agent/papers/items/2026-2607-21273-the-dark-room-in-the-reward-channel-dense-prediction-rewards-collapse-grpo-train.md

Paper: The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works

---

type: paper title: "The Dark Room in the Reward Channel: Dense Prediction Rewards Collapse GRPO-Trained LLM Agents -- and What Actually Works" authors: Yu Wang year: 2026 venue: arXiv url: https://arxiv.org/abs/2607.21273 code_url: source: arxiv collected_at: 2026-07-27 published_at: 2026-07-23 updated_at: 2026-07-23 status: queued relevance: high topics:

• memory
• planning
• tool-use methods:

benchmarks:

models:

datasets:

• cs.LG related_concepts:

related_jobs:

related_experiments:

related_projects:

collection_score: 14 collection_queries: llm-agent

One-line Takeaway

Auto-collected from arXiv because it matched the Agent collection queries. Needs human skim.

Why Collected

• matched queries: llm-agent
• inferred topics: memory, planning, tool-use
• arXiv categories: cs.LG
• collection score: 14

Review Checklist

• Does this paper directly inform Agent architecture, evaluation, memory, tools, safety, coding agents, GUI/browser agents, or multi-agent workflows?
• Does it include a benchmark, dataset, code, or reproducible experimental setup?
• Should it be promoted from queued to skimmed or summarized?

Links

• arXiv: https://arxiv.org/abs/2607.21273