GRPO normalization breaks dense prediction rewards for LLM agents
Researchers identified a critical failure mode in GRPO-trained language model agents using dense prediction rewards. When Qwen models were trained to predict next observations as a reward signal on ALFWorld tasks, the optimizer converged to a degenerate state where prediction accuracy maximized while task success collapsed to zero. The root cause traces to GRPO's standard deviation normalization, which creates pathological incentives in failure-dominated groups. Removing this single component restored baseline performance, suggesting that reward shaping for long-horizon agent training requires careful interaction with normalization schemes. This finding has immediate implications for anyone scaling RL-based agent training and highlights how optimizer design choices can silently break seemingly reasonable reward structures.
Modelwire context
ExplainerThe critical finding isn't just that dense prediction rewards fail on GRPO agents, but that the failure mechanism traces to a single component (standard deviation normalization) that creates perverse incentives when most trajectories end in failure. This is a sharp, actionable diagnosis, not a vague 'reward shaping is hard' observation.
This work sits in a different layer than recent agent-training coverage. While GRADRAG (from this week) addresses multi-agent coordination in RAG pipelines and the multi-task learning paper focuses on prediction efficiency, this paper identifies a fundamental pathology in how RL optimizers interact with reward signals. The closer parallel is the moral-value dataset work, which also flags that training signal design matters as much as scale. Both papers suggest that practitioners building systems at scale need to audit their training mechanics, not just their data or architecture choices.
If the GRPO standard deviation fix generalizes to other dense-reward RL setups beyond ALFWorld (e.g., robotic manipulation or code generation benchmarks), it confirms the root cause is optimizer-agnostic. If other groups report similar collapse patterns with GRPO on their own tasks within the next 3-4 months without applying this fix, that would validate the diagnosis and accelerate adoption of the patch.
Coverage we drew on
- A Unified Moral-Value Dataset for Instruction Tuning · arXiv cs.CL
This analysis is generated by Modelwire’s editorial layer from our archive and the summary above. It is not a substitute for the original reporting. How we write it.
MentionsQwen3 · GRPO · ALFWorld · arXiv
Modelwire Editorial
This synthesis and analysis was prepared by the Modelwire editorial team. We use advanced language models to read, ground, and connect the day’s most significant AI developments, providing original strategic context that helps practitioners and leaders stay ahead of the frontier.
Modelwire summarizes, we don’t republish. arXiv cs.LG originally reported this story as “The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works”. The full content lives on arxiv.org. If you’re a publisher and want a different summarization policy for your work, see our takedown page.