# The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and The Channel, Not the Content, Decides What Works

arXiv: https://arxiv.org/abs/2607.21273 · HTML: https://mlsanity.com/paper/2607.21273/

## Key facts

- Cited by 5 papers in the map (data through 14 Oct 2026).
- Placed in Reinforcement learning for reasoning, one of 17 directions in LLM reasoning.
- Builds on 25 of its 66 references that are in the map.
- First posted on arXiv 23 Jul 2026.

## Placed in

- LLM reasoning / [Reinforcement learning for reasoning](https://mlsanity.com/lr/reinforcement-learning-for-reasoning/)
- LLM agents and tool use / [Reinforcement learning for LLM agents](https://mlsanity.com/agents/reinforcement-learning-for-llm-agents/)

## Builds on (most cited first, 25 of 25)

- [DeepSeekMath: Pushing the Limits of Mathematical Reasoning in Open Language Models](https://mlsanity.com/paper/2402.03300/) (2024)
- [DeepSeek-R1: Incentivizing Reasoning Capability in LLMs via Reinforcement Learning](https://mlsanity.com/paper/2501.12948/) (2025)
- [DAPO: An Open-Source LLM Reinforcement Learning System at Scale](https://mlsanity.com/paper/2503.14476/) (2025)
- [Understanding R1-Zero-Like Training: A Critical Perspective](https://mlsanity.com/paper/2503.20783/) (2025)
- [Process Reinforcement through Implicit Rewards](https://mlsanity.com/paper/2502.01456/) (2025)
- [The Entropy Mechanism of Reinforcement Learning for Reasoning Language Models](https://mlsanity.com/paper/2505.22617/) (2025)
- [RAGEN: Understanding Self-Evolution in LLM Agents via Multi-Turn Reinforcement Learning](https://mlsanity.com/paper/2504.20073/) (2025)
- [Fine-Tuning Pretrained Language Models: Weight Initializations, Data Orders, and Early Stopping](https://mlsanity.com/paper/2002.06305/) (2020)
- [REINFORCE++: Stabilizing Critic-Free Policy Optimization with Global Advantage Normalization](https://mlsanity.com/paper/2501.03262/) (2025)
- [ArCHer: Training Language Model Agents via Hierarchical Multi-Turn RL](https://mlsanity.com/paper/2402.19446/) (2024)
- [GDPO: Group reward-Decoupled Normalization Policy Optimization for Multi-reward RL Optimization](https://mlsanity.com/paper/2601.05242/) (2026)
- [Small-scale proxies for large-scale Transformer training instabilities](https://mlsanity.com/paper/2309.14322/) (2023)
- [Reinforcement Learning with Verifiable Rewards: GRPO's Effective Loss, Dynamics, and Success Amplification](https://mlsanity.com/paper/2503.06639/) (2025)
- [Pref-GRPO: Pairwise Preference Reward-based GRPO for Stable Text-to-Image Reinforcement Learning](https://mlsanity.com/paper/2508.20751/) (2025)
- [Reward Shaping to Mitigate Reward Hacking in RLHF](https://mlsanity.com/paper/2502.18770/) (2025)
- [Scaf-GRPO: Scaffolded Group Relative Policy Optimization for Enhancing LLM Reasoning](https://mlsanity.com/paper/2510.19807/) (2025)
- [BNPO: Beta Normalization Policy Optimization](https://mlsanity.com/paper/2506.02864/) (2025)
- [Reinforcement World Model Learning for LLM-based Agents](https://mlsanity.com/paper/2602.05842/) (2026)
- [MO-GRPO: Mitigating Reward Hacking of Group Relative Policy Optimization on Multi-Objective Problems](https://mlsanity.com/paper/2509.22047/) (2025)
- [Uncalibrated Reasoning: GRPO Induces Overconfidence for Stochastic Outcomes](https://mlsanity.com/paper/2508.11800/) (2025)
- [ECHO: Terminal Agents Learn World Models for Free](https://mlsanity.com/paper/2605.24517/) (2026)
- [Policy and World Modeling Co-Training for Language Agents](https://mlsanity.com/paper/2606.02388/) (2026)
- [When does Self-Prediction help? Understanding Auxiliary Tasks in Reinforcement Learning](https://mlsanity.com/paper/2406.17718/) (2024)
- [Potential-Based Reward Shaping For Intrinsic Motivation](https://mlsanity.com/paper/2402.07411/) (2024)
- [GRPO, Dr. GRPO, and DAPO Are Three Operations on One Number: The Group-Standard-Deviation Identity](https://mlsanity.com/paper/2607.00152/) (2026)

## Built on by (most cited first, 5 of 5)

- [Hints, Critics, and Teachers: Prior Injection for Sparse-Reward RL in Vision-Language Math Reasoning](https://mlsanity.com/paper/2608.21811/) (2026)
- [The Filter Metric is Safety-Critical: Phantom Advantages in Group-Relative RL under Shaped Rewards](https://mlsanity.com/paper/2609.13866/) (2026)
- [ProCredit: From Outcome Rewards to Progress Credit in Agentic Reinforcement Learning](https://mlsanity.com/paper/2609.27532/) (2026)
- [RELACE: retrospective likelihood-based action credit estimation for long-horizon language agents](https://mlsanity.com/paper/2610.07349/) (2026)
- [Tev: Scaling Long-Horizon Agentic Environments with Dense Progress Rewards](https://mlsanity.com/paper/2610.18868/) (2026)

## Abstract (arXiv)

Dense per-step supervision is the standard remedy for sparse-reward long-horizon LLM agents: reward the policy for predicting its next observation, which looks provably safe under potential-based shaping. Published prediction-reward and auxiliary-loss variants report both successes and instabilities; we supply the controlled account: 74 preregistered arms dissect one fixed prediction signal under GRPO across ALFWorld, WebShop, a synthetic POMDP, and Qwen3-1.7B/4B/8B, varying only the delivery mechanism. (1) Every run sustaining this difference-form reward under untouched std normalization (no filtering, dynamic-sampling, or decoupling mitigations) collapses: eleven runs across scales, coefficients, group sizes, and groupings (the floor-bound synthetic environment stalls instead); ALFWorld runs end in an absorbing state (prediction accuracy -> 1.0, success -> 0): the optimizer builds the "dark room". The algebra is one line: in all-fail groups z-scoring cancels the shaping coefficient; removing only std normalization restores baseline parity. (2) A signal's danger is set by its within-group variance trajectory, plus hackability as a second axis; it retrodicts every reward-channel collapse and survives preregistered prospective tests. (3) The same signal as a teacher-forced auxiliary loss is harmless on ALFWorld at 4B, but the gain is not the signal's: content-free placebos as a class match or beat gold at both matched seeds (s0: 78.8 vs 68.6; s42: 67.9 vs 57.9); the auxiliary update is the regularizer. (4) At 8B the recipe turns bistable: gold full-weight locks two of three seeds; every content-free or reduced-weight arm stays healthy. No ALFWorld or WebShop reward-channel variant measurably beats its matched-normalization baseline and no gold signal measurably outperforms its content-free placebo: the delivery channel, not the content, decides; which channel is safe is regime-dependent.