본문으로 건너뛰기

#Large Language Models

374개의 포스트

[논문리뷰] ThinkTwice: Jointly Optimizing Large Language Models for Reasoning and Self-Refinement

본 논문은 Reasoning 최적화와 Self-Refinement 최적화를 하나의 GRPO 프레임워크 안에서 결합한 ThinkTwice를 제안합니다. ThinkTwice는 각 훈련 단계에서 모델이 먼저 Reasoning 문제를 풀고, 동일한 문제에 대해 자신의 이전 답변을 개선하는(Thinking twice) 과정을 연속적으로 수행합니다 .
#Review#Large Language Models#Reinforcement Learning#Reasoning#Self-Refinement#RLVR#Policy Optimization#Implicit Curriculum
댓글 수 로딩 중

[논문리뷰] Paper Circle: An Open-source Multi-agent Research Discovery and Analysis Framework

본 논문은 문헌 탐색부터 분석, 비평까지 연구 전 과정을 지원하는 다중 에이전트 플랫폼인 Paper Circle을 제안한다. 시스템은 Discovery Pipeline과 Analysis Pipeline으로 구성되며, 연구자는 CodeAgent(CoA) 기반의 오케스트레이션 프레임워크를 통해 결정론적이고 추적 가능한 결과를 얻을 수 있다 , .
#Review#Multi-agent Systems#Research Discovery#Knowledge Graph#Large Language Models#Academic Literature Review
댓글 수 로딩 중

[논문리뷰] MegaTrain: Full Precision Training of 100B+ Parameter Large Language Models on a Single GPU

MegaTrain은 파라미터와 옵티마이저 상태를 호스트 메모리에 영구 저장하고, GPU는 레이어별 연산 시에만 파라미터를 스트리밍하는 구조를 채택한다. 저자들은 CPU-GPU 대역폭 병목을 극복하기 위해 Pipelined Double-Buffered Execution Engine을 도입하여 연산과 데이터 이동을 비동기적으로 중첩시켰다 .
#Review#Large Language Models#Memory-Centric#CPU-GPU Offloading#Parameter Streaming#Stateless Execution#Pipelined Execution#Single-GPU Training
댓글 수 로딩 중

[논문리뷰] LightThinker++: From Reasoning Compression to Memory Management

저자들은 암시적 압축에서 시작하여 명시적 행동 수준의 관리로 진화하는 LightThinker 계열 모델을 제안한다. LightThinker는 gist tokens와 특수 설계된 attention mask를 활용하여 긴 사고 과정을 컴팩트한 표현으로 변환한다 .
#Review#Large Language Models#Reasoning Compression#Memory Management#Agentic Reasoning#Context Optimization
댓글 수 로딩 중

[논문리뷰] DataFlex: A Unified Framework for Data-Centric Dynamic Training of Large Language Models

본 논문은 LLaMA-Factory의 모델 관리 및 최적화 기능을 보존하면서, 데이터 최적화 전략을 모듈형으로 통합한 DataFlex를 제안합니다. DataFlex는 7개의 데이터 선택, 2개의 데이터 혼합, 1개의 재가중치 알고리즘을 통합하여 단일 인터페이스 하에서 연구 및 실행할 수 있도록 설계되었습니다 .
#Review#Data-Centric AI#Large Language Models#Dynamic Training#Data Selection#Data Mixture#Data Reweighting#LLaMA-Factory
댓글 수 로딩 중

[논문리뷰] Reasoning Shift: How Context Silently Shortens LLM Reasoning

본 연구는 모델이 동일한 문제를 Baseline, Subtask, Long input, Multi-turn 등 서로 다른 4가지 Context 조건에서 해결하도록 설계하여 추론 성능과 생성된 토큰 수를 측정했습니다. 실험 결과, 모델은 고립된 환경(Baseline) 대비 다른 조건들에서 최대 50%까지 짧은 추론 트레이스를 생성하는 현상을 보였습니다.
#Review#Large Language Models#Chain-of-Thought#Test-time Scaling#Context Management#Reasoning Shift#Self-verification#Overthinking
댓글 수 로딩 중

[논문리뷰] MemRerank: Preference Memory for Personalized Product Reranking

저자들은 사용자 구매 이력을 within-category와 cross-category 선호도로 구분하여 추출하는 Memory Extractor 모델을 설계하였다. 이 추출기는 재순위화 작업의 성능을 극대화하기 위해 GRPO 기반의 강화학습으로 post-training 되며, 이를 통해 별도의 Gold Label 없이도 재순위화 보상을 직접 최적화한다 .
#Review#Recommender Systems#Large Language Models#Product Reranking#Preference Memory#Reinforcement Learning
댓글 수 로딩 중

[논문리뷰] Conv-FinRe: A Conversational and Longitudinal Benchmark for Utility-Grounded Financial Recommendation

arXiv에 게시된 'Conv-FinRe: A Conversational and Longitudinal Benchmark for Utility-Grounded Financial Recommendation' 논문에 대한 자세한 리뷰입니다.
#Review#Financial Recommendation#Conversational AI#Large Language Models#Utility-Grounded Evaluation#Behavioral Finance#Stock Recommendation#Longitudinal Benchmark#Inverse Optimization
댓글 수 로딩 중

[논문리뷰] WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning

arXiv에 게시된 'WideSeek-R1: Exploring Width Scaling for Broad Information Seeking via Multi-Agent Reinforcement Learning' 논문에 대한 자세한 리뷰입니다.
#Review#Multi-Agent Reinforcement Learning#Width Scaling#Large Language Models#Information Seeking#Task Decomposition#Parallel Execution#Lead-Agent-Subagent Framework#Orchestration
댓글 수 로딩 중

[논문리뷰] A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5

Yutao Wu이 arXiv에 게시한 'A Safety Report on GPT-5.2, Gemini 3 Pro, Qwen3-VL, Doubao 1.8, Grok 4.1 Fast, Nano Banana Pro, and Seedream 4.5' 논문에 대한 자세한 리뷰입니다.
#Review#AI Safety#Large Language Models#Multimodal LLMs#Benchmark Evaluation#Adversarial Robustness#Multilingual Evaluation#Regulatory Compliance#Image Generation Safety
댓글 수 로딩 중

[논문리뷰] Are LLMs Vulnerable to Preference-Undermining Attacks (PUA)? A Factorial Analysis Methodology for Diagnosing the Trade-off between Preference Alignment and Real-World Validity

Chi Zhang이 arXiv에 게시한 'Are LLMs Vulnerable to Preference-Undermining Attacks (PUA)? A Factorial Analysis Methodology for Diagnosing the Trade-off between Preference Alignment and Real-World Validity' 논문에 대한 자세한 리뷰입니다.
#Review#Large Language Models#Preference Alignment#Preference-Undermining Attacks#Factorial Analysis#Sycophancy#Prompt Engineering#Truth-Deference Trade-off
댓글 수 로딩 중

[논문리뷰] EpiQAL: Benchmarking Large Language Models in Epidemiological Question Answering for Enhanced Alignment and Reasoning

Guanchen Wu이 arXiv에 게시한 'EpiQAL: Benchmarking Large Language Models in Epidemiological Question Answering for Enhanced Alignment and Reasoning' 논문에 대한 자세한 리뷰입니다.
#Review#Epidemiological Question Answering#Large Language Models#Benchmark#Multi-step Inference#Evidence Grounding#LLM Evaluation#Public Health AI#Chain-of-Thought
댓글 수 로딩 중

[논문리뷰] Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem

Wei Gao이 arXiv에 게시한 'Let It Flow: Agentic Crafting on Rock and Roll, Building the ROME Model within an Open Agentic Learning Ecosystem' 논문에 대한 자세한 리뷰입니다.
#Review#Agentic Learning Ecosystem#Large Language Models#Reinforcement Learning#Agentic Crafting#Tool Use#ROME Model#Policy Optimization#Sandbox Environment
댓글 수 로딩 중

[논문리뷰] Multi-LLM Thematic Analysis with Dual Reliability Metrics: Combining Cohen's Kappa and Semantic Similarity for Qualitative Research Validation

arXiv에 게시된 'Multi-LLM Thematic Analysis with Dual Reliability Metrics: Combining Cohen's Kappa and Semantic Similarity for Qualitative Research Validation' 논문에 대한 자세한 리뷰입니다.
#Review#Thematic Analysis#Large Language Models#Qualitative Research#Cohen's Kappa#Semantic Similarity#Reliability Metrics#Ensemble Validation#Prompt Engineering
댓글 수 로딩 중

[논문리뷰] Can LLMs Estimate Student Struggles? Human-AI Difficulty Alignment with Proficiency Simulation for Item Difficulty Prediction

Hong Jiao이 arXiv에 게시한 'Can LLMs Estimate Student Struggles? Human-AI Difficulty Alignment with Proficiency Simulation for Item Difficulty Prediction' 논문에 대한 자세한 리뷰입니다.
#Review#Large Language Models#Item Difficulty Prediction#Human-AI Alignment#Proficiency Simulation#Metacognition#Curse of Knowledge#Educational Assessment#Zero-shot Learning
댓글 수 로딩 중

[논문리뷰] Beyond Token-level Supervision: Unlocking the Potential of Decoding-based Regression via Reinforcement Learning

Jiacheng Chen이 arXiv에 게시한 'Beyond Token-level Supervision: Unlocking the Potential of Decoding-based Regression via Reinforcement Learning' 논문에 대한 자세한 리뷰입니다.
#Review#Decoding-based Regression#Reinforcement Learning#Numerical Prediction#Large Language Models#Policy Gradient#Tokenization#Sequence Generation
댓글 수 로딩 중

[논문리뷰] From Imitation to Discrimination: Toward A Generalized Curriculum Advantage Mechanism Enhancing Cross-Domain Reasoning Tasks

Yang Li이 arXiv에 게시한 'From Imitation to Discrimination: Toward A Generalized Curriculum Advantage Mechanism Enhancing Cross-Domain Reasoning Tasks' 논문에 대한 자세한 리뷰입니다.
#Review#Reinforcement Learning#Large Language Models#Curriculum Learning#Advantage Function#Reasoning Tasks#Multimodal AI#Policy Optimization#Generalization
댓글 수 로딩 중

[논문리뷰] MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling

cyyang822이 arXiv에 게시한 'MiroThinker: Pushing the Performance Boundaries of Open-Source Research Agents via Model, Context, and Interactive Scaling' 논문에 대한 자세한 리뷰입니다.
#Review#Research Agent#Tool-Augmented Reasoning#Interaction Scaling#Large Language Models#Reinforcement Learning#Context Management#Open-Source AI
댓글 수 로딩 중

[논문리뷰] BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping

Junrui Shen이 arXiv에 게시한 'BAPO: Stabilizing Off-Policy Reinforcement Learning for LLMs via Balanced Policy Optimization with Adaptive Clipping' 논문에 대한 자세한 리뷰입니다.
#Review#Off-Policy Reinforcement Learning#Large Language Models#Adaptive Clipping#Policy Optimization#PPO#Entropy Preservation#RL Stabilization
댓글 수 로딩 중

[논문리뷰] PokeeResearch: Effective Deep Research via Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold

arXiv에 게시된 'PokeeResearch: Effective Deep Research via Reinforcement Learning from AI Feedback and Robust Reasoning Scaffold' 논문에 대한 자세한 리뷰입니다.
#Review#Deep Research Agent#Reinforcement Learning from AI Feedback#RLOO Algorithm#Large Language Models#Tool Use#Self-Correction#Reasoning Scaffold#Agent Alignment
댓글 수 로딩 중

[논문리뷰] Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain

Lingxi Lu이 arXiv에 게시한 'Hierarchical Frequency Tagging Probe (HFTP): A Unified Approach to Investigate Syntactic Structure Representations in Large Language Models and the Human Brain' 논문에 대한 자세한 리뷰입니다.
#Review#Large Language Models#Syntactic Structure#Human Brain#Frequency Tagging#Neuroscience#Model Interpretability#Representational Similarity Analysis#Intracranial EEG
댓글 수 로딩 중

[논문리뷰] Recycling Pretrained Checkpoints: Orthogonal Growth of Mixture-of-Experts for Efficient Large Language Model Pre-Training

Peng Cheng이 arXiv에 게시한 'Recycling Pretrained Checkpoints: Orthogonal Growth of Mixture-of-Experts for Efficient Large Language Model Pre-Training' 논문에 대한 자세한 리뷰입니다.
#Review#Mixture-of-Experts#Large Language Models#Checkpoint Recycling#Model Growth#Efficient Pretraining#Depth Growth#Width Growth#Sunk Cost
댓글 수 로딩 중

[논문리뷰] Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints

Huazhe Xu이 arXiv에 게시한 'Entropy Regularizing Activation: Boosting Continuous Control, Large Language Models, and Image Classification with Activation as Entropy Constraints' 논문에 대한 자세한 리뷰입니다.
#Review#Entropy Regularization#Activation Functions#Continuous Control#Large Language Models#Image Classification#Reinforcement Learning#Policy Stochasticity#Entropy Constraints
댓글 수 로딩 중

[논문리뷰] EvolProver: Advancing Automated Theorem Proving by Evolving Formalized Problems via Symmetry and Difficulty

Xuanwu Wang이 arXiv에 게시한 'EvolProver: Advancing Automated Theorem Proving by Evolving Formalized Problems via Symmetry and Difficulty' 논문에 대한 자세한 리뷰입니다.
#Review#Automated Theorem Proving#Data Augmentation#Large Language Models#Formal Mathematics#Symmetry#Difficulty Evolution#Abstract Syntax Tree#Generalizability
댓글 수 로딩 중

[논문리뷰] DIWALI - Diversity and Inclusivity aWare cuLture specific Items for India: Dataset and Assessment of LLMs for Cultural Text Adaptation in Indian Context

Maunendra Sankar Desarkar이 arXiv에 게시한 'DIWALI - Diversity and Inclusivity aWare cuLture specific Items for India: Dataset and Assessment of LLMs for Cultural Text Adaptation in Indian Context' 논문에 대한 자세한 리뷰입니다.
#Review#Cultural Adaptation#Large Language Models#Indian Culture#Dataset Creation#CSI#Human Evaluation#LLM Evaluation#Cultural Bias
댓글 수 로딩 중

[논문리뷰] Multimodal Reasoning for Science: Technical Report and 1st Place Solution to the ICML 2025 SeePhys Challenge

Wentao Zhang이 arXiv에 게시한 'Multimodal Reasoning for Science: Technical Report and 1st Place Solution to the ICML 2025 SeePhys Challenge' 논문에 대한 자세한 리뷰입니다.
#Review#Multimodal Reasoning#Science AI#Caption-assisted Reasoning#SeePhys Challenge#Large Language Models#Visual Question Answering#Physics Problems#Cross-modal Alignment
댓글 수 로딩 중

[논문리뷰] AMBEDKAR-A Multi-level Bias Elimination through a Decoding Approach with Knowledge Augmentation for Robust Constitutional Alignment of Language Models

Rahul Karthikeyan이 arXiv에 게시한 'AMBEDKAR-A Multi-level Bias Elimination through a Decoding Approach with Knowledge Augmentation for Robust Constitutional Alignment of Language Models' 논문에 대한 자세한 리뷰입니다.
#Review#Bias Mitigation#Large Language Models#Speculative Decoding#Constitutional AI#Fairness#Inference-Time Control#Indian Sociocultural Context
댓글 수 로딩 중

[논문리뷰] TreePO: Bridging the Gap of Policy Optimization and Efficacy and Inference Efficiency with Heuristic Tree-based Modeling

Zhoufutu Wen이 arXiv에 게시한 'TreePO: Bridging the Gap of Policy Optimization and Efficacy and Inference Efficiency with Heuristic Tree-based Modeling' 논문에 대한 자세한 리뷰입니다.
#Review#Reinforcement Learning#Policy Optimization#Large Language Models#Inference Efficiency#Tree Search#Segment-level Decoding#Advantage Estimation#Reasoning
댓글 수 로딩 중

[논문리뷰] Breaking the Exploration Bottleneck: Rubric-Scaffolded Reinforcement Learning for General LLM Reasoning

Jiale Zhao이 arXiv에 게시한 'Breaking the Exploration Bottleneck: Rubric-Scaffolded Reinforcement Learning for General LLM Reasoning' 논문에 대한 자세한 리뷰입니다.
#Review#Reinforcement Learning#Large Language Models#Exploration Bottleneck#Instructional Scaffolding#Rubric-based Rewards#General Reasoning#RL with Verifiable Rewards#Policy Optimization
댓글 수 로딩 중

[논문리뷰] On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting

Guoyin Wang이 arXiv에 게시한 'On-Policy RL Meets Off-Policy Experts: Harmonizing Supervised Fine-Tuning and Reinforcement Learning via Dynamic Weighting' 논문에 대한 자세한 리뷰입니다.
#Review#Large Language Models#Reinforcement Learning#Supervised Fine-Tuning#On-Policy RL#Off-Policy Experts#Dynamic Weighting#LLM Alignment#Reasoning
댓글 수 로딩 중

[논문리뷰] Can LLM-Generated Textual Explanations Enhance Model Classification Performance? An Empirical Study

Gjergji Kasneci이 arXiv에 게시한 'Can LLM-Generated Textual Explanations Enhance Model Classification Performance? An Empirical Study' 논문에 대한 자세한 리뷰입니다.
#Review#Explainable NLP#Natural Language Explanations#Large Language Models#Pre-trained Language Models#Natural Language Inference#Model Performance Enhancement#Text Generation
댓글 수 로딩 중

[논문리뷰] Visual Document Understanding and Question Answering: A Multi-Agent Collaboration Framework with Test-Time Scaling

Ruolin Shen이 arXiv에 게시한 'Visual Document Understanding and Question Answering: A Multi-Agent Collaboration Framework with Test-Time Scaling' 논문에 대한 자세한 리뷰입니다.
#Review#Visual Document Understanding#Visual Question Answering#Multi-Agent System#Test-Time Scaling#Self-Correction#Mixed Reward Modeling#Large Language Models
댓글 수 로딩 중

[논문리뷰] RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization

Kechi Zhang이 arXiv에 게시한 'RL-PLUS: Countering Capability Boundary Collapse of LLMs in Reinforcement Learning with Hybrid-policy Optimization' 논문에 대한 자세한 리뷰입니다.
#Review#Large Language Models#Reinforcement Learning#Capability Collapse#Hybrid Policy Optimization#Multiple Importance Sampling#Exploration#Math Reasoning#Out-of-Distribution
댓글 수 로딩 중