Tag
本文提出 MRTPO，通过基于回放的相关性度量将每个奖励的优势只分配给相关 token，从而改进多奖励强化学习训练，加速收敛并提升在 text-to-SQL、多约束指令遵循和工具调用任务上的性能。
The paper introduces Density-Aware Reward Aggregation (DARA), which corrects reward-wise normalized multi-reward RL with an inverse-square-root density weighting based on active-group density, reducing signal imbalance between rewards. Experiments on tool calling and math reasoning show DARA reaches high format/length compliance in up to 26% and 65% fewer training steps than GDPO.
CorrGRPO proposes correlation-normalized GRPO that converts pairwise reward covariances into Pearson correlation coefficients, preventing large-scale rewards from dominating multi-reward normalization and improving performance on code generation, tool calling, and agent security across 0.5B-8B parameter models.
提出 Adaptive Reward Routing，在联合音频-视频扩散模型的前向过程强化学习（DiffusionNFT）中动态调整更新位置与奖励权重：利用双向跨注意力响应定位跨模态影响，并以偏好先验加残差修正的方式协调多个竞争奖励，从而提升模态质量、语义一致性与音视频同步。
This paper identifies aggregation-induced reward hacking in multi-reward reinforcement learning for LLMs and proposes an adaptive projection method, AMRP, to dynamically adjust weights for better reward balance and performance.
SA-MRPO introduces a saturation-aware advantage reweighting technique for multi-reward policy optimization in reinforcement learning, improving performance on mathematical reasoning, adaptive reasoning, and coding benchmarks by focusing optimization on under-optimized objectives.
GD^2PO introduces a conflict-aware filtering mechanism to mitigate multi-reward conflicts in reinforcement learning for large language models, preventing signal cancellation and accelerating training efficiency.
DVAO adaptively weights objectives based on reward variance to improve multi-reward RL training stability and multi-objective performance.
UniDoc-RL presents a reinforcement learning framework for Large Vision-Language Models that optimizes retrieval, reranking, and visual reasoning through hierarchical decision-making and dense multi-reward supervision, achieving up to 17.7% improvements over prior RL-based methods on visual RAG tasks.