Ideas
Techniques that work in other fields and are still rare in materials science.Techniques that work in other fields and are still rare in materials science.Techniques that work in other fields and are still rare in materials science.Techniques that work in other fields and are still rare in materials science.Techniques that work in other fields and are still rare in materials science.Techniques that work in other fields and are still rare in materials science.
결과가 아니라 시도해볼 방향입니다. 근거 링크는 그 기법이 다른 분야에서 통했다는 것만 보여줍니다. 우리 분야에 맞는지는 연구자가 판단할 일입니다.
학습 없는 후보 압축 검색 · Training-free candidate compression for huge tool/skill libraries
This approach compresses thousands of tools or skills into only the top candidates that match the user request. Before the final LLM decision, it reduces the candidate set to improve speed and selection accuracy.
코퍼스 수준 테스트시점 검색 확장 · Corpus-level test-time scaling for retrieval via metadata/keyword augmentation
This method uses an LLM to add metadata and keywords to documents so that the existing search system can find them better. Rather than switching to a stronger search model, it focuses on improving the searchability of the corpus itself.
검색 루브릭 기반 문서 재랭킹과 강화학습 · Rubric-guided document reranking with reinforcement learning
This approach explicitly defines the criteria that a document set should satisfy as a rubric, and trains the reranker to meet those criteria. It focuses on finding a “balanced evidence bundle” rather than simple relevance.
네이티브 텍스트-비전 상호작용 · native text-vision interaction
This is a method that treats images as continuously maintained context rather than as one-time attachments during retrieval or reasoning. It combines with training and rollout settings that recover from partial failures to improve robustness in multimodal tasks.
파일 기반 장기 멀티모달 검색 · long-horizon agentic multimodal search
Instead of keeping images and diagrams in the context, store them in an external file system and keep only short identifiers during the conversation. This is an approach for reducing context bloat by reloading visual materials when needed in long interactions.
검색 기반 선택형 다중 LLM 협업 · Scalable multi-LLM collaboration with retrieval-based selection
This is a method of placing multiple LLMs and selecting the model best suited to each question or subtask through search and exploration-exploitation strategies. It is a design intended to improve task-specific suitability compared with a single model or simple majority voting.
제한된 LLM 판단의 에이전틱 프로그램 · Agentic programs with bounded LLM judgment
This is a software pattern in which deterministic code takes the center of the pipeline, and the LLM makes limited judgments only at specified gates. It is an approach intended to improve verifiability and operational stability compared with delegating the entire system to an agent.
통합 테스트타임 스케일링 프레임워크 · Unified test-time compute scaling framework with pluggable scorers and cost-aware benchmark
This method places multiple test-time scaling methods and scorers into one framework and measures quality and computational cost together. It is well suited to comparing cost-effectiveness rather than simple accuracy.
불확실성 인식 예산 배분 · Uncertainty-aware budget allocation at test time
This method estimates uncertainty in an initial pass, then assigns fewer samples to easy inputs and more samples or larger computation to difficult inputs. The key is not to use a fixed budget identically for every case.
타입드 에이전트 기반 코디자인 워크플로 · Agentic co-design workflow with typed tools and normalized experiment data
This is a method in which an agent that directly calls tools repeatedly performs experimental and simulation procedures, records the results in a structured repository, and then queries them again. It emphasizes a data-centric iterative loop over conversational memory.
그래프 기반 에이전틱 커널 생성과 실시간 컴파일러 피드백 · Graph-based autonomous search for kernel generation with real-time compiler feedback
This is a method in which an agent generates kernel code and incrementally improves it based on compiler and runtime feedback. Using graph search can allow a wider candidate space to be explored.
종단간 에이전틱 하드웨어/소프트웨어 공동설계 벤치마크 · End-to-end agentic hardware/software co-design benchmark
This is an evaluation style in which an agent must modify multiple components together to produce executable outputs while satisfying system-level constraints. The key point is to examine actual execution and integration, not static code quality.