzerland can file an infringement lawsuit in France against a company headquartered in France, not only for the infringement of its French patent but also for its German or Swiss patents.

 

3.2. 복수 피고가 유럽 특허를 침해한 경우 (Infringement of a European Patent by Multiple Defendants)

1) Brussels I bis Regulation (Regulation (EU) No 1215/2012) – Article 8(1)

8(1) – 복수 피고 관련재판적 (Jurisdiction in respect of multiple defendants)

Article 8(1)
A person domiciled in a Member State may also be sued:
(1) where he is one of several defendants, in the courts for the place where any one of them is domiciled, provided the claims are so closely connected that it is expedient to hear and determine them together to avoid the risk of irreconcilable judgments resulting from separate proceedings.

규정은 브뤼셀 I 규정(EC No 44/2001) 6(1) 내용을 그대로 유지하면서 조항 번호가 8(1) 변경되었습니다.

피고가 여러 명인 경우, 이들 명이 거주하는 회원국에서 모든 피고를 상대로 소송을 제기할 있으며, 이는 피고에 대한 청구가 밀접하게 연결되어 있어야 하고(연결 요건), 별도로 소송을 진행할 경우 모순된 판결이 나올 위험이 있는 경우에만 적용됩니다.

This provision maintains the same principle as Article 6(1) of the Brussels I Regulation (EC No 44/2001), with its numbering changed to Article 8(1).

When multiple defendants are involved, a lawsuit can be brought against all of them in the Member State where at least one of them is domiciled, provided that the claims are closely connected and that separate proceedings could result in contradictory judgments.

 

연결 요건의 기준은 일명 거미줄 이론(a spider-in-the-web rule)’으로 불리며, 적용 기준은 다음과 같습니다.

                       i.            복수의 피고 회사가 존재할 ,

                    ii.            이들 피고 회사가 동일한 기업집단에 속할 ,

                  iii.            피고 회사들이 공통의 정책에 따라 행위를 하였을 ,

                   iv.            이러한 공통의 정책이 피소 법원이 위치한 국가의 회사에 의해 기획되었을 .

The criteria for determining this connection, commonly referred to as the “spider-in-the-web rule,” are as follows:

i.                    There must be multiple defendant companies.

ii.                  The defendant companies must belong to the same corporate group.

iii.                The defendant companies must have acted according to a common policy.

iv.                This common policy must have been formulated by a company located in the jurisdiction where the lawsuit is filed.

이와 관련하여, Expandable Grafts 사건에서는 사업 계획을 기획한 본점 소재지 국가의 법원에서 모든 피고에 대한 소송을 제기해야 한다고 판단 반면, Roche 사건에서는 공동 피고의 행위에 대한 공통의 정책이 피소 국가 어느 피고 회사에 의해 기획되었어야 한다는 점을 보다 엄격하게 요구하였습니다.

원칙은 주로 특허, 상표, 저작권, 반독점법과 관련된 다국적 소송에서 적용됩니다.

In the Expandable Grafts case, the court ruled that the lawsuit should be filed in the jurisdiction where the headquarters responsible for planning the business strategy is located. However, in the Roche case, the court imposed a stricter requirement, ruling that the common policy must have been formulated by at least one of the defendants domiciled in the jurisdiction where the case was filed.

This principle is primarily applied in multinational disputes involving patents, trademarks, copyrights, and antitrust law.

 

2) 예제를 통한 해설 (Explanation Through Examples)

예를 들어, A 국가에 본사를 기업과 B 국가에 본사를 기업이 계열사이거나 밀접한 사업상 관계를 맺고 있으며, 공동으로 특허를 침해한 경우, 피해자는 A 국가 또는 B 국가 법원 곳에서 회사를 동시에 제소할 있습니다.

또한, 명의 원고가 여러 국가에 거주하는 피고들을 상대로 소송을 제기할 경우, 해당 조항을 활용하여 피고들 명이 주소를 회원국 법원에서 모든 피고를 대상으로 소송을 진행할 있습니다.

For instance, if a company headquartered in Country A and another company headquartered in Country B are either affiliated or maintain a close business relationship and have jointly infringed a patent, the patent holder may sue both companies in the courts of either Country A or Country B.

Additionally, if a single plaintiff wishes to file a lawsuit against multiple defendants domiciled in different countries, this provision allows the case to be brought before the court of the Member State where at least one of the defendants is domiciled, thereby consolidating the claims against all defendants in a single jurisdiction.

 

3) 로슈(Roche) 사례: Roche Case (2006):

Roche Nederland BV v. Primus, Case C-539/03 [2006] ECR I-6535

사건은 네덜란드 법원이 10개국에 등록된 유럽 특허에 대해 각국의 8 기업을 상대로 특허 침해 판단을 내릴 있는지 여부 쟁점이었습니다.

This case centered on whether a Dutch court could rule on patent infringement claims involving European patents registered in ten different countries and brought against eight separate companies operating in those jurisdictions.

 

<사건 개요> Case Overview

프리무스(Primus) 박사 골든버그(Goldenberg) 오스트리아, 벨기에, 독일, 프랑스, 네덜란드, 영국, 스위스, 스웨덴, 리히텐슈타인, 룩셈부르크 등록된 유럽 특허(EP) 보유하고 있었습니다.

1997, 이들은 네덜란드에 설립된 Roche Nederland BV 미국, 영국, 프랑스, 스웨덴, 스위스, 독일, 오스트리아, 벨기에에 설립된 Roche 그룹의 8 계열사 상대로 네덜란드 법원에 특허 침해 소송을 제기하였습니다.

원고들은 피고들이 각각 설립된 국가에서 의료 기기를 판매함으로써 해당 국가의 유럽 특허를 침해했다 주장했습니다. 이에 대해 피고들은 침해 사실을 부인하는 한편, 특허의 유효성에 대한 무효 항변을 제기하였으며, 특히 네덜란드 밖에서 설립된 Roche 그룹 계열사들은 네덜란드 법원의 관할권을 부인하였습니다.

Case Overview

Dr. Primus and Goldenberg held European patents (EPs) registered in Austria, Belgium, Germany, France, the Netherlands, the United Kingdom, Switzerland, Sweden, Liechtenstein, and Luxembourg.

In 1997, they filed a patent infringement lawsuit in the Netherlands against Roche Nederland BV, as well as eight other subsidiaries of the Roche Group established in the United States, the United Kingdom, France, Sweden, Switzerland, Germany, Austria, and Belgium.

The plaintiffs alleged that each defendant had infringed their respective European patents by selling medical devices in the countries where they were established.

In response, the defendants denied the infringement allegations and filed a counterclaim challenging the validity of the patents. Additionally, the Roche Group subsidiaries based outside the Netherlands contested the Dutch court’s jurisdiction over them.

 

<법적 쟁점 ECJ 판단>

네덜란드 대법원은 브뤼셀 협약(Brussels Convention) 8(1) 해석과 관련하여, 유럽사법재판소(ECJ) 다음과 같은 법적 질문을 제기하였습니다.

                    i.            국가에서 이루어진 개별적인 특허 침해 행위가 동일한 기업 그룹의 공통 정책에 따라 이루어진 경우, 단일 회원국 법원에서 모든 피고를 상대로 소송을 진행할 있는지?

                  ii.            브뤼셀 협약 8(1) 적용 범위가 국경을 초월한 특허 침해 소송에도 적용될 있는지?

ECJ 이에 대해 유럽 특허(EP) 개별 국가에서 독립적으로 등록된 특허의 총합(a bundle of national patents)이며, 단일한 법이 적용되는 것이 아니므로 동일한 사실관계에 해당하지 않는다 판단하였습니다.

Legal Issues and ECJ's Judgment

The Dutch Supreme Court referred the case to the Court of Justice of the European Union (ECJ) for a preliminary ruling on the interpretation of Article 8(1) of the Brussels Convention. The key legal questions were:

i. If patent infringement acts committed in different countries follow a common policy set by the same corporate group, can a single Member State court hear the case against all defendants?
ii. Does Article 8(1) of the Brussels Convention apply to cross-border patent infringement cases?

The ECJ ruled that European patents (EPs) are a bundle of national patents, each registered and enforced independently in different countries, rather than a single unified legal entity. As a result, a uniform legal framework does not apply to all European patents, and the cases did not share an identical set of facts.

 

, 피고가 서로 다른 체약국에서 서로 독립적이고 개별적으로 특허를 침해한 경우, 각각의 소송이 동일한 사실관계를 공유하지 않으므로, 브뤼셀 협약 8(1) 따라 단일 국가 법원이 모든 피고를 관할할 없다고 판시하였습니다. 따라서, 국가별 법원이 서로 다른 결론을 내리더라도 모순된 판결로 간주되지 않는다 보았습니다.

또한, 피고의 개별 행위마다 적용되는 법률이 다르기 때문에 브뤼셀 협약 8(1) 요건을 충족하지 못하며, 이에 따라 거미줄 이론(a spider-in-the-web rule)’ 역시 적용될 없다고 판단하였습니다.

결국, 로슈 사건에서는 브뤼셀 협약 6(1) 적용이 전제가 되는 거미줄 이론 부정되었으며, 다국적 기업을 대상으로 단일 국가 일괄 소송 전략이 제한되었습니다.

Consequently, the court held that when multiple defendants infringe the same European patent in separate contracting states, their acts are legally and factually independent of one another. As such, under Article 8(1) of the Brussels Convention, a single court in one Member State cannot assert jurisdiction over all defendants. The ECJ further clarified that different national courts reaching different conclusions on infringement did not constitute contradictory rulings.

Additionally, because each defendant's actions were governed by the laws of different jurisdictions, the ECJ found that the requirements for Article 8(1) were not met. As a result, the "spider-in-the-web" theory, which would have allowed jurisdiction consolidation, was rejected.

Ultimately, in the Roche case, the ECJ refused to apply the "spider-in-the-web" theory, limiting the ability of plaintiffs to bring a single multinational lawsuit in one jurisdiction against multiple corporate defendants.

 

3.3. 특별재판적, Exclusive jurisdiction (전속관할)

3.3.1. 브뤼셀 I bis 규정 24(4) Article 24(4) of the Brussels I bis Regulation

24(4) – 특허, 상표, 디자인 지식재산권 유효성 관련 전속관할 (Exclusive jurisdiction for IP validity disputes)

Article 24(4)
In proceedings concerned with the registration or validity of patents, trade marks, designs, or other similar rights required to be deposited or registered, the courts of the Member State in which the deposit or registration has been applied for, has taken place, or is under the terms of an international convention deemed to have taken place, shall have exclusive jurisdiction, regardless of domicile.

Without prejudice to the jurisdiction of the European Patent Office under the Convention on the Grant of European Patents, the courts of each Member State shall have exclusive jurisdiction in proceedings concerning the registration or validity of any European patent granted for that State.

조항은 브뤼셀 I 규정(EC No 44/2001) 22(4) 원칙을 그대로 유지하면서, 24(4) 변경되었습니다.

특허, 상표, 디자인 기타 등록이 필요한 지식재산권의 유효성 또는 등록과 관련된 소송은 해당 권리가 등록된 회원국 법원이 전속 관할권을 갖습니다. 또한, 유럽특허청(EPO) 관할권(뮌헨 협약, EPC) 영향을 미치지 않으며, 국가에서 등록된 유럽 특허의 유효성과 관련된 소송은 해당 국가 법원의 전속 관할에 속합니다.

This provision maintains the same principle as Article 22(4) of the Brussels I Regulation (EC No 44/2001), with its numbering changed to Article 24(4).

Under this rule, disputes concerning the registration or validity of patents, trademarks, designs, and other registered intellectual property rights fall under the exclusive jurisdiction of the courts of the Member State where the right was registered.

Additionally, this provision does not affect the jurisdiction of the European Patent Office (EPO) under the Munich Convention (EPC). However, for disputes concerning the validity of a European patent (EP) granted for a specific country, the national courts of that country retain exclusive jurisdiction.

 

3.3.2. 브뤼셀 I bis 규정 24(5) Article 24(5) of the Brussels I bis Regulation

24(5) – 판결 집행 관련 전속관할 (Exclusive jurisdiction for enforcement of judgments)

Article 24(5)
In proceedings concerned with the enforcement of judgments, the courts of the Member State in which the judgment has been or is to be enforced shall have exclusive jurisdiction.

조항은 브뤼셀 I 규정(EC No 44/2001) 22(5) 대체하는 규정입니다.

판결이 이미 집행되었거나 집행될 예정인 회원국의 법원이 해당 판결 집행에 대한 전속 관할권을 가집니다.

예를 들어, A 법원이 판결을 내렸고 이를 B국에서 집행해야 하는 경우, B 법원이 해당 판결 집행 절차를 전속적으로 담당하게 됩니다.

This provision replaces Article 22(5) of the Brussels I Regulation (EC No 44/2001).

Under this rule, the courts of the Member State where a judgment has already been enforced or is to be enforced have exclusive jurisdiction over enforcement-related proceedings.

For example, if a court in Country A issues a judgment that needs to be enforced in Country B, then the courts of Country B have exclusive jurisdiction over the enforcement process.

 

3.3.3. 사례를 통한 해설 Explanation Through Case Analysis

지식재산권 침해 소송에서 해당 권리의 유효성이 쟁점으로 제기된 경우, 브뤼셀 I bis 규정 24(4) 따라 유효성 판단은 해당 지식재산권이 등록된 국가의 법원이 전속 관할권을 가진다는 입장이 적용됩니다.

따라서, 피고 주소지 국가의 관할 법원에서 주소지 A 국가 특허는 물론 다른 국가 B, C 특허 유효성이 다투어지는 경우, 피고의 주소지를 관할하는 법원의 특허 침해 소송 절차는 특허가 등록된 B, C, D, E, F국에서의 유효성 관련 소송이 종료될 때까지, 중지되어야 한다는 의미를 갖습니다.

In intellectual property infringement litigation, when the validity of the IP right is challenged, Article 24(4) of the Brussels I bis Regulation establishes that the courts of the country where the right is registered have exclusive jurisdiction over the validity determination.

Accordingly, if a court in the defendant's country of domicile (Country A) is handling a patent infringement case that also involves the validity of patents registered in other countries (e.g., Countries B and C), the infringement proceedings in Country A must be stayed until the validity disputes in Countries B, C, D, E, and F are resolved by the respective national courts.

 

<독일의 특허 침해 무효 소송 절차>

Patent Infringement and Nullity Proceedings in Germany

독일에서는 특허 침해 소송과 특허 무효 소송이 분리되어 심리되며, 각각 다른 법원이 담당하기 때문에 침해 소송을 담당하는 민사 법원에서 원칙적으로 특허 무효 항변을 제기하는 것이 불가능합니다.

이는 특허 침해 소송이 일반 민사 법원에서 진행되는 반면, 특허의 유효성에 대한 심리는 일반 민사 법원이 아닌, 뮌헨(Munich) 위치한 독일 연방 특허 법원(Bundespatentgericht) 전문 기술판사(specialist judges) 의해 다뤄지기 때문입니다.

그러나, 독일에서는 특허 무효 항변이 제기되더라도 연방법원의 최종 판단을 기다리지 않고 침해 여부에 대한 결정을 신속하게 진행하는 경향이 강했습니다. 이러한 절차적 특성 덕분에, 독일은 이론상 무효 항변으로 인해 발생하는 관할권 논란에서 상대적으로 자유로울 있었습니다.

In Germany, patent infringement and patent nullity (invalidity) proceedings are handled separately, with different courts responsible for each matter. As a result, it is generally not possible to raise a patent invalidity defense before the civil court handling an infringement case.

This separation exists because patent infringement cases are heard in ordinary civil courts, whereas the validity of a patent is determined not by civil courts but by the Federal Patent Court (Bundespatentgericht), located in Munich, where cases are adjudicated by specialist judges.

Despite this division, German courts tend to proceed swiftly with infringement cases without waiting for a final decision from the Federal Patent Court on a pending invalidity challenge. Due to this procedural characteristic, Germany has been relatively insulated from jurisdictional disputes arising from invalidity defenses.

 

<독일 GAT 사건의 영향>

그럼에도 불구하고, 독일 GAT 사건(Case C-4/03 Gesellschaft für Antriebstechnik mbH & Co KG (GAT) v. Lamellen und Kupplungsbau Beteiligungs KG (LuK) [2006] ECR I-6509)에서, 브뤼셀 I bis 규정 24(4) 전속 관할 규정은 특허 무효 쟁점이 항변으로 제기되었는지, 별도의 무효 소송으로 제기되었는지에 관계없이 적용된다고 판시되었습니다.

, 특허의 등록 유효성과 관련된 모든 소송은 해당 특허가 등록된 국가의 법원이 전속 관할권을 가지며, 이에 따라 특허 침해 소송이 제기된 법원은 해당 등록국 법원의 무효 판단이 나올 때까지 침해 소송을 중지해야 한다고 판결하였습니다.

Impact of the GAT Case in Germany

Despite Germany’s procedural separation between patent infringement and invalidity proceedings, the GAT case (Case C-4/03 Gesellschaft für Antriebstechnik mbH & Co KG (GAT) v. Lamellen und Kupplungsbau Beteiligungs KG (LuK) [2006] ECR I-6509) significantly influenced jurisdictional rules concerning patent validity disputes.

In this case, the Court of Justice of the European Union (CJEU) ruled that the exclusive jurisdiction rule under Article 24(4) of the Brussels I bis Regulation applies regardless of whether a patent validity issue is raised as a defense in an infringement case or through a separate invalidity action.

Accordingly, any legal proceedings concerning the registration or validity of a patent fall under the exclusive jurisdiction of the courts of the country where the patent is registered. As a result, a court hearing an infringement case must stay the proceedings until the competent court in the country of registration rules on the patent’s validity.

 

<영국의 접근 방식>

한편, 영국은 동일한 민사 법원에서 특허 침해 소송과 특허 무효 소송을 함께 심리할 있도록 허용하고 있습니다. 그러나, 소송에서 외국 특허의 유효성이 쟁점이 되는 경우, 일반적으로 소를 각하하는 입장을 취해 왔습니다 (Coin Controls Ltd v. Suzo International (UK) Ltd and Others [1999] Ch 33, Coin Control 사건).

The UK’s Approach

In contrast, the United Kingdom allows patent infringement and patent validity disputes to be heard together in the same civil court.

However, when a case involves the validity of a foreign patent, UK courts have generally taken the approach of declining jurisdiction and dismissing the case. This principle was affirmed in Coin Controls Ltd v. Suzo International (UK) Ltd and Others [1999] Ch 33 (Coin Controls case).

  

4. BSH Hausgeräte v. Electrolux 사건에서의 유럽사법재판소 해석(2025)

Interpretation of the BSH Hausgeräte v. Electrolux Case by the Court of Justice of the European Union (2025)

2025 유럽사법재판소(CJEU) 판결한 BSH Hausgeräte v. Electrolux 사건 EU 국경을 초월한 지식재산권 분쟁에서의 관할권 규칙 해석에 중요한 의미를 갖는 판결입니다.

사건은 브뤼셀 I bis 규정(Regulation (EU) No 1215/2012) 특허 침해 소송에서 다수의 피고가 포함된 경우의 관할권 배분 문제 다루었으며, 침해 소송과 무효 항변이 동시에 제기된 경우 유럽 법원이 관할권을 어떻게 해석해야 하는지 명확히 하였습니다.

특히, 판결은 여러 국가에서 등록된 유럽 특허의 침해를 다루는 소송에서 어떤 국가의 법원이 관할권을 가지는지 브뤼셀 I bis 규정에 따라 다수의 피고를 포함하는 청구를 단일 국가 법원에서 진행할 있는지 여부 대한 법적 기준을 확립하는 중요한 역할을 하였습니다.

The BSH Hausgeräte v. Electrolux case, decided by the Court of Justice of the European Union (CJEU) in 2025, provided a crucial interpretation of jurisdictional rules in cross-border intellectual property disputes within the EU.

This case examined the interaction between the Brussels I bis Regulation (Regulation (EU) No 1215/2012) and the allocation of jurisdiction in cases involving multiple defendants in patent infringement disputes. The ruling further clarified how European courts should approach jurisdictional issues when both infringement and validity challenges are raised in transnational litigation.

The CJEU’s interpretation in this case had significant implications for cross-border patent enforcement, particularly in determining which national courts have jurisdiction over European patents registered in multiple countries and how claims involving multiple defendants should be consolidated under the Brussels I bis framework.

 

4.1 사건의 개요 (Case Overview)

 

4.1.1. 원고의 주장 (Plaintiff’s Claims)

원고 BSH 진공청소기 기술과 관련된 유럽 특허(EP 1 434 512) 권리자로서, 해당 특허는 독일, 그리스, 스페인, 프랑스, 이탈리아, 네덜란드, 오스트리아, 스웨덴, 영국, 튀르키예에서 유효하게 등록(validated)되었으며, 이에 따라 각국에서 국가 특허(national patents) 인정받고 있다.

2020 2 3, BSH Electrolux 특허를 침해했다는 이유로 스웨덴 특허·상업법원(Patent- och marknadsdomstolen, Patent and Commercial Court) 소송을 제기하였다.

BSH 소장에서 Electrolux 해당 특허가 유효하게 등록된 모든 국가에서 특허 기술을 사용하는 행위를 즉시 중단하도록 법원이 명령할 것을 요청하였다. 또한, Electrolux 해당 발명을 부당하게 사용한 것에 대한 공정한 보상(equitable remuneration) 손해배상을 지급하도록 법원이 판결할 것을 요구하였다.

The plaintiff, BSH, is the rights holder of European Patent EP 1 434 512, which pertains to vacuum cleaner technology. This patent has been validated in Germany, Greece, Spain, France, Italy, the Netherlands, Austria, Sweden, the United Kingdom, and Türkiye, and as a result, it has been recognized as a national patent in each of these countries.

On February 3, 2020, BSH filed a patent infringement lawsuit against Electrolux before the Swedish Patent and Commercial Court (Patent- och marknadsdomstolen), alleging that Electrolux had infringed upon its patent.

In its complaint, BSH requested that the court order Electrolux to immediately cease using the patented technology in all countries where the patent has been validly registered. Additionally, BSH sought a ruling requiring Electrolux to pay equitable remuneration and damages for the unauthorized use of the patented invention.

 

4.1.2. 피고의 주장 (Defendant’s Arguments)

피고 Electrolux BSH 청구가 기각되어야 하며, 스웨덴 이외의 국가에서 유효화된 특허(외국 특허, foreign patents) 대한 침해 소송은 허용될 없다 주장하였습니다.

Electrolux 먼저 외국 특허가 무효이며, 따라서 스웨덴 법원은 해당 특허의 침해 여부를 판단할 관할권이 없다고 주장하였습니다. 또한, 브뤼셀 I bis 규정(Brussels I bis Regulation) 24(4) 따르면, 특허 침해 소송은 본질적으로 해당 특허의 유효성과 불가분(indissociable) 관계를 가지므로, 특허가 유효화된 국가의 법원만이 침해 여부를 판단할 관할권을 가진다 해석되어야 한다고 주장하였습니다. 이에 따라, 스웨덴 법원은 외국 특허의 침해 여부를 판단할 없다 결론지었습니다.

Electrolux 추가적으로, 스웨덴 특허법 61 2항이 특허 침해 소송과 별도로 특허 무효 소송을 진행해야 한다고 규정하고 있으나, 조항은 스웨덴 특허에만 적용되며 외국 특허에는 적용되지 않는다 강조하였습니다. 따라서, 스웨덴 법원은 침해 소송에서 외국 특허의 무효 여부를 판단할 없으며, BSH 외국 특허의 침해 소송을 해당 특허가 유효화된 국가에서 각각 제기해야 한다 주장하였습니다.

The defendant, Electrolux, argued that BSH’s claims should be dismissed and that patent infringement lawsuits concerning patents validated outside of Sweden (foreign patents) should not be allowed.

First, Electrolux contended that the foreign patents in question were invalid and, therefore, that the Swedish court lacked jurisdiction to determine whether these patents had been infringed. Furthermore, the company asserted that under Article 24(4) of the Brussels I bis Regulation, patent infringement actions are inherently and indissociably linked to the validity of the patent. Consequently, only the courts of the country where the patent has been validated should have jurisdiction over infringement matters. Based on this interpretation, Electrolux concluded that the Swedish court could not rule on the alleged infringement of foreign patents.

Additionally, Electrolux highlighted that Section 61(2) of the Swedish Patent Act requires patent invalidity actions to be pursued separately from infringement proceedings. However, Electrolux emphasized that this provision applies only to Swedish patents and does not extend to foreign patents. As a result, the Swedish court lacks the authority to determine the validity of foreign patents in an infringement case, and BSH must file separate infringement lawsuits in each country where the patents have been validated.

 

4.2. 스웨덴 법원의 결정 및 BSH의 항소 (Decision of the Swedish Court and BSH’s Appeal)

2020 12 21, 스웨덴 특허·상업법원(Patent- och marknadsdomstolen) 브뤼셀 I bis 규정 24(4) 27조를 근거로 스웨덴 국가에서 유효하게 등록된 특허에 대한 침해 소송을 심리할 관할권이 없다고 판시하였습니다.

또한, 튀르키예(Türkiye)에서 유효화된 특허(‘튀르키예 특허’) 대한 침해 소송 역시 법원이 관할권을 가지지 않는다고 판단하였습니다. 법원은 근거로 브뤼셀 I bis 규정 24(4) 국제적으로 인정되는 관할권 원칙을 반영하고 있다는 점을 제시하였습니다.

On December 21, 2020, the Swedish Patent and Commercial Court (Patent- och marknadsdomstolen) ruled that it lacked jurisdiction to hear the infringement claims concerning patents that had been validly registered outside of Sweden, citing Articles 24(4) and 27 of the Brussels I bis Regulation as the legal basis for its decision.

Additionally, the court determined that it also lacked jurisdiction over the infringement claims related to the patent validated in Türkiye (the "Turkish Patent"). The court justified its ruling by emphasizing that Article 24(4) of the Brussels I bis Regulation reflects internationally recognized principles of jurisdiction.

 

이에 대해 BSH 스웨덴 항소법원(Svea hovrätt, 특허·상업법원 항소부) 항소하였습니다.

BSH 항소장에서 다음과 같은 논거를 제시하였습니다.

·        브뤼셀 I bis 규정 24(4)순수한 특허 침해 소송(pure patent infringement action)’에는 적용되지 않으므로,

-      4(1) 근거하여 관할권을 가진 법원은 해당 특허의 무효 여부를 판단할 권한이 없어도, 침해 여부는 판단할 있다고 주장하였습니다.

·        또한, 스웨덴 특허·상업법원은 비회원국(: 튀르키예)에서 유효화된 특허에 대한 침해 소송을 심리할 있다 주장하였습니다.

-      이는 브뤼셀 I bis 규정 4(1) 따른 국제적 관할권 원칙에 따른 것이라고 설명하였습니다.

·        결론적으로, 해당 법원은 튀르키예 특허를 포함하여 전체 침해 소송을 심리할 관할권을 가진다 주장하였습니다.

·        또한, 피고의 거주지 법원이 관할권을 가진다는 원칙은 국제법상 인정되는 원칙이라는 점을 강조하였습니다.

 

BSH’s Appeal to the Swedish Court of Appeal

Following this ruling, BSH filed an appeal with the Svea Court of Appeal (Svea hovrätt, Patent and Commercial Court of Appeal), arguing that the lower court’s decision was incorrect.

In its appeal, BSH presented the following arguments:

·        Article 24(4) of the Brussels I bis Regulation does not apply to a ‘pure patent infringement action.’

-      Therefore, under Article 4(1) of the Brussels I bis Regulation, a court that has jurisdiction over the infringement claim may rule on infringement even if it lacks the authority to assess the validity of the patent.

·        The Swedish Patent and Commercial Court has jurisdiction to hear infringement claims concerning patents validated in non-EU countries, such as Türkiye.

-      This jurisdiction is based on the principle of international jurisdiction established in Article 4(1) of the Brussels I bis Regulation.

·        As a result, the court has jurisdiction over the entire infringement lawsuit, including the Turkish Patent.

·        Additionally, BSH emphasized that the principle that a defendant’s domicile determines jurisdiction is a well-established principle in international law.

 

4.3. 피고 Electrolux의 반론 및 스웨덴 항소법원의 법적 의문

Defendant Electrolux’s Counterarguments and the Swedish Court of Appeal’s Legal Questions

4.3.1. 피고의 반박

Defendant’s Counterarguments

피고 Electrolux 1심에서 주장한 입장을 유지하며, 스웨덴 법원이 전체 소송을 심리할 관할권을 가질 없다는 점을 강조하며 다음과 같이 반박하였습니다.

먼저, 브뤼셀 I bis 규정 24(4) 특허 침해 소송에서 피고가 특허 무효를 방어 논거로 제기한 경우에도 적용되어야 한다고 해석되어야 한다고 주장하였습니다. 따라서, 특허 침해 여부와 특허 유효성 판단은 불가분의 관계를 가지므로, 침해 여부를 판단하기 위해서는 반드시 해당 특허의 유효성 또한 검토되어야 한다는 입장입니다. 그러나 특허의 유효성 판단은 해당 특허가 유효화된 국가에서만 가능하므로, 스웨덴 법원은 외국 특허의 침해 여부를 판단할 있는 관할권을 가질 없다고 주장하였습니다.

Electrolux maintained its position from the first instance and strongly contested the Swedish court’s jurisdiction over the entire lawsuit. The defendant presented the following counterarguments:

·        Article 24(4) of the Brussels I bis Regulation should apply even when the defendant raises a patent invalidity defense in an infringement action.

-      According to Electrolux, patent infringement and patent validity are indissociable issues, meaning that determining infringement inherently requires an assessment of the patent’s validity.

-      Since only the courts of the country where the patent has been validated have jurisdiction to rule on its validity, the Swedish court cannot have jurisdiction to decide on the infringement of foreign patents.

 

4.3.2. 스웨덴 항소법원의 법적 의문 제기

Legal Questions Raised by the Swedish Court of Appeal

 

이와 같은 Electrolux 주장에 대해, 스웨덴 항소법원(송부 법원, referring court) 스웨덴 법원이 관할권을 가질 있는지 여부에 대해 법적 의문을 제기하였습니다.

In response to Electrolux's arguments, the Swedish Court of Appeal (the referring court) raised legal questions regarding whether the Swedish court could have jurisdiction over the case.

 

1) 브뤼셀 I bis 규정 24(4) 해석 문제

우선, 브뤼셀 I bis 규정 24(4)특허 등록 또는 유효성과 관련된 소송이라는 표현이 특허 침해 소송에서 피고가 무효 항변을 제기한 경우까지 포함하는지 여부 쟁점이 되었습니다.

만약 조항이 무효 항변이 제기된 특허 침해 소송에도 적용된다고 해석될 경우, 스웨덴 법원은 외국 특허의 침해 여부를 판단할 관할권을 가질 없게 됩니다. 그러나 이러한 해석이 인정될 경우, 원고는 여러 회원국에서 개별적으로 침해 소송을 제기해야 하는 부담을 가지게 되며, 이는 절차적 비효율성을 초래할 가능성이 있다 문제가 발생합니다.

1) Interpretation of Article 24(4) of the Brussels I bis Regulation

The first key issue was whether the phrase “proceedings concerned with the registration or validity of patents” in Article 24(4) of the Brussels I bis Regulation also applies when the defendant raises a patent invalidity defense in an infringement lawsuit.

·        If Article 24(4) is interpreted as applying to infringement cases where a validity defense is raised, then the Swedish court would lack jurisdiction to rule on the infringement of foreign patents.

·        However, such an interpretation would require plaintiffs to file separate infringement lawsuits in multiple jurisdictions where the patents have been validated, leading to procedural inefficiencies and potential fragmentation of litigation across different courts.

The Swedish Court of Appeal, therefore, questioned whether Article 24(4) should be read as broadly as Electrolux suggested or whether a more restrictive interpretation was warranted to prevent unnecessary procedural burdens on the plaintiff.

 

2) 다른 해석 가능성

Alternative Interpretation Possibilities

이에 따라, 다른 해석 가능성에 대한 검토도 이루어졌습니다. , 특허 침해 소송을 심리하는 법원이 피고가 제기한 무효 항변을 심리할 권한은 없지만, 침해 여부 자체는 판단할 있는지 여부 주요 논점이 되었습니다.

이와 같은 해석은 브뤼셀 I bis 규정 24(4) 가지는 예외적 성격을 고려한 제한적 해석(restrictive interpretation) 부합하며, 또한 특허 유효성에 관한 분쟁은 등록 국가의 법원에서만 심리해야 한다는 기본 원칙을 유지할 있다는 점에서 법리적으로 타당성이 있을 있습니다.

Given these concerns, the Swedish Court of Appeal also examined the possibility of an alternative interpretation. The key question was whether a court hearing a patent infringement case could rule on the infringement claim while lacking the authority to assess the validity of the patent if the defendant raised a validity defense.

·        Under this interpretation, the Swedish court would not have jurisdiction to rule on the validity of the foreign patents but could still determine whether infringement had occurred.

·        This approach aligns with a restrictive interpretation of Article 24(4) of the Brussels I bis Regulation, recognizing its exceptional nature while maintaining the fundamental principle that patent validity disputes should be adjudicated only by the courts of the country where the patent is registered.

The Swedish Court of Appeal considered whether this narrower reading of Article 24(4) would allow infringement cases to proceed without requiring courts to rule on validity issues beyond their jurisdiction, thereby preserving procedural efficiency while respecting international jurisdictional principles.

 

3) 추가적인 법적 고려사항

스웨덴 항소법원은 브뤼셀 I bis 규정 24(4) 해석과 관련하여 다음과 같은 법적 쟁점들을 추가적으로 검토할 필요가 있다고 판단하였습니다.

A.       만약 브뤼셀 I bis 규정 24(4)무효 항변이 제기된 경우, 침해 소송 전체가 독점적 관할권의 대상이 된다 해석될 경우,

-       스웨덴 특허법 61 2(무효 판단은 별도의 소송에서 이루어져야 한다는 규정) 적용될 있는지 여부가 문제가 있습니다.

B.        본안 사건의 유럽 특허가 EU 국가(튀르키예)에서 유효화된 사실이 관할권 판단에 영향을 미치는지 여부도 검토할 필요가 있습니다.

-       브뤼셀 I bis 규정 24(4) EU 국가 법원에도 적용될 있는가?

-       반면, 브뤼셀 I bis 규정 33 34조에서는 EU 국가 법원을 명시적으로 언급하고 있다는 점이 고려되어야 합니다.

-       또한, 2005 3 1 Owusu 판결(C-281/02) 따르면, 브뤼셀 I bis 규정 4조가 EU 국가 법원에도 적용될 있다는 해석이 가능합니다.

결론적으로, 스웨덴 항소법원은 사건이 브뤼셀 I bis 규정의 해석에 있어 중요한 법적 문제를 포함하고 있다고 판단하였으며, 이에 따라 스웨덴 법원이 외국 특허의 침해 여부를 판단할 있는 관할권을 가질 있는지에 대한 법적 검토가 필요하다는 입장을 보였습니다.

3) Additional Legal Considerations

The Swedish Court of Appeal identified further legal issues that required examination in the interpretation of Article 24(4) of the Brussels I bis Regulation.

A. If Article 24(4) is interpreted to mean that "when a validity defense is raised, the entire infringement lawsuit falls under exclusive jurisdiction," does this affect the application of Swedish patent law?

·        Could Section 61(2) of the Swedish Patent Act—which mandates that validity determinations be made in separate proceedings—still apply in this context?

·        This raised the question of whether Swedish procedural law and the Brussels I bis Regulation could be harmonized in cases where validity challenges arise in infringement disputes.

B. Does the validation of the European patent in a non-EU country (Türkiye) affect the jurisdictional analysis?

·        Can Article 24(4) of the Brussels I bis Regulation be applied to non-EU courts?

·        In contrast, Articles 33 and 34 of the Brussels I bis Regulation explicitly reference non-EU courts, which suggests that their role in jurisdictional matters must be considered.

·        Additionally, in the March 1, 2005 Owusu ruling (C-281/02), the Court of Justice of the European Union (CJEU) suggested that Article 4 of the Brussels I bis Regulation could potentially apply to non-EU courts.

Conclusion

Ultimately, the Swedish Court of Appeal recognized that this case raises significant legal questions regarding the interpretation of the Brussels I bis Regulation. Given these complexities, the court determined that a thorough legal review was necessary to establish whether the Swedish court could exercise jurisdiction over the alleged infringement of foreign patents.

 

4.4. 스웨덴 항소법원의 유권해석 요청과 유럽사법재판소 판단

The Swedish Court of Appeal’s Request for a Preliminary Ruling and the ECJ’s Judgment

4.4.1. 스웨덴 항소법원의 선결적 판결(유권해석) 요청

The Swedish Court of Appeal’s Request for a Preliminary Ruling

이에 따라, 스웨덴 항소법원(Svea hovrätt, Patent- och marknadsöverdomstolen) 절차를 중단하고 다음과 같은 질문을 ECJ 제출했습니다.

(1) 브뤼셀 I bis 규정 24(4) 특허 침해 소송에서 피고가 무효 항변을 제기한 경우, 침해 소송 전체에 대해 관할권을 배제하는가?

(2) 스웨덴 특허법 61 2항과 같은 국내법 규정이 해석에 영향을 미칠 있는가?

(3) 브뤼셀 I bis 규정 24(4) EU 국가(: 튀르키예) 법원에도 적용되는가?

In light of the legal uncertainties surrounding jurisdiction in this case, the Swedish Court of Appeal (Svea hovrätt, Patent- och marknadsöverdomstolen) decided to suspend proceedings and refer the following questions to the Court of Justice of the European Union (ECJ) for a preliminary ruling:

·        Does Article 24(4) of the Brussels I bis Regulation exclude jurisdiction over the entire infringement action when the defendant raises a validity defense?

·        Can a domestic legal provision, such as Section 61(2) of the Swedish Patent Act, influence the interpretation of Article 24(4)?

·        Does Article 24(4) of the Brussels I bis Regulation apply to non-EU courts, such as those in Türkiye?

 

4.4.2. 유럽사법재판소의 판단

The ECJ’s Judgment

유럽사법재판소는 스웨덴 항소법원의 요청을 받아들여, 항소법원이 요청한 브뤼셀 I bis 규정 24(4) 해석에 대한 번째 번째 질문 함께 심리하고 다음과 같이 판시하였습니다.

브뤼셀 I bis 규정 24(4) 해석과 관련하여,

- 피고가 특허 무효성을 방어 논거로 주장하더라도,

- 피고의 거주지 회원국 법원이 특허 침해 소송을 심리할 관할권을 유지한다고 해석해야 한다.

- 다만, 해당 특허의 유효성 문제는 특허 부여국 법원의 독점적 관할권에 속한다.

The Court of Justice of the European Union (ECJ) accepted the Swedish Court of Appeal’s request and ruled on the first and second questions together, providing the following interpretation of Article 24(4) of the Brussels I bis Regulation:

·        A defendant’s assertion of patent invalidity as a defense does not deprive the court of the defendant’s domicile of its jurisdiction to hear the infringement action.

·        The court in the defendant’s Member State retains jurisdiction over the patent infringement case, even if the validity of the patent is challenged.

·        However, the question of a patent’s validity falls exclusively within the jurisdiction of the courts of the country where the patent was granted.

Thus, the ECJ reaffirmed the distinction between infringement and validity disputes, clarifying that while a Member State court may hear infringement cases, it cannot rule on the validity of a foreign patent, which remains under the exclusive jurisdiction of the granting country.

 

유럽사법재판소(ECJ) 브뤼셀 I bis 규정(Brussels I bis Regulation) 24(4) 3 법원에도 적용되는지 여부 대한 번째 질문을 받아들여 심리한 다음과 같이 판시하였습니다.

1)     브뤼셀 I bis 규정 24(4) 회원국이 아닌 3국에서 부여되거나 유효화된 특허에는 적용되지 않는다.

·       규정은 EU 회원국 내에서 부여되거나 유효화된 특허의 유효성 관련 분쟁에 대해 전속 관할권을 부여하는 조항이며, 3국에서 부여되거나 유효화된 특허에는 적용되지 않는다.

·       따라서, 3(: 튀르키예)에서 부여되거나 유효화된 특허의 유효성을 판단할 전속 관할권이 해당 3 법원에 자동적으로 부여된다고 없으며, 브뤼셀 I bis 규정에 의해 해당 법원이 그러한 권한을 갖는다고 해석해서는 된다.

·       , 브뤼셀 I bis 규정 24(4) 3 법원에 적용되지 않으며, 해당 법원에 특허의 유효성을 판단할 배타적 관할권을 부여하지 않는다.

2)     스웨덴 법원의 관할권 문제는 브뤼셀 I bis 규정 4(1) 일반 국제법 원칙(비간섭 원칙 ) 따라 따로 검토해야 한다.

·       4(1) 따르면, 3국에서 부여되거나 유효화된 특허의 침해를 이유로 제기된 소송에서, 피고가 방어 논거로 특허 무효성을 주장하는 경우에도, 피고의 거주지 회원국 법원은 원칙적으로 해당 침해 소송을 심리할 관할권을 가진다.

·       , 피고가 특허 무효성을 방어 논거로 주장하더라도, 스웨덴 법원은 해당 침해 소송을 심리할 있는 관할권을 유지한다.

·       다만, 피고의 무효 항변이 침해 여부를 판단하는 있어 불가피하게 고려되어야 하는 경우에도, 스웨덴 법원은 해당 특허의 유효성을 직접 판단하기보다는, 침해 여부에 대한 판단의 일환으로 간접적으로 다룰 있다.

3)     회원국 피고의 거주지 법원이 특허 무효성까지 관할할 있는 원칙은, 브뤼셀 I bis 규정 73조에서 규정한 특별 규칙(special rules) 의해 제한될 있다.

·       예를 들어, 73(1) 따르면, 루가노 협약(Lugano Convention) 가입한 국가의 법원은, 해당 협약 당사국에서 부여된 특허의 유효성을 판단할 독점적 관할권을 가진다.

·       그러나, 튀르키예는 루가노 협약의 당사국이 아니며, 법원이 검토한 사건 기록에서는 스웨덴과 튀르키예 간에 관련 협약이 존재한다는 근거를 찾을 없다.

4)     국가 주권 비간섭 원칙의 적용

·       각국은 특허권을 부여, 등록 유효화할 있는 독립적인 권한을 가지며, 이는 국가 주권(national sovereignty) 행사로 간주된다.

·       이러한 비간섭 원칙(the principle of non-interference, 71항에서 언급됨) 따르면,

·       특허를 무효화하는 사법적 결정은 오직 해당 특허가 부여되거나 유효화된 국가의 법원만이 내릴 있으며,

·       해당 판결에 따라 해당 국가의 특허 등록부(national register) 수정될 있다.

·       결론적으로, EU 회원국 법원은 3국에서 부여되거나 유효화된 특허의 유효성을 판단하여 등록된 특허의 무효를 선언할 권한이 없으며, 그러한 판단은 해당 3 법원의 전속 관할권에 속한다.

ECJ’s Ruling on the Applicability of Article 24(4) of the Brussels I bis Regulation to Non-EU Courts

The Court of Justice of the European Union (ECJ) reviewed the third question regarding whether Article 24(4) of the Brussels I bis Regulation applies to non-EU courts and rendered the following judgment:

1) Article 24(4) of the Brussels I bis Regulation does not apply to patents granted or validated in non-EU countries.

o  This provision grants exclusive jurisdiction over disputes concerning the validity of patents granted or validated within EU Member States and does not extend to patents issued or validated in non-EU jurisdictions.

o  Consequently, courts in non-EU countries (e.g., Türkiye) do not automatically acquire exclusive jurisdiction over patent validity disputes simply because the patent was granted or validated there, nor should Article 24(4) be interpreted as conferring such authority on those courts.

o  In other words, Article 24(4) of the Brussels I bis Regulation does not apply to non-EU courts and does not grant them exclusive jurisdiction over patent validity assessments.

2) Jurisdiction of the Swedish Court Must Be Assessed Separately Under Article 4(1) of the Brussels I bis Regulation and General Principles of International Law.

o  According to Article 4(1), in cases where an infringement lawsuit is filed based on a patent granted or validated in a non-EU country, the court of the defendant’s domicile retains jurisdiction over the infringement case, even if the defendant raises a validity defense.

o  Thus, the Swedish court maintains jurisdiction over the infringement case, even if the defendant argues that the patent is invalid.

o  However, even if the defendant’s invalidity defense must inevitably be considered in determining infringement, the Swedish court should address it indirectly as part of the infringement analysis rather than ruling on the patent's validity directly.

3) The Scope of Jurisdiction of the Court of the Defendant’s Domicile May Be Restricted by Special Rules Under Article 73 of the Brussels I bis Regulation.

o    For example, under Article 73(1), the courts of a country that is a party to the Lugano Convention have exclusive jurisdiction to determine the validity of patents granted within that country.

o    However, Türkiye is not a party to the Lugano Convention, and based on the records examined by the Swedish court, there is no indication of a relevant agreement between Sweden and Türkiye.

4) Application of National Sovereignty and the Principle of Non-Interference

o    Each country possesses independent authority to grant, register, and validate patents, which is considered an exercise of national sovereignty.

o    According to the principle of non-interference (as referenced in paragraph 71 of the judgment):

-       Only the courts of the country where a patent was granted or validated have the authority to annul that patent through judicial decision.

-       A ruling on patent invalidity by the courts of the granting country may result in modifications to that country’s national patent register.

o  In conclusion, courts in EU Member States do not have the authority to declare a patent invalid if it was granted or validated in a non-EU country. Such determinations fall exclusively within the jurisdiction of the courts of the respective non-EU country.

  

5. 유럽연합 내 특허 소송 관할권에 대한 최근 유럽사법재판소 판결의 의미

The Significance of the Recent ECJ Ruling on Patent Litigation Jurisdiction in the European Union

 

결과적으로, 피고가 EU 회원국 법원의 관할권 (또는 UPC 관할 ) 거주하는 경우, 해당 피고 연결성이 인정된 공동 피고에 대해서는 전통적으로 해당 법원이 관할하는 영토를 벗어난 특허 침해 행위에 대해서도 손해배상 지급 또는 침해 중지 명령을 내릴 있습니다

이는 피고가 특허 무효 항변을 제기한 경우에도 동일하게 적용됩니다. 다만, 특허의 무효성 여부는 해당 특허가 등록된 국가의 전속 관할 사항이므로, 피고의 무효 항변은 침해 여부 판단의 일환으로 간접적으로 심리 있을 , 해당 법원이 직접 특허를 무효화할 권한은 없습니다

또한, 이번 판결의 효력은 당사자 (inter partes) 한정되며, 다른 국가의 특허 등록부에서 해당 외국 특허가 삭제되거나 특허 청구범위가 수정되는 결과를 초래하지 않습니다

따라서, 유럽연합 회원국이 아닌 국가, 예를 들어 미국, 중국, 브라질 등의 피고 기업이 EU 내에 영업소 소재지 등의 주소를 두거나, EU 본사를 다른 피고들과 함께 브뤼셀 I bis 규정 8(1) 연결 요건을 충족하는 경우, 해당 법원은 비록 국가의 특허를 무효화할 수는 없더라도, EU 회원국 (또는 UPC 가입국 ) 물론 미국, 중국, 브라질 3국에서의 침해 행위에 대해서도 책임을 물을 있게 되었습니다.

앞으로 많은 사례를 통해 유럽 특허 소송 관할이 어떻게 확립되어 가는지 지켜볼 필요가 있습니다.

As a result of the ECJ's recent ruling, when a defendant resides within the jurisdiction of an EU Member State court (or within the jurisdiction of the Unified Patent Court, UPC), the court traditionally has the authority to issue damages awards and injunctions against that defendant even for patent infringement acts committed outside the territorial jurisdiction of the court.

This principle applies even when the defendant raises a patent invalidity defense. However, because patent validity falls within the exclusive jurisdiction of the country where the patent was granted, the court hearing the infringement case may only assess the validity issue indirectly as part of its infringement determination but has no authority to declare the patent invalid.

Additionally, the effect of this ruling is limited to inter partes disputes—it does not result in the removal of a foreign patent from the corresponding national register or the modification of its claims.

Furthermore, in cases involving non-EU defendants, such as companies based in the United States, China, or Brazil, EU courts may still assert jurisdiction over infringement claims if:

·        The non-EU company has a place of business or commercial presence within the EU, or

·     The non-EU company is closely connected to another co-defendant domiciled in the EU, thereby satisfying the "closely connected claims" requirement under Article 8(1) of the Brussels I bis Regulation.

Even though the EU court cannot invalidate patents registered in non-EU countries, it may still hold defendants liable for infringement occurring both within EU Member States (or UPC countries) and in non-EU jurisdictions, such as the U.S., China, or Brazil.

As this ruling sets an important precedent, further case law developments will be crucial in shaping the evolving landscape of patent litigation jurisdiction within the European Union.

 

3D 프린팅 대중화 시대의 도래, 특허법 제127조 간접침해 규정을 다시 고민하다

우연히 3D 프린터 Bamboo-Lab A1의 영상을 보게 되었습니다. 너무 매력적이라 하마터면 쿠팡에 들어가 주문 버튼을 누를 뻔 했습니다.


몇 년 전, 특허법 제127조(간접침해) 개정 위원회에 참여했을 때, 제4차 산업혁명의 시대는 획일형 대량생산에서 맞춤형 개인생산의 시대로 전환될 것이며, 데이터 전송을 통한 특허품의 개인적 복제(생산)가 일상화될 가능성이 크다고 언급한 적이 있었습니다.


당시에는 3D 프린팅이 모델링 기법이나 엔지니어링에 대한 배경지식을 요구했기 때문에, 일반 대중이나 법률 전문가들에게 다소 먼 이야기처럼 들렸고, 이에 대한 공감대도 부족했던 것 같습니다.


그러나 이제, 이 제품의 사용기를 보면 생각이 많이 달라질 수 있습니다.


3D 프린터는 그 특허품의 생산만을 위한 전용 장치가 아닙니다. 반면, 플랫폼에서 제공하는 3D 모델링 데이터는 특정 특허품을 제작하는 데 사용되는 전용 데이터입니다. 더욱이, 특허품을 반복적으로 생산하지 않는 한, 개인적 실시가 ‘업(業)’으로 평가되기도 어렵습니다.


이미 많은 분들이 느끼시겠지만, 침해를 회피할 수 있는 법적 허점이 많은 특허법 제127조에 더욱 큰 구멍이 뚫린 셈입니다.


사실상 특허발명의 본질을 그대로 채택한 경우에도, 특허 침해를 ‘업으로 하는 행위’로 한정하거나, 간접침해를 ‘생산’에만 국한하는 나라는 우리나라를 빼면 이제 쉽게 찾아보기 어렵습니다.


특허법이 발명가에게 ‘방법’과 ‘물건’이라는 두 갈래의 물길만 허락하고서, 정작 어느 하나의 물길은 목적지에 도달하기 전에 끊어 놓고, 계속 선택은 자유라고 말하는 것은 너무 무책임한 구조입니다.


우리는 인터넷-first 시대를 지나 모바일-first 시대를 거쳐, 이제 AI-first 시대로 진입하고 있습니다. 더 늦기 전에, 간접침해 규정 개정 논의를 다시 시작해야 할 때입니다.


<특허법> 제127조(침해로 보는 행위) [전문개정 2014.6.11.]

다음 각 호의 구분에 따른 행위를 업으로서 하는 경우에는 특허권 또는 전용실시권을 침해한 것으로 본다.

1. 특허가 물건의 발명인 경우:

그 물건의 생산에만 사용하는 물건을 생산·양도·대여 또는 수입하거나,

그 물건의 양도 또는 대여의 청약을 하는 행위.

2. 특허가 방법의 발명인 경우:

그 방법의 실시에만 사용하는 물건을 생산·양도·대여 또는 수입하거나,

그 물건의 양도 또는 대여의 청약을 하는 행위.


<관련 영상>: Bamboo-Lab A1 사용기


※ 법적 고지 (Legal Notice) ※
본 블로그 포스트는 일반적인 정보 제공을 목적으로 하며, 특정 사안에 대한 법률적 자문을 대체할 수 없습니다. 개별적인 법률 문제에 대해서는 반드시 전문가와 상담하시기 바랍니다.
This blog post is for general informational purposes only and cannot substitute for legal advice on specific matters. Please be sure to consult with a professional regarding individual legal issues.

Sunday, February 9, 2025

누구나 쉽게 이해할 수 있는 강화학습 원리(The Principles of Reinforcement Learning Made Easy)

I. 들어가는 말(Background)

얼마 전 한 변호사님이 강화학습이 무엇인지 질문하셨습니다. 딥러닝 및 인공지능 기술을 수학적 배경 지식이 부족한 분께 설명하는 것은 쉽지 않은 일입니다. 이를 해결하기 위해, "역와쿠이 요시유키. (2020). 엑셀로 배우는 순환 신경망·강화학습 초(超)입문"이라는 책을 참고하여, 수학적 용어를 최소화하고 논리적으로 설명을 시도하였습니다. 이 책을 바탕으로, 수학적 기초 지식이 부족한 분들도 인공지능 컴퓨팅 기술의 기본을 이해할 수 있도록 이 글을 작성하게 되었습니다. 이 과정은 저에게도 관련 내용을 체계적으로 정리할 수 있는 좋은 기회가 되었습니다.

Not long ago, a lawyer asked me what reinforcement learning is. Explaining deep learning and artificial intelligence technologies to someone without a strong mathematical background is no easy task. To address this challenge, I referred to the book Excel de Manabu RNN & Reinforcement Learning: A Super Introductory Guide by Yoshiyuki Yakuwa (2020). Drawing from this resource, I attempted to provide a logical explanation while minimizing the use of mathematical terminology. Based on this book, I wrote this piece to help those with limited mathematical knowledge grasp the fundamentals of AI computing technologies. This process also provided me with an excellent opportunity to systematically organize my understanding of the subject.

Friday, February 7, 2025

특허출원 건수가 기술 혁신리더의 증거인가 (Is the number of patent applications evidence of technological innovation leadership)?

우리나라에서는 미국 내 특허 출원 건수를 근거로 우리가 첨단 기술 주도권을 확보하고 있다며 안도하는 경향이 있습니다.

하지만 미국은 빅테크 기업들의 출원 건수나 순위 자체에는 큰 의미를 두지 않는 것으로 보입니다. 물론 동아시아 국가들이 미국 내 특허 출원에서 높은 점유율을 차지하는 부분은 미국내 미치는 영향 때문에 경계하고는 있지만, 미국은 기술 혁신의 강점을 빅테크가 아닌 분산된 리틀테크(스타트업, 중소기업, 대학, 개인 발명가 등)의 네트워크에서 찾고 있는 것이 인상적입니다.

In Korea, there is a tendency to feel reassured about having secured leadership in advanced technologies by focusing on the number of U.S. patent applications filed by Korean companies. However, the U.S. doesn’t seem to place much emphasis on the application numbers or rankings of big tech companies. While the high share of U.S. patent applications from East Asian countries—such as Korea and China—raises concerns due to its potential impact on the domestic market, it’s notable that the U.S. views its strength in technological innovation as stemming not from big tech but from a decentralized network of “little tech” players (startups, small businesses, universities, and individual inventors).

PatentlyO 블로그를 운영하는 Dennis 교수의 LinkedIn 글과 블로그를 보면, 한국과 중국 등 동아시아 빅테크 기업들이 미국 빅테크보다 높은 특허 출원 순위를 기록하고 있는 것이 사실이지만, 이것이 미국이 기술 혁신의 주도권을 상실했다는 증거는 아니라는 점을 강조합니다.

빅테크 기업들의 양적 순위는 낮을 수 있지만, 미국의 혁신은 대기업이 아닌 리틀테크의 분산된 생태계에서 활발히 이루어지고 있다고 합니다. 저도 그 주장에 동의합니다.

According to Dennis Crouch, the professor behind the PatentlyO blog, East Asian big tech companies like those from Korea and China rank higher than U.S. big tech companies in terms of patent filings. However, he emphasizes that this is not evidence of the U.S. losing its leadership in innovation. While the quantitative rankings of U.S. big tech companies may fall behind, the core of U.S. innovation is driven by the vibrant, decentralized ecosystem of small tech players. I fully agree with this perspective.

따라서 개별 기업의 순위를 완전히 무시할 필요는 없지만, 미국의 전반적인 기술 혁신력은 여전히 강력하다는 메시지를 전달하고 있습니다.

결국, 이는 과거 빅테크 중심으로 왜곡되었던 특허 정책을 리틀테크(발명자) 중심으로 전환해, 특허제도가 본래 지향해야 할 균형과 혁신 생태계를 회복해야 한다는 함의를 담고 있습니다.

While individual company rankings shouldn’t be completely ignored, the broader message is that the overall U.S. innovation capacity remains robust. Ultimately, this implies that the patent system, which has become skewed toward big tech in the past, should shift back toward supporting little tech players and restore its role in fostering a balanced, innovation-driven ecosystem.

저는 미국의 강점을 더 강화하기 위해서는 스타트업과 중소기업에 대한 투자 확대와 빅 테크 기업과 리틀 테크 기업 간의 협력을 장려하는 정책이 매우 효과적일 것이며, 또한, 특허 출원 건수나 순위에 초점을 맞추기보다는 전 세계적으로 혁신의 질과 상업적 성공을 더 잘 측정할 수 있는 새로운 지표를 개발해야 한다고 댓글을 달았습니다.

In my response, I noted that to further enhance the U.S.’s strengths, it would be highly effective to increase investments in startups and small businesses and encourage collaboration between big tech and little tech companies. Additionally, I suggested that developing new metrics to better assess the quality of innovation and commercial success globally, rather than focusing solely on patent filing numbers or rankings, would be crucial.


Sunday, February 2, 2025

DeepSeek model V3와 R1의 모든 것

DeepSeek: V2에서 R1까지, 실용적 AI 혁신의 여정

DeepSeek은 V2에서 V3, R1-Zero를 거쳐 R1 개발에 성공하기까지 수많은 도전과 변화를 겪어왔습니다. 지난해 말 V3가 발표될 당시만 해도 "아직 갈 길이 멀다"는 회의적인 시각이 많았지만, 올해 1월 R1이 발표되면서 AI 업계뿐 아니라 다양한 산업군에서 DeepSeek에 대한 관심이 급격히 높아졌습니다. V3 기술보고서와 R1 논문을 읽어보면, 기존 연구들을 기반으로 효과적인 최적화를 수행하면서도 실제 AI 시스템에 적용할 수 있는 방법론을 정교하게 발전시켜 왔음을 알 수 있습니다.
V3와 R1의 관계
DeepSeek을 둘러싼 평가와 현실을 먼저 짚어봅니다. DeepSeek의 기술에 대해 "새로운 패러다임을 제시한 것은 아니다", "시장 기술 대비 월등한 성능을 보이는 것도 아니다"라는 평가도 존재합니다. 하지만 DeepSeek이 거인의 어깨 위에 올라선 방식은 매우 실용적이며, AI 연구에서 중요한 발전을 이루었다는 점을 간과해서는 안 됩니다. 논문과 소스코드를 보면, DeepSeek의 접근 방식은 단순해 보일 수 있습니다 — "별것 아닌 것 같은데?"라는 안도감과, "이 아이디어를 이렇게 적용하면 효과가 있나?"라는 의구심이 동시에 듭니다. 하지만 실제로 오픈소스를 분석해 보면 그러한 안도감과 의구심은 곧 경외감으로 바뀝니다. DeepSeek이 공개한 코드는 놀라울 정도로 단순하게 구현되었습니다. 복잡한 이론은 기존 연구를 활용하면서도 Hugging Face 모듈을 적극적으로 활용했고, 논문에서 다루지 않았지만 실용적인 최적화 아이디어를 적용하여 더욱 효과적인 모델을 구축했습니다. AI 연구에서 "자금이 부족해서 할 수 없다", "전폭적인 지원이 부족하다", "인재가 없다"는 변명이 더 이상 통하지 않을지도 모릅니다. DeepSeek의 접근 방식은 누구나 생각해볼 수 있는 아이디어들이지만, 결국 그 아이디어들을 제대로 구현하고 실용화한 것이 차별점입니다.

기술적으로도 두 모델은 서로 얽혀 있습니다. DeepSeek-V3는 671억 개의 총 매개변수(MoE)로 구성되며, 지도 미세조정(SFT) 및 강화학습(RL)이 적용된 DeepSeek-R1로부터 지식 증류(Knowledge Distillation)를 받아 사후학습(post-training)을 대신할 수 있습니다. 즉 R1은 DeepSeek-V3-Base를 기반으로 강화학습을 적용해 만들어졌고, 그렇게 얻어진 R1의 추론 능력이 다시 V3의 최종 사후학습 단계로 증류되어 들어가는 순환 구조입니다. 다만 V3 기술보고서가 참조하는 R1은 2025년 1월 정식 공개된 최종 R1이 아니라 그 이전 단계의 R1 계열 선행 버전이라는 점은 유의할 부분입니다.

1. DeepSeek-V3

DeepSeek은 최근 DeepSeek-V3 및 DeepSeek-R1의 두 가지 고급 모델을 오픈소스로 공개함으로써 AI 커뮤니티에 상당한 기여를 했습니다. 이 모델들은 비용 효율적으로 고성능을 제공하도록 설계되어 OpenAI의 주요 모델에 대한 강력한 대안으로 떠올랐습니다.

DeepSeek-V3는 671억 개의 총 매개변수(parameter)로 구성된 MoE(Mixture-of-Experts) 아키텍처를 사용하며 토큰당 370억 개가 활성화됩니다. 14.8조 개의 다양하고 고품질 토큰으로 사전학습(pre-training)되었습니다. 이 모델을 학생모델(student model)로 하여 DeepSeek-R1의 지식을 전수받으면 질문 응답, 코드 생성, 수학 문제 해결, 법률 리서치 등 다양한 작업에도 최적화할 수 있습니다. 특히 DeepSeek-V3는 H800 칩에서 2.788백만 GPU 시간으로 학습을 완료하여 학습 효율성을 강조했습니다. 평가에 따르면 다른 오픈소스 모델보다 성능이 우수하고 주요 폐쇄소스(Closed Source) 모델의 기능과 일치합니다.

가. 아키텍처(Architecture)

1) Transformer with MHLA and MoE

DeepSeek-V3 모델은 Transformer 프레임워크를 사용합니다. 전통적인 Transformer(Vaswani et al., 2017)와 다른 점이 있다면, 효율적인 추론을 위해 Attention 블럭에 입력 토큰 간의 관계를 직접 계산하는 Multi-Head Self-Attention(MHSA) 방식 대신 다중 헤드 잠재 주의 메커니즘(Multi-head Latent Attention, MHLA)(DeepSeek-AI, 2024c) 아키텍처를 적용하였고, 경제적인 학습을 위해 모든 노드를 활성화하는 밀집(dense) FFN 대신 입력마다 다른 subset의 뉴런을 활성화하는 Mixture-of-Experts(MoE)(Dai et al., 2024) 아키텍처를 적용했다는 것입니다.

이러한 특징은 DeepSeek의 종전 모델인 V2와 같습니다. V2보다 좋아진 점이 있다면, DeepSeekMoE에서 부가 손실(auxiliary-loss)을 사용하지 않는 로드 밸런싱(load balancing) 전략(Wang et al., 2024a)을 추가 도입했다는 점입니다. 이는 로드 밸런스를 보장하려는 과정에서 발생할 수 있는 성능 저하 문제를 완화하기 위한 것입니다.

DeepSeek-V3의 MLA와 DeepSeekMoE 기본 아키텍처 도식
DeepSeek-V3 기본 아키텍처: MLA + DeepSeekMoE

2) 기존 Attention과 차이점: Multi-Head Latent Attention (MHLA)

Attention 블럭에 채용된 MHLA 아키텍처는 기존 attention이 입력 토큰 간의 관계를 직접 계산하는 MHSA 방식에서 발전하여, 잠재(latent) 공간에서 추가적인 추론을 수행하는 메커니즘을 추가합니다. 이 방식은 일반적인 Self-Attention보다 계산량이 줄어들어(O(N²) → O(N log N) 수준으로 최적화 가능), 대용량 데이터에서 더 안정적으로 학습이 가능해지고 보다 강력한 문맥 이해 및 추론 능력을 확보하게 합니다.

3) 기존 FFN과의 차이점: DeepSeekMoE(Mixture of Experts)

가) 기존 MoE 모델과의 차이점
FFN 블럭은 종래 MoE(Mixture of Experts)에서 개량된 DeepSeekMoE 아키텍처를 기반으로 작동합니다. 기존 MoE보다 더 정밀한 전문가(Finer-grained Experts)를 사용하며, 일부 전문가를 공유(Shared Experts)로 설정하여 일관된 성능을 유지합니다. 입력에 따른 전문가 선택(gating)은 토큰과 전문가 간 연관도를 시그모이드(Sigmoid) 함수로 계산하고, 정규화를 적용하여 최적의 전문가를 선택합니다. 이러한 개선 덕분에 추론 및 학습 효율성이 증가하고, MoE 아키텍처의 단점(불균형 로드, 성능 저하)을 완화합니다.

나) Auxiliary-Loss-Free Load Balancing (보조 손실 없는 부하 균형 조정)
기존 MoE 모델은 전문가 간 부하 불균형을 해결하기 위해 보조 손실(Auxiliary Loss)을 사용했지만, DeepSeek-V3는 보조 손실 없이도 부하를 동적으로 균형 맞추는 기법을 도입했습니다. 각 전문가에 편향(Bias) 값을 추가하여 토큰-전문가 매칭을 조정하고, 과부하된 전문가의 편향을 줄이며 저부하 전문가의 편향을 증가시켜 자동으로 부하를 균형 있게 조정합니다.

다) Sequence-Wise Auxiliary Loss (시퀀스 단위 부하 균형)
부하 균형을 더욱 정교하게 유지하기 위해, 각 시퀀스 내에서 전문가 간의 부하가 특정 전문가에 집중되지 않도록 보완적인 부하 균형 손실을 적용했습니다. 기존 Auxiliary Loss처럼 성능을 저하시키지 않으면서도, 필요한 범위 내에서 최소한의 손실만 추가하여 균형을 조절합니다.

라) 노드 제한 라우팅 (Node-Limited Routing)
토큰이 특정 전문가에게 과도하게 몰리는 것을 방지하고, 통신 비용을 줄이기 위해 토큰을 제한된 개수의 노드에만 할당하는 기법을 도입했습니다. 이를 통해 연산-통신 병목을 최소화하여 모델 학습 속도를 최적화할 수 있습니다.

마) 토큰 드롭 방지 (No Token-Dropping)
기존 MoE 모델에서는 부하 불균형 문제로 인해 일부 토큰이 학습 과정에서 제거되었지만, DeepSeek-V3는 모든 토큰을 유지하여 정보 손실 없이 안정적인 학습이 가능하도록 설계되었습니다. 추론(Inference) 과정에서도 토큰 드롭 없이 균형을 유지하는 배포 전략을 적용하여 성능 저하를 방지합니다.

4) 다중 토큰 예측 (Multi-Token Prediction, MTP)

DeepSeek-V3는 Gloeckle et al.(2024) 연구에 영감을 받아, 기존의 단일 토큰 예측을 넘어 여러 개의 토큰을 동시에 예측하여(MTP 기법), 학습 신호 밀도를 높이고 데이터 효율성을 극대화하는 방식으로 설계되었습니다.

각 MTP 모듈은 임베딩 레이어(Embedding Layer), 선형 투영 행렬(Projection Matrix), Transformer 블록, 출력 헤드(Output Head)로 구성되어 있으며, 이전 단계에서 출력 헤드로 입력되는 값이 다음 단계의 선형 투영 행렬로 들어가 각 모듈은 Transformer 블록과 출력 헤드의 값을 공유하며, 모든 예측이 인과적 체인(Causal Chain)을 유지하도록 설계되었습니다.

다중 토큰 예측(MTP) 모듈의 순차적 구조 도식
다중 토큰 예측(MTP) 모듈 구조

이러한 공유 구조 덕분에 모델의 학습 비용을 증가시키지 않으면서도 추가적인 예측 능력을 확보할 수 있으며, 순차적 예측 방식을 채택하여 인과적 체인을 유지함으로써 모델이 보다 자연스럽고 논리적인 문맥을 이해할 수 있도록 돕습니다. 학습 과정에서는 교차 엔트로피 손실(Cross-Entropy Loss)을 활용하여 최적화함으로써 모델의 전반적인 성능을 향상시키며, 추론 과정에서는 Speculative Decoding을 적용하여 생성 속도를 최적화하는 실용적인 기법으로 활용될 수 있습니다.

나. DeepSeek-V3의 사전 학습 (Pre-training)

1) FP8 Training

DeepSeek-V3에서 새로 채택한 FP8 Training은 8비트 부동소수점(FP8) 정밀도를 활용하여 계산 효율성을 극대화하는 학습 기법입니다. 학습에서 FP32 → BF16 → FP8 순으로 정밀도를 낮춰가며 연산 최적화와 메모리 절약을 추구하였습니다.

FP32-BF16-FP8 정밀도 변환 흐름도
FP8 혼합 정밀도 학습 흐름도
  • 입력 처리: BF16(BFloat16) → FP8 변환
  • Forward Propagation(FP32): FP8 → FP32 연산 → BF16 변환 후 출력
  • Backward Propagation(FP32): FP8 Gradient → FP32 계산 → FP8 저장
  • Weight Update(FP32): FP32 Master Weight로 업데이트 → FP8 저장

FP8 Training은 특히 곱셈 연산을 단순화하여 GPU 활용도를 극대화하면서도, 기존 FP32 및 BF16 대비 메모리 사용량과 계산 비용을 줄일 수 있습니다. 이를 통해 동일한 하드웨어에서도 더 큰 모델을 학습할 수 있는 확장성이 확보됩니다. 또한 FP8 사용 시 발생할 수 있는 활성화(Activation), 가중치(Weight), 그래디언트(Gradient) 값의 이상치(Outlier) 문제를 해결하기 위해 정밀한 양자화 전략을 적용하였습니다. FP32 Master Weight를 활용한 혼합 정밀도(Mixed Precision) 학습 방식을 채택하여 FP32 수준의 정밀도를 유지하면서도 FP8의 연산 최적화를 극대화하였습니다. MoE 모델 학습에서는 FP8 캐싱을 활용하여 메모리 및 통신 비용을 절감하여 대규모 모델 학습 환경에서도 효율적인 운영이 가능합니다.

2) 데이터 구축 (Data Construction)

DeepSeek-V3는 DeepSeek-V2 대비 사전학습 데이터의 최적화를 위해 수학 및 프로그래밍 샘플의 비율을 증가시키고, 다국어 데이터 범위를 영어와 중국어를 넘어 확장하였습니다. 데이터 처리 파이프라인을 개선하여 말뭉치의 중복을 최소화하면서도 다양성을 유지하였습니다. 데이터의 일관성을 보장하기 위해 문서 패킹(Document Packing) 기법을 도입하였으며, 학습 과정에서는 교차 샘플 어텐션 마스킹(Cross-Sample Attention Masking)은 적용하지 않음으로써 처리 속도와 효율성을 극대화하였습니다. 최종적으로 DeepSeek-V3는 14.8조(14.8T)개의 고품질 및 다양한 토큰을 포함하는 학습 데이터셋을 사용하였습니다.

가) Fill-in-Middle(FIM) 전략 적용
DeepSeek-V3의 사전학습 과정에 적용한 Fill-in-Middle(FIM) 전략은 이미 DeepSeekCoder-V2의 학습 과정에서 다음 토큰 예측 성능을 저하시키지 않으면서, 문맥을 기반으로 중간 텍스트를 정확히 예측하는 데 효과적이라는 점이 관찰되었습니다. Prefix-Suffix-Middle(PSM) 프레임워크를 활용하여 데이터 구조(<|fim_begin|>f_pre<|fim_hole|>f_suf<|fim_end|>f_middle<|eos_token|>)를 구성하였으며, FIM 전략은 PSM 프레임워크와 일관되게 10%(0.1)의 확률로 적용되었습니다.

나) 토크나이저(Tokenization) 최적화
DeepSeek-V3는 Byte-level BPE(Byte Pair Encoding) 방식을 활용하는 128K 크기의 확장된 어휘(Vocabulary)를 도입하였습니다. 다국어 압축 효율성을 최적화하기 위해 프리토크나이저 및 학습 데이터를 조정하였고, Few-shot 평가 시 토큰 경계 편향(Token Boundary Bias) 문제를 완화하기 위해 훈련 과정에서 결합된 토큰 데이터의 일부를 랜덤하게 분할하였습니다.

3) 하이퍼파라미터 (Hyper-Parameters)

가) 모델 하이퍼파라미터

  • Transformer 구조: 총 61개 레이어(Layers), 히든 차원(Hidden Dimension) = 7168, 모든 학습 가능한 파라미터는 표준편차 0.006으로 랜덤 초기화
  • MLA 설정: 어텐션 헤드 개수 n_h = 128, 헤드당 차원 d_h = 128; KV 압축 차원 d_c = 512, 쿼리 압축 차원 d'_c = 1536; 디커플드(Decoupled) 쿼리·키의 헤드당 차원 = 64
  • MoE 적용: 특정 레이어의 FFN을 MoE 레이어로 대체, 각 MoE 레이어는 1개의 공유 전문가(Shared Expert)와 256개의 라우팅 전문가(Routed Experts)로 구성, 각 전문가의 총 활성화 가능한 유닛 수 = 2048, 각 토큰은 8개의 전문가를 활성화
  • MTP 설정: MTP 깊이(D) = 1, 각 토큰이 다음 토큰 외 하나의 추가 토큰을 예측
  • 모델 크기: 총 671B(6,710억) 파라미터 보유, 토큰당 활성화 파라미터 수 37B(370억)

나) 학습 하이퍼파라미터

  • 최적화 기법: AdamW 옵티마이저, β1=0.9, β2=0.95, Weight Decay=0.1
  • 시퀀스 길이·데이터: 최대 시퀀스 길이 4K(4096 토큰), 사전학습 데이터 14.8T 토큰
  • 학습률 스케줄링: 초기 2K 스텝 동안 0에서 2.2×10⁻⁴까지 선형 증가 후 10K 토큰 동안 유지 → 이후 43T 토큰 동안 코사인 감소(Cosine Decay)로 2.2×10⁻⁵까지 감소 → 최종 500B 토큰 동안 2.2×10⁻⁵에서 7.3×10⁻⁶로 감소
  • 기타 학습 설정: 그래디언트 클리핑 1.0; 배치 크기는 초기 469B 토큰 동안 3072→15360으로 증가 후 15360 유지
  • 병렬 학습: 파이프라인 병렬화로 8개 노드에 분산된 64개 GPU가 균등하게 작업 수행
  • 부하 균형 설정: 토큰당 최대 전송 노드 수 M=4; 부하 균형 속도는 초기 14.3T 토큰 동안 0.001, 마지막 500B에서 0.0002로 점진적 감소; 균형 손실 한계 0.0001; MTP 손실 가중치는 초기 0.2에서 이후 0.1로 감소(4.8T 토큰 기준)

4) 장문 컨텍스트 확장 (Long Context Extension)

DeepSeek-V3는 DeepSeek-V2에서처럼 장문 컨텍스트 처리 능력을 확장하기 위해 YaRN(Yet another RoPE eXtension) 기법을 활용하여 사전학습 이후 두 단계의 추가 학습을 진행하였습니다(1단계 4K→32K, 2단계 32K→128K). 배치 크기와 학습률을 조정하여 안정적인 학습을 진행하였고, 이 두 단계의 학습을 통해 DeepSeek-V3는 최대 128K 길이의 입력을 처리할 수 있는 능력을 획득하였습니다. Supervised Fine-Tuning 이후 "Needle In A Haystack(NIAH)" 테스트에서 우수한 성능을 기록하여, 컨텍스트 윈도우가 최대 128K까지 확장된 상태에서도 강력한 성능을 유지함을 보여주었습니다.

다. DeepSeek-V3의 사후 학습 (Post-training)

1) 지도 학습 미세 조정 (Supervised Fine-Tuning, SFT)

DeepSeek-V3는 SFT와 RL을 효과적으로 결합하여 모델의 추론 능력과 일반 작업 성능을 최적화하였습니다. 다양한 도메인에서 150만 개(1.5M)의 인스턴스를 포함하는 SFT 데이터셋을 구성하였습니다. 추론 데이터는 DeepSeek-R1을 기반으로 생성되었으며, 강화학습과 Rejection Sampling을 활용하여 데이터의 정확도를 유지하면서도 응답 형식을 최적화하였습니다. 특히 수학, 프로그래밍, 논리 퍼즐 등에서 R1의 강점을 살리면서도 과도한 사고(Overthinking)와 불필요한 길이를 줄이는 방향으로 조정되었습니다.

비추론 데이터는 DeepSeek-V2.5를 활용하여 생성한 후 인간 주석자의 검토를 거쳐 품질을 보장하였습니다. 창의적 글쓰기, 역할극(Role-Play), 일반 질의응답(QA) 작업에서도 신뢰성 있고 유용한 출력을 제공할 수 있도록 데이터를 정제하였습니다. 학습 과정에서 코사인 감쇠 학습률과 샘플 마스킹 전략을 적용하여 학습 안정성을 강화하였습니다.

2) 강화 학습 (Reinforcement Learning, RL)

DeepSeek-V3는 정확한 보상 모델(Reward Model)과 효율적인 GRPO(Group Relative Policy Optimization)를 활용하여 강화학습을 진행하였습니다. 규칙 기반 보상 모델과 모델 기반 보상 모델을 병행하여 활용하고, Critic 모델 없이 그룹 점수를 기반으로 최적화하는 GRPO 기법을 적용하였습니다.

정확한 정답이 존재하는 문제(수학, 프로그래밍)에서는 규칙 기반 보상 모델(Rule-Based RM)을 사용하여 응답의 정확성을 검증했고, LeetCode 문제의 경우 컴파일러를 활용한 테스트 케이스 검증으로 자동 판단하였습니다. 반면 정답이 없는 창의적 글쓰기 같은 작업에서는 모델 기반 보상 모델(Model-Based RM)을 적용했는데, SFT 체크포인트 기반 보상 모델이 단순 점수만이 아니라 보상 값이 도출된 과정(Chain-of-Thought)까지 평가에 포함해 보상 해킹(Reward Hacking) 가능성을 줄였습니다.

결과적으로 DeepSeek-V3는 정교한 보상 모델과 GRPO 기반의 효율적인 학습 최적화 기법을 결합하여, RL을 통해 높은 품질의 응답을 생성하는 강력한 언어 모델로 발전하였습니다.

2. DeepSeek-R1

OpenAI의 o1 모델은 Inference-Time Scaling(추론 시간 확장) 기법을 도입하여 여러 중간 단계를 거쳐 깊이 있는 논리를 형성함으로써 수학, 코딩, 과학적 문제 해결에서 큰 성능 향상을 이루었습니다. 그러나 이 모델조차도 AGI에 도달하기까지는 여전히 한계가 있습니다. 많은 연구자들이 과정 기반 보상 모델, 강화학습 알고리즘, 탐색 기법(Monte Carlo Tree Search, Beam Search) 등을 모색해 왔지만, 일반적인 추론 성능을 완벽히 구현하는 데는 아직 도달하지 못했습니다.

본 연구("DeepSeek-R1: Incentivizing Reasoning Capability in Large Language Models via Reinforcement Learning")에서는 지도학습 없이 순수 강화학습(RL)만을 사용하여 AI의 추론 능력을 향상하려는 첫 시도를 했습니다. DeepSeek-V3-Base 모델과 GRPO 알고리즘을 적용하여 수천 번의 강화학습을 거친 결과, DeepSeek-R1-Zero는 강력한 추론 능력을 획득하였으며, 이 연구는 AI가 사람이 제공하는 정답 데이터에 의존하지 않고도 강화학습을 통해 추론 능력을 크게 향상할 수 있음을 입증하였습니다. 다만 순수 강화학습만으로는 낮은 문장 가독성과 언어 혼합 문제가 있었으며, 이를 해결하기 위해 소량의 지도학습 데이터와 다단계 학습을 추가한 DeepSeek-R1 모델이 개발되어 더 정교하고 안정적인 성능을 제공하게 되었습니다.

본 연구의 활용 시사점
  • 이미 완성된 다양한 모델에 대규모 강화학습을 사후학습으로 적용하면 추론 능력이 급격히 향상될 수 있음을 입증했습니다. 기존 모델의 성능을 향상시키는 데 이 접근법을 적용하면 더 높은 성능을 기대할 수 있습니다.
  • 고성능 모델의 지식을 소형 모델에 증류하여 주입하면, 강화학습 같은 사후학습 없이도 소형 모델이 고성능 모델에 근접한 성능을 낼 수 있음을 보여주었습니다.
  • 기술 자료와 소스를 공개해 AI 오픈 커뮤니티로부터 지속적으로 피드백을 받는 개방형 혁신 연구 방식의 유용성을 다시 한번 확인시켜 주었습니다.

가. DeepSeek-R1 개발 과정 및 특징

1) DeepSeek-R1-Zero (중간 모델): 순수 강화학습(RL) 기반 추론 모델

  • 순수 강화학습(Pure RL)만을 이용하여 언어 모델의 추론 능력을 향상시킨 최초 모델
  • Supervised Data 없이 RL만으로 reasoning 능력을 습득하는 데 성공
  • DeepSeek-V3-Base를 기반 모델로 사용, GRPO 알고리즘을 활용하여 RL을 적용
  • 학습 과정에서 자연스럽게 강력한 reasoning 능력이 나타남
  • 다만 가독성 저하, 언어 혼합(Language Mixing) 등의 문제가 발생
DeepSeek-R1 개발의 동기
R1-Zero가 RL만으로 강력한 추론 능력을 획득하는 데 성공하자, DeepSeek 연구팀은 두 가지 핵심 질문을 바탕으로 R1 개발에 착수했습니다. 첫째, 소량의 고품질 SFT 데이터를 Cold Start에 도입하면 추론 성능이 더 향상되거나 수렴 속도가 빨라질 수 있는가? 둘째, 명확하고 일관된 Chain of Thought(CoT)를 생성하면서도 전반적인 일반화 성능까지 겸비한 사용자 친화적 모델을 어떻게 설계할 수 있는가? 이를 위해 지도학습과 강화학습이 교번하는 다단계 학습 파이프라인을 설계하여 R1을 개발하였습니다.

2) DeepSeek-R1 (최종 모델)

  • 다단계 학습 파이프라인(Multi-Stage Training Pipeline)과 Cold-Start 데이터를 활용한 SFT로 R1-Zero의 문제를 해결
  • 추론 성능에서 OpenAI의 o1-1217 모델과 대등한 성능을 기록
  • 보다 일관된 문장 구조와 명확한 응답을 제공하도록 최적화

나. 오픈소스 버전

  • DeepSeek-R1에서 학습된 고급 추론 패턴을 다양한 크기의 밀집 모델(Dense Models)로 증류(Distillation)
  • 1.5B에서 70B까지의 다양한 모델 크기로 제공, 개발자들이 확장성과 성능을 고려하여 선택 가능
  • 강화학습을 활용한 최첨단 추론 능력을 갖춘 오픈소스 모델을 누구나 활용할 수 있도록 공개

다. DeepSeek-R1의 다단계 학습 과정 (Multi-Stage Training Pipeline)

DeepSeek-R1은 강화학습을 활용한 사후학습, 다단계 학습 파이프라인 구축, 그리고 지식 증류를 통해 대규모 언어 모델의 추론 능력을 극대화하는 데 기여하였습니다.

DeepSeek-R1의 다단계 학습 파이프라인 전체 구조
R1의 다단계 학습 파이프라인 개요

1) Cold-Start 데이터를 지도학습된 데이터로 사용

수천 개의 기본 데이터(cold-start data)를 수집하여 DeepSeek-V3-Base 모델을 초기 미세조정(fine-tuning)합니다. 목적은 RL 전에 모델의 기초 성능을 보완하여 가독성을 향상시키는 것입니다.

2) DeepSeek-R1-Zero 방식의 RL 적용 (추론 능력 극대화)

기존 DeepSeek-R1-Zero와 동일하게 추론 중심 RL(reasoning-oriented RL)을 적용하여 논리적 사고 능력을 강화합니다.

3) 새로운 지도학습 데이터(SFT) 생성 및 추가 학습

RL 과정이 수렴할(converge) 때쯤, 모델 성능을 추가로 개선하기 위해 새로운 SFT 데이터를 생성합니다. RL 체크포인트(강화학습 과정에서 저장된 특정 시점의 모델 상태)에서 거부 샘플링(rejection sampling)을 통해 더 나은 데이터를 추출합니다 — 즉 모델이 생성한 여러 출력 중 질 좋은 데이터만 선별합니다. DeepSeek-V3의 SFT 데이터도 함께 사용하며, 적용 분야는 글쓰기(Writing), 사실 기반 QA(Factual QA), 자기 인식(Self-Cognition)입니다. 이렇게 만든 SFT 데이터로 DeepSeek-V3-Base를 다시 훈련합니다.

4) 최종 RL 적용 (모든 시나리오에서 최적화)

모든 유형의 프롬프트를 고려하여 RL을 다시 적용, 최적의 모델로 개선합니다.

라. Distillation(지식 증류)

DeepSeek-R1에서 발견된 고급 추론 패턴을 더 작은 밀집 모델(Dense Models)에 전이(Distill)하여, 효율성과 성능을 동시에 향상시키는 지식 증류 기법을 적용하였습니다. 이 과정에서 Qwen2.5-32B 및 Llama 시리즈를 기반으로 1.5B부터 70B까지 다양한 크기의 모델을 성공적으로 증류하였습니다. 특히 Distilled 14B 모델이 QwQ-32B-Preview를 능가하는 성능을 기록하면서, 소형 모델에서도 높은 성능을 유지할 수 있음을 입증하였습니다.

마. 개발 단계 설명

기존 강화학습(RL) 모델은 SFT 후 RL을 적용하는 방식이 일반적이지만, DeepSeek-R1은 새로운 접근법을 채택하여 SFT 없이 순수 강화학습만으로 DeepSeek-R1-Zero를 먼저 학습하였습니다. R1-Zero는 GRPO를 적용해 자기 검증(Self-Verification), 반성(Reflection), 긴 CoT 생성과 같은 고급 추론 능력을 학습하도록 설계되었습니다. 이후 소량의 고품질 SFT 데이터를 활용해 R1-Zero를 미세조정한 뒤, 추가 RL을 통해 성능을 최적화하였습니다. 마지막으로 R1에서 학습한 고급 추론 패턴을 지식 증류를 통해 소형 모델에 전이하여 소형 모델에서도 효율적이고 강력한 성능을 발휘할 수 있도록 설계하였습니다.

R1-Zero에서 R1로 이어지는 개발 단계별 흐름도
R1-Zero → Cold Start → RL → SFT → 2차 RL로 이어지는 개발 단계

1) DeepSeek-R1-Zero: 순수 강화학습

가) Group Relative Policy Optimization (GRPO) 알고리즘 활용
GRPO는 기존 강화학습(RL) 방법에서 사용하는 Critic 모델(평가 모델)을 제거하고, 대신 그룹 내 보상 값들로부터 기준점을 추정하여 정책을 최적화하는 방식입니다. 이를 통해 학습 비용을 줄이면서도 효과적으로 모델을 개선할 수 있습니다.

GRPO(Group Relative Policy Optimization) 알고리즘 구조도
GRPO 알고리즘 구조: Critic 모델 없이 그룹 내 상대 평가로 정책을 업데이트

나) GRPO 알고리즘의 주요 개념

GRPO의 목적함수(손실함수)는 다음 순서로 작동합니다: ① 모델이 데이터셋 P(Q)에서 질문 q를 가져오는 질문 샘플링, ② 이전 정책 모델 π_θold에서 질문 q에 대한 여러 응답(출력 그룹) {o1,o2,…,oG}을 샘플링하는 이전 정책에서 응답 샘플링, ③ 샘플링된 응답을 바탕으로 새로운 정책 π_θ를 최적화하는 새로운 정책 업데이트입니다.

GRPO 목적함수: 질문 샘플링과 이전 정책 기반 응답 샘플링
GRPO의 질문·응답 샘플링 및 정책 업데이트 구조

목적함수는 세 가지 주요 항으로 구성됩니다.

GRPO 목적함수의 세 가지 구성 요소
GRPO 목적함수 구성: 정책 비율항 · 클리핑항 · KL 발산 제약항
  • 정책 비율항(Policy Ratio Term): 새로운 정책 π_θ가 기존 정책 π_θold보다 얼마나 변화했는지를 나타냅니다. 비율이 1보다 크면 새로운 정책이 해당 응답을 더 선호한다는 의미이고, 1보다 작으면 덜 선호한다는 의미입니다. Advantage 값(A_i)과 곱해져서 보상이 높은 응답일수록 정책을 더 크게 업데이트하도록 유도합니다.
  • 정책 변경 제한항(Clipping Term): 정책이 한 번에 너무 크게 변경되지 않도록 방지합니다. 비율이 (1−ϵ, 1+ϵ) 범위를 벗어나면 잘라서(clip) 급격한 정책 변화를 막고, RL의 흔한 문제인 과도한 최적화(overfitting)를 방지합니다.
  • KL 발산 제약항(KL Divergence Regularization Term): 새로운 정책 π_θ와 기준 정책 π_ref 사이의 차이를 측정하는 KL 발산으로, 정책이 기준 정책과 너무 달라지지 않도록 패널티를 부여합니다. β는 KL 발산의 영향력을 조절하는 하이퍼파라미터입니다.

Advantage(A_i) 계산 (보상 기준점 조정): 기존 RL에서는 Critic 모델이 Advantage 값을 계산하지만, GRPO에서는 그룹 내 응답들의 보상 값을 사용하여 계산합니다. 각 출력의 보상을 평균과 비교한 후 표준편차로 정규화하여 상대적으로 얼마나 좋은지를 평가합니다 — A_i가 양수면 그룹 평균보다 좋은 결과, 음수면 나쁜 결과를 의미합니다. 이를 통해 Critic 모델 없이도 학습이 가능합니다.

Advantage(A_i) 계산 수식
Advantage(A_i) 계산: 그룹 내 보상을 정규화해 산출

GRPO는 그룹 내 여러 샘플을 비교해 상대적으로 좋은 응답을 학습하도록 유도하며, 개별 출력을 최적화하는 대신 전체적으로 균형 잡힌 학습이 가능하도록 설계됩니다. 따라서 기존 RL 방식보다 학습 비용을 줄이면서도 모델을 효과적으로 최적화하는 방법으로 평가되고 있습니다.

[참고] GRPO는 왜 개별 보상을 직접 사용하지 않으면서도 학습 신호로 보상을 활용하는가?
GRPO는 개별 출력에 직접 보상을 적용하지 않지만, 그룹 내 상대적인 보상 차이를 학습 신호로 활용합니다. 각 출력 o_i의 보상 r_i는 계산되지만, 이 값을 그대로 정책 최적화에 쓰는 대신 그룹 내 상대적 차이를 통해 Advantage(A_i)를 산출해 정책 방향을 결정합니다. 즉 GRPO는 기존 RL처럼 보상을 직접 적용하지는 않지만, 학습 방향을 정하는 데 보상 정보를 활용하는 방식이라고 이해하면 됩니다.

2) 보상 모델링 (Reward Modeling)

강화학습에서 보상(Reward)은 모델이 어떤 방향으로 최적화될지를 결정하는 핵심 신호입니다. DeepSeek-R1-Zero를 학습시키기 위해 규칙 기반 보상 시스템(rule-based reward system)을 채택하였으며, 정확도 보상(Accuracy Rewards)과 형식 보상(Format Rewards)의 두 유형으로 구성됩니다.

정확도 보상 (Accuracy Rewards): 모델이 정확한 정답을 생성하도록 유도합니다. 수학 문제처럼 결과가 정해진 결정적(Deterministic) 문제는 정답을 박스(☐) 안에 작성하도록 하여 규칙 기반 검증이 가능하게 했고, LeetCode 코딩 문제는 컴파일러와 미리 정의된 테스트 케이스로 코드 실행 결과를 평가해 통과 여부에 따라 보상을 부여했습니다.

형식 보상 (Format Rewards): 모델이 논리적 사고 과정을 명확하게 표현하도록 유도합니다. 추론 과정(Reasoning Steps)은 <think> 태그 안에 정리하고, 최종 응답(Answer)은 <answer> 태그로 표시하도록 학습시켜, 사람이 모델의 논리 과정을 쉽게 이해하고 디버깅할 수 있게 했습니다.

정확도 보상과 형식 보상(think/answer 태그) 예시
규칙 기반 보상 시스템의 정확도·형식 보상 예시

DeepSeek-R1-Zero가 신경망 기반 보상 모델(Neural Reward Model) 대신 규칙 기반 시스템을 택한 이유는, 신경망 기반 보상 모델이 보상 해킹(Reward Hacking)(모델이 실제 목표 대신 보상 시스템 자체를 조작해 높은 점수를 노리는 문제)과 추가 학습 비용(보상 모델 자체를 지속적으로 재학습해야 하는 부담) 문제를 초래하기 때문입니다. 규칙 기반 시스템을 채택함으로써 더 간단하고 안정적인 방식으로 학습하며 이런 문제들을 방지할 수 있었습니다.

3) 자기 진화(Self-Evolution)와 Aha Moment 기법

자기 진화(Self-Evolution Process)는 모델이 스스로 더 복잡한 추론 작업을 처리하도록 학습하는 단계로, 강화학습을 통해 모델이 외부 개입 없이 스스로 추론 능력을 향상시키는 과정입니다. SFT 없이 RL만으로 학습을 진행하므로 순수한 강화학습을 통한 모델의 발전 양상을 관찰할 수 있다는 점이 핵심입니다. 이 과정에서 나타나는 주요 현상은 다음과 같습니다.

  • Test-Time Computation(테스트 시간 연산) 확장: 모델이 한 번의 추론에서 더 많은 토큰(reasoning tokens)을 생성할 수 있도록 허용되며, 이를 통해 더 깊이 사고할 기회를 얻습니다. RL 훈련 과정에서 더 많은 reasoning tokens을 생성하는 것이 더 높은 보상을 받도록 학습되기 때문에, 초기의 짧고 단순한 응답이 점차 단계별 논리 과정을 포함하는 방향(Chain-of-Thought)으로 발전합니다.
  • Reflection(반성): 모델이 자신이 생성한 답변을 다시 검토하고 평가하는 능력이 자발적으로 생겨납니다. 초기에는 답을 바로 출력하지만, 학습이 진행될수록 "이 답이 맞는지", "더 나은 답변이 있는지"를 스스로 판단하는 과정이 추가됩니다.

이러한 행동들이 명시적으로 프로그래밍되지 않았음에도 나타나는 이유는, 강화학습이 정확한 답이나 논리적으로 정교한 사고 과정에 높은 보상을 부여하도록 설계되었기 때문입니다. 모델이 보상을 극대화하려는 과정에서 사고 시간 확장과 반성 같은 고차원적 행동이 자연스럽게 발현되며, Test-Time Computation이 늘어날수록 더 정교한 사고 과정이 필요해지고, Self-Evolution 과정에서 학습된 행동 패턴이 강화되는 것입니다. 즉 이는 사람이 직접 코딩한 것이 아니라, 모델이 RL 환경에서 더 나은 보상을 얻기 위해 스스로 발전한 결과입니다.

Aha Moment는 DeepSeek-R1-Zero가 강화학습 과정에서 문제 해결 방식을 근본적으로 다시 평가하는 순간을 의미합니다. 기존 방식으로 문제를 풀다가 특정 지점에서 "이 방식으로는 안 되겠다"는 깨달음을 얻고 해결 전략을 다시 설계하는 현상으로, RL 보상 시스템과 Test-Time Computation 증가가 결합되면서 자연스럽게 발생합니다.

예시: 방정식의 해를 구하는 문제에서 초기 모델은 양변을 제곱하는 방식으로 풀려고 시도하지만, 이 접근은 오히려 계산하기 까다로운 복잡한 방정식을 만들어 잘못된 결론으로 이어질 가능성이 생깁니다. 모델은 자신의 풀이가 비효율적임을 깨닫고 사고를 다시 정리하는데("잠깐만, 무언가 잘못됐다!", "이 방식이 너무 복잡하다. 좀 더 직관적으로 접근해 보자"), 이것이 바로 Aha Moment입니다. 예를 들어 내부 제곱근을 새로운 변수로 치환해 방정식을 단순화하는 식으로, 문제를 단계적으로 해결하면서 계산 복잡도를 낮추고 더 정확한 해를 도출합니다.

'Aha Moment' 발생 예시
'Aha Moment' 예시: 비효율적인 풀이 전략을 스스로 재평가하고 전환

Aha Moment는 다음과 같은 방식으로 인식되고 학습됩니다.

  • 인식: 초기 접근 방식이 낮은 보상을 받으며 실패 패턴으로 학습되고, Reflection과 Test-Time Computation 증가 과정에서 이전보다 더 많은 reasoning tokens을 생성하는 패턴 변화가 감지되며, 특정 문제 유형에서 사고 시간이 증가하는 현상이 나타납니다.
  • 재평가: 비효율적인 해결 방식에는 낮은 보상이 부여되고, 모델은 보상을 극대화하려는 특성 때문에 새로운 접근 방식을 탐색하며 Self-Consistency와 Reflection을 강화합니다.
  • 전환: 탐색(Exploration)과 활용(Exploitation) 개념에 따라, 모델은 기존 방식의 실패를 인식하고 새로운 접근을 탐색하는 방향으로 학습하며, 체크포인트 기반 비교를 통해 더 높은 보상을 받는 새 방식을 강화하고, 더 많은 reasoning tokens을 사용해 해결 방식을 개선합니다.

GRPO는 개별 출력에 직접 보상을 주지 못하는 구조이지만, Aha Moment 이후 새로운 접근법을 강화하는 것이 여전히 가능합니다. 같은 질문에 대해 기존 방식과 새로운 방식(Aha Moment 이후)을 함께 그룹에 포함해 상대 비교하면, reasoning tokens이 더 많고 논리적인 응답이 상대적으로 더 높은 평가를 받도록 조정되어 점진적으로 강화됩니다. 이처럼 GRPO는 직접 보상 없이도 그룹 내 상대 평가를 통해 Aha Moment 같은 예측 불가능한 정교한 문제 해결 전략의 등장을 학습 신호로 반영할 수 있다는 점이, 강화학습의 강력한 특성을 보여주는 사례입니다.

4) DeepSeek-R1: Cold Start와 다단계 학습

가) Cold Start란?
Cold Start는 초기 강화학습 과정이 불안정해지는 문제를 방지하기 위해, 사전에 일정량의 데이터를 활용하여 모델을 미세조정하는 방법입니다. R1-Zero는 완전한 RL 과정에서 출발했지만 초기에 불안정한 상태를 겪었고, 이를 해결하기 위해 R1에서는 먼저 소규모의 Long CoT 데이터를 수집해 초기 RL Actor(RL 학습을 본격적으로 시작하기 전에 미세조정된 모델)를 구성했습니다.

Cold Start의 이점은 두 가지입니다. 가독성: R1-Zero의 주요 한계는 응답 가독성이 낮다는 점(여러 언어 혼합, 마크다운 형식 부족)이었는데, R1은 초기 데이터 수집 시 가독성을 고려한 패턴을 설계해 각 응답 끝에 요약을 포함하고 읽기 어려운 응답은 필터링했습니다. 잠재력: 인간의 사전 지식을 반영한 초기 데이터 패턴을 신중하게 설계한 결과, R1이 R1-Zero보다 향상된 성능을 보였습니다.

Long CoT Cold Start Data 수집 방법: (A) Few-shot Prompting을 활용해 긴 CoT 데이터를 예시로 제공, (B) Reflection & Verification 기반 프롬프팅으로 모델이 스스로 답을 재검토하도록 유도, (C) R1-Zero의 출력을 가독성 높은 형식으로 변환, (D) 인간 주석자의 후처리(Post-processing)로 품질 개선.

출력의 Readable Format: Cold Start 데이터는 |special_token|<추론 과정>|special_token|<요약> 형식을 표준으로 사용합니다.

Cold Start 데이터의 가독성 높은 출력 포맷 예시
Cold Start 데이터의 가독성 높은 출력 포맷 정의

이 포맷이 바람직한 이유는 (i) 모든 답변이 동일한 형식을 따르도록 강제해 응답 일관성을 높이고, (ii) 추론 과정과 최종 답을 명확히 구분해 사람이 이해하기 쉽게 하며, (iii) 모델이 더 구조적으로 정보를 처리하도록 도와 정교한 CoT 추론을 가능하게 하고, (iv) Iterative Training(반복 학습)과의 궁합이 좋아 이후 더 개선된 데이터셋을 만들어낼 수 있기 때문입니다.

데이터 설계 및 구성: R1은 Cold Start 데이터, Reasoning 데이터, Non-Reasoning 데이터의 세 가지 유형을 활용합니다. Reasoning 데이터는 Rejection Sampling과 규칙 기반 평가(수학 정답 검증, 코드 컴파일 테스트 등)로 정확한 응답만 선별하고, DeepSeek-V3를 활용해 혼합 언어와 불필요한 코드 블록을 필터링하였습니다. Non-Reasoning 데이터는 창의적 글쓰기, 사실 기반 QA, 자기 인식, 번역 등을 다루며 DeepSeek-V3 파이프라인을 재사용하여 효율성을 극대화합니다. 최종적으로 Reasoning 데이터(약 60만 개)와 Non-Reasoning 데이터(약 20만 개)를 통합해 약 80만 개의 데이터셋을 구성하고, 2 에포크 동안 Fine-Tuning을 진행합니다.

5) 추론 중심 강화학습 (Reasoning-Oriented Reinforcement Learning, RoRL)

Cold Start로 미세조정된 DeepSeek-V3-Base 모델을 기반으로 대규모 RL 훈련을 적용해, 수학·코딩·과학·논리 추론과 같은 고난이도 reasoning-intensive task에서 성능을 강화하는 것이 목표입니다. DeepSeek-R1은 단순한 정답 예측이 아니라 문제 해결 과정을 체계적으로 표현하도록 학습되며, 추론 과정을 명확히 나타내는 긴 Reasoning Tokens을 생성하도록 최적화됩니다.

이 과정에서는 언어 혼합(Language Mixing) 문제가 도전 과제로 등장했습니다 — RL 훈련 중 여러 언어가 혼합된 프롬프트를 제공하면 CoT 출력에서 언어적 일관성이 깨지고 가독성과 논리 흐름이 왜곡되는 현상입니다. 해결책으로 언어 일관성 보상(Language Consistency Reward)을 도입해, CoT 내 목표 언어(예: 영어 또는 중국어)의 단어 비율을 보상 점수로 활용했습니다. 최종 보상 값은 추론 정확도 점수와 언어 일관성 점수를 합산하여 결정됩니다. 실험 결과 이 방식이 정확도를 약간 낮출 수도 있지만, 가독성이 향상되고 사람이 더 선호하는 응답을 제공하는 방향으로 정렬되었습니다. 결과적으로 DeepSeek-R1은 수학·코딩·과학·논리 추론 같은 고난도 과제에서 더욱 강력한 성능을 발휘하며, 언어적 일관성을 유지하는 능력까지 갖춘 모델로 발전하였습니다.

6) 거부 샘플링(Rejection Sampling)과 SFT

Reasoning-Oriented RL이 수렴한 이후, 생성된 Resulting Checkpoint를 이용해 SFT 데이터를 추가로 수집하고 모델을 미세조정합니다. 이는 (A) 이전 단계에서 주로 다루지 못한 범용적 태스크(문서 작성, 역할 수행 등) 성능을 보완하고, (B) 모델이 생성한 데이터 중 품질 높은 응답만 선별(Rejection Sampling)해 데이터 신뢰성을 높이기 위함입니다. Reasoning 데이터와 Non-Reasoning 데이터를 성격이 다르다는 이유로 분리해 수집·학습합니다 — Reasoning 태스크는 정확성과 추론 능력이, Non-Reasoning 태스크는 일관성과 가독성이 중요하기 때문입니다.

7) 모든 시나리오를 위한 2차 강화학습 (Reinforcement Learning for All Scenarios)

1단계 RL에서 논리적 사고를 강화했지만 그대로 쓰면 사용자 경험이 최적이 아닐 수 있고, 정답이 명확하지 않은 복잡한 태스크(창의적 글쓰기, 철학적 논의, 윤리적 문제)에서는 규칙 기반 평가가 불가능하며, AI가 편향되거나 위험한 내용을 생성하지 않도록 안전성 개선도 필요합니다. 이를 위해 2차 강화학습에서는 다음을 적용했습니다.

  • 추론 관련 보상은 정답이 명확한 태스크에서 기존 Rule-Based 보상 모델을 유지
  • 일반 데이터 관련 보상은 정답이 명확하지 않은 질문에서 사용자 선호를 반영하는 보상 모델(DeepSeek-V3 파이프라인 기반 선호 쌍 활용)을 사용
  • 다양한 Prompt Distributions을 활용해 일상적 질문부터 고급 추론, 민감한 윤리적 질문까지 폭넓게 학습
  • Helpfulness(유용성): 추론 과정 자체는 그대로 두고 최종 요약(Final Summary)의 유용성과 관련성만 평가해 보상 부여
  • Harmlessness(무해성): 추론 과정과 요약을 포함한 전체 응답을 평가해 편향·위험 요소를 제거

이를 통해 DeepSeek-R1은 강력한 논리적 사고 능력을 유지하면서도 사용자 친화적인 응답을 생성하는, Reasoning + Helpfulness + Harmlessness를 갖춘 모델로 완성되었습니다.

8) 증류(Distillation)를 통한 소형 모델의 추론 능력 강화

R1의 추론 능력을 소형 모델로 증류하는 과정 도식
DeepSeek-R1의 추론 능력을 소형 모델로 증류하는 과정

대형 모델이 학습한 고급 추론 패턴과 능력을 소형 모델에 전달해, 리소스 효율적이면서도 강력한 추론 성능을 가진 모델을 구축하는 것이 목표입니다. 800k개의 데이터셋(DeepSeek-R1에서 수집된 데이터)을 활용해 오픈소스 모델에 미세조정을 진행했습니다. 증류 과정에서는 강화학습(RL) 없이 지도 미세조정(SFT)만 적용했는데, 이는 증류 기법 자체만으로도 소형 모델이 고급 추론 능력을 상당히 계승할 수 있음을 순수하게 검증하기 위해서였습니다. 실험 결과, 가벼운 모델에 추가로 강화학습을 적용하는 것보다 SFT만으로 증류하는 편이 더 우수한 성능을 보였습니다. DeepSeek 팀은 RL 단계를 추가하면 더 정교하고 고차원적인 추론 성능을 제공할 수 있을 것으로 추정하며, 이를 오픈 커뮤니티의 후속 연구 과제로 제안했습니다.

3. V3와 R1, 무엇이 다르고 언제 무엇을 쓰나

두 모델은 종종 나란히 언급되지만 지향점이 다릅니다. R1이 OpenAI o1과 어떻게 비교되는지는 널리 알려져 있는 반면, V3가 어떤 모델과 경쟁하는지는 상대적으로 덜 언급됩니다.

구분DeepSeek-V3DeepSeek-R1
성격 효율적인 범용 MoE 언어모델 순수 RL 기반 추론 특화 모델
주요 비교 대상 GPT-4o, Claude-3.5-Sonnet 등 범용 폐쇄형 모델 OpenAI o1 계열 추론 모델
강점 영역 지식(MMLU 등), 중국어 사실 지식(Chinese SimpleQA), 코딩 경진 벤치마크 수학·코딩·논리 추론에서 긴 사고과정을 통한 정밀한 문제 해결
학습 방식 사전학습 + SFT + RL(+ R1 증류) SFT 없이 순수 RL(R1-Zero) → Cold Start + 다단계 RL/SFT(R1)
적합한 용도 일반 대화, 문서 작성, 다목적 업무 처리 긴 추론 과정이 필요한 수학·과학·복잡한 논리 문제

V3는 GPT-4o·Claude-3.5-Sonnet과 견줄 만한 성능을 경제적인 비용으로 구현한 범용 모델이고, R1은 긴 사고과정을 강화학습만으로 학습해 o1급 추론 성능을 오픈소스로 구현한 특화 모델입니다. 두 모델은 경쟁 관계가 아니라, R1의 추론 능력이 V3로 환류되는 상호보완적 관계에 있다는 점이 이 프로젝트 전체를 이해하는 핵심입니다.

4. 최신 업데이트: Nature 게재와 수치 정정

DeepSeek-R1 논문의 Nature 게재
DeepSeek-R1 논문은 2025년 1월 arXiv 프리프린트로 처음 공개된 이후 동료심사를 거쳐 Nature지(2025, Vol. 645, pp. 633–638)에 정식 게재되었습니다. 외부 전문가 검증을 통과했다는 점에서 연구의 신뢰도를 뒷받침하는 사실입니다.
AIME 2024 성능 수치 정정
R1-Zero의 AIME 2024 pass@1 성능은 2025년 1월 arXiv 프리프린트(v1) 기준 "15.6% → 71.0%"였고, 다수결(majority voting) 적용 시 86.7%까지 상승해 OpenAI-o1-0912와 유사한 성능을 보였다고 보고되었습니다. 이 수치는 여전히 대부분의 자료에서 인용되고 있지만, 2026년 1월 갱신된 Nature 게재판(arXiv v2)에서는 해당 수치가 15.6% → 77.9%로 소폭 상향 조정되었습니다(self-consistency 적용 시 86.7%는 동일하게 유지). 최신 공식 수치를 인용할 때는 77.9%를 기준으로 삼는 것이 정확합니다.
핵심 요약
V3와 R1의 관계: R1은 V3-Base에서 파생되고, R1의 추론 능력은 다시 V3로 증류되는 순환 구조입니다.
V3의 위치: GPT-4o·Claude-3.5-Sonnet급 범용 성능을 훨씬 낮은 비용으로 구현한 MoE 모델입니다.
R1의 위치: 지도학습 없이 순수 RL만으로 o1급 추론 능력을 얻어낸 오픈소스 추론 모델입니다.
최신 정정: R1 논문은 Nature에 정식 게재되었고, AIME 2024 수치는 77.9%로 갱신되었습니다.

DeepSeek의 연구는 "특별한 자금이나 인재가 없어서 못 한다"는 변명이 더 이상 통하지 않을 수 있음을 보여줍니다. 논문과 소스코드를 함께 공개하며 지속적으로 커뮤니티의 피드백을 받는 개방형 연구 방식 자체가, 이 프로젝트가 남긴 또 하나의 중요한 유산이라 할 수 있습니다.

참고자료

  • DeepSeek-AI et al. (2024). DeepSeek-V3 Technical Report. arXiv:2412.19437. arxiv.org/abs/2412.19437
  • DeepSeek-AI et al. (2025). DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning. arXiv:2501.12948 (Nature 645, 633–638, 2025). arxiv.org/abs/2501.12948
개정이력
2026.071차 검증(핵심 수치의 원문 대조)과 2차 검증(V3·R1 관계 설명 보강, V3의 GPT-4o·Claude-3.5-Sonnet 비교 추가, 결론 비교표 신설)을 통합 반영. Nature 게재 사실 및 AIME 2024 수치 정정(71.0%→77.9%) 반영. 원문의 아키텍처·하이퍼파라미터·GRPO 수식·Cold Start·증류 등 기술적 세부 내용을 원문 그대로 충실히 복원하되, 동일 내용을 반복하던 문답(Q&A) 형식 일부는 중복 없이 통합 정리. 원문에 삽입된 모든 도식(총 12개 이미지)을 화면 크기에 따라 자동 조정되는 반응형 이미지로 포함.
※ 본 포스트는 공개된 기술 논문을 바탕으로 한 요약·해설이며, 세부 하이퍼파라미터 및 수식은 원 논문을 기준으로 재확인하시기 바랍니다.

Friday, January 31, 2025

Patents and National Power

"특허는 역사적으로 국가 권력 구축에 어떤 역할을 해왔으며, 21세기 들어 미국, 중국, EU는 지식재산(IP)과 관련하여 각각 어떻게 다른 전략을 펼쳐왔는가?"

"How have patents historically contributed to building national power, and how have the U.S., China, and the EU taken different approaches to intellectual property (IP) in the 21st century?"

이 주제로 ChinaTalk (Jordan Schneider)가 조지메이슨대학교 안토닌 스칼리아 법대의 애덤 모소프(Adam Mossoff) 교수와 진행한 인터뷰가 공개되었습니다.

This is the topic of a recent ChinaTalk (Jordan Schneider) interview with Adam Mossoff, professor at George Mason University’s Antonin Scalia Law School.

이 인터뷰에는 특허 제도와 관련된 주옥같은 통찰이 가득합니다. 특허 제도에 관심 있는 분이라면 꼭 시간을 내어 읽어보시길 강력히 추천드립니다.

This conversation is packed with insightful discussions on the patent system, making it a must-read for anyone interested in intellectual property. I strongly encourage you to take the time to go through it.

주요 내용 Key Takeaways:

- 미국 특허제도의 역사와 특허권의 자연권적 성격

  The history of the U.S. patent system and its natural rights foundation

- 발명을 실시하지 않는 발명가(NPE)의 라이선스 수익화가 특허 제도의 본질적 기능인 이유

  Why the monetization of patents by an inventor (NPEs) is an essential function of the patent system

- 발명(lab or garage invention)혁신(market innovation)의 차이

  The distinction between invention (lab or garage invention) and innovation (market innovation)

- 미국과 영국의 차이: 발명을 혁신(innovation)으로 전환하는 과정

  How the U.S. and the U.K. differ in turning inventions into real-world innovations

- 야셍 황(Yasheng Huang) 교수의 분석: 중국 문명이 특허제도와 비교할 때 과거제(科擧制度)를 선택한 것이 결정적 실수였던 이유

  Yasheng Huang’s analysis: Why China’s reliance on the imperial exam system (Keju) instead of a patent-based system was a historical misstep

- 미국 특허 시스템과 문화적 요인이 어떻게 혁신을 가능하게 했는가

  How the U.S. patent system and cultural factors enabled innovation

- 미국 실용주의와 발명의 민주화(Democratization of Invention)

  The role of American pragmatism and the democratization of invention

- 산업혁명과 기술혁신의 중심이 영국에서 미국으로 이동한 이유

  How the center of industrial and technological innovation shifted from the U.K. to the U.S.

- 현대 특허 시스템의 문제점과 도전 과제

  Modern challenges facing the patent system

- 일론 머스크(Elon Musk)의 특허제도에 대한 인식 변화와 특허의 진정한 역할

  Elon Musk’s evolving stance on patents and the true role of patents in innovation

- 트럼프 행정부의 특허 정책과 향후 전망

  Patent policy under the Trump administration and potential future shifts

- 중국의 지식재산권 정책의 양면성

  China’s two-faced approach to intellectual property—both strengthening its own system and engaging in IP theft

- 표준필수특허(SEP) 전쟁에 중국과 EU의 개입

  The ongoing Standard Essential Patent (SEP) battles involving China and the EU

- 화웨이의 역할 변화

  Huawei’s shifting role in global tech standards

- 중국 법원이 특허 시스템을 정치적으로 활용하는 방식

  How China’s court system is being weaponized to suppress foreign patent holders

- 법치주의(rule of law)특허 시스템의 성패

  Why the rule of law matters for the success of a patent system

- 중국 IP정책에 대한 미국 정부의 대응 전략

  U.S. strategies for countering China’s IP policies

- 미국 특허 시스템의 도덕적·법적·경제적 리더십 회복 필요성

  The urgent need to restore the moral, legal, and economic leadership of the U.S. patent system


📌 혁신과 특허 정책에 대한 깊이 있는 논의를 담고 있는 만큼, 꼭 읽어보시길 바랍니다. 🚀

📌 If you're interested in innovation, intellectual property, and global technology policy, this interview is a must-read! 🚀

균등론에도 경계선은 있다

미국 특허 균등론(Doctrine of Equivalents)에서 '구성요소의 역할'과 '장점의 결여'가 비침해 판단에 미치는 영향 ...