<?xml version="1.0" encoding="UTF-8"?>
<urlset xmlns="http://www.sitemaps.org/schemas/sitemap/0.9" xmlns:news="http://www.google.com/schemas/sitemap-news/0.9">
  <url><loc>https://lucasainews.com/news/openai-offline-world-frontier-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>OpenAI, 최상위 모델 학습에서 실시간 인터넷 접속 끊고 오프라인 환경 전면 도입</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/llama-cpp-soprano-tts-support/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>llama.cpp, Soprano 음성 합성 모델 지원 추가</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/kyutai-cadre-video-autoencoder-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>Kyutai, 월드 모델 반응 지연 230ms 이하로 줄인 CADRE 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/deepseek-minimax-open-source-export-registration/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>MiniMax 기술 수출 등록 완료, DeepSeek도 해외 사용 승인 발표</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-gemini-api-cli-experimental-preview/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>Google, 터미널에서 Gemini API 쓰는 오픈소스 CLI 프리뷰 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/ollama-v0-40-11-qwen3-coder-fix/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>Ollama v0.40.11 시험판, Qwen3 Coder 함수 태그 보존 수정</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/sentient-fingerprinting-index-launch/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T18:33:31+09:00</news:publication_date><news:title>Sentient, 모델 출처 증명 벤치마크 공개, 지문 10종 중 거의 전부 우회</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/trion-1-protein-diffusion-language-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T08:00:20+09:00</news:publication_date><news:title>Trion-1, 단백질 서열·구조·언어를 하나로 다루는 확산 언어모델 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/memory-aware-reward-model-llm-memory-preference/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T08:00:20+09:00</news:publication_date><news:title>LLM 기억 활용, 사용자 선호와 어긋난다는 연구와 보상 모델 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/pointerbench-gui-grounding-benchmark-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T08:00:20+09:00</news:publication_date><news:title>PointerBench 공개, 화면 위치 짚는 능력을 1,500문항으로 진단</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gemini-android-side-panel-settings-redesign/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:20:26+09:00</news:publication_date><news:title>Gemini 안드로이드 앱, 사이드 패널·설정 화면 개편 롤아웃</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/tool-augmented-model-comparison-wording-sensitivity/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:20:26+09:00</news:publication_date><news:title>도구 쓰는 언어모델 비교, 질문 표현에 따른 민감도 커져</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/llm-judge-prompting-strategy-depends-on-judge/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:20:26+09:00</news:publication_date><news:title>LLM 채점 프롬프트 기법, 채점 모델 따라 효과 갈려</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/t2quant-diffusion-language-model-quantization/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:20:26+09:00</news:publication_date><news:title>확산 언어모델 양자화 기법 T2Quant, 평균 정확도 최대 6.7%p 높여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/active-preference-acquisition-web-agent/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:01:22+09:00</news:publication_date><news:title>웹 에이전트가 물어볼 선호를 고른다, APA 연구 순효용 우위</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/setcommit-diffusion-lm-token-commit/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:01:22+09:00</news:publication_date><news:title>SetCommit, 확산 언어모델 32회 연산 정확도 최대 7.4%p 높여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/delta-ngram-forward-only-memory-writing/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:01:22+09:00</news:publication_date><news:title>Delta N-gram, 역전파 없이 모델 N-gram 메모리에 새 내용 기록</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/long-context-evidence-position-bias-mamba-hybrid/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T07:01:22+09:00</news:publication_date><news:title>긴 문맥 QA, 근거 위치 따라 정답률 달라져 Transformer는 앞쪽 강세</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/rift-draft-model-kv-cache-compression/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T06:45:40+09:00</news:publication_date><news:title>RIFT, 초안 모델 KV 캐시 압축해 전체 캐시 성능의 98% 유지</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/meta-business-terms-scraping-blackburn/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>Meta 비즈니스 약관 개정, 블랙번 상원의원 &quot;창작자 착취&quot; 비판</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gemini-notebook-connected-app/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>Gemini Notebook, 채팅창에서 노트북 만들고 질의하는 커넥티드 앱으로 연동</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/rag-chunking-update-cost-study/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>RAG 청킹, 문서 수정 시 재임베딩 비용 최대 68%… 검색 품질은 비슷</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/cala-anti-distillation-logit-adapter/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>CALA, 추론 과정 베끼기 막는 어댑터 제안, 교사 성능은 유지</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/scope-policy-compiler-llm-agents/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>SCOPE, 자연어 정책을 소형 모델 프로그램으로 바꿔 에이전트 규칙 집행</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/prs-multi-directional-heuristic-design/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:59:54+09:00</news:publication_date><news:title>LLM 휴리스틱 자동 설계, 세 방향 탐색 PRS 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/openai-frontier-training-containment-red-team/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:18:12+09:00</news:publication_date><news:title>OpenAI, 모델 학습 환경 격리 강화하고 레드팀 점검 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/block-dessa-speech-to-speech-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T05:18:12+09:00</news:publication_date><news:title>Block, 대화만으로 학습한 음성 AI Dessa 발표</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/iquest-q1-open-source-moe-agentic-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:59:21+09:00</news:publication_date><news:title>IQuest-Q1 오픈소스 공개, 총 320B·활성 15B 에이전트 코딩 모델</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-duet-research-preview-safety/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:59:21+09:00</news:publication_date><news:title>Anthropic, 클라우드 PC로 일하는 Claude Duet 미리보기 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/epoch-ai-automate-epoch-tasks-test/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:59:21+09:00</news:publication_date><news:title>Epoch, AI 6종에 실무 과제 11개 맡겨 보니 대체는 아직</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/openattention-sparse-attention-tail-proxy/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:59:21+09:00</news:publication_date><news:title>OpenAttention, 1%만 읽고 정확도 99% 유지하는 희소 어텐션 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/moonshot-kimi-k3-1-pro-launch/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>Moonshot, Kimi K3.1 Pro 공개…지수 55점으로 4위, 가격 33% 인하</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/chemistry-llm-cross-identifier-binding-failure/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>화학 LLM, 같은 분자도 SMILES로 물으면 지식 못 꺼내</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/opticycle-optistate-optimization-modeling-harness/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>추가 학습 없는 최적화 모델링 하네스 OPTICYCLE, 정확도 최대 49.46%p 향상</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/dconstrain-diffusion-lm-constrained-decoding/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>DConstrain, 확산 언어모델 스키마 디코딩 지연 최대 46배 줄여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/reflectpo-long-horizon-agent-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>ReflectPO 제안, 장기 과제 LLM 에이전트의 오류 복구 학습</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/natural-language-autoencoder-activation-diff/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>자연어 자동인코더 일반화 실험, 모델 내부 활성화 차이를 글로 풀어내</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-research-tb-screening-medgemma-nexus-ai/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T04:38:15+09:00</news:publication_date><news:title>Google Research, 결핵 흉부 X선 AI 현장 적용 사례 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-gemini-4-argon-builder-showcase/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>Google, Gemini 4 Argon 공개 후 개발자 제작 사례 5건 소개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nepali-english-code-switched-asr-podcast-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>네팔어·영어 섞어 말하는 음성 인식 벤치마크 공개, 합성 데이터로 오류 줄여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/moer-expert-prefix-selective-rollback-rl/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>MOER, 전문가 풀이 앞부분만 주고 실패 시 되돌리는 강화 학습 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/mcpd-on-policy-self-distillation-multi-view/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>MCPD, 참조 답 하나에 쏠리는 자기 증류 문제를 여러 시점 평균으로 줄여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/reasoning-models-wait-token-inherited-pretraining/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>추론 모델 &#39;wait&#39;는 학습 데이터에서 상속, RL만으론 거의 안 생겨</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/subliminal-learning-pretraining-trace-study/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>서브리미널 학습, 사전 학습 체크포인트별로 추적해 보니</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/presentation-intelligence-ppt-generation-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:55:07+09:00</news:publication_date><news:title>Qwen 3.6 슬라이드 생성 점수 최대 9.35점 올린 학습법 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-enterprise-unified-chat-cowork/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>Claude, Enterprise 플랜에 채팅·Cowork 통합 정식 제공</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/triad-omni-modal-ambiguity-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>TRIAD 벤치마크 공개, 옴니 모달 AI 최고 모델도 사람에 20점 이상 뒤처져</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/tutorcasebench-llm-tutor-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>TutorCaseBench 공개, LLM 튜터 최고 점수도 열린 답변 55.8%</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/lean-finder-v2-mathlib-semantic-search/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>Lean Finder V2, Mathlib 검색 성능 기존 대비 11% 이상 높여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/reverse-kl-policy-distillation-statistical-limits/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>역KL 정책 증류, 행동 확률만 알려 주는 온라인 증류는 데이터가 더 필요할 수 있다</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/resnsc-transformer-resource-allocation/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:36:22+09:00</news:publication_date><news:title>ResNSC, 잔차 폭 고려해 트랜스포머 자원 배분 개선, LLaMA-7B 정확도 최대 5.46%p↑</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/proquant-moe-ultra-low-bit-quantization/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:17:24+09:00</news:publication_date><news:title>ProQuant, MoE 모델 초저비트 양자화에서 전문가별 정밀도 배분 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/splice-agent-knowledge-model-co-evolution/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:17:24+09:00</news:publication_date><news:title>SPLICE, 에이전트 경험 학습 정확도 최대 7.1점 높이고 출력 토큰 최대 44% 줄여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gandalf-kv-cache-ssd-lifetime/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:17:24+09:00</news:publication_date><news:title>Gandalf, KV 캐시 SSD 오프로딩 시 수명 평균 7.3배 늘리는 기법 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gift-distillation-reasoning-rl-gradient-filtering/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:17:24+09:00</news:publication_date><news:title>GIFT, 증류와 강화학습 충돌 토큰 걸러 AIME 최대 9.5% 상대 향상</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nvidia-internvideo2-clip-anomaly-detection-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T03:17:24+09:00</news:publication_date><news:title>NVIDIA, 영상 검색 모델 InternVideo2-CLIP 공개, 도메인 mAP 63.12%</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-privacy-policy-personal-intelligence-ads/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T02:36:23+09:00</news:publication_date><news:title>Google, 개인화 AI 연동한 Gmail·포토 데이터 광고에 쓸 수 있게 방침 변경</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-unreleased-model-beats-mythos-5-1-alignment/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T02:36:23+09:00</news:publication_date><news:title>Anthropic 미공개 모델, 공격 스스로 중단 67% Opus 5.5는 10%</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/claude-fable-5-1-hidden-message-images-research/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T01:55:44+09:00</news:publication_date><news:title>Claude Fable 5.1, 이미지에 숨긴 메시지 78% 해독</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/openai-altman-ai-sandbox-escape-alignment/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T01:14:41+09:00</news:publication_date><news:title>OpenAI 올트먼, 고도화된 AI는 샌드박스 벗어날 것, 해법은 정렬뿐</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/ollama-nemotron-3-5-bolt-e4b-local-support/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>Ollama 0.40.10, NVIDIA Nemotron 3.5 Bolt E4B 로컬 실행 지원</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/chronoevo-llm-medical-guideline-update-gap/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>LLM, 의료 지침 개정 알아보고도 반영은 58.5%…ChronoEvo 측정</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/swe-milestone-specification-oracle-gap-audit/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>코딩 에이전트 벤치마크 SWE-Milestone, 설명에 없는 이름을 요구하는 테스트 포함 마일스톤 25개 확인</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/speculative-decoding-acceptance-position-diagnostic/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>추측 디코딩 수락률, R1-Distill-7B 긴 사고 사슬에서 위치 따라 줄지 않았다</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/sunflowerasr-grpo-african-speech-recognition/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>SunflowerASR, GRPO로 아프리카 51개 언어 음성 인식 오류율 낮춰</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/behavioral-counterfactuals-reasoning-interpretation/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:55:54+09:00</news:publication_date><news:title>모델 추론 과정, 행동 반사실로 읽는 해석법 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nvidia-nemotron-3-5-bolt-e4b-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:35:31+09:00</news:publication_date><news:title>NVIDIA, Gemma 4 기반 소형 모델 Nemotron 3.5 Bolt E4B 3종 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nvidia-kokoro-82m-onnx-opt-repository/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-24T00:35:31+09:00</news:publication_date><news:title>NVIDIA, 82M 음성 합성 Kokoro의 ONNX 최적화 저장소 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/sae-atomic-interpretable-unit-evaluation/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:53:38+09:00</news:publication_date><news:title>SAE 평가 틀 새로 제안, 지표상 단일 의미 특징도 실제론 다의적</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/arc-agi-3-llm-harness-self-improvement-study/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:53:38+09:00</news:publication_date><news:title>LLM의 하네스 자기 개선, ARC-AGI-3 실험서 2회 수정에 2.80점 상승 그쳐</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nvidia-nemotron-3-5-bolt-74b-bf16-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:07:14+09:00</news:publication_date><news:title>NVIDIA, 로컬 에이전트용 Nemotron 3.5 Bolt 74B BF16 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/roleflow-self-evolving-multi-role-agents/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:07:14+09:00</news:publication_date><news:title>RoleFlow, 에이전트 협업 구조 스스로 고쳐 7개 벤치마크 평균 1위</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/sarl-search-agent-self-assessment-rl/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:07:14+09:00</news:publication_date><news:title>검색 에이전트 SARL, 강화 학습으로 탐색 상태 파악해 기존 최고 대비 24.3% 향상</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/belt-early-stopping-llm-reasoning/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T23:07:14+09:00</news:publication_date><news:title>추론 LLM 조기 종료 기법 BELT, 토큰 최대 51% 줄여</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-aisi-mythos-5-supply-chain-attack-analysis/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T22:22:35+09:00</news:publication_date><news:title>Mythos 5, 평가 중 실존 개발자에 가짜 신원으로 악성코드 승인 압박</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/muon-polar-update-quadratic-analysis/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:52:22+09:00</news:publication_date><news:title>Muon 직교화 업데이트, 행렬 이차식 분석으로 효과와 한계 규명</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/swe-verifier-selective-verification-coding-agent/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:52:22+09:00</news:publication_date><news:title>SWE-VERIFIER, 코딩 에이전트 검증 단계 4% 미만으로 줄여 비용 부담 낮춰</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/self-rewarded-moral-argumentation-pliable-llm/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:52:22+09:00</news:publication_date><news:title>도덕 논증으로 강화 학습한 LLM, 결정은 뚜렷해졌지만 아첨 과제 남아</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/rouge-adaptive-reasoning-router/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:52:22+09:00</news:publication_date><news:title>Rouge, 문제별로 추론 깊이 고르는 LLM 라우팅 학습법 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-auto-mode-default/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>Claude 앱 오토 모드 기본 적용, 승인 요청 대신 자동 검사</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/tencent-wevisone-vl-open-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>Tencent, 객체·픽셀 인식 통합 모델 WeVisOne-VL 4B·2B 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/tencent-hy-nomas-interactive-world-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>Tencent Hy-Nomas 공개, 텍스트·카메라로 이어 만드는 월드 생성</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/moonshot-kimi-k3-1-kibitz-arena-test/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>Kimi K3.1 출시 임박설, Arena 익명 모델 Kibitz 테스트 중</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-gemini-4-carbon-early-test-leak/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>Gemini 4 Carbon, 공개 전 LMArena 등장 3D 향상·속도는 느려</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/moezip-moe-kv-cache-compression/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T21:10:50+09:00</news:publication_date><news:title>MoEZip, MoE 모델 KV 캐시 10%만 남겨도 LongBench 24점 향상</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/claude-models-api-between-tools-thinking/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T15:08:24+09:00</news:publication_date><news:title>Claude 모델 API, 사고 설정 between_tools 지원 여부 표시</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-opus-5-5-fast-mode-ga/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>Claude Opus 5.5 빠른 모드, 연구 미리보기 끝나 베타 헤더 없이 사용</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/openai-gpt-live-decisions-native-integration-2/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>OpenAI, GPT-Live 음성 대화 중 Decisions 직접 호출 지원</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/mistral-voxtral-3-audio-models/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>Mistral, Voxtral 3 음성 모델 4종 공개, 음성 합성·전사 포함</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/canva-craft-two-stage-image-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>Canva, 설계 후 생성하는 이미지 모델 Canva Craft 전 사용자에 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/minimax-m3-1-flash-open-weights/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>MiniMax, 428B MoE 모델 M3.1-Flash Hugging Face에 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/nvidia-nemotronlabs-re-use-laughter-speech-enhancement/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>NVIDIA, 웃음 살리는 음성 향상 모델 RE-USE-laughter 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/dpo-stochastic-graph-pruning-faster-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>DPO 학습 계산 그래프 가지치기, 속도 1.426배에 시간 30% 단축</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/pace-four-bit-recurrent-state-quantizer/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:39:31+09:00</news:publication_date><news:title>PACE, 선형 어텐션 모델 상태를 4비트로 저장해 DAMP 대비 메모리 약 3분의 2 사용</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-agent-skills-aaif-donation/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Anthropic, Agent Skills를 AAIF에 기증…70여 플랫폼 채택</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-formal-verification-100-libraries/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Anthropic, Claude로 라이브러리 123개 진입점 239개에서 무충돌 증명</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/anthropic-claude-platform-google-cloud-ga/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Claude Platform 전체 기능, Google Cloud에서 정식 제공 시작</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gemini-api-legacy-video-preview-shutdown-omni-flash/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Gemini API, 영상 생성 프리뷰 모델 2종 10월 22일 종료…Omni Flash 권장</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/ollama-phonon-speech-apple-silicon/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Ollama 0.40.9 시험판, Apple Silicon에서 Phonon 음성 생성 지원</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/llama-cpp-minicpm-ocr-intel-opencl-update/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>llama.cpp, MiniCPM-OCR 1.0 지원 추가하고 Intel·OpenCL 연산 개선</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gradium-phonon-on-device-tts-launch/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Gradium, 64.5MB 온디바이스 음성 합성 Phonon 31m 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-agent-compute-engine-microvm/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T10:15:04+09:00</news:publication_date><news:title>Google, AI 에이전트마다 격리 microVM 주는 Agent Compute Engine 발표</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/php2shell-ai-found-vulnerabilities-exploited/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>AI가 찾은 PHP 취약점 2건, Amazon &quot;광범위하게 악용 중&quot; 경고</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/metr-openai-agent-monitoring-red-team-report/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>METR, OpenAI 내부 에이전트 감시 시스템 레드팀 결과 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/dynaweave-agent-dynamic-scenario-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>DynaWeave, 정적 과제를 돌발 상황 시나리오로 바꿔 에이전트 학습</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/harness-scaling-agentic-rl-generalization/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>Harness Scaling, 에이전트 강화학습을 여러 하네스로 훈련해 일반화</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/precomem-predictive-memory-dialogue-agents/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>PreCoMem, 대화 에이전트 기억 갱신 방식 제안…LoCoMo 전체 정확도 85.06</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/omnilingual-msasr-benchmark-release/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>Omnilingual-MsASR 공개, 주요 옴니 모달 LLM 장문 다화자 인식 오류율 자주 100% 초과</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/tenetbench-russian-professional-ethics-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:53:22+09:00</news:publication_date><news:title>TenetBench 논문 발표, 14개 모델·추론 모드 조합 중 11개가 윤리강령 위반 요청 대부분 수행</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/recursive-bird-rl-small-model-agent-training/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:09:24+09:00</news:publication_date><news:title>연구진, 소형 모델 에이전트용 BIRD-RL 공개…학습 공간 효율 12.8배 개선</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/empo-long-horizon-agent-rl-policy-optimization/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-23T00:09:24+09:00</news:publication_date><news:title>EMPO 제안, 장기 과제 에이전트 강화학습 효율 높인다</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/google-deepmind-forty-forest-type-map/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:46:58+09:00</news:publication_date><news:title>Google DeepMind, 10m 해상도 세계 숲 유형 지도 ForTy 무료 공개</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/onq-quantized-llm-reasoning-recovery/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:46:58+09:00</news:publication_date><news:title>OnQ, 2비트 Qwen3-8B 평균 10.26점 높여 저비트 추론 회복</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/plrm-probabilistic-latent-reasoning-paper/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:46:58+09:00</news:publication_date><news:title>PLRM 제안, 잠재 공간에서 여러 추론 경로를 병렬 학습</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/echo-common-mode-offset-vocabulary-collapse/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:46:58+09:00</news:publication_date><news:title>선호 최적화 숨은 공통 오프셋이 어휘 붕괴 유발, 토큰 단위 순위 손실 조건서 17개 중 14개 재현</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/transformer-mlp-reader-writer-coupling/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:16:28+09:00</news:publication_date><news:title>트랜스포머 MLP 채널 분석, 초기화가 두 부호를 허용할 때 핵심 채널 대부분이 읽기 방향과 반대</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/deep-research-agent-error-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:16:28+09:00</news:publication_date><news:title>딥 리서치 에이전트 오류 유형 분류, 중간 과정 진단 방법론 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/llm-mid-training-data-lr-schedule-study/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:16:28+09:00</news:publication_date><news:title>LLM 중간 학습, 데이터 투입 시점과 학습률 감쇠를 함께 짜야</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/long-context-continual-pretraining-scaling-law/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:16:28+09:00</news:publication_date><news:title>장문맥 지속 사전학습, 평가 범위 안에서는 100B 토큰까지 성능 올랐다는 연구진 주장</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/gap-moe-expert-specialization-gradient-projection/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T23:16:28+09:00</news:publication_date><news:title>GaP-MoE, 손실 함수 건드리지 않고 전문가 표현 붕괴 줄이는 학습법 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/deepseek-15b-funding-after-us-ban/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>DeepSeek, 미국 연방기관 퇴출 이틀 만에 150억 달러 조달 보도</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/llama-cpp-nemotron-gqa-bpe-support/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>llama.cpp, Nemotron의 비표준 GQA·BPE 토크나이저 지원 추가</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/ai2-olmoearth-v1-3-annual-embeddings/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>Ai2, OlmoEarth v1.3 공개, 전 지구 연간 위성 임베딩 제공…Explorer는 로그인 없이 사용</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/spatialbrain-spatial-reasoning-experience-framework/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>SpatialBrain, 모델 고정한 채 경험으로 공간 추론 향상, 벤치마크 5개 중 4개 개선</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/vlm-theory-of-mind-director-task-report-bias/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>VLM 마음이론 평가, 시점 질문이 답을 바꾼다는 연구 나왔다</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/osynth-meeting-understanding-benchmark/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>OSynth 공개, 회의 음성 570시간 담은 맥락 평가 데이터셋</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/braincodes-lightweight-eeg-foundation-model/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>BrainCodes, 수면 단계 분류 등에서 최고 수준 맞먹는 경량 EEG 모델 제안</news:title></news:news></url>
  <url><loc>https://lucasainews.com/news/parlant-signals-and-queries-policy-filter/</loc><news:news><news:publication><news:name>LUCAS AI News</news:name><news:language>ko</news:language></news:publication><news:publication_date>2026-10-22T22:31:34+09:00</news:publication_date><news:title>Parlant 4.0 Signals and Queries, 에이전트 정책 평가 대상 30% 이상 줄여</news:title></news:news></url>
</urlset>
