(논문 요약) iGRPO: Self-Feedback–Driven LLM Reasoning (paper)
핵심 내용
- Iterative Group Relative Policy Optimization
- draft 여러개 생성후 verifiable reward 가 가장 높은 draft 선택
- selected draft 를 prompt 에 추가하여 다시 생성
- GRPO 로 학습
Table of contents
- (개념 요약) Tokenization
- (논문 요약) Agent Learning via Early Experience
- (논문 요약) Chat Vector; A Simple Approach to Equip LLMs with Instruction Following and Model Alignment in New Languages
- (논문 요약) Compute Optimal Scaling of Skills; Knowledge vs Reasoning
- (논문 요약) Compute as Teacher; Turning Inference Compute Into Reference-Free Supervision
- (논문 요약) Conditional Memory via Scalable Lookup; A New Axis of Sparsity for Large Language Models
- (논문 요약) DOES RLHF SCALE? EXPLORING THE IMPACTS FROM DATA, MODEL, AND METHOD
- (논문 요약) DR Tulu; Reinforcement Learning with Evolving Rubrics for Deep Research
- (논문 요약) Direct Preference Optimization; Your Language Model is Secretly a Reward Model
- (논문 요약) Disentangling the Factors of Convergence between Brains and Computer Vision Models
- (논문 요약) Do NOT Think That Much for 2+3=? On the Overthinking of o1-Like LLMs
- (논문 요약) EMERGENT HIERARCHICAL REASONING IN LLMS THROUGH REINFORCEMENT LEARNING
- (논문 요약) Evolution Strategies at the Hyperscale
- (논문 요약) GDPO; Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization
- (논문 요약) Generative Representational Instruction Tuning
- (논문 요약) Geometric-Mean Policy Optimization
- (논문 요약) Grounded Language-Image Pre-training
- (논문 요약) Group Sequence Policy Optimization
- (논문 요약) Guided Self-Evolving LLMs with Minimal Human Supervision
- (논문 요약) HOW TO TEACH LARGE MULTIMODAL MODELS NEW SKILLS?
- (논문 요약) Hunyuan-Large; An Open-Source MoE Model with 52 Billion Activated Parameters by Tencent
- (논문 요약) Inference-Time Scaling for Generalist Reward Modeling
- (논문 요약) LLAMA-OMNI; SEAMLESS SPEECH INTERACTION WITH LARGE LANGUAGE MODELS
- (논문 요약) Learning to Discover at Test Time
- (논문 요약) Learning to Reason in 13 Parameters
- (논문 요약) MAGPIE; Alignment Data Synthesis from Scratch by Prompting Aligned LLMs with Nothing
- (논문 요약) META-REWARDING LANGUAGE MODELS; Self-Improving Alignment with LLM-as-a-Meta-Judge
- (논문 요약) Make Your LLM Fully Utilize the Context
- (논문 요약) Maximum Likelihood Reinforcement Learning
- (논문 요약) One Initialization to Rule them All; Fine-tuning via Explained Variance Adaptation
- (논문 요약) Quantization-Aware Distillation for NVFP4 Inference Accuracy Recovery
- (논문 요약) QwenLong-L1.5; Post-Training Recipe for Long-Context Reasoning and Memory Management
- (논문 요약) Reinforcement Learning via Self-Distillation
- (논문 요약) RouteLLM; Learning to Route LLMs with Preference Data
- (논문 요약) SERA; Soft-Verified Efficient Repository Agents
- (논문 요약) SFT Memorizes, RL Generalizes; A Comparative Study of Foundation Model Post-training
- (논문 요약) STEM; Scaling Transformers with Embedding Modules
- (논문 요약) Scaling Exponents Across Parameterizations and Optimizers
- (논문 요약) Scaling Laws for Data Filtering—Data Curation cannot be Compute Agnostic
- (논문 요약) Scaling Laws for Precision
- (논문 요약) Self-Distillation Enables Continual Learning
- (논문 요약) Self-Taught Evaluators
- (논문 요약) SimPO; Simple Preference Optimization with a Reference-Free Reward
- (논문 요약) Simple and Scalable Strategies to Continually Pre-train Large Language Models
- (논문 요약) Smaller, Weaker, Yet Better; Training LLM Reasoners via Compute-Optimal Sampling
- (논문 요약) Soft Adaptive Policy Optimization
- (논문 요약) Stabilizing Reinforcement Learning with LLMs; Formulation and Practices
- (논문 요약) THINKING LLMS; GENERAL INSTRUCTION FOLLOWING WITH THOUGHT GENERATION
- (논문 요약) TTRL; Test-Time Reinforcement Learning
- (논문 요약) Teaching Models to Teach Themselves; Reasoning at the Edge of Learnability
- (논문 요약) Textbooks Are All You Need
- (논문 요약) The Path Not Taken; RLVR Provably Learns Off the Principals
- (논문 요약) Training Language Models to Self-Correct via Reinforcement Learning
- (논문 요약) Why Language Models Hallucinate
- Diversity or Precision? A Deep Dive into Next Token Prediction
- Gated Attention for Large Language Models; Non-linearity, Sparsity, and Attention-Sink-Free
- Miscellaneous Finetuning Methods in Large Language Models