Uncovering and Mitigating Aggregation-Induced Reward Hacking in Multi-Reward Reinforcement Learning

• URL: http://arxiv.org/abs/2609.00213v1
• Date: Mon, 31 Aug 2026 18:24:22 GMT
• Title: Uncovering and Mitigating Aggregation-Induced Reward Hacking in Multi-Reward Reinforcement Learning
• Abstract summary: Adaptive Multi-Reward Projection (AMRP) is a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress.<n>AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines.
• Score: 10.293423808492678
• License: http://creativecommons.org/licenses/by/4.0/
• Abstract: Reinforcement learning fine-tuning of large language models increasingly adopts multiple reward dimensions, including verifiable rules, task-specific evaluators, and learned reward models, to provide richer supervision across diverse capabilities. These dimensions are commonly scalarized with fixed aggregation weights. We identify a failure mode in which aggregation itself induces reward hacking: static projection aliases qualitatively different reward profiles into a single scalar, steering optimization toward whichever dimensions are easiest, densest, or systematically favored by the reward signal. Over training, this traps the policy in suboptimal profiles and prevents convergence to better-balanced ones that would yield higher task performance. To address this, we propose Adaptive Multi-Reward Projection (AMRP), a lightweight online method that reallocates aggregation weights using three signals, relative shortfall, reward volatility, and recent progress, increasing pressure on lagging, unstable, or stagnant dimensions while relieving saturated ones. Across structured reasoning, citation-grounded generation, and open-ended alignment under GRPO, AMRP consistently improves reward-profile balance and downstream performance over fixed and dynamic weighting baselines; it also remains effective with GDPO and PPO, supporting compatibility across RL algorithms. Our code is available at https://github.com/yyhappier/AMRP.git.

Related papers

• GD$^2$PO: Mitigating Multi-Reward Conflicts via Group-Dynamic reward-Decoupled Policy Optimization [54.596224644751565]
  Post-training reinforcement learning (RL) increasingly relies on multi-dimensional rewards to cultivate comprehensive capabilities.<n>To address this, existing methods such as Group reward-Decoupled Policy Optimization (GDPO) decompose the overall score into independent reward groups, then compute the RL loss separately within each group.<n>We propose Group-Dynamic reward-Decoupled Policy Optimization (GD$2$PO), which employs a conflict-aware filtering mechanism to mask out rollouts suffering from severe reward-wise disagreement.
  arXiv  Detail & Related papers  (2026-06-15T14:19:28Z)
• SAW: Stage-Aware Dynamic Weighting for Multi-Objective Reinforcement Learning in Large Language Models [66.92079232778707]
  We propose Stage-Aware Dynamic Weighting (SAW) as a lightweight, algorithm-agnostic dynamic weighting mechanism.<n>SAW reweights each dimension's reward or advantage contribution by its relative informativeness within the batch.<n> Experiments on tool-calling and text summarization tasks demonstrate that SAW consistently improves both training efficiency and final performance.
  arXiv  Detail & Related papers  (2026-06-05T10:00:19Z)
• Power Reinforcement Post-Training of Text-to-Image Models with Super-Linear Advantage Shaping [66.25536973294726]
  Post-training methods for text-to-image (T2I) models are prone to reward hacking.<n>Super-Linear Advantage Shaping (SLAS) reshapes the local policy space.<n>SLAS consistently surpasses the DanceGRPO baseline across multiple backbones and benchmarks.
  arXiv  Detail & Related papers  (2026-05-11T17:59:25Z)
• POETS: Uncertainty-Aware LLM Optimization via Compute-Efficient Policy Ensembles [35.88280776565273]
  POETS is a novel framework that bridges uncertainty quantification and policy optimization.<n>Our approach is grounded in the insight that policies trained with Kullback-Leibler (KL) regularization implicitly encode an underlying reward function.<n>We demonstrate POETS achieves state-of-the-art sample efficiency across diverse scientific discovery domains.
  arXiv  Detail & Related papers  (2026-05-08T14:16:32Z)
• MAESTRO: Meta-learning Adaptive Estimation of Scalarization Trade-offs for Reward Optimization [56.074760766965085]
  Group-Relative Policy Optimization has emerged as an efficient paradigm for aligning Large Language Models (LLMs)<n>We propose MAESTRO, which treats reward scalarization as a dynamic latent policy, leveraging the model's terminal hidden states as a semantic bottleneck.<n>We formulate this as a contextual bandit problem within a bi-level optimization framework, where a lightweight Conductor network co-evolves with the policy by utilizing group-relative advantages as a meta-reward signal.
  arXiv  Detail & Related papers  (2026-01-12T05:02:48Z)
• Implicit Actor Critic Coupling via a Supervised Learning Framework for RLVR [25.56828724912418]
  Recent advances in Reinforcement Learning with Verifiable Rewards (RLVR) have empowered large language models (LLMs) to tackle challenging reasoning tasks such as mathematics and programming.<n>Despite its promise, the RLVR paradigm poses significant challenges, as existing methods often suffer from sparse reward signals and unstable policy gradient updates.<n>We propose $textbfPACS$, a novel RLVR framework that achieves im$textbfP$licit $textbfA$ctor $textbfC$ritic coupling.
  arXiv  Detail & Related papers  (2025-09-02T17:22:46Z)
• Not All Rollouts are Useful: Down-Sampling Rollouts in LLM Reinforcement Learning [55.15106182268834]
  Reinforcement learning with verifiable rewards (RLVR) has emerged as the leading approach for enhancing reasoning capabilities in large language models.<n>It faces a fundamental compute and memory asymmetry: rollout generation is embarrassingly parallel and memory-light, whereas policy updates are communication-heavy and memory-intensive.<n>We introduce PODS (Policy Optimization with Down-Sampling), which decouples rollout generation from policy updates by training only on a strategically selected subset of rollouts.
  arXiv  Detail & Related papers  (2025-04-18T17:49:55Z)
• Optimizing Safe and Aligned Language Generation: A Multi-Objective GRPO Approach [2.8626097661711394]
  Reinforcement Learning from Human Feedback has achieved notable success in steering models, but is complex and can be unstable.<n>Recent approaches such as Direct Preference Optimization (DPO) simplify preference-based fine-tuning but may introduce bias or trade-off certain objectives.<n>We propose a Group Relative Policy Optimization framework with a multi-label reward regression model to achieve safe and aligned language generation.
  arXiv  Detail & Related papers  (2025-03-26T05:50:33Z)
• MPO: An Efficient Post-Processing Framework for Mixing Diverse Preference Alignment [14.541973333460149]
  Mixing Preference Optimization (MPO) is a post-processing framework for aggregating single-objective policies.<n>MPO achieves balanced performance across diverse preferences, outperforming existing models with significantly reduced computational costs.
  arXiv  Detail & Related papers  (2025-02-25T23:22:12Z)
• REBEL: Reinforcement Learning via Regressing Relative Rewards [59.68420022466047]
  We propose REBEL, a minimalist RL algorithm for the era of generative models.<n>In theory, we prove that fundamental RL algorithms like Natural Policy Gradient can be seen as variants of REBEL.<n>We find that REBEL provides a unified approach to language modeling and image generation with stronger or similar performance as PPO and DPO.
  arXiv  Detail & Related papers  (2024-04-25T17:20:45Z)

This list is automatically generated from the titles and abstracts of the papers in this site.

This site does not guarantee the quality of this site (including all information) and is not responsible for any consequences.