Asymmetric REINFORCE for off-Policy Reinforcement Learning: Balancing positive and negative rewards
원문 초록
Reinforcement learning (RL) is increasingly used to align large language models (LLMs). Off-policy methods offer greater implementation simplicity and data efficiency than on-policy techniques, but often result in suboptimal performance. In this work, we study the intermediate range of algorithms between off-policy RL and supervised fine-tuning by analyzing a simple off-policy REINFORCE algorithm, where the advantage is defined as $A=r-V$, with $r$ a reward and $V$ some tunable baseline. Intuitively, lowering $V$ emphasizes high-reward samples, while raising it penalizes low-reward ones more heavily. We first provide a theoretical analysis of this off-policy REINFORCE algorithm, showing that when the baseline $V$ lower-bounds the expected reward, the algorithm enjoys a policy improvement guarantee. Our analysis reveals that while on-policy updates can safely leverage both positive and negative signals, off-policy updates benefit from focusing more on positive rewards than on negative ones. We validate our findings experimentally in a controlled stochastic bandit setting and through fine-tuning state-of-the-art LLMs on reasoning tasks.
함께 살펴볼 논문
공통 태그를 근거로 연결했습니다. 실제 인용 관계나 논문의 우수성을 뜻하지 않습니다.
High-Throughput Distributed Reinforcement Learning via Adaptive Policy Synchronization
Personalized Exercise Recommendation with Semantically-Grounded Knowledge Tracing
Bridging the Gap in Vision Language Models in Identifying Unsafe Concepts Across Modalities
Exploring the robustness of TractOracle methods in RL-based tractography
Reasoning or Memorization? Unreliable Results of Reinforcement Learning Due to Data Contamination
Deep Reinforcement Learning with Gradient Eligibility Traces
토론
이 논문과 연결된 포럼 글이 아직 없습니다.