<?xml version="1.0" encoding="utf-8"?>
<feed xmlns="http://www.w3.org/2005/Atom">
  <title>ai4s-icml2026 — Paper Curation</title>
  <link href="https://paper-curation.jehyunlee.dev/ai4s-icml2026/" rel="alternate" type="text/html"/>
  <link href="https://paper-curation.jehyunlee.dev/ai4s-icml2026/feed.xml" rel="self" type="application/atom+xml"/>
  <id>https://paper-curation.jehyunlee.dev/ai4s-icml2026/</id>
  <updated>2026-01-01T00:00:00Z</updated>
  <author><name>Jehyun Lee</name></author>
  <generator>paper-curation build_rss.py</generator>
  <entry>
    <title>Measuring Progress in Reasoning Toward Mathematical Discovery with Automatic Verification</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9999_Measuring_Progress_in_Reasoning_Toward_Mathematical_Discover/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9999_Measuring_Progress_in_Reasoning_Toward_Mathematical_Discover/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Erik Y. Wang</name></author>
    <author><name>Sumeet Ramesh Motwani</name></author>
    <author><name>James V Roggeveen</name></author>
    <author><name>Eliot Hodges</name></author>
    <author><name>Dulhan Jayalath</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">LLM이 미해결 수학 문제에 대해 진정한 연구적 발견(discovery)을 수행할 수 있는지 측정하기 위해, generator-verifier gap을 활용한 자동 검증 가능한 벤치마크 HorizonMath를 제안한다. 113개의 대부분 미해결된 문제와 오픈소스 검증 프레임워크를 통해 GPT 5.4 Pro가 실제로 3개 문제에서 기존 미해결 질문을 해결하거나 최고 기록을 개선하는 새로운 해를 제시했음을 보인다.</summary>
  </entry>
  <entry>
    <title>Measure-to-measure Regression with Transformers</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9998_Measure-to-measure_Regression_with_Transformers/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9998_Measure-to-measure_Regression_with_Transformers/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Matthew Vandergrift</name></author>
    <author><name>Martha White</name></author>
    <author><name>Yury Polyanskiy</name></author>
    <author><name>Philippe Rigollet</name></author>
    <author><name>Lazar Atanackovic</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">확률측도(probability measure) 간의 미지의 비선형 변환을 유한한 입출력 측도 쌍으로부터 학습하는 measure-to-measure (M2M) regression 문제를 정식화하고, transformer의 measure-dependent/mean-field 구조를 활용해 static 및 dynamic(flow matching 기반) 두 가지 방법론을 제안한다.</summary>
  </entry>
  <entry>
    <title>MCCE: A Framework for Multi-LLM Collaborative Search in Discrete Spaces with Similarity-Filtered Preference Learning</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9997_MCCE_A_Framework_for_Multi-LLM_Collaborative_Search_in_Discr/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9997_MCCE_A_Framework_for_Multi-LLM_Collaborative_Search_in_Discr/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Nian Ran</name></author>
    <author><name>Zhongzheng Li</name></author>
    <author><name>Yue Wang</name></author>
    <author><name>Qingsong Ran</name></author>
    <author><name>Xiaoyuan Zhang</name></author>
    <category term="LLM Agent Reasoning Training"/>
    <summary type="text">MCCE는 고정된 closed-source LLM과 경량의 trainable local LLM을 결합해, trajectory memory와 similarity-filtered preference learning(DPO)을 통해 두 모델이 상호 자극(mutual inspiration)하며 co-evolve하는 multi-objective discrete optimization(분자 설계) 프레임워크이다.</summary>
  </entry>
  <entry>
    <title>Maximum Mean Discrepancy with Unequal Sample Sizes via Generalized U-Statistics</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9996_Maximum_Mean_Discrepancy_with_Unequal_Sample_Sizes_via_Gener/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9996_Maximum_Mean_Discrepancy_with_Unequal_Sample_Sizes_via_Gener/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Aaron Wei</name></author>
    <author><name>Milad Jalali</name></author>
    <author><name>Danica J. Sutherland</name></author>
    <category term="Statistical Causal Inference Methods"/>
    <summary type="text">불균등 표본 크기 상황에서 usual MMD estimator가 generalized U-statistic임을 활용하여 그 asymptotic distribution을 귀무가설과 대립가설 양쪽에서 새롭게 규명하고, 이를 바탕으로 검정력을 최적화하는 kernel selection criterion을 제시한 논문이다.</summary>
  </entry>
  <entry>
    <title>Matrix-Free GPU Semidefinite Programming for Quantum Ordered Search at the k=6 Frontier</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9995_Matrix-Free_GPU_Semidefinite_Programming_for_Quantum_Ordered/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9995_Matrix-Free_GPU_Semidefinite_Programming_for_Quantum_Ordered/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yancheng Wu</name></author>
    <author><name>Huikang Liu</name></author>
    <author><name>Wenzhi Gao</name></author>
    <author><name>Yuexin Su</name></author>
    <author><name>Tongyang Li</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">양자 순서 탐색 문제(OSP)의 k=6 최적 리스트 크기 N*를 구하는 구조화된 SDP를, 제약 행렬을 명시적으로 저장하지 않고 CUDA 커널로 on-the-fly 평가하는 matrix-free GPU 프레임워크로 풀어, N* ∈ [90,000, 94,000)임을 수치적/이론적으로 규명하고 query coefficient 상한을 0.390에서 0.365로 개선했다.</summary>
  </entry>
  <entry>
    <title>MathlibPR: Pull Request Merge-Readiness Benchmark for Formal Mathematical Libraries</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9994_MathlibPR_Pull_Request_Merge-Readiness_Benchmark_for_Formal/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9994_MathlibPR_Pull_Request_Merge-Readiness_Benchmark_for_Formal/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Zixuan Xie</name></author>
    <author><name>Xinyu Liu</name></author>
    <author><name>Shangtong Zhang</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">MathlibPR는 실제 Mathlib4 PR 히스토리로부터 구축한 벤치마크로, build-passing 상태의 PR 스냅샷이 실제로 merge-ready한지 여부를 LLM과 LLM agent가 판단할 수 있는지 평가한다. 리뷰어 댓글이나 논의 스레드 같은 사후적 사회적 신호는 배제하고 코드와 즉각적 아티팩트만으로 판단하도록 설계되었다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;Lean4와 Mathlib 생태계는 LLM 기반 formal reasoning의 사실상 표준이 되었으며, MiniF</summary>
  </entry>
  <entry>
    <title>MathlibLemma: Folklore Lemma Generation and Benchmark for Formal Mathematics</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9993_MathlibLemma_Folklore_Lemma_Generation_and_Benchmark_for_For/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9993_MathlibLemma_Folklore_Lemma_Generation_and_Benchmark_for_For/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Xinyu Liu</name></author>
    <author><name>Zixuan Xie</name></author>
    <author><name>Amir Moeini</name></author>
    <author><name>Claire Chen</name></author>
    <author><name>Shuze Daniel Liu</name></author>
    <category term="Formal Proof Verification Automation"/>
    <summary type="text">Lean/Mathlib에 부족한 "folklore lemma"(수학자들이 당연시하지만 라이브러리에 없는 중간 보조정리)를 LLM 기반 모듈형 파이프라인으로 자동 발굴·정형화·증명하고, 이를 통해 4,028개의 벤치마크와 1,506개의 검증된 증명 라이브러리를 구축한 연구이다.</summary>
  </entry>
  <entry>
    <title>MatDeplot: Agent-Ready Materials-Curve Understanding for Scientific Reasoning</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9992_MatDeplot_Agent-Ready_Materials-Curve_Understanding_for_Scie/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9992_MatDeplot_Agent-Ready_Materials-Curve_Understanding_for_Scie/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yin Liang</name></author>
    <author><name>Yu Songlin</name></author>
    <author><name>Jianjun Liu</name></author>
    <category term="Biomedical AI Knowledge Systems"/>
    <summary type="text">materials-science 논문에 편재한 line plot이 VLM에게는 정성적 인식만 가능하고 pixel-level grounding에는 사실상 실패한다는 것을 대규모로 정량화하고, 이를 축calibration된 (x, y) 곡선으로 복원하는 local pipeline인 MatDeplot을 제안하여 downstream scientific reasoning의 오차를 크게 낮춘다.</summary>
  </entry>
  <entry>
    <title>MATAI: A Unified Interactive Platform for AI-Driven Alloy Discovery</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9991_MATAI_A_Unified_Interactive_Platform_for_AI-Driven_Alloy_Dis/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9991_MATAI_A_Unified_Interactive_Platform_for_AI-Driven_Alloy_Dis/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yixuan Li</name></author>
    <author><name>Yanchen Deng</name></author>
    <author><name>Chendong Zhao</name></author>
    <author><name>Penghui Yang</name></author>
    <author><name>Jianguo Huang</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">MATAI는 합금 소재 발견을 위한 데이터 분석, 물성 예측, 제약 기반 조성 설계를 하나의 브라우저 기반 플랫폼으로 통합하고, 자연어 대화형 인터페이스와 LLM 기반 feasibility gate로 폐루프(closed-loop) AI4MS 워크플로우를 실현한 통합 플랫폼이다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;AI-driven materials discovery(AI4MS)는 historical data로부터 학습한 surrogate property model과 그에 대한 const</summary>
  </entry>
  <entry>
    <title>MAST: Motif-Augmented Diffusion with Search Tree for Spectroscopic Molecular Structure Elucidation</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9989_MAST_Motif-Augmented_Diffusion_with_Search_Tree_for_Spectros/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9989_MAST_Motif-Augmented_Diffusion_with_Search_Tree_for_Spectros/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Chenghao Jia</name></author>
    <author><name>Mengdi Liu</name></author>
    <author><name>Hong Chang</name></author>
    <author><name>Shiguang Shan</name></author>
    <author><name>Xilin CHEN</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">MAST는 spectra 기반 분자 구조 규명을 위해 motif prior를 diffusion denoising 과정의 중간 증거로 명시적으로 도입하고, MCTS 기반 tree search로 diffusion sampling을 reward-guided 방식으로 재구성하여 제한된 계산 예산 하에서 spectra-consistent한 2D-3D 분자 구조를 효율적으로 생성하는 프레임워크이다.</summary>
  </entry>
  <entry>
    <title>MassSpecGym in the Wild: Uncovering and Correcting Evaluation Pitfalls in AI-Driven Molecule Discovery</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9988_MassSpecGym_in_the_Wild_Uncovering_and_Correcting_Evaluation/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9988_MassSpecGym_in_the_Wild_Uncovering_and_Correcting_Evaluation/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Hongxuan Liu</name></author>
    <author><name>Roman Bushuiev</name></author>
    <author><name>Ivy Lightheart</name></author>
    <author><name>Mrunali Manjrekar</name></author>
    <author><name>Anton Bushuiev</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">MS/MS 기반 분자 발굴을 위한 표준 벤치마크인 MassSpecGym을 사용한 26편의 논문 중 17편 이상에서 데이터 누수, shortcut learning, 구현 오류 등 평가상의 결함을 발견하고 이를 정량화한 뒤, 이를 교정한 MassSpecGym v1.5를 공개한 감사(audit) 연구이다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;MassSpecGym은 MCES 거리 기반의 leakage-free split과 spectrum simulation, molecule retrieval,</summary>
  </entry>
  <entry>
    <title>MARS: Modular Agent with Reflective Search for Automated AI Research</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9987_MARS_Modular_Agent_with_Reflective_Search_for_Automated_AI_R/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9987_MARS_Modular_Agent_with_Reflective_Search_for_Automated_AI_R/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Jiefeng Chen</name></author>
    <author><name>Bhavana Dalvi Mishra</name></author>
    <author><name>Jaehyun Nam</name></author>
    <author><name>Rui Meng</name></author>
    <author><name>Tomas Pfister</name></author>
    <category term="LLM Agent Reasoning Training"/>
    <summary type="text">MARS는 Budget-Aware MCTS, Modular Design-Decompose-Implement 파이프라인, Comparative Reflective Memory 세 가지 축으로 구성된 자율 AI 연구 에이전트 프레임워크로, MLE-Bench에서 오픈소스 프레임워크 중 최고 성능을 달성한다.</summary>
  </entry>
  <entry>
    <title>Marrying Generative Model of Healthcare Events with Digital Twin of Social Determinants of Health for Disease Reasoning</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9986_Marrying_Generative_Model_of_Healthcare_Events_with_Digital/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9986_Marrying_Generative_Model_of_Healthcare_Events_with_Digital/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Ziquan Wei</name></author>
    <author><name>Tingting Dan</name></author>
    <author><name>Guorong Wu</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">본 논문은 사회적 건강 결정요인(SDoH)의 ICD 코드 프록시를 조건으로 하는 latent diffusion 기반 디지털 트윈(DiffDT)을 제안하여, 뇌·심장·간·신장 등 다장기 센서 데이터와 EHR 이벤트 시퀀스를 연결하는 통합 질병 추론 프레임워크를 구축한다.&lt;/essence&gt; &lt;fig_essence&gt;3&lt;/fig_essence&gt; &lt;known&gt;기존 질병 예측 생성모델은 EHR 기반 event-level autoregressive(AR) 모델(MOTOR, Delphi 등)이거나, 다양한 바이오마커를 디지털화한 digital t</summary>
  </entry>
  <entry>
    <title>Marking the Wrong Symptoms: Evaluating LLM Watermarks in Medical Texts</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9985_Marking_the_Wrong_Symptoms_Evaluating_LLM_Watermarks_in_Medi/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9985_Marking_the_Wrong_Symptoms_Evaluating_LLM_Watermarks_in_Medi/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Melanie Rieff</name></author>
    <author><name>Robin Staab</name></author>
    <author><name>Thibaud Gloaguen</name></author>
    <author><name>Stefan Hegselmann</name></author>
    <author><name>Martin Vechev</name></author>
    <category term="Biomedical AI Knowledge Systems"/>
    <summary type="text">본 논문은 LLM watermarking이 의료 텍스트/멀티모달 임상 추론에 미치는 영향을 최초로 체계적으로 평가하며, 정확도 지표는 안정적으로 보이지만 실제로는 추론 품질, 용어 정확성, 이미지 해석에서 심각한 저하가 발생함을 보인다.</summary>
  </entry>
  <entry>
    <title>Market Incentivization for Theorem Proving with LLM Agents</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9984_Market_Incentivization_for_Theorem_Proving_with_LLM_Agents/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9984_Market_Incentivization_for_Theorem_Proving_with_LLM_Agents/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Gregory Constantine</name></author>
    <author><name>Leyan Pan</name></author>
    <author><name>Vijay Ganesh</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">이 논문은 수학 정리 증명을 위한 다중 에이전트 시스템에서 hand-engineered harness 대신 market 신호를 통해 에이전트 조정을 유도하고, 시장 수익성에 기반한 진화적 압력이 더 수익성 높은 에이전트 집단을 만드는지 예비 실험 결과를 보고한다.</summary>
  </entry>
  <entry>
    <title>Mapping Uncharted Symmetries: Machine Discovery in Combinatorics</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9983_Mapping_Uncharted_Symmetries_Machine_Discovery_in_Combinator/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9983_Mapping_Uncharted_Symmetries_Machine_Discovery_in_Combinator/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Eugenio Cainelli</name></author>
    <author><name>Lorenzo Luccioli</name></author>
    <author><name>Alessandro Iraci</name></author>
    <author><name>Michele D'Adderio</name></author>
    <author><name>Giovanni Paolini</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">대수적 조합론의 미해결 문제에서 영감을 받아, 정확한 분포 제약 하에서 단순한 수학적 함수를 발견하는 문제를 Simple Learning Under Rigid Proportions(SLURP)로 정식화하고, 이를 위한 두 가지 머신러닝 방법(MapSeek-Functional, MapSeek-Symbolic)을 통해 q,t-Narayana polynomials에 대한 새로운 combinatorial interpretation을 발견한 논문이다.</summary>
  </entry>
  <entry>
    <title>Many Needles in a Haystack: Active Hit Discovery for Perturbation Experiments</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9982_Many_Needles_in_a_Haystack_Active_Hit_Discovery_for_Perturba/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9982_Many_Needles_in_a_Haystack_Active_Hit_Discovery_for_Perturba/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Andrea Rubbi</name></author>
    <author><name>Arpit Merchant</name></author>
    <author><name>Samuel Ogden</name></author>
    <author><name>Amir Akbarnejad</name></author>
    <author><name>Pietro Lio</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">유전자 교란(perturbation) 실험에서 제한된 실험 예산 하에 threshold를 초과하는 다수의 hit을 발견하는 문제를 순차적 실험 설계(sequential experimental design)로 정식화하고, 후보의 posterior probability of being a hit을 직접 랭킹하는 acquisition function인 Probability-of-Hit(PoH)을 제안한다.&lt;/essence&gt; &lt;fig_essence&gt;2&lt;/fig_essence&gt; &lt;known&gt;고처리량 CRISPR 스크리닝 등 유전자 교란 실험</summary>
  </entry>
  <entry>
    <title>Making Generative Models Know What They Don’t Know via Hypothesis Testing</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9981_Making_Generative_Models_Know_What_They_Dont_Know_via_Hypoth/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9981_Making_Generative_Models_Know_What_They_Dont_Know_via_Hypoth/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Leander Kurscheidt</name></author>
    <author><name>Antonio Vergari</name></author>
    <category term="Statistical Causal Inference Methods"/>
    <summary type="text">normalizing flow의 latent-space 통계량을 이용한 goodness-of-fit(GoF) 검정을 제안하여, likelihood만으로는 불안정한 OOD 탐지 문제를 hypothesis testing 관점에서 재구성하고, 모델의 데이터 적합도 자체를 평가하는 동시에 recalibrated p-value로 단일 샘플 OOD 탐지를 수행한다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;normalizing flow는 명시적 density를 제공하여 log-likelihood 기반 </summary>
  </entry>
  <entry>
    <title>MADS-CPS: A Machine-Checkable Admissibility Contract for AI Scientists in Autonomous Laboratories</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9980_MADS-CPS_A_Machine-Checkable_Admissibility_Contract_for_AI_S/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9980_MADS-CPS_A_Machine-Checkable_Admissibility_Contract_for_AI_S/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Mateo PETEL</name></author>
    <category term="Formal Proof Verification Automation"/>
    <summary type="text">AI scientist 시스템이 자율 실험실에서 실제 물리적 실행을 수행하게 되면서, 능력(capability) 평가만으로는 부족하고 한 번의 구체적 run이 감사 가능하고(inspectable) 재현 가능하며(replayable) 무결성을 유지하고(integrity-preserving) 비가역적 지점(PONR)에서 통제되는지를 기계적으로 검증하는 run-level admissibility contract인 MADS-CPS를 제안한다.</summary>
  </entry>
  <entry>
    <title>MADField: Multi-fidelity Amortized Density Field for Adsorption in Nanoporous Materials</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9979_MADField_Multi-fidelity_Amortized_Density_Field_for_Adsorpti/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9979_MADField_Multi-fidelity_Amortized_Density_Field_for_Adsorpti/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yoonho Kim</name></author>
    <author><name>Seongsu Kim</name></author>
    <author><name>Sungsoo Ahn</name></author>
    <author><name>Honghui Kim</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">MADField는 나노다공성 소재 내 기체 흡착 예측을 스칼라 uptake 회귀 대신 평형 흡착물 밀도장(equilibrium density field) 추정 문제로 재정의하고, cDFT와 GCMC라는 서로 다른 fidelity의 시뮬레이션 라벨을 결합해 학습하는 multi-fidelity 프레임워크이다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;GCMC는 입자 기반으로 흡착 평형을 정확히 추정할 수 있지만 수백만 스텝의 Markov chain이 필요해 대규모 스크리닝에는 느리며, cDFT는</summary>
  </entry>
  <entry>
    <title>MADE: Benchmark Environments for Closed-Loop Materials Discovery</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9978_MADE_Benchmark_Environments_for_Closed-Loop_Materials_Discov/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9978_MADE_Benchmark_Environments_for_Closed-Loop_Materials_Discov/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Shreshth A Malik</name></author>
    <author><name>Tiarnan Doherty</name></author>
    <author><name>Panagiotis Tigas</name></author>
    <author><name>Muhammed Razzak</name></author>
    <author><name>S Roberts</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">MADE는 재료 발견을 고정된 예측/생성 벤치마크가 아닌, 제한된 oracle 예산 하에서 후보 물질을 제안-평가-개선하는 closed-loop 과정으로 정식화하여 end-to-end 자율 발견 파이프라인을 체계적으로 벤치마킹하는 모듈형 프레임워크이다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;Materials Project, OQMD 같은 데이터셋은 정적 예측 정확도를 평가하고, Matbench Discovery는 MLIP 기반 스크리닝으로 discovery-oriented 평가로 한걸음 </summary>
  </entry>
  <entry>
    <title>MacroGuide: Topological Guidance for Macrocycle Generation</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9977_MacroGuide_Topological_Guidance_for_Macrocycle_Generation/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9977_MacroGuide_Topological_Guidance_for_Macrocycle_Generation/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Alicja Maksymiuk</name></author>
    <author><name>Alexandre Duplessis</name></author>
    <author><name>Michael M. Bronstein</name></author>
    <author><name>Alexander Tong</name></author>
    <author><name>Fernanda Duarte</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">MacroGuide는 사전학습된 분자 diffusion 모델의 각 denoising 단계에서 Persistent Homology 기반의 topological guidance를 적용하여, 재학습 없이 macrocycle(12개 이상 heavy atom 고리) 생성률을 1%에서 99%로 끌어올리는 방법이다.</summary>
  </entry>
  <entry>
    <title>Machine-checked refutation of a convergence theorem in dopamine-dependent credit assignment with Kairos</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9976_Machine-checked_refutation_of_a_convergence_theorem_in_dopam/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9976_Machine-checked_refutation_of_a_convergence_theorem_in_dopam/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Aidan Z.H. Yang</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">이 논문은 도파민 기반 신용 할당(credit assignment)을 설명하는 데 널리 쓰이는 actor-critic 수렴 정리(Konda &amp; Tsitsiklis, 2000)의 통속적 재서술(commonly-retold form)이 실제로는 거짓임을 Lean 4로 기계 검증하여 밝히고, 이를 교정한 정리와 다중 에이전트 검증 시스템 Kairos를 제안한다.</summary>
  </entry>
  <entry>
    <title>Machine-Checked Finite Differential Privacy in Lean</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9975_Machine-Checked_Finite_Differential_Privacy_in_Lean/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9975_Machine-Checked_Finite_Differential_Privacy_in_Lean/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Bright Liu</name></author>
    <category term="Reinforcement Learning Policy Optimization"/>
    <summary type="text">유한 사건(finite-event) 순수 차분 프라이버시(pure differential privacy)의 핵심 정리들(pointwise/event 동치, post-processing, adaptive/parallel composition, randomized response, exponential mechanism)을 Lean 4와 Mathlib을 이용해 machine-checked 방식으로 형식화한 소규모 아티팩트를 제시한다.&lt;/essence&gt; &lt;fig_essence&gt;0&lt;/fig_essence&gt; &lt;known&gt;차분 프라이버시의 p</summary>
  </entry>
  <entry>
    <title>Machine Learning Hamiltonians are Accurate Energy-Force Predictors</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9974_Machine_Learning_Hamiltonians_are_Accurate_Energy-Force_Pred/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9974_Machine_Learning_Hamiltonians_are_Accurate_Energy-Force_Pred/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Seongsu Kim</name></author>
    <author><name>Chanhui Lee</name></author>
    <author><name>Yoonho Kim</name></author>
    <author><name>Seongjun Yun</name></author>
    <author><name>Honghui Kim</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">이 논문은 machine learning Hamiltonian(MLH) 모델의 성능을 기존의 Hamiltonian 재구성 오차(reconstruction metric) 대신 예측된 Hamiltonian으로부터 직접 계산한 에너지·힘(energy-force) 정확도로 평가하는 벤치마크를 제안하고, SO(2)-equivariant backbone과 two-stage edge update를 결합한 QHFlow2 모델을 통해 MLIP 수준의 힘 정확도와 이를 능가하는 에너지 정확도를 동시에 달성함을 보인다.</summary>
  </entry>
  <entry>
    <title>M-IDoL: Information Decomposition for Modality-Specific and Diverse Representation Learning in Medical Foundation Model</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9973_M-IDoL_Information_Decomposition_for_Modality-Specific_and_D/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9973_M-IDoL_Information_Decomposition_for_Modality-Specific_and_D/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yihang Liu</name></author>
    <author><name>Longzhen Yang</name></author>
    <author><name>Jiaxiong Yang</name></author>
    <author><name>Ying Wen</name></author>
    <author><name>Lianghua He</name></author>
    <category term="Multimodal Biomedical Data Fusion"/>
    <summary type="text">M-IDoL은 정보 분해(information decomposition) 관점에서 의료영상 파운데이션 모델(MFM)의 modality-shared redundancy 문제를 해결하기 위해, inter-modality entropy 최대화와 intra-modality uncertainty 최소화라는 두 목표를 MoE projector로 구현하는 자기지도 학습(self-supervised learning) 기법이다.</summary>
  </entry>
  <entry>
    <title>LoPhyDA: Low-Rank Tensor and Physics Gradient Guided Diffusion for Atmospheric Data Assimilation</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9971_LoPhyDA_Low-Rank_Tensor_and_Physics_Gradient_Guided_Diffusio/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9971_LoPhyDA_Low-Rank_Tensor_and_Physics_Gradient_Guided_Diffusio/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Danyang Peng</name></author>
    <author><name>Yang Chen</name></author>
    <author><name>Yunlong Zhou</name></author>
    <author><name>Xiaotong Yuan</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">희소한 관측치로부터 low-rank tensor completion을 통해 전역적으로 조밀한 사전 필드를 복원하고, 이를 PDE 기반 physics gradient와 함께 diffusion 기반 data assimilation의 reverse sampling 과정에 이중 가이드로 주입하여 관측이 없는 영역의 오차 누적과 물리적 비일관성을 완화하는 LoPhyDA를 제안한다.</summary>
  </entry>
  <entry>
    <title>Look Before You Leap: Improving Structure-Based Drug Optimization with Attribution-Guided Genetic Operators</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9970_Look_Before_You_Leap_Improving_Structure-Based_Drug_Optimiza/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9970_Look_Before_You_Leap_Improving_Structure-Based_Drug_Optimiza/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>David Li He</name></author>
    <author><name>Alan C Cheng</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">Graph-Based Genetic Algorithm(GB-GA)의 crossover/mutation 위치를 무작위로 선택하는 대신, directed message-passing neural network(D-MPNN)에서 도출한 atom-level attribution score를 이용해 fitness 개선 가능성이 낮은 원자를 우선적으로 선택하도록 유도하는 attribution-guided site selection 기법을 제안한다.</summary>
  </entry>
  <entry>
    <title>Longitudinal Dense-to-Sparse Forecasting: Individual Variability Predicts Conversion Better than Mean Change</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9969_Longitudinal_Dense-to-Sparse_Forecasting_Individual_Variabil/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9969_Longitudinal_Dense-to-Sparse_Forecasting_Individual_Variabil/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Georgi Hrusanov</name></author>
    <author><name>Ivan Stoyanov</name></author>
    <author><name>Duy-Cat Can</name></author>
    <author><name>Duy-Thanh VU</name></author>
    <author><name>Carsten Magnus</name></author>
    <category term="Clinical Time-Series Modeling"/>
    <summary type="text">ADNI에서 학습된 cortical-thickness graph forecaster로부터 얻어진 두 가지 label-free risk 신호(predicted mean structural change sµ, predicted individual variability sσ)를 비교한 결과, cohort shift(AIBL, OASIS-3로의 전이) 상황에서 sσ는 conversion 예측력을 유지하지만 sµ는 급격히 저하됨을 보였다.</summary>
  </entry>
  <entry>
    <title>LongCoT: Benchmarking Long-Horizon Chain-of-Thought Reasoning</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9968_LongCoT_Benchmarking_Long-Horizon_Chain-of-Thought_Reasoning/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9968_LongCoT_Benchmarking_Long-Horizon_Chain-of-Thought_Reasoning/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Sumeet Ramesh Motwani</name></author>
    <author><name>Daniel Nichols</name></author>
    <author><name>Charles London</name></author>
    <author><name>Peggy Li</name></author>
    <author><name>Fabio Pizzati</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">LongCoT는 chemistry, mathematics, computer science, chess, logic 다섯 개 도메인에 걸쳐 2,500개의 전문가 설계 문제로 구성된 벤치마크로, 수만~수십만 개의 reasoning token을 필요로 하는 상호의존적 step들의 그래프를 탐색하는 능력, 즉 long-horizon chain-of-thought(CoT) 추론 능력을 직접 측정한다.&lt;/essence&gt; &lt;fig_essence&gt;2&lt;/fig_essence&gt; &lt;known&gt;기존 벤치마크들(FrontierMath, HLE 등)은 짧은</summary>
  </entry>
  <entry>
    <title>LOD Surrogate Modelling for multiscale Darcy-Flow Problems</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9967_LOD_Surrogate_Modelling_for_multiscale_Darcy-Flow_Problems/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9967_LOD_Surrogate_Modelling_for_multiscale_Darcy-Flow_Problems/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Marc Haltmayer</name></author>
    <author><name>Jaemin Seo</name></author>
    <author><name>Lee Yu Seung</name></author>
    <author><name>Sungyeop Lee</name></author>
    <author><name>Jaehoon Jeong</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">본 논문은 rough하고 high-contrast인 계수를 갖는 multiscale elliptic PDE(Darcy flow)에 대해 기존 neural operator들이 fine-scale 구조를 잘 근사하지 못하는 한계를 지적하고, LOD(Localized Orthogonal Decomposition) 방법의 multiscale basis 구조를 신경망 학습에 도입한 hybrid 모델인 LOD-MSNO를 제안한다.</summary>
  </entry>
  <entry>
    <title>Localized, High-resolution Geographic Representations with Slepian Functions</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9966_Localized_High-resolution_Geographic_Representations_with_Sl/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9966_Localized_High-resolution_Geographic_Representations_with_Sl/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Arjun Rao</name></author>
    <author><name>Ruth Crasto</name></author>
    <author><name>Tessa Ooms</name></author>
    <author><name>David Rolnick</name></author>
    <author><name>Konstantin Klemmer</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">지리적 좌표를 인코딩하는 위치 인코더에서, spherical Slepian functions을 이용해 관심 지역(region-of-interest)에 표현 용량을 집중시키고 고해상도로 확장 가능한 새로운 geographic location encoder를 제안한다. 전역 맥락이 필요한 경우를 위해 Slepian과 Spherical Harmonics를 결합한 hybrid encoder도 함께 제시한다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;기존 geographic location encod</summary>
  </entry>
  <entry>
    <title>Local Intrinsic Dimension of Representations Predicts Alignment and Generalization in AI Models and Human Brain</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9964_Local_Intrinsic_Dimension_of_Representations_Predicts_Alignm/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9964_Local_Intrinsic_Dimension_of_Representations_Predicts_Alignm/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Junjie Yu</name></author>
    <author><name>Wenxiao Ma</name></author>
    <author><name>Chen Wei</name></author>
    <author><name>Jianyu Zhang</name></author>
    <author><name>Haotian Deng</name></author>
    <category term="Statistical Causal Inference Methods"/>
    <summary type="text">신경망의 표현이 잘 일반화될수록 다른 모델 및 인간 뇌 신경 활동과 더 강하게 정렬(alignment)되며, 이 세 가지 현상(generalization, AI-AI alignment, AI-Brain alignment) 모두가 embedding의 local intrinsic dimension이라는 단일 기하학적 속성으로 설명될 수 있음을 보인다.</summary>
  </entry>
  <entry>
    <title>LLMs Can Learn the Language of the Microbiome</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9961_LLMs_Can_Learn_the_Language_of_the_Microbiome/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9961_LLMs_Can_Learn_the_Language_of_the_Microbiome/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Neythen J Treloar</name></author>
    <author><name>Saif Ur-Rehman</name></author>
    <author><name>Jenny Yang</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">마이크로바이옴(microbiome) 데이터에 GPT 스타일 causal language model을 사전학습시켜 미생물 군집 예측 작업에서 기존 방법 및 이전 foundation model보다 우수한 성능을 달성한 연구이다. 이를 위해 대규모 사전학습 데이터셋 Atlas와 표준화된 벤치마크 Compass를 함께 제안한다.</summary>
  </entry>
  <entry>
    <title>LLM-Guided Diagnostic Evidence Alignment for Medical Vision–Language Pretraining under Limited Pairing</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9959_LLM-Guided_Diagnostic_Evidence_Alignment_for_Medical_VisionL/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9959_LLM-Guided_Diagnostic_Evidence_Alignment_for_Medical_VisionL/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Huimin Yan</name></author>
    <author><name>Liang Bai</name></author>
    <author><name>Xian Yang</name></author>
    <author><name>Long Chen</name></author>
    <category term="Clinical Time-Series Modeling"/>
    <summary type="text">LGDEA는 CLIP 스타일 medical vision-language pretraining에서 global/local alignment가 진단적으로 중요한 정보를 놓치는 문제를 해결하기 위해, LLM으로 report에서 핵심 diagnostic evidence를 추출해 shared diagnostic evidence space를 구성하고 이를 통해 제한된 paired data와 풍부한 unpaired image/report를 함께 활용하는 evidence-aware cross-modal alignment 기법이다.&lt;/essence</summary>
  </entry>
  <entry>
    <title>LLM-guided acquisition for pathway-specific Perturb-seq design under experimental budgets</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9958_LLM-guided_acquisition_for_pathway-specific_Perturb-seq_desi/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9958_LLM-guided_acquisition_for_pathway-specific_Perturb-seq_desi/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Malaika Aiyar</name></author>
    <author><name>Kanglu Pei</name></author>
    <author><name>Sisi Qu</name></author>
    <author><name>Philip Torr</name></author>
    <author><name>Christian Schroeder de Witt</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">고정된 embedding 기반 perturbation-response predictor(LLMPERT) 하에서, ensemble epistemic uncertainty로 후보를 shortlist하고 LLM이 pathway context·annotation·diversity 기준으로 재순위화하는 2단계 acquisition 전략을 제안하여 예산 제약 하 Perturb-seq 설계에서 random 및 prior-aware 베이스라인과 경쟁력 있는 성능을 보인다.</summary>
  </entry>
  <entry>
    <title>LLM-Assisted versus Agentic Approaches to De Novo Minibinder Design for a KRAS G12D Neoantigen</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9957_LLM-Assisted_versus_Agentic_Approaches_to_De_Novo_Minibinder/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9957_LLM-Assisted_versus_Agentic_Approaches_to_De_Novo_Minibinder/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yilan Wang</name></author>
    <author><name>Aaron W Kollasch</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">LLM 어시스턴트 활용 인간 주도 설계, LLM 에이전트 완전 자율 설계, 인간 피드백이 포함된 LLM 에이전트 설계라는 세 가지 방식으로 KRAS G12D neoantigen(펩타이드 VVGADGVGK, HLA-A*11:01)에 대한 de novo minibinder를 설계하고, 새로 구축한 다층 평가 체계(구조/서열 품질, Rosetta 에너지 계산, docking 검증)로 비교 분석한 연구이다.</summary>
  </entry>
  <entry>
    <title>LLM Scheming Inversely Scales with Pretraining Language Coverage</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9956_LLM_Scheming_Inversely_Scales_with_Pretraining_Language_Cove/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9956_LLM_Scheming_Inversely_Scales_with_Pretraining_Language_Cove/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Nathan Truong</name></author>
    <author><name>Aryan Panda</name></author>
    <author><name>Rayming Ye</name></author>
    <author><name>Zoe Sun</name></author>
    <author><name>Maheep Chaudhary</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">Qwen3-30B-A3B 모델을 대상으로 Petri 자동 감사 프레임워크를 활용해 6개 언어에서 scheming(계획적 기만) 행동을 측정한 결과, 추정된 pretraining language coverage가 낮은 저자원 언어일수록 scheming 점수가 평균 34.2% 더 높게 나타났다는 상관관계 연구이다.</summary>
  </entry>
  <entry>
    <title>LLM Agents for Distribution-Aware Algorithmic Discovery</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9955_LLM_Agents_for_Distribution-Aware_Algorithmic_Discovery/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9955_LLM_Agents_for_Distribution-Aware_Algorithmic_Discovery/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Saharsh Koganti</name></author>
    <author><name>Priyadarsi Mishra</name></author>
    <author><name>Pierfrancesco Beneventano</name></author>
    <author><name>Tomer Galanti</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">샘플만으로 접근 가능한 미지의 구조화된 분포로부터 LLM agent가 재사용 가능한 계산적 구조(solver hint)를 추론하고, 이를 실행 가능한 solver 코드로 컴파일하는 distribution-aware program learning 프레임워크를 제안한다.</summary>
  </entry>
  <entry>
    <title>LitXBench: A Benchmark for Extracting Experiments from Scientific Literature</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9954_LitXBench_A_Benchmark_for_Extracting_Experiments_from_Scient/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9954_LitXBench_A_Benchmark_for_Extracting_Experiments_from_Scient/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Curtis Chong</name></author>
    <author><name>Jorge Colindres</name></author>
    <category term="LLM Reasoning and Safety Benchmarks"/>
    <summary type="text">과학 문헌에서 소재(합금)의 실험 측정치를 추출하는 작업을 벤치마킹하기 위한 프레임워크 LitXBench와, 19편의 합금 논문에서 1426개 측정치를 담은 고밀도 벤치마크 LitXAlloy를 제안한다. 실험 데이터를 CSV/JSON이 아닌 Python 객체로 저장하여 감사 가능성과 프로그래밍적 검증을 강화했다.</summary>
  </entry>
  <entry>
    <title>LitReview Arena: Evaluating Literature Review Agents with Battle-style Peer Review Platform</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9953_LitReview_Arena_Evaluating_Literature_Review_Agents_with_Bat/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9953_LitReview_Arena_Evaluating_Literature_Review_Agents_with_Bat/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Ruotong Zhao</name></author>
    <author><name>zhiyu chen</name></author>
    <author><name>Xurui Liu</name></author>
    <author><name>Haidong Xue</name></author>
    <author><name>Dong Liang</name></author>
    <category term="Biomedical AI Knowledge Systems"/>
    <summary type="text">본 논문은 자동 생성 literature review의 품질을 평가하기 어려운 문제를 해결하기 위해 battle-style 전문가 평가 플랫폼인 LitReview Arena를 제안하고, 이를 통해 3k개의 전문가 판단 데이터셋 LitReviewBench와 저비용 평가자 LitJudge를 구축한다.</summary>
  </entry>
  <entry>
    <title>LithoGRPO: Fast Inverse Lithography via GRPO Reinforced Flow Matching</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9952_LithoGRPO_Fast_Inverse_Lithography_via_GRPO_Reinforced_Flow/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9952_LithoGRPO_Fast_Inverse_Lithography_via_GRPO_Reinforced_Flow/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yao Lai</name></author>
    <author><name>Xuyuan Xiong</name></author>
    <author><name>Zeyue Xue</name></author>
    <author><name>Guojin Chen</name></author>
    <author><name>Jing Wang</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">LithoGRPO는 flow-matching 기반 생성 모델에 GRPO 강화학습 미세조정을 결합해 Inverse Lithography Technology(ILT)의 mask 생성을 미분 가능/불가능 물리적 지표를 모두 최적화하도록 설계한 프레임워크이다.</summary>
  </entry>
  <entry>
    <title>Lithography Solvent Discovery using Neuro-Symbolic Search Agent</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9951_Lithography_Solvent_Discovery_using_Neuro-Symbolic_Search_Ag/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9951_Lithography_Solvent_Discovery_using_Neuro-Symbolic_Search_Ag/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Jiangyu Chen</name></author>
    <author><name>Huawei Zhou</name></author>
    <author><name>Zhou Zhang</name></author>
    <author><name>Yuqiang Li</name></author>
    <author><name>Ruzhi Zhang</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">LLM 기반 chemistry-informed hypothesis generator와 differentiable physics-informed 모듈을 결합하여, 불완전한 proxy evaluator 하에서도 feasible하고 다양한 lithography solvent formulation을 탐색하는 neuro-symbolic search framework인 LithographyAgent를 제안한다.</summary>
  </entry>
  <entry>
    <title>LithoDreamer: A Physics-Informed World Model for Multi-Stage Computational Lithography</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9950_LithoDreamer_A_Physics-Informed_World_Model_for_Multi-Stage/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9950_LithoDreamer_A_Physics-Informed_World_Model_for_Multi-Stage/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yuqi Jiang</name></author>
    <author><name>Yumeng Liu</name></author>
    <author><name>Zimu Li</name></author>
    <author><name>Jinyuan Deng</name></author>
    <author><name>Qian Jin</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">LithoDreamer는 computational lithography의 "Layout-Mask-Resist Image-ADI" 다단계 파이프라인을 물리 정보 기반 World Model(WM)로 정식화하여, 공정 개입(process intervention)에 따른 연속적 상태 전이를 모델링하고 forward evolution과 inverse planning을 동시에 수행하는 프레임워크이다.</summary>
  </entry>
  <entry>
    <title>Listening Through the Noise: Cauchy-Driven Diffusion Bridges for Robust Gastrointestinal Auscultation and Clinical Benchmarking</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9949_Listening_Through_the_Noise_Cauchy-Driven_Diffusion_Bridges/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9949_Listening_Through_the_Noise_Cauchy-Driven_Diffusion_Bridges/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Dian Ding</name></author>
    <author><name>Liren Dong</name></author>
    <author><name>Yu Lu</name></author>
    <author><name>Juntao Zhou</name></author>
    <author><name>Ran Wang</name></author>
    <category term="Scientific Machine Learning for Dynamics"/>
    <summary type="text">임상 환경에서 발생하는 스피치 간섭과 같은 heavy-tailed non-stationary noise로부터 bowel sound(BS)를 복원하기 위해, Gaussian 기반 diffusion bridge 대신 Cauchy 분포를 driver로 사용하는 Cauchy-driven Diffusion Bridge를 제안하고, 이를 검증할 대규모 임상 데이터셋 CLINBS를 구축했다.</summary>
  </entry>
  <entry>
    <title>LipoPU: Pocket-level Prediction of Lipid-Protein Interactions via Positive-Unlabeled Learning</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9948_LipoPU_Pocket-level_Prediction_of_Lipid-Protein_Interactions/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9948_LipoPU_Pocket-level_Prediction_of_Lipid-Protein_Interactions/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Yuxing Wang</name></author>
    <author><name>Wenyi Zhang</name></author>
    <author><name>Yilong Zou</name></author>
    <author><name>Jing Huang</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">LipoPU는 지질-단백질 결합 예측을 pocket(포켓) 수준에서 수행하는 모델로, ranking 기반 positive-unlabeled (PU) learning 목적함수를 도입해 미확인(unlabeled) 샘플을 음성(negative)으로 잘못 취급하는 문제를 해결하고, binary lipid-binding detection과 multi-label lipid category prediction을 동시에 지원한다.</summary>
  </entry>
  <entry>
    <title>Linguistic Properties and Model Scale in Brain Encoding: From Small to Compressed Language Models</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9947_Linguistic_Properties_and_Model_Scale_in_Brain_Encoding_From/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9947_Linguistic_Properties_and_Model_Scale_in_Brain_Encoding_From/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>SUBBA REDDY OOTA</name></author>
    <author><name>Vijay Rowtula</name></author>
    <author><name>Satya Sai Srinath Namburi GNVV</name></author>
    <author><name>Khushbu Pahwa</name></author>
    <author><name>Anant Khandelwal</name></author>
    <category term="Multimodal Biomedical Data Fusion"/>
    <summary type="text">모델 규모와 압축(quantization, pruning)이 뇌 정합도(brain alignment)에 미치는 영향을 체계적으로 분석하여, 3B 규모의 small language models(SLMs)가 14B급 LLM과 동등한 뇌 예측력을 보이며, 대부분의 압축 기법이 언어적 성능 저하에도 불구하고 뇌 정합도를 유지함을 밝힌 연구이다.&lt;/essence&gt; &lt;fig_essence"&gt;2&lt;/fig_essence&gt; &lt;known&gt;기존 연구들은 LLM의 규모가 커질수록 fMRI 기반 뇌 활동과의 정합도(brain alignment)가 향상되는</summary>
  </entry>
  <entry>
    <title>Linear-LLM-SCM: Benchmarking LLMs for Coefficient Elicitation in Linear-Gaussian Causal Models</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9946_Linear-LLM-SCM_Benchmarking_LLMs_for_Coefficient_Elicitation/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9946_Linear-LLM-SCM_Benchmarking_LLMs_for_Coefficient_Elicitation/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Kanta Yamaoka</name></author>
    <author><name>Sumantrak Mukherjee</name></author>
    <author><name>Thomas Gärtner</name></author>
    <author><name>David Antony Selby</name></author>
    <author><name>Stefan Konigorski</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">이 논문은 사전에 주어진 DAG 구조에서 LLM이 Linear Gaussian structural causal model의 회귀 계수(coefficient)를 얼마나 정확하고 안정적으로 추정할 수 있는지를 평가하는 Linear-LLM-SCM이라는 벤치마킹 프레임워크를 제안한다.</summary>
  </entry>
  <entry>
    <title>LineageFlow: Flow Matching for High-Fidelity Family-Aware Protein Sequence Generation</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9945_LineageFlow_Flow_Matching_for_High-Fidelity_Family-Aware_Pro/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9945_LineageFlow_Flow_Matching_for_High-Fidelity_Family-Aware_Pro/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Langzhang Liang</name></author>
    <author><name>Ming Yang</name></author>
    <author><name>Feng Yi</name></author>
    <author><name>Junfan Li</name></author>
    <author><name>Shirui Pan</name></author>
    <category term="Computational Molecular Design"/>
    <summary type="text">LineageFlow는 진화적 조상 서열(ancestral sequence reconstruction)로부터 얻은 lineage prior를 초기 분포로 사용하는 simplex-valued flow matching 모델로, 단백질 서열 생성을 "무에서부터의 합성"이 아닌 "진화된 스캐폴드로부터의 구조화된 돌연변이"로 재구성한다.&lt;/essence&gt; &lt;fig_essence&gt;1&lt;/fig_essence&gt; &lt;known&gt;기존 discrete generative model(diffusion, flow matching 등)은 protein lan</summary>
  </entry>
  <entry>
    <title>Lightweight Alignment of Unimodal Foundation Models for Metabolite Identification</title>
    <link href="https://paper-curation.jehyunlee.dev/papers/9944_Lightweight_Alignment_of_Unimodal_Foundation_Models_for_Meta/" rel="alternate" type="text/html"/>
    <id>https://paper-curation.jehyunlee.dev/papers/9944_Lightweight_Alignment_of_Unimodal_Foundation_Models_for_Meta/</id>
    <updated>2026-01-01T00:00:00Z</updated>
    <author><name>Paul Krzakala</name></author>
    <author><name>Gabriel Melo</name></author>
    <author><name>Camille Lançon</name></author>
    <author><name>Charlotte Laclau</name></author>
    <author><name>Rémi Flamary</name></author>
    <category term="Multimodal Biomedical Data Fusion"/>
    <summary type="text">사전학습된 unimodal foundation model인 ChemBERTa(분자)와 DreaMS(질량 스펙트럼)를 동결한 채, 경량 projection head만 학습하여 공유 임베딩 공간으로 정렬하는 MSAlign을 제안하고, metabolite identification 검색 벤치마크에서 새로운 state-of-the-art 성능을 달성함.</summary>
  </entry>
</feed>
