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.
에이전트 기반 코드 진화와 사이클 정확도 시뮬레이션 · Agentic code evolution + cycle-accurate simulation
This is a 반복 loop in which an LLM agent mutates design code or control code, and the simulator score drives the next revision. Its essence is automated exploration through generation, evaluation, and modification, and it is easy to incorporate physics-based scores.
강화학습 기반 임베더 다단계 검색 · RL-trained Embedder for Multi-step Retrieval
This is an approach that adjusts the embedder itself with reinforcement learning to carry out multi-stage retrieval, rather than a Q&A-oriented LLM. For complex questions, it updates the retrieval path multiple times and gathers the needed evidence instead of finding the answer in one step.
배포 정합 저정밀 신경망 구조 탐색 · Deployment-Aligned Low-Precision Neural Architecture Search
This approach explores and trains structures under the low-precision computational conditions actually used in deployment, rather than focusing on high-precision performance at training time. Instead of quantizing afterward, it reflects the numerical conditions of the target hardware from the start.
검색증강 정책 최적화 · Retrieval-augmented policy optimization
This is a method that, during policy learning, retrieves past step-by-step trajectories, off-policy examples, and failure/success paths and reflects them in the update. The goal is to broaden the search space compared with looking only at the current rollout.
테스트타임 계산 예산 할당 · Test-time compute allocation
This is an approach that does not fix the token, retrieval, retry, and tool-call budget used at inference time, but allocates them according to problem difficulty. The goal is to use less on easy subproblems and more on subproblems with high uncertainty.
블랙보드형 멀티에이전트 오케스트레이션 · Blackboard-style multi-agent orchestration
This is a structure in which multiple agents post partial results to a shared workspace, and whichever agent needs them next reads them and continues the task. The goal is more flexible collaboration than a single coordinator or a fixed master-worker setup.
집합 수준 도구 검색 · Set-Level Tool Retrieval
This technique predicts a set of tools to use together in the next step instead of selecting just one tool at a time. It is designed to be especially useful in agent environments where there are too many tool candidates to fit into the prompt.
에이전트 궤적 기반 검색 학습 · Learning to Retrieve from Agent Trajectories
This technique trains a retriever not only from human relevance judgments, but also from learning signals based on which queries, snippets, and documents an agent actually passed through and whether the outcome was successful. In other words, it aligns search with task success rather than with user satisfaction.
에이전트 기억 시스템 · Agent memory systems
This technique has an agent store prior tool use, failures, and successful subroutines over the long term and reuse them on the next task. It focuses not on one-off conversations, but on continuous workflows spanning multiple tasks.
시험시 루브릭 기반 검증 · Test-time rubric-guided verification
Instead of the model outputting an answer directly, it checks its own output against an explicit rubric and revises it by reflecting structured feedback. It is suited to tasks that must satisfy multiple criteria at once rather than a single correct answer.
파일시스템 기반 장기 추론 에이전트 · File-system-based long-horizon agents
The agent persistently stores materials it has read, intermediate notes, and extracted results in a folder structure, and another agent writes the final answer based on that workspace. An external file system serves as the memory for long-term work.
필드 기반 불리언 검색 · Fielded Boolean retrieval for agents (“search–inspect–fetch”)
This is a search interface that first filters documents by conditions on fields such as title, abstract, sections, tables, and captions, then ranks only the remaining set and retrieves only the needed parts. Rather than reading long documents in full, it lets the agent search while specifying evidence types precisely.