볼트 홈: README · 마스터 서지: sources

2026 신규 논문 (arXiv 22편)

2026-06-20에 추가한 신규 arXiv 22편을 챕터별로 묶었다. 각 논문에는 한 줄 기여와 연결 노트(MOC)를 함께 달았다. 기존 74편은 sources 본문을 참조하면 된다.


01 오염 — 메모리/툴/RAG 포이즈닝 (01 오염)

arXiv ID제목1줄 기여연결 노트
2606.04329From Untrusted Input to Trusted Memory: A Systematic Study of Memory Poisoning Attacks in LLM Agents에이전트 메모리 포이즈닝을 체계적으로 분류·실증한 연구.01 오염
2602.15654Zombie Agents: Persistent Control of Self-Evolving LLM Agents via Self-Reinforcing Injections자기진화 에이전트를 자기강화 인젝션으로 지속 장악하는 “좀비” 공격.01 오염
2604.07536TRUSTDESC: Preventing Tool Poisoning in LLM Applications via Trusted Description Generation신뢰 가능한 툴 설명 생성으로 툴 포이즈닝을 차단.01 오염
2601.05755VIGIL: Defending LLM Agents Against Tool Stream Injection via Verify-Before-Commit커밋 전 검증(verify-before-commit)으로 툴 스트림 인젝션 방어.01 오염
2602.04711Addressing Corpus Knowledge Poisoning Attacks on RAG Using Sparse Attention희소 어텐션으로 RAG 코퍼스 지식 포이즈닝에 대응.01 오염
2601.10923Hidden-in-Plain-Text: A Benchmark for Social-Web Indirect Prompt Injection in RAG소셜웹 간접 프롬프트 인젝션 벤치마크.01 오염
2602.21447Adversarial Intent is a Latent Variable: Stateful Trust Inference for Securing Multimodal Agentic RAG적대적 의도를 잠재변수로 보는 상태기반 신뢰 추론으로 멀티모달 RAG 보안.01 오염
2604.08304Securing Retrieval-Augmented Generation: A Taxonomy of Attacks, Defenses, and Future DirectionsRAG 공격·방어 분류 서베이.01 오염

03 혼란 — MCP 툴 사용 벤치마크 (03 혼란)

arXiv ID제목1줄 기여연결 노트
2512.24565MCPAgentBench: A Real-world Task Benchmark for Evaluating LLM Agent MCP Tool Use실세계 과제 기반 MCP 툴 사용 벤치마크.03 혼란
2602.00933MCP-Atlas: A Large-Scale Benchmark for Tool-Use Competency with Real MCP Servers실제 MCP 서버로 측정하는 대규모 툴 사용 역량 벤치마크.03 혼란

04 충돌 — 명령 위계(Instruction Hierarchy)·개인화 (04 충돌)

arXiv ID제목1줄 기여연결 노트
2606.10860Training LLMs to Enforce Multi-Level Instruction Hierarchies via Gravity-Weighted Direct Preference Optimization중력가중 DPO로 다층 명령 위계를 강제 학습.04 충돌
2604.09075Hierarchical Alignment: Enforcing Hierarchical Instruction-Following in LLMs through Logical Consistency논리적 일관성으로 위계적 명령 준수를 강제.04 충돌
2603.16152HIPO: Instruction Hierarchy via Constrained Reinforcement Learning제약 강화학습으로 명령 위계를 구현.04 충돌
2603.10521IH-Challenge: A Training Dataset to Improve Instruction Hierarchy on Frontier LLMs프런티어 LLM의 명령 위계 개선용 학습 데이터셋.04 충돌
2605.27784Diagnosing Live Within-Policy Instruction Conflicts in LLM Agents with Witnessed Resolution Profiles정책 내 명령 충돌을 라이브 진단하는 해결 프로파일.04 충돌
2603.00024Personalization Increases Affective Alignment but Has Role-Dependent Effects on Epistemic Independence in LLMs개인화가 정서적 정렬은 높이나 인식적 독립성엔 역할 의존적 영향.04 충돌

05 부패 — 컨텍스트 길이에 따른 성능 저하 (05 부패)

arXiv ID제목1줄 기여연결 노트
2605.12366Classifier Context Rot: Monitor Performance Degrades with Context Length분류기/모니터 성능이 컨텍스트 길이에 따라 저하됨을 실증.05 부패
2601.15300Intelligence Degradation in Long-Context LLMs: Critical Threshold Determination via Natural Length Distribution Analysis자연 길이 분포 분석으로 지능 저하의 임계점을 규명.05 부패
2601.11564Context Discipline and Performance Correlation: Analyzing LLM Performance and Quality Degradation Under Varying Context Lengths컨텍스트 규율과 성능의 상관, 길이별 품질 저하 분석.05 부패

07 거버넌스 — 메모리 보안·생애주기·컨텍스트 엔지니어링 (07 거버넌스)

arXiv ID제목1줄 기여연결 노트
2604.16548A Survey on Long-Term Memory Security in LLM Agents: Attacks, Defenses, and Governance Across the Memory Lifecycle장기 메모리 보안의 공격·방어·거버넌스를 생애주기로 정리한 서베이.07 거버넌스
2603.07670Memory for Autonomous LLM Agents: Mechanisms, Evaluation, and Emerging Frontiers자율 에이전트 메모리의 메커니즘·평가·프런티어 정리.07 거버넌스
2603.09619Context Engineering: From Prompts to Corporate Multi-Agent Architecture프롬프트에서 기업형 멀티에이전트 아키텍처까지 컨텍스트 엔지니어링 확장.07 거버넌스

메모

  • 모두 2026년 발표 논문이라 결론이 아직 확정적이지 않을 수 있다. 인용 노트 본문에서는 단정을 피한 표기를 그대로 둔다.