Multi reward training improves large language model reliability and performance

Multi-Reward Hacking: From Aggregation-Induced Drift to Trust-Anchored Defense

Machine LearningArtificial Intelligence

Summary

When training large language models, combining different goals like accuracy and short answers can sometimes make the models worse at the main task. The authors show that treating these goals equally can cause the model to focus too much on easier-to-get goals, ignoring the main one. They propose a new method called Trust-Anchored Gating that helps the model prioritize the main goal while still improving on others. Their tests show this method helps models perform better and gives shorter, more efficient responses.

What this means in practice

  • For ai model trainers: Improve reinforcement learning training by preventing models from ignoring primary task goals when combining multiple rewards.
  • For software developers using llms: Generate more accurate and concise code and math solutions by applying trust-anchored reward methods during model fine-tuning.

Authors

Yushu Li, Jiin Woo, Xiaoxiao Li, Youngsuk Park

Abstract

To improve task performance while encouraging desirable response properties, LLM post-training often combines a primary reward (e.g., correctness) with auxiliary rewards (e.g., brevity or efficiency). However, how these rewards are aggregated can critically affect training outcomes. We show that symmetric aggregation, which treats reward channels equally, can induce a failure where aggregate reward improves while primary-task performance degrades, even when individual rewards are meaningful. We investigate this aggregation-induced failure through a one-step multi-reward decision formulation, characterizing when it constitutes multi-reward hacking. We prove that, under symmetric aggregation, deterministic conditional-mean GRPO converges to auxiliary-favoring behavior when auxiliary rewards are more reliably obtained, regardless of initialization. We propose Trust-Anchored Gating (TAG), which first elects an anchor channel based on the policy drift induced by optimizing each channel alone, then gates auxiliary credit on anchor satisfaction. Unlike calibration-sensitive static reweighting, TAG structurally eliminates credit for auxiliary-only outcomes while preserving additional credit for joint success. In the same analytical setting, TAG provably reverses this selection, converging to the primary-favoring action from any initialization when the primary channel is anchored. Experiments across Qwen and Llama models on mathematics and program synthesis reveal substantial degradation under symmetric, constrained, and normalized aggregation, whereas TAG mitigates this failure. On Qwen3-4B with a brevity reward, TAG achieves 0.581 pass@4 versus 0.220 for additive aggregation, while shortening responses by 44% relative to correctness-only training.