Image de couverture de livre

Mirage or Method? How Model-Task Alignment Induces Divergent RL Conclusions

Fuente: arXiv

Enregistré dans:

Détails bibliographiques
Auteurs principaux:	Wu, Haoze, Wang, Cheng, Zhao, Wenshuo, He, Junxian
Format:	Preprint
Publié:	2025
Sujets:	Machine Learning Computation and Language
Accès en ligne:	Acceder al recurso
Tags:	Ajouter un tag Pas de tags, Soyez le premier à ajouter un tag!

Documents similaires

Internalizing World Models via Self-Play Finetuning for Agentic RL
par: Chen, Shiqi, et autres
Publié: (2025)

SimpleRL-Zoo: Investigating and Taming Zero Reinforcement Learning for Open Base Models in the Wild
par: Zeng, Weihao, et autres
Publié: (2025)

Unveiling Causal Reasoning in Large Language Models: Reality or Mirage?
par: Chi, Haoang, et autres
Publié: (2025)

What Makes Good Data for Alignment? A Comprehensive Study of Automatic Data Selection in Instruction Tuning
par: Liu, Wei, et autres
Publié: (2023)

ToolSample: Dual Dynamic Sampling Methods with Curriculum Learning for RL-based Tool Learning
par: Feng, Zihao, et autres
Publié: (2025)

TACO-RL: Task Aware Prompt Compression Optimization with Reinforcement Learning
par: Shandilya, Shivam, et autres
Publié: (2024)

Objective Metrics for Evaluating Large Language Models Using External Data Sources
par: Du, Haoze, et autres
Publié: (2025)

Is Chain-of-Thought Reasoning of LLMs a Mirage? A Data Distribution Lens
par: Zhao, Chengshuai, et autres
Publié: (2025)

Evaluating Evaluation Metrics -- The Mirage of Hallucination Detection
par: Kulkarni, Atharva, et autres
Publié: (2025)

RL Grokking Recipe: How Does RL Unlock and Transfer New Algorithms in LLMs?
par: Sun, Yiyou, et autres
Publié: (2025)

DBR: Divergence-Based Regularization for Debiasing Natural Language Understanding Models
par: Li, Zihao, et autres
Publié: (2025)

Preserving Long-Tailed Expert Information in Mixture-of-Experts Tuning
par: He, Haoze, et autres
Publié: (2026)

Mirage Persistent Kernel: A Compiler and Runtime for Mega-Kernelizing Tensor Programs
par: Cheng, Xinhao, et autres
Publié: (2025)

Rewarding the Rare: Uniqueness-Aware RL for Creative Problem Solving in LLMs
par: Hu, Zhiyuan, et autres
Publié: (2026)

Learning from Failures: Understanding LLM Alignment through Failure-Aware Inverse RL
par: Patel, Nyal, et autres
Publié: (2025)

From Lists to Emojis: How Format Bias Affects Model Alignment
par: Zhang, Xuanchang, et autres
Publié: (2024)

HAL: Inducing Human-likeness in LLMs with Alignment
par: Hasan, Masum, et autres
Publié: (2026)

Better and Worse with Scale: How Contextual Entrainment Diverges with Model Size
par: Kukreja, Dikshant, et autres
Publié: (2026)

Capability-Oriented Training Induced Alignment Risk
par: Zhou, Yujun, et autres
Publié: (2026)

Synthetic Data RL: Task Definition Is All You Need
par: Guo, Yiduo, et autres
Publié: (2025)

Distributional Clarity: The Hidden Driver of RL-Friendliness in Large Language Models
par: Sun, Shaoning, et autres
Publié: (2026)

From Parameters to Data: A Task-Parameter-Guided Fine-Tuning Pipeline for Efficient LLM Alignment
par: Chen, Hao, et autres
Publié: (2026)

B-STaR: Monitoring and Balancing Exploration and Exploitation in Self-Taught Reasoners
par: Zeng, Weihao, et autres
Publié: (2024)

How Does Code Pretraining Affect Language Model Task Performance?
par: Petty, Jackson, et autres
Publié: (2024)

From Accuracy to Robustness: A Study of Rule- and Model-based Verifiers in Mathematical Reasoning
par: Huang, Yuzhen, et autres
Publié: (2025)

Weight-Inherited Distillation for Task-Agnostic BERT Compression
par: Wu, Taiqiang, et autres
Publié: (2023)

Attention as a Compass: Efficient Exploration for Process-Supervised RL in Reasoning Models
par: Liu, Runze, et autres
Publié: (2025)

Genius: A Generalizable and Purely Unsupervised Self-Training Framework For Advanced Reasoning
par: Xu, Fangzhi, et autres
Publié: (2025)

Mortgage Language Model: Domain-Adaptive Pretraining with Residual Instruction, Alignment Tuning, and Task-Specific Routing
par: Jain, Manish, et autres
Publié: (2025)

CodeMirage: A Multi-Lingual Benchmark for Detecting AI-Generated and Paraphrased Source Code from Production-Level LLMs
par: Guo, Hanxi, et autres
Publié: (2025)

Baichuan Alignment Technical Report
par: Lin, Mingan, et autres
Publié: (2024)

On-Policy RL with Optimal Reward Baseline
par: Hao, Yaru, et autres
Publié: (2025)

Noise Contrastive Alignment of Language Models with Explicit Rewards
par: Chen, Huayu, et autres
Publié: (2024)

How Contaminated Is Your Benchmark? Quantifying Dataset Leakage in Large Language Models with Kernel Divergence
par: Choi, Hyeong Kyu, et autres
Publié: (2025)

Towards Unified Task Embeddings Across Multiple Models: Bridging the Gap for Prompt-Based Large Language Models and Beyond
par: Wang, Xinyu, et autres
Publié: (2024)

ToolRL: Reward is All Tool Learning Needs
par: Qian, Cheng, et autres
Publié: (2025)

Hidden States Know Where Reasoning Diverges: Credit Assignment via Span-Level Wasserstein Distance
par: Chen, Xinzhu, et autres
Publié: (2026)

How to Leverage Demonstration Data in Alignment for Large Language Model? A Self-Imitation Learning Perspective
par: Xiao, Teng, et autres
Publié: (2024)

Convergence and Divergence of Language Models under Different Random Seeds
par: Fehlauer, Finlay, et autres
Publié: (2025)

Mirage: A Multi-Level Superoptimizer for Tensor Programs
par: Wu, Mengdi, et autres
Publié: (2024)